EMB 20100301 Mar 2010

IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
Magazine
VOLUME 29
v
NUMBER 2 n http://magazine.embs.org n MARCH/APRIL 2010
Contents | Zoom in | Zoom out Search Issue | Next Page For navigation instructions please click here
Contents | Zoom in | Zoom out Search Issue | Next Page For navigation instructions please click here
Digital Object Identifier 10.1109/MEMB.2010.936462
Previous Page | Contents | Zoom in | Zoom out | Front Cover | Search Issue | Next Page
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
VOLUME 29
v
NUMBER 2
MARCH/APRIL 2010
http://magazine.embs.org
IEEE ENGINEERINGIN
MEDICINE AND BIOLOGY
Magazine
Navigating Technology
Transfer
Themes
17 Health-Care Technology
Nicolas Chbat, Emilio Sacristan, and Andrew Laine
18 The eICUResearch Institute
Michael McShea, Randy Holl, Omar Badawi, Richard R. Riker, and Eric Silfen
26 Nerve Conduction Studies
Xuan Kong, Eugene A. Lesser, and Shai N. Gozani
37 ANovel Phototherapy Device
Robert Malkin and Vijay Anand
44 Continuous Monitoring of Respiratory Flowand CO
2
Michael B. Jaffe and Joseph A. Orr
53 Quest for the Artificial Pancreas
Rebecca A. Harvey, Youqing Wang, Benyamin Grosman, Matthew W. Percival,
Wendy Bevier, Daniel A. Finan, Howard Zisser, Dale E. Seborg, Lois Jovanovic,
Francis J. Doyle, III, and Eyal Dassau
63 Optical Coherence Tomography
Freddy T. Nguyen, Adam M. Zysk, Eric J. Chaney, Steven G. Adie,
Jan G. Kotynek, Uretz J. Oliphant, Frank J. Bellafiore, Kendrith M. Rowland,
Patricia A. Johnson, and Stephen A. Boppart
71 Deep-Tissue Dynamic Monitoring of Decubitus Ulcers
Rohin Moza, J. Michael Dimaio, and Jose Melendez
78 Robotic Systemfor Transrectal Biopsy of the Prostate
Haytham Elhawary, Zion Tsz-ho Tse, Marc Rea, Alex Zivanovic,
Brian L. Davies, Colin Besant, Nandita M. deSouza, Donald McRobbie,
Ian Young, and Michael U. Lamperth
87 Image-Based Motion Detection
Eyal Zadicario, Shlomi Rudich, Ghassan Hamarneh, and Daniel Cohen-Or
95 Electronic Data-Capturing Technology for Clinical Trials
Zhengwu Lu
103 Wireless Sensing Systems in Clinical Environments
JeongGil Ko, Tia Gao, Richard Rothman, and Andreas Terzis
110 Real-Time Analysis for Intensive Care
Marion Blount, Maria R. Ebling, J. Mikael Eklund, Andrew G. James,
Carolyn MCGregor, Nathan Percival, Kathleen P. Smith, and Daby Sow
119 Clustering Techniques
Antonio Miguel Cruz, Sandra Patricia Usaquen Perilla, and
Nidia Nelly Vanegas Pabon
127 Clinical Appointment Process
Jie Lian, Kori DiStefano, Somer D. Shields, Cindy Heinichen,
Melissa Giampietri, and Lian Wang
135 Dysphonic Speech Reconstruction
H. Irem Turkmen and M. Elif Karsligil
ARTVILLE, EYEWIRE, & MASTERSERIES
IEEE ENGINEERING IN MEDICINE AND BIOLOGY MAGAZINE MARCH/APRIL 2010 1
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Features
12 Technology Transfer
Jesse Jayne Rutherford
139 Engineers Find Electric Connection
Jesse Jayne Rutherford
Departments
3 Fromthe Editor
A Tale of Two Flemings
5 Presidents Message
Grand Challenges in Biomedical Engineering
6 Letter to the Editor
Threshold for Plagiarism
7 Students Corner
Looking Ahead! Tips on Putting Together
a Five-Year Career Plan
9 Student Activities
Blogging . . . Again . . . Yes, Again?
10 GOLD
Technology for the Benefit of Humanity
141 Senior Design
Developing Teamwork Skills in Capstone
Design Courses
144 Point of View
Medical Devices in Cinema
145 Book Reviews
150 Conference Calendar
152 Et Cetera
Mission Statement
The Engineering in Medicine and Biology Society of the
IEEE advances the application of engineering sciences and
technology to medicine and biology, promotes the profes-
sion, and provides global leadership for the benefit of its
members and humanity by disseminating knowledge, setting
standards, fostering professional development, and recogniz-
ing excellence.
NOTES FOR CONTRIBUTORS
IEEE Engineering in Medicine and Biology Magazine is a theme-article publica-
tion that covers the full range of fields within biomedical engineering (BME),
with each issue covering one theme. Articles are written for technically knowl-
edgeable readers who are not necessarily specialists in the theme topic. A sam-
ple list of theme topics of interest includes: biochemical engineering,
biocontrols, bioinformatics, biomems, biomaterials, biomechanics, biosignal
processing, biotechnology, cellular and tissue engineering, clinical engineering,
imaging and image processing, information technology, instrumentation, sensors
and measurements, micro- and nanotechnology, neural systems and engineering,
physiological systems modeling, proteomics, radiology, rehabilitation engineer-
ing, robotics in surgery, and telemedicine. In addition to the theme articles,
which are invited contributions, the magazine also publishes unsolicited features
that are of interest to a broad segment of IEEE Engineering in Medicine and
Biology Magazine readers.
IEEE Engineering in Medicine and Biology Magazine also publishes over 20
regularly scheduled columns for readers interested in industry, academia, and
government. All articles are peer reviewed and written by experts in the field.
On occasion, the magazine publishes comprehensive, in-depth review, tutorial,
and survey articles. Letters to the editor, notes, commentaries, and other pieces
of personal opinion will be published as such. We also seek press releases
related to activities in your company, organization, or school.
Manuscripts are ONLY accepted in electronic format through Manuscript
Central at the Web site http://embs-ieee.manuscriptcentral.com. Instructions for
creating an account and how to electronically submit a manuscript are available at
the Web site. Do not send original submissions or revisions directly to the editor-
in-chief. If you are unable to submit your contribution electronically or have ques-
tions on manuscripts style, please contact the editor-in-chief: Michael R. Neuman,
Department of Biomedical Engineering, Michigan Technical University, 1400
Townsend Dr, Houghton, MI 49931-1295, USA. Voice: +1 906 487 1949. E-mail:
mneuman@mtu.edu.
As per IEEE policy on standards for publications, review and editorial
handling of manuscripts are part of the paper submission process to guarantee
quality control.
Make the Deadline
In order to have your news published in the magazine in a timely fashion,
please submit your notices to the editor by the following dates:
Issue Deadline
Jan/Feb September 1
Mar/Apr November 1
May/June January 1
July/Aug March 1
Sept/Oct May 1
Nov/Dec July 1
IEEE Engineering in Medicine and Biology Magazine (ISSN 0739-5175) (IEMBDE) is
published bimonthly by The Institute of Electrical and Electronics Engineers, Inc., IEEE
Headquarters: 3 Park Ave., 17th Floor, New York, NY 10016-5997. NY Telephone 1 212
419 7900. NY Telecopier 1 212 752 4929. NY Telex: 236-411 (International messages
only). IEEE Service Center (for orders, subscriptions, address changes, Educational
Activities, Region/Section/Student Services): 445 Hoes Lane, Piscataway, NJ 08854. NJ
Telephone: 1 732 981 0060. Price/Publication Information: Individual copies: IEEE
Members $20.00 (first copy only), nonmembers $88.00 per copy. Subscriptions: $5.00 per
year (included in Society fee) for each member of the IEEE Engineering in Medicine and
Biology Society. Nonmember subscription prices available on request. Available in micro-
fiche and microfilm. Copyright and Reprint Permissions: Abstracting is permitted with
credit to the source. Libraries are permitted to photocopy beyond the limits of U.S.
Copyright Law for private use of patrons: 1) those post-1977 articles that carry a code at
the bottom of the first page, provided the per-copy fee indicated in the code is paid through
the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923 USA; 2) pre-1978
articles without fee. For all other copying, reprint, or republication information, write
to: Copyrights and Permission Department, IEEE Publishing Services, 445 Hoes Lane,
Piscataway, NJ 08854 USA. Copyright ' 2010 by the Institute of Electrical and Electronics
Engineers, Inc. All rights reserved. Printed in U.S.A. Periodicals postage paid at New York,
NY and at additional mailing offices. Postmaster: Send address changes to IEEE
Engineering in Medicine and Biology Magazine, IEEE, 445 Hoes Lane, Piscataway, NJ
08854 USA.
Printed in USA
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_________
From the Editor
atale of two flemings
Michael R. Neuman
O
n a recent trip to London,
my wife and I spent a
sunny, autumn Sunday af-
ternoon taking a guided Lon-
don walk (http://www.walks.com). We
try to do this each time we are in the
U.K. capitol, for these tours explore
parts of the city that we would never
find on our own. The guides are every
bit as interesting as the sights. We have
had guides who were actors, historians,
professors, archaeologists, or just plain
interesting people. Their presentations
are as entertaining as they are educa-
tional. On a recent tour, we visited
Dickens and Shakespeares London,
with a guide who recited excerpts from
the writings of these giants of English
literature at sites where the authors had
visited, lived, or were prominent in
their work. These inexpensive excur-
sions are a great way to spend about
two hours on the streets and lanes of
London and are highly recommended
for any reader who finds himself or
herself in that part of the world.
On our last visit, our Sunday afternoon
tour was to the Little Venice section of
the city not far from St. Marys Hospital
in Paddington. Our guide, Shaughan,
pointed out the many upscale homes
along the Regents Canal and regaled us
by reciting poetry and singing music hall
ditties about the canal boatmen and the
neighborhood of the days gone by. He
also pointed out the historic sites and
homes of the rich and famous. One of the
The pineapple was
considered a
symbol of
wealth in
Victorian England.
IEEE ENGINEERING IN MEDICINE AND BIOLOGY MAGAZINE
Editor-in-Chief
Michael R. Neuman
Michigan Technological University
Deputy Editor-in-Chief
Silvestro Micera
Scuola Superiore SantAnna
Editorial Board
Hojjat Adeli
The Ohio State University
Howard I. Bassen
Food and Drug Administration
Krzysztof J. Cios
Virginia Commonwealth University
Pouran Faghri
University of Connecticut
Limin Luo
Southeast University, Nanjing
Jasjit Suri
Biomedical Technologies Inc.
Eugene Veklerov
Lawrence Berkeley Laboratory
Associate Editors
A Look At
Jean-Louis Coatrieux
University of Rennes, France
Book Reviews
Paul King
Vanderbilt University
Cellular &Tissue Engineering
Nenad Bursac
Duke University
Clinical Engineering
Stephen L. Grimes
GENTECH
COMAR
Richard A. Tell
Richard Tell Associates, Inc.
Emerging Technologies
Dorin Panescu
St. Jude Medical
Faces and Places
Andrew Szeto
San Diego State University
Genomics
Harold (Skip) Garner
University of Texas
Southwestern Medical Ctr.
Government Affairs
Luis Kun
National Defense University
Industry Affairs
Semahat Demir
National Science Foundation
Issues in Ethics
John Fielder
Villanova University
International News
John Webster
University of Wisconsin, Madison
Patents
Maurice M. Klee
Fairfield, CT
Point of View
Gail D. Baura
Keck Graduate Institute,
Claremont, CA
Regulatory Issues
Robert Munzner
DoctorDevice.com
Grace Bartoo
Instrumentation for Science
and Medicine
Retrospectroscope
L.A. Geddes
Purdue University
Senior Design
Jay Goldberg
Marquette University
Society News
Jorge E. Monzon
Universidad Nacional del
Nordeste
Student Activities
Cristian A. Linte
University of Western Ontario
IEEE PERIODICALS MAGAZINES DEPARTMENT
Managing Editor
Debby Nowicki
Senior Art Director
Janet Dudar
Assistant Art Director
Gail A. Schnitzer
Business Development Manager
Susan Schneiderman
+1 732 562 3946
ss.ieeemedia@ieee.org
Fax: +1 732 981 1855
Advertising Production Manager
Felicia Spagnoli
Production Director
Peter M. Tuohy
Editorial Director
Dawn Melley
Staff Director, Publishing Operations
Fran Zappulla
IEEE prohibits discrimination, harassment, and bullying. For more information, visit http://www.ieee.org/web/aboutus/whatis/policies/p9-26.html.
Editorial Correspondence: Address to Michael R. Neuman, Department of Biomedical Engineering, Michigan Technical University, 1400 Townsend Dr, Houghton, MI 49931-
1295, USA. Voice: +1 906 487 1949. E-mail: mneuman@mtu.edu.
Indexed in: Current Contents (Clinical Practice), Engineering Index (Bioengineering Abstracts), Inspec, Excerpta Medica, Index Medicus, MEDLINE, RECAL Information Services,
and listed in Citation Index.
All materials in this publication represent the views of the authors only and not those of the EMBS or IEEE.
_________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
___________
streets we visited was named Fleming
Court in honor of Sir Alexander Fleming
who discovered penicillin at nearby St.
Marys Hospital. Shaughan then told us
the story of how this Nobel prize-win-
ning discovery was made quite by acci-
dent. By the way, we later learned that
Dr. Flemings laboratory at St. Marys is
nowa museumthat is open to the public.
We then continued on the tour and
learned many more interesting facts
about the attractive Little Venice section
of London. For example, the pineapple
was considered a symbol of wealth in
Victorian England. No wonder we saw
so many representations of it in this
neighborhood. It also reminded us that
that it was time for tea as the guided
walk ended, so we decided to find a
place to have a cup. As we walked along
the residential Clifton Gardens toward a
cafe recommended by Shaughan, we
passed the house at number nine. I
noticed a blue plaque on the front. This
is the way that London recognizes places
associated with famous Londoners, and
there are many of these plaques on build-
ings scattered throughout the city. This
one was recognized the home of John
Ambrose Fleming. Fleming was an
electrical engineer and physicist and
knighted for his work as the developer of
the first vacuum tube diode, which later
became known as the Fleming valve.
Many attribute him to be the founder of
modern electronics, and in addition to
the blue plaque, he was awarded the
IEEEMedal of Honor in 1933.
So what is the point here? There have
been many famous Flemings in several
different fields; even our own. I doubt
that either Sir Alexander or Sir Ambrose
had any idea of biomedical engineering
(BME); yet, I think that they can remind
us of what is special about this field. They
represent a bridge between medicine and
engineering, even if only through their
name, that BME came to signify so many
years after their two important discov-
eries. I do not know if they ever person-
ally met one another, but if they did, they
might have discussed how their two
seemingly different disciplines could be
brought together to better understand and
improve the health of mankind. Indeed
in their lifetimes, early developments in
medical electronics were just beginning
to appear. Perhaps, Sir Ambrose could
have developed instrumentation to quan-
tify the growth of Sir Alexanders bacte-
rial cultures or monitor his patients. Sir
Alexander might have tried to justify
Darwins theory of evolution to Sir
Ambrose, since the latter had some
misgivings with the theory [1].
Actually, I have no idea if they ever
met or what they would have talked
about if they did get together, but one
can imagine that they had the ingredients
to become great BMEs. We as BMEs
come from a variety of backgrounds, for
our field is one of the truly interdiscipli-
nary engineering specialties. Most of my
generation studied something else as stu-
dents, for BME did not exist during those
years. This has helped to make BME an
exciting area in which to work because of
the varying backgrounds that we all bring
together; even those who have been more
recently trained as BMEs. It is not just
the varying backgrounds, but the fact that
we all talk together and share our experi-
ence as we address problems in the life
sciences and clinical fields. It is through
this interaction that we can cross-fertilize
ideas and bring new approaches to our
investigations. Perhaps, the two Flemings
would have done the same.
Our department at Michigan Tech is
special to me because we have faculty
with different backgrounds, ranging
fromelectrical engineering to analytical
chemistry with a lot of applied biology
in between. It has given me great pleas-
ure to see these young people interact-
ing and collaborating on their research
and teaching activities. They and their
students are the future of our profes-
sion, and they are off to a fantastic start
by minimizing disciplinary barriers.
Perhaps, cross-disciplinary collabora-
tion was not as frequent in the Flem-
ings time, but it has made a difference
today. Sir Alexander and Sir Ambrose
made quantum changes to their fields,
and BMEs have the potential to do the
same in our time. We have an additional
advantage: our ever-expanding oppor-
tunities to interact and collaborate across
disciplines. Let us use this opportunity to
its greatest advantage as our profession
moves forward. Maybe the two Flem-
ings did not have the occasion to in-
teract, but we still do. Lets take full
advantage of it.
Reference
[1] J. A. Fleming, Evolution or Creation? London:
Marshall Morgan and Scott, 1938.
Sir Alexander
and Sir Ambrose
represent a
bridge between
medicine and
engineering.
4 IEEE ENGINEERING IN MEDICINE AND BIOLOGY MAGAZINE MARCH/APRIL 2010
Fromthe Editor (continued)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Presidents Message
grandchallenges in biomedical engineering
S
tarting in 2010, the IEEE Engi-
neering in Medicine and Biology
Society (EMBS) is launching a
new series of forums addressing
the grand challenges in biomedical engi-
neering. The forum series will review
the significant progress we have made in
the past decade and identify the grand
challenges facing the scientific commu-
nity in a specific discipline within the
biomedical engineering field in the next
ten years.
The EMBS has a strong portfolio of
conferences, including its flagship an-
nual conference and several special
topic conferences. All these conferences
accept unsolicited submissions to pro-
vide a platform for authors to present
original R&D work in parallel oral or
poster sessions. Such technical confer-
ences provide an important platform to
exchange original research ideas.
With the ever accelerated pace of
R&D, there is a need to hold public
debates to identify the grand challenges
we are facing in a scientific discipline
and the future trends of the field. Such
debates would provide important guid-
ance to the scientific community to focus
the limited resources by tackling signifi-
cant issues to make big progress in a
field. In the new forum series, leading
intellectuals from academia, govern-
ment, and industry will be invited to
present their visions. All the participants
will be encouraged to join the interactive
panel discussions that will follow the
invited presentations. The forum is
planned to have a single-session format
to maximize the discussions and debates.
The first IEEE EMBS Grand Chal-
lenge Forum will address the grand
challenges in neuroengineering. Bridg-
ing engineering and neuroscience, neu-
roengineering is an emerging field that
translates research discoveries into neu-
rotechnologies that provide new and
powerful tools for basic and clinical
neuroscience research and lead to
enhanced patient care. Exploration of
neural systems has long focused on
understanding how neural systems work
at the molecular, cellular, circuitry, and
system levels. Engineering methodolo-
gies have always played an important
role in the study of neural systems, pro-
viding tools needed to detect, process,
and model neural signals. Recently,
tremendous progress has been made in
the field of neuroengineering, both in
the application of engineering concepts
and methodologies to the study of neural
systems and in interfacing neural sys-
tems with external devices for restora-
tion of lost neural function. The rapid
progress and tremendous translational
potential of neuroengineering has been
well recognized for the past several
years. The EMBS has played a leader-
ship role in shaping the development of
neuroengineering through excellent
representation in its annual conference
and by starting a special topic confer-
enceIEEE EMBS International Con-
ference on Neural Engineering in 2003,
which has continued to provide a plat-
form for the presentation of original
research work.
The IEEE EMBS forum on grand
challenges in neuroengineering will
further EMBS efforts in promoting this
important emerging field. The aim is to
conduct strategic discussions and de-
bates open to and engaging the scientific
community. The outcome of the forum
shall answer questions, including: Where
are we? What are the major obstacles
and challenges? and Where should the
field go in the next decade? The organ-
izing committee has initially identified
the following grand challenges in neuro-
engineering: 1) How can we decode
brain signalsfromsingle neuron to the
whole brain? 2) How far can we inter-
face the brain with the machine? 3) Can
we image brain function with high
spatial and temporal resolution? 4) What
are the mechanisms underlying the ef-
fects of neurostimulation and neuromo-
dulation? 5) Howdoes the brain compute
and process information? 6) Is it possible
to reverse engineer the brain? 7) How
can we repair or reverse damage to the
nervous system? and 8) How does ge-
netic information control the function of
the brain?
The scientific community in neuro-
engineering is invited to participate in
this first EMBS Forum on Grand Chal-
lenges in Biomedical Engineering,
which will be held 78 May 2010 in
Washington, D.C. Send your sugges-
tions on grand challenges in neuroen-
gineering or your interest to participate
in the forum to EMBS Executive Office
at emb-conferences@ieee.org.
Bin He
Engineering
methodologies
have always played
an important role in
the study of neural
systems, providing
tools needed to
detect, process,
and model neural
signals.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_________________
Letter to the Editor
thresholdfor plagiarism
T
he editorial on plagiarism (Dont
Play It Again, Sam) in the Sep-
tember/October issue of IEEE
Engineering in Medicine and Bi-
ology Magazine has hit on several sub-
jects for which I have much interest. I
have served for many years on the
Student Honor Boards at the University
of Maryland.
The student dishonesty cases are
mostly concerned with either exam
cheating or plagiarism. I found myself
sympathetic toward freshers accused of
plagiarism, when it was clear that high-
school teachers had encouraged them to
include verbatim excerpts in their own
papers, and no instructor in the univer-
sity had yet explained to them what con-
stituted plagiarism. On the other hand,
upper-class students who plagiarized
certainly should have known better, and
consequently, they were given grades of
XF in the offending courses (F for fail-
ure and X indicating dishonesty). They
could have the X removed from their
transcript by attending an academic
integrity seminar. Thus, the emphasis
for the punishment was to teach them to
modify their behavior for the better.
Once the seminar was successfully com-
pleted, they could retake the course if
they wished. The F would still remain
for the first time they took the course.
I think that plagiarism needs to be
defined differently for science and engi-
neering compared with arts. In the arts,
words are used to convey feelings or
particular interest in the arrangement of
words; science and engineering use
words to describe products or proc-
esses. It is difficult to achieve clear and
adequate descriptions of products or
processes without using the same words
as previously used. The value of these
words lies in the rational interpretation
rather than irrational emotion. There-
fore, I would want to see lines from a
poem included in quotation marks and
given an acknowledgment of their
source, but I would want neither quota-
tion marks nor source attribution for the
description of a hemostat.
It is similar to a brand-name usage.
Some brands such as Jello have become
part of the common use vernacular.
Other brands have not, because their
owners have taken great pains to protect
the integrity of their brands. Once they
transition into common usage, brands
are no longer considered exclusive.
Scientific and engineering descriptions
usually carry with them the assumption
of common usage.
That is why one must use duplica-
tion-search software with caution. There
is no clear limit as to how many words
can be duplicated before it is defined
as plagiarism. Identical strings of two,
three, or more words are not at all rare
in descriptive writings. Identical strings
of 100 words would clearly indicate
plagiarism. Somewhere between these
limits is the threshold for plagiarism.
In the arts, an identical pair of words
not in quotes could well be plagia-
rism. It is hardly ever so in science
and engineering.
There are other issues, too, that are
more difficult to detect. Figures and
drawings, as well as words, can be pla-
giarized. How much different does one
drawing have to be from another in
order that the two not be considered
legally identical? My publisher says
40%, but it is really hard to recognize a
40% difference in a drawing. This is
especially true, again, for a common-
use, vanilla-type drawing.
So, it is good that we do not look too
closely at the issue of plagiarismin IEEE
Engineering in Medicine and Biology
Magazine. There is only quicksand on
the path to the muck, and it is always bet-
ter to begin with an assumption of hon-
esty rather than with a suspicion of guilt.
Arthur T. Johnson
The student
dishonesty cases
are mostly
concerned with
either exam
cheating or
plagiarism.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Students Corner
lookingahead! tips on putting
together afive-year career plan
M
any of you browsing through
this column may find your-
selves a few months away
from finishing your engi-
neering degree, putting the finishing
touches on your masters or doctoral
thesis, or wrapping up your last couple
of publications in your postdoctoral
training, and, at the same time, asking
yourselves what comes next.
Although finishing up a degree is a
great achievement and a huge relief, it is
stressful. You are no longer in the sweet
spot where you knew what you are
going to do for the next two to four years
but rather find yourself having to face
the real world and make decisions about
your future. Not only do you have to
worry about submitting and defending
your thesis, but you also need to have
your next fewyears all figured out.
Take a graduate students case for
instance, for whom a common question
is what are you going to do next? It is not
easy, and it does not really get any eas-
ier, especially considering that career
development is not usually a topic that is
introduced or taught as a fundamental
component of ones undergraduate or
graduate curriculum, but it is neverthe-
less a critical component of ones entire
career. The point is, in the big picture, no
one is going to look after your career for
you, but you, as Erin Malone stated in
her 2004 article on Planning Your
Future, nor will anyone wait for you to
get your pieces together! So what is the
best way to go about it? Plan ahead one
step at a time!
Start Planning! The
Earlier the Better!
When you are in school, you may not have
felt the pressure to think a lot about your
long-term career goals. You just walked
into new opportunities that taught you
new skills and kept you challenged, and
staying challenged has been enough of a
plan and a goal. However, often times, as
Graeme Edwards once said, its not the
plan that is important, its the planning.
For most of you still in training, a good
strategy to start thinking about your
future is to draft out a short-term, five-
year career plan, which you can revisit
on a timely basis, readjust your goals,
and identify the tools and decisions you
need to make to achieve those goals.
Strategic Career Building
Putting together your five-year career
plan is much like planning anything else
in your day-to-day activities. First you
need to identify what you want to
achieve, evaluate yourself and your
career now, and reflect upon what you
have done so far toward achieving your
goals. Second, you need to develop a
strategy that will help you obtain what
you aim for, eventually breaking down
what lies ahead into several milestones.
Finally, somewhere along the way, you
must take the time to reevaluate your
goals in the context of your achieve-
ments to date and modify your plan
accordingly. From an engineering per-
spective, this process involves several
iterations: you identify and analyze,
strategize accordingly, then you revisit,
reevaluate, and adjust.
Identify, Evaluate, and Analyze
There are a few key facts that need to be
considered when drafting out your plan.
You need to assess where you are in your
career at present and reflect upon your
passions and priorities to help you
identify your goals five years from now.
As you know your strengths and weak-
nesses, you should be able to come up
with an answer to what makes me
unique? No matter which way you
choose to go, academia or private sector,
you need to gather as much information
as you require about your foreseen posi-
tion. Think of it as joining a new sports
team, one that you have always wanted
to play on. As a rookie, you will need to
familiarize yourself with the rules of the
game. You are about to venture into a
future direction that you need to know
more about before getting into it. Also, it
helps to know who is with you and who
is against you. You will need to provide
your prospective employer with several
references from past supervisors, team
leaders, or colleagues; so you should
have a good idea who you can rely on
for a good reference. Also, before you
start your five-year plan draft, make sure
you include the date. When you reflect
back on this document, it will act as a
reminder of where you were when you
first started and what was important for
you at that time.
Strategize
Now that you have an idea of where you
are, what you are good at, and what you
want to achieve in the next five years, so
you can start thinking of what approach to
take. One thing is for sure: you will prob-
ably work your way through several differ-
ent milestones to reach your final goal.
First, start small (plan 36 months
ahead) and think of each milestone as a
short-term tangible goal that can be
easily achieved and help you move
toward the long- term goals. For exam-
ple, if you are an undergraduate student
interested in biomechanics, plan to
include at least a couple of biomechanics
courses in your curriculumas electives in
your senior year. Similarly, as a graduate
student, aim to have part of your work
submitted to a high-impact conference in
the field. These kinds of deadlines help
you complete parts of your research and
write it up, making it easier when you
have to compile your entire thesis. Keep
in mind that it is important to include
tangible goals that are achievable in the
available time; otherwise, you would not
be able to accomplish your short-term
goals, which will lead to disappointment
and then panic.
Now it is time to start thinking bigger
(612 months ahead), e.g., aiming to learn
a new skill or exploring another subtopic
of your research project. Get together with
your adviser and collaborators, share your Digital Object Identifier 10.1109/MEMB.2009.935719
Cristian A. Linte
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
ideas, plan it out, and aim for a paper sub-
mission at the end of the year.
As an undergraduate student, if you
plan to look for a job as a software devel-
oper, you might need to seek and get
involved in a software development pro-
ject for your undergraduate thesis. If you
are a graduate student aiming to remain in
academia, it may be a good time to look
into the requirements of getting a teaching
certificate while working as a teaching
assistant for a couple of courses.
Thinking Beyond
the One-Year Mark
Nowthings are getting a bit more serious.
You have planned your next half year,
you have planned past your next year,
and it is time to look even further ahead.
For most of you, the two-year mark may
involve getting out of your current status
and getting into something else, such as
academia, industry, or even becoming
self-employed. Some of you will have
finished your degree by that time and
may be working for that software devel-
opment company you were aiming for,
whereas others are getting ready to finish
up their doctoral degree to join another
laboratory soon for postdoctoral training.
Although you are looking two years or so
ahead, you need to be realistic about
when you expect to complete your degree
and are ready to move on.
Use the next two years to make con-
tacts and get to know your peers. Attend
conferences, visit collaborators, make
yourself known, and let the world know
who you are and what you do. Also, do
not forget to maintain a good relation-
ship with your previous colleagues; it is
always good to keep in touch with your
team members. You never know how
they can help you!
Long-Term Goals
Now you are finally getting to that goal
you had in mind all along and you are
also aware that you will most likely be in
a different professional environment than
the one you are currently in. It is not unre-
alistic to portray yourself as a project
manager for a biomedical engineering
company or an assistant professor in the
faculty of engineering at some university.
You may find this planning stage
somewhat uncertain, mainly because it is
too far down the road. Despite this uncer-
tainty, you should include these long-
term goals in your plan, not only for the
record of your intent to pursue them but
also as incentives to help you make cer-
tain decisions and changes to make it
happen. Although you may not know
exactly where you are going, you do have
a generic idea of where and what you
would like to be. We are all allowed to
have dreams, so do not be afraid to put
down a goal that may seem to be a
stretch. After all, the higher we aim, the
more we achieve!
Revisiting and
Adjusting Your Plans
You have managed to put together a
plan, but how do you make use of it? It is
after all an iterative process. You need to
go back to it, revisit your milestones,
reevaluate yourself and your progress
against your goals, and analyze your
achievements. Be flexible and proactive,
and feel free to revise your milestones if
things have changed. However, do not be
too quick to adjust the timeline and keep
pushing your deadlines by another cou-
ple of months; rather, try to figure out
whether the initial plan was too stringent
or perhaps too many things came up at
the same time delaying your plans. Be
honest with yourself, admit you could
have worked harder, and readjust your
milestone timeline, but learn from this
experience and do not let it repeat!
When you revisit your plan, do not
forget to balance out achievements,
such as professional and technical
skills, training and coaching opportuni-
ties, and other relationships or contacts
that can or need to be cultivated at each
stage in your career. Moreover, learn
from your colleagues and superiors,
identify the skills that make them suita-
ble for their roles, and strategize on
how you can include those skills in your
professional development. In addition
to achieving your goals, try to maxi-
mize your collateral accomplishments
that will contribute to you becoming a
more well-rounded and effective candi-
date or asset for a particular career goal.
What If I Change My Mind?
Currently you are still in training, so
there is a good chance you will change
your mind, either because of personal
aspirations (you discover during your
graduate degree that you are not meant
for academia and instead decide to work
in industry) or external factors (perhaps
the job market is poor, so instead of
looking for a job in the private sector,
you decide to apply for a post-doctoral
position). Do not be afraid of change,
often it happens for a good reason, one
which you may not have thought of ini-
tially. If you revisit your five-year
career plan often, you will be able to
adjust your plans to prepare yourself for
a slightly different career path.
To Conclude. . .
Planning your career is not a task that you
can carry out once and then sleep on it for
the rest of your life. It is a back-and-forth
process, where the initial plan will almost
never be the final one. Every new oppor-
tunity you encounter will open up your
eyes to bigger and greater ones. However,
you need a starting point, and the first five
years will define the next few down the
road. Similar strategies outlined here can
be implemented to develop a life-long
career plan: revisit your initial plan on a
regular basis, modify it according to your
achievements, set new goals . . . and start
all over again!
In the end, I would like to thank Dr.
Elizabeth Krupinksy (University of Ari-
zona) for her excellent course on Early
Career Professional Development in
Medical Imaging, parts of which I
have used as reference in this article,
offered at the SPIE Medical Imaging
Symposium held 1621 February 2008,
San Diego, California, USA.
Cristian A. Linte received his B.A.Sc.
degree in mechanical and materials engi-
neering from the University of Windsor
in 2004 and M.E.Sc. degree in BME
from the University of Western Ontario
in 2006. He is currently studying the Bio-
medical Engineering Graduate Program
at the University of Western Ontario,
London, Canada, toward a Ph.D. degree
in biomedical engineering.
Students Corner (continued)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Student Activities
blogging . . . again . . . yes, again?
B
logging. Yes, people are still
blogging, expounding upon their
feelings daily to a friendly and
receptive audience, sharing jokes,
and exchanging thoughts. My question is:
out of the zillion of blogs online (a slight
exaggeration, perhaps), how do people
find these fabulous blogs that enthrall
them on a continual and regular basis?
Personally, I have yet to discover a blog
that can hold my attention for more than a
few weeks. I have never created my own
cyber diary, although perhaps the columns
I write for this magazine could be seen as
a blogging variant (Mag-blogging? Mag-
ging? Ag-ging?). I would cherish a new
online haunt to while away the hours
on a topic I am interested inreading,
perusing, clicking, laughing at, laughing
with, and contributing to. The new Web
2.0 sites including Facebook and Twit-
ter are great for socializing when time
permits, but I lose interest when my aca-
demic side twinges for intellectual fod-
der. Already, I spend many honest hours
a week reading the IEEE Engineering
and Biology Society (EMBS)-related
e-mails, e-journals, e-magazines, e-news-
letters, and more traditional and ink-
filled sections of newspapers (shocking,
I know). There are times when I wish I
could visit just one Web site and get my
fill of EMB newspast, present, and
futurewith a smattering of entertain-
ment, advertisements, and a chance to
interact with peers. Do you see where I
amheaded with all of this?
Yes, the EMBS has started a young
members (YM) blog, which sits com-
fortably on the EMBS home page
(www.embs.org/ymblog). This blog in-
cludes hot topics, news items, articles of
interest, reminders, advertisements, cur-
rent opinions, questions, and concerns
all posted regularly (by the minute/hour/
day/week). Subscription to the YM blog
is available via e-mail and really simple
syndication (RSS) feed. The site is main-
tained by both the student and EMBS
Graduates of the Last Decade (GOLD)
representatives, and all of our EMBS
members (both young and old) are
encouraged to visit on a regular basis.
Yes, reader, that means you. The YM
blog is an informal, personable, sociable,
creative, and fun method of promoting
dialog and encouraging relationships
between EMBS members.
Interestingly, this is not the first time
that EMBS has started a blog. In 2006,
our student representative, Dr. Jennifer
Flexman, created the EMBS student
blog . . . for students of the EMBS and
all peeps interested in biomedical engi-
neering. Always having done her
research, she even provided a compre-
hensive list of EMB blogs that she had
come across, printed again here for your
interest (Wednesday, 23 August 2006
entryOther Biomedical Engineering
Blogs). However, four years later, it is
the blogs on the list that have survived,
whereas the EMBS-sponsored site failed
to launch over the long term. Unfortu-
nately, the EMBS turned its attention
elsewhere, i.e., to other Web 2.0 technol-
ogies, to the creation of more professio-
nal and consistent e-newsletters, and to
the newdesign of the EMBS Web site.
What makes EMBS think that this
newblogging endeavor will be different
than that of 2006? How will it succeed
and last more than just a few months?
First, the YM blog has been created in
response to members request. The
EMBS blog will act as your guide to the
fast-moving world of biomedical engi-
neering, and there is no organization
better equipped to support this effort. In
2006, perhaps our members were not
quite ready for a blog or did not know
about it or forgot to go to the link. Sec-
ond, the site will be overseen by a team
of enthusiastic and e-savvy EMBS vol-
unteers sharing the responsibilities of
monitoring the site and looking out for
its longevity, namely, the editor-in-
chief of IEEE Engineering and Biology
Magazine as well as the student and
GOLD representatives and other mem-
bers of the Member and Student Activ-
ities Committee.
Finally and most importantly, hope-
fully you (as a reader) will graciously
support the site and visit at least once,
perhaps even making a contribution or
comment along the waymaybe not
today and tomorrow, but definitely
sometime this month. If blogging is not
the right forum for you, then please do
share the Web site with colleagues,
peers, and students. This is an exciting
new project where membership interac-
tion holds the key to its success. I, too,
promise to visit the YM blog often and
even contribute when the mood strikes.
Perhaps after all of these years, I will
have found my much-sought after and
kindred blog. Perhaps you will too!
Lisa Lazareck has completed her
doctorate of philosophy in electrical
engineering in the Engineering Science
Department, University of Oxford. She
is currently working as a research
assistant at the City e-Health Research
Center (CeRC), City University Lon-
don, U.K., developing educational (sci-
ence/hygiene) Web games for children.
References
[1] A blogring about biomedical engineering and
NASA [Online]. Available: http://www.xanga.com/
groups/group.aspx?id=540895
[2] Internet Journal of Emerging Medical Technolo-
gies (MedGadget) [Online]. Available: http://www.
medgadget.com/
[3] Neurobot, a blog about computational neuro-
science and bioengineering [Online]. Available:
http://neurobot.bio.auth.gr/
[4] Chaaraka: Medical devices, biotechnology, bio-
engineering and the like [Online]. Available: http://
chaaraka.blogspot.com/
[5] Purdue Universitys BME blog [Online]. Avail-
able: http://weldonbme.org/
[6] BMESource, a resource for the biomedical
technology design community [Online]. Available:
http://englibrary.blogspot.com/2005/11/bmesource-
design-resource-for.html Digital Object Identifier 10.1109/MEMB.2009.935715
Lisa Lazareck
______________
___
_________________________
____________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
GOLD
technology for the benefit of humanity
T
he IEEEs mission statement
says, IEEEs core purpose is to
foster technological innovation
and excellence for the benefit of
humanity. The essence of biomedical
engineering is to develop technology
for the benefit of the patient and human-
ity. It is not surprising that this years
IEEE Presidents Change the World
Competition had many projects
awarded that were related to biomedical
engineering. The task of the competi-
tion was to tackle real-world problems
using engineering, science, computing,
and leadership skills to benefit the
community and/or humanity (www.
ieee125.org/change-the-world/). The
IEEE Student Humanitarian Supreme
Prize of US$10,000 was awarded to
Drew Hall and Richard Gaster, two stu-
dents at Stanford University, Califor-
nia, for their development of NanoLab:
a hand-held diagnostic laboratory
that was designed to create a portable
protein-detection platform. Its impact
on society ranges from life-saving clini-
cal diagnostics in the third world to
rapid over-the-counter tests sold in
pharmacies (www.ieee125.org/change-
the-world/top-competitors/nanolab-hand-
held-diagnostic-labratory.html). The
IEEEs Distinguished Student Humani-
tarian Prize of US$5,000 was awarded
to students at B.V. Bhoomaraddi Col-
lege of Engineering and Technology,
India. The team developed electronic
aids for physically/mentally handi-
capped children. The aim was to use
electronics to help physically/mentally
handicapped children of USHAS Center
for Exceptional Children, Hubli, India
(www.ushas.org), by developing games/
devices/toys, etc. that will create excite-
ment/interest in the children to play
them and hence undergo physical/men-
tal exercise, which is prescribed to over-
come them from their disability but
which they normally tend to neglect due
to lack of enthusiasm and interest
(www.ieee125.org/change-the-world/top-
competitors/electronic-aids-physically-
mentally-handicapped-children.html).
Another two projects with biomedi-
cal engineering background were each
awarded an Outstanding Student Human-
itarian Prize (US$1,000): Information
System on Human and Health Services
(ISHHS) by Sampathkumar Veerara-
ghavan, Medford, and NIDAANAn
e-Healthcare Solution for the Under
Privileged by Sumit Pandey, Gandhina-
gar, India. Sampathkumar Veeraragha-
van also received the Peoples Choice
Prize for the same project with which he
was able to help millions of people in
India. He and his project are also featured
in ieee.tvs Profiles in Volunteering
(www.ieee.org/portal/site/ieeetv/menuitem.
6ce799f946c20d660374ca695bac26c8/
index.jsp?&pName=ieee.tv.viewer&
path=membport/ieee_tv&file=CNT_
profiles_Veera.xml&vid=111877&play=
true).
Altogether, biomedical research proj-
ects received five of the nine awards in
this years IEEE Presidents Change the
WorldCompetition. We wouldlike tocon-
gratulate the winners whole heartedly and
encourage them to continue their work for
the benefit of humanity. The 2010 IEEE
Presidents Change the World Competi-
tion was open for project submission until
31 January 2010, and the awards will be
presented in person on 26 June 2010
(www.ieeechangetheworld.org).
Together with the United Nations
Foundation, the IEEE and especially
Graduates of the Last Decade (GOLD)
members are working on the Humanitar-
ian Technology Challenge (HTC). Here,
the IEEEis sponsoring a Regional Student
Design Competition for solutions to one
of the three humanitarian problems as part
of the joint IEEE United Nations Founda-
tion HTC. The competition starts from
October 2009 to May 2010. Again, two of
the three humanitarian problems presented
are related to the engineering in medicine
and biology field: Data Connectivity of
Rural District Health Offices and Indi-
vidual ID Tied to Health Records. The
third project concerns reliable electricity
in resource-constrained environments.
Both the IEEE Presidents Change the
World Competition as well as the IEEE
HTC target the student and GOLD mem-
bers to work in teams to develop technol-
ogy for the benefit of humanity. Many of
the challenges and problems identified
are closely related to biomedical engi-
neering. I would like to encourage all our
GOLD members to participate in these
challenges and invite everyone to be-
come part of the EMBS GOLD Global
Community. We will hold conference
calls regularly to discuss our involve-
ment and your engagement to tackle the
challenges at hand. The EMBS GOLD
Global Communitys aim is to help you
around the world to realize your goals in
your region. Please contact Matthias
Reumann (mreumann@ieee.org) if you
would like to get involved and for more
information on howyou can participate. Digital Object Identifier 10.1109/MEMB.2009.935710
Matthias Reumann
The essence of
biomedical
engineering is to
develop
technology for the
benefit of the
patient and
humanity.
________________________
________________________
_________________________
____
________________
_________________________
_____________________
_________________________
___
_________________________
_________________________
_________________________
_____________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
____
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
STOCKBYTE
What Every Biomedical
Engineer Should Know
T
he means of transforming what began as university
research into an industrial product for clinical
application, known as technology transfer, is a long
and difficult process that is best undertaken with
the early support of several specialized professionals who
understand the pitfalls and obstacles the inventors typically
encounter along the way, and even then many inventions never
make it to marketsome of them perfectly safe, effective, much-
needed inventions that could have been transferred successfully,
had an inventor changed strategies or had more support. If the road
to successful technology transfer, characterized as it is by potholes,
appears frustratingly pointless to the mission of science, consider that the
process has many parallels to science and that its ultimate destination is
the validation of science. In science, although one makes hypotheses and
bases new research on careful analysis of previous studies, a researcher does not
know for certain the outcome of an experiment until it is complete. Technology transfer is
a process of moving forward with the belief that a particular invention has value, and while the process may not
have the result the scientist hypothesized, on the whole technology transfer results in contributions to science
and to the world in general. If engineers hope their inventions can change their field, improve health and well-
being, and even change the world, then technology transfer cannot be overlooked.
According to Emilio Sacristan, a visiting professor at Stanford University and founder of Innovamedica,
Abdeo Medical, and Critical Perfusion, Inc., who has navigated the waters of bringing inventions to industry
multiple times, there are three pillars of successful technology transfer: intellectual property (IP) strategy,
regulatory strategy, and the business model. Other considerations, including the larger implications of the
transfer of your technology and the ethical questions it raises, must be made as well.
The technology transfer office (TTO) at your institution is crucial to helping you with these three pillars,
and any researcher who believes an invention could have a practical application should contact their TTO
immediately. Many TTO directors are directly affiliated with the Association of University Technology Man-
agers (AUTM), a global network of technology transfer professionals who promote, improve, and support
technology transfer. Besides universities, its members come from industry, legal, and government settings.
AUTM is not a governing board but an organization that reports on technology transfer, informs its members
of news and proposed changes to laws affecting their work, and supports efforts to make transferred technol-
ogy accessible to developing nations. Your university or research institution may or may not be affiliated with
AUTM, and as each TTOis unique, it is helpful to understand the three pillars of technology transfer for your-
self and to be aware of where you might need to ask for additional support if it is not discussed by the TTO.
Intellectual Property Strategy
Protecting your IP is one step in a successful transfer of technology, and it means more than filing for patents.
Experts agree that the most important aspect of protecting your IP is that you refrain from public disclosure of
inventions until they are protected. This means, obviously, that researchers not publish papers divulging the inven-
tion. Dr. Ashley Stevens, executive director of the TTO at Boston University and president-elect of AUTM,
BY JESSE JAYNE RUTHERFORD
12 IEEE ENGINEERING IN MEDICINE AND BIOLOGY MAGAZINE 0739-5175/10/$26.002010IEEE MARCH/APRIL 2010
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
remarks that one of the saddest occurrences in the business of
technology transfer is when a researchers own work is cited
against him or her. The published articles frequently end with a
section titled Recommendations for Further Study, and a
phrase along the lines of, This research suggests that . . . which
is followed by an allusion to the authors next project, perhaps
one that has already been funded. Unfortunately, publishing such
a phrase makes the subsequent research obvious, in the sense that
any reader can reach the same conclusion; therefore, the idea is
not exclusive to the researcher and can put IP at risk.
Cecily ORegan is an attorney with a background in engineer-
ing who works in the Intellectual Property & Technology Group
at Greenberg Traurig law firm performing medical device pat-
enting. She agrees with Stevens, saying that writing or even talk-
ing about an invention at a conference can jeopardize an
inventors ability to apply for a patent, especially with regards to
filing overseas. In the United States, there is a grace period, but
since the grace period varies fromcountry to country and is non-
existent in some, she says, As a general rule, we like people to
think, Dont talk before you file. She also adds that unpro-
tected IP is unlikely to be picked up by industry because the
financial risk of commercializing an invention that can be imple-
mented by any competitor is too great, so leaving your technol-
ogy unprotected to free it up for anyone to acquire is not an
effective method for seeing your inventions commercialized.
This advice may appear, at first glance, to be at odds with sur-
viving in a publish or perish academic community, where
publishing and presenting can lend the researcher job security.
Although universities are beginning to appreciate the revenue
streams generated from licensing technologies developed in
their laboratories, there is not much professional recognition
bestowed on scientists who keep mum about exciting discov-
eriesyet. Increasingly, youll see an appreciation among
tenure review committees about an individuals accomplish-
ments regarding patented technology and practical applications,
much more so in than in the past, notes Dr. Andrew Laine,
professor of biomedical engineering and radiology and vice
chair of the Department of Biomedical Engineering at Colum-
bia University. But I think itll take another five years, maybe
ten years, before it becomes on par with published literature.
Attorney ORegan urges every inventor she works with to
obtain signed confidentiality agreements when discussing their
inventions and to keep official, up-to-date lab notebooks and get
them countersigned. Years ago, such chores seemed like admi-
nistrivia to her, but her experience illustrates the importance of
taking care of these seemingly small details: at one point,
ORegan had a client, Dr. McMann, who discussed an unpub-
lished invention with another engineer, Howard Brown (names
have been changed). Brown liked the idea so much he enlisted his
own attorney to file a patent application on the invention, which
was published just as ORegan was working on the conversion of
McManns nonpublished case. Coincidentally, McMann was just
about to close negotiations on US$5 million of funding. When
ORegan heard of Browns case, she called it to McManns atten-
tion, who recognized Browns name. Crushed, McMann told his
investors what had happened and called ORegan on Christmas
Eve to tell her they had pulled out and that his wife was probably
going to have to get a job. However, because of his careful plan-
ning, McMann was able to present the signed confidentiality
agreement he had obtained fromBrown and three years worth of
his well-kept lab notebooks, countersigned by a noninventor who
could serve as a witness. Brown, on the other hand, had no
documentation and finally admitted to having stolen the idea;
everything eventually worked out in McManns favor.
As evident from ORegans story, developing a sound IP
strategy requires early planning and support from a patent law
attorney who keeps abreast of your field. Wading through patent
lawis not a task most researchers are equipped for, and at any rate,
most would prefer to spend their time working on research that is
compelling enough to warrant patenting in the first place. How-
ever, it is important that the inventor be involved in the process
and plan for IP protection early on by not disclosing unprotected
inventions, by obtaining signed confidentiality agreements, and by
keeping up-to-date, countersigned lab notebooks. It is also helpful
during this process for the attorney to explain what the patent
means, howit works, and what is next on the inventors to-do list.
Regulatory Strategy
The key to effective regulatory strategy is to understand the
Food and Drug Administration (FDA) approval process, includ-
ing how a device will be classified, what types of data will be
required to prove it is safe and effective, and how to maintain
FDA approval once the device is approved. The FDAs Center
for Devices and Radiological Health (CDRH) provides resour-
ces and support for this process, including an extensive Web
site with a section for education called CDRH Learn. We
spend a lot of time holding peoples hands and trying to help
them work through the process, explains Dr. Donna-Bea Till-
man, the CDRHDirector and an IEEEMember.
Dr. Bruce Burnett, director of regulatory affairs for the Duke
Translational Medicine Institute at Duke Universitys School of
Medicine, advises inventors to take advantage of the opportunity
to meet with reviewers and ask questions about the acceptability
of the clinical trial design before submitting it, decreasing the
likelihood of a bad surprise. There is no fee for these meetings,
C
O
P
Y
R
I
G
H
T
2
0
1
0
I
S
T
O
C
K
P
H
O
T
O
L
P
C
O
P
Y
R
I
G
H
T
2
0
1
0
I
S
T
O
C
K
P
H
O
T
O
L
P
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
unlike in Europe, where some similar agencies charge a fee. Dr.
Burnett has supported researchers in the FDA approval process
for many years. I think the FDA sees themselves as wanting to
approve good devices and drugs, so people should take advant-
age of the potential to interact with themearly in the process and
thereby make sure that you dont run into a situation where
through lack of some understandingassuming that this a good
device[approval] is delayed. Getting involved and utilizing the
procedures that are there, meaning having contact with the FDA, I
think would be very helpful [in making] the process much more
efficient and thereby [in getting] good devices and drugs
approved earlier and available for patients, Burnett concludes.
At the CDRH, Dr. Tillman says the most common stumbling
block for inventors is underestimating what it takes to bring a
product to market, which she believes is rooted in a lack of under-
standing of the regulatory process. Dr. Tillman is pleased that bio-
medical engineering students today are being trained on the FDA
and the FDAs regulatory mandates, which should help future
inventors develop more effective regulatory strategies. In the
meantime, she advises todays inventors to distance themselves
from their passion for an invention and take the mindset of a neu-
tral third party while making a thorough analysis, asking them-
selves what kinds of evidence would be needed to prove their
device is safe and effective. The people that work here . . . we
are scientists, too, Tillman explains, and . . . we expect to see
good science and a solid foundation in terms of what evidence is
provided to support marketing of devices . . . if [inventors] . . .
think about what they need to do wearing their scientist hat and
try to factor in the regulatory framework, that may be the best path
forward, because the same kinds of questions that they might ask
are probably going to be the questions we would ask.
Business Model
Like those who work closely with IP protection and FDA
approval, professionals who support the business model aspect
of technology transfer advise inventors to plan early and to reach
out for support from qualified professionals who specialize in
this area. If the premise of an endeavor is to bring the technology
to the consumer at a profitor, in a nonprofit setting, at a price
that enables the organization to continue providing the technol-
ogythen several factors must be considered and aligned to
provide the market opportunity, says Bill Cawley, senior vice
president of UTEK Corporation, an IP licensing and innovation
services company based in Florida. UTEK connects its client
companies with universities and research laboratories that have
technology ready for licensing. What Cawley says is often
missing from technology coming out of universities is the driv-
ing force of unmet consumer needs, a disconnect he frequently
encounters when research fromuniversities is organized around
what the government will fund and what researchers are inter-
ested in, rather than what is economically attractive for
commercial buyers and the end users they sell to.
Jerry Keusch, professor of international health at Boston Univer-
sity and author of When you reach a fork in the road, take it: Sci-
ence and product development as linked paths, (American
Journal of Law & Medicine, 2008), writes Researchers are, by
nature, entrepreneurial, albeit not necessarily in a commercial
sense. Speaking by phone, he explains further: The problemthat
I see is not that researchers arent inherently entrepreneurialyou
have to be, to succeed in a research career where youre dependent
onwritingandgettinggrant fundingbut theres a lackof exposure
tothelarger worldwhereresearchfindings aretranslatedintopoten-
tial products and where the issues around patenting, licensing, and
development of products for actual use byreal people comes up.
Those real people and their need for, and ability to purchase,
products are not part of the way research is typically set up at
universities. The federal government is not driving their deci-
sions as to what to invest in based on consumer unmet needs,
UTEKs Cawley explains, citing funding of alternative energy
solutions by governments around the world: Theyre thinking
about energy needs and geopolitical issues that affect those
areas, but theyre missing the other components, which are eco-
nomic: Is an organization able to make money off of it? Are
they able, with current technology, to manufacture products that
are at the price point that consumers can afford and will there-
fore buy? And the answer is no. So, while the business model
is not ready for the technology, since funding is available,
research proceeds in that direction, and it is not until it is time to
transfer the technology that this disconnect becomes apparent.
Another point to consider is how well developed the technol-
ogy is. Commercial buyers are loath to invest in a technology that
is still in the early stages or that does not dovetail well with the
manufacturing and marketing of their existing products, because
the financial risk they must incur to finalize the technology and
prepare it for sales is too great. The final selling price of a device
has less to do with the cost of manufacturing and more to do with
recovering the initial investment before the first sale is ever
made. Dr. Wolfgang Henggeler, a technology transfer manager
and head of the Physical Sciences Group at Unitectra, the
technology transfer organization of the Universities of Bern and
Zurich in Switzerland, agrees. He reports that Unitectra success-
fully transferred a dynamic vision sensor they patented and cites
the concurrent development of another device by the acquiring
company, a device in which the sensor fit perfectly, as the reason
for the success. This case demonstrates the greater likelihood of
success when the costs to the company for adoption are low, but
the potential benefits are high. Most companies have many
C
O
P
Y
R
I
G
H
T
2
0
1
0
I
S
T
O
C
K
P
H
O
T
O
L
P
Dr. Sacristans Considerations for the Business Model
Hereareafewcrucial points toconsider early inbusiness
model development, accordingtoDr. EmilioSacristan:
How are you going to distribute/sell this?
Whos goingtopay?Whos goingtoget paid?How?
(For services) Are you going to sell the service or
rent the service?
(For products) Are you going to sell disposable or
consumable products?
What are the costs associated with bringing the
product tomarket, andhowwill you recover them?
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
internal development projects in the pipeline which compete for
resources, he explains via e-mail. If the technology can
provide benefit to any of those projects, adoption is feasible. To
start a completely new development/commercialization project
within a company is usually only possible if there is an internal
champion who is convinced of its economic potential and ready
to strongly promote the technology internally.
Inaddition, other business planningaspects suchas analyzingthe
market and your competitors before proceeding can be time con-
sumingandexpensive onthe front endbut maypayoff wheninven-
tors realize that a slight change in the business model or shift in
market focus can vastly improve a products chances in the market-
place. Another strategy for ensuring that technology is transferred
successfully, when a spinoff company is formed, is to understand
your skill area as a professional and to hire others to assist you with
business development. This is especially crucial when it comes to
running a company if business leadership is not within your skill
set. Just as most engineers arenot equippedtowrite their ownpatent
application and need to hire an attorney, they may not be equipped
with the marketing skills and business vision necessary for running
a successful company and should look for a qualified leader they
trust. Cawley points out that it is difficult to acquire funding from
investors if youare tryingtosell a business with yourself, the inven-
tor, named as chief executive officer (CEO). Naming another CEO
does not mean you must erase yourself fromcompany leadership; a
positionsuchas chief technologyofficer couldbe a better fit.
Larger Implications and Ethical Considerations
Perhaps the weightiest consideration for biomedical engineers
is accessibility of their inventions once the technology has
been transferred. It is all fine and well to improve global health
in theory, but if in practice, only a few people have access to
the technology, whether by geography, knowledge, or finan-
ces, what has science really accomplished? Dr. Keusch of Bos-
ton University has personally faced this question on projects of
his own. Now, having a greater understanding of the process,
he explains that when you patent, you have some degree of
control in licensing that you can use to create access to your
technology in impoverished countries. One of the things that
I think our tech transfer office has done has been to work on
the language for a socially responsible licensing policy, where
there are inventions that have been patented that are of impor-
tance for health in resource-poor countries . . . the licensing
agreement does not give up [the right] to push for access and
affordable products for those populations, whether its through
non-exclusive licensing or arrangements with a company thats
primarily involved in exclusive license to be able to market at
an appropriate-level price for that level of population.
Next, there are ethical considerations for your research insti-
tution, for competitors, and for the public that contributes to
the funding of the research. Obviously, allowing technology to
be licensed and generate income for the university or other
institution is a reciprocation of the support the institution gave
the research; income can even benefit a department directly,
paying for new lab space, endowing chairs, and so on, ulti-
mately furthering the field and facilitating research. This is
another reason leaving technology unpatented is not quite the
altruistic move some inventors would like to believe; patenting
it and using the income it generates to further research may
actually be in alignment with an inventors goals.
Understanding where funding comes from, what motivates it,
and the trends that shape science may also help the inventors plan
for the long term. As Dr. Keusch explains, expanding on his article,
a privatefoundationcandistributegrants as it sees fit, withthebene-
fit to grantees being that it is unlikely to change its vision in a few
years. The downside is that a private foundation may not fully
understand howto prioritize funding at high levels of a specialized
field. A governmental organization, on the other hand, has much
more oversight and typically strives for informed decisions made
withamorediversepanel of advisers, but it canbeheavilyimpacted
by a change in political power, distributing funds only to programs
that adhere to current ideology. The benefit of governmental over-
sight, however, is greater accountability: if an expensive program
funded by a governmental organization is mismanaged, an outcry
is likely; if an expensive programfunded by a private foundation is
mismanaged, few outside of the foundation and researchers who
could have better used the funds will be upset, and the situation
could recur. Although it may not be reasonable to turn down fund-
ingfor a particular project, understanding the likelihood of continu-
ing funding based on the funding source and what could influence
it canhelpengineers prepare for possible changes.
Besides a technology not reaching intended patients, there is
another unfortunate consequence that can occur after licensing:
often, exclusive licensees will use the patents they acquire to
obliterate preexisting competitors with products already on the
market, some of those competitors being clinical laboratories at
other universities, according to Bob Cook-Deegan, director of
the IGSP Center for Genomics, Law, and Policy at Duke
What Happens When There Is a Monopoly
on a Patented Health Service?
In May of 2009, the American Civil Liberties Union and
the Public Patent Foundation filed suit against the U.S.
Patent and Trademark Office (PTO), Myriad Genetics,
and the University of Utah Research Foundation,
charging that the PTO granted them unconstitutional
and invalid patents on two human genes associated
with breast cancer, BRCA1 and BRCA2. The lawsuit,
Association for Molecular Pathology, et al., v. United
States Patent and Trademark Office, et al., was filed
because Myriads lab is now the only place in the
country where diagnostic testing for mutations of the
genes can be performed. No other laboratory can
test the genes or work on alternative tests, and Myriad
charges more than US$3,000 for their tests. In Novem-
ber, the court upheld the plaintiffs right to challenge
the patents in court. While much of the focus in this
upcoming trial is associated with the legality of pat-
enting a part of the human body, monopolies like this
could also impact biomedical engineering because
the patent prohibits research at other labs and
restricts patient access to care.
C
O
P
Y
R
I
G
H
T
2
0
1
0
I
S
T
O
C
K
P
H
O
T
O
L
P
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
University, who has studied the subject. I dont think thats a
terribly good use of intellectual property that was created at pub-
lic expense, he remarks. The question of the ethicality of pub-
licly funded inventions being placed in the hands of licensees
who then knock other publicly funded laboratories out of the
market warrants considering, as it is not unusual. It does seem
to be fairly frequent. Its against the norms of [AUTMs] Nine
Points document, Cook-Deegan explains, referring to the asso-
ciations Nine Points to Consider in Licensing University
Technology. However, since AUTM is not a governing board,
endorsement and abidance by the nine points are at will, even
among universities that are AUTM members. That raises the
question, What should the purpose of licensing be: to ensure
something gets to market, or to destroy competitors? Its nice
to have aspirational statements; its less credible to have those
aspirational statements if nobody abides by them, Cook-Dee-
gan points out. Being aware of who is acquiring your license and
what their business practices have been could help prevent pat-
ents on your IPfrombeingused inthis way.
Arguments for transfer of technologytobenefit science andyour
institutions donot necessarilymean that federally fundedprograms
ought to reimburse the federal government for grants that result in
lucrative inventions; that is not how grants are designed, for good
reason, and this is an important point to understand. According to a
2009 Congressional Research Service report on the Bayh-Dole
Act, federal funding is designed to stimulate research because of
the benefit tothe public as a whole, bothdirectlyfromthe invention
and from the economic stimulus that results from commercializa-
tion, not because the government can recoup the expenses from
any one company. Although industry can and does provide spon-
sorship to universities, the federal government is still the largest
source of academic research funding, in large part for the very rea-
son the Bayh-Dole Act works so well: industry, on the whole, can-
not afford to extend itself completely into research, and so needs to
partner with universities to have access to truly innovative technol-
ogytocommercialize.
Conclusions
Technology transfer has served the field of biomedical engi-
neering well. Although the process is fraught with obstacles
and may appear to be a distraction from more important work
in the laboratory, application of technology is crucial to the fur-
thering of the field and to public health in general. A given
inventor may not want to take over the administrative tasks of
protecting IP, developing a regulatory strategy, and developing
a business model, and he or she does not necessarily have to;
however, the inventor needs to at least know about the strat-
egies and know there are people to turn to for leadership and
guidance outside of the laboratory early in the process.
And that can be all an inventor wants to do or can do: to sim-
ply turn the invention over to someone else and move on to
more research in the hopes that the invention will help improve
patients health and perhaps afford the inventor and the
research institution some financial reward. However, in turn-
ing an invention over completely, an inventor also loses power
over how, for whom, and at what price the invention will be
offered, and that is a reason to become more informed about
the process, to know what you can ask for, and to stand by your
inventions side in whatever capacity you are equipped to do
so throughout the entire technology transfer process.
Jesse Jayne Rutherford is an award-winning ghostwriter
and a freelance writer. She lives near San Diego, California.
For more information, please visit her Web site at www.
JesseRutherford.com.
Spotlight on Boston University
Boston University has one of the largest biomedical
engineering departments in the United States. It is
home to Dr. Jerry Keusch, who has worked on several
transfers of technology, and Dr. Ashley Stevens, its
executive director of technology transfer and presi-
dent-elect of AUTM. Among the developments at BU:
the US$200 million, 200,000 ft
2
National Emerging
Infectious Diseases Laboratory slated to open in
late 2010
the founding of the Consortium of Universities for
Global Health, borne of a meeting Dr. Keusch
organized and supported by the Bill and Melinda
Gates Foundation andthe Rockefeller Foundation
to standardize and improve global health educa-
tion and coordinate efforts between resource-rich
institutions and resource-poor countries
innovative graduate education that connects
students in biomedical engineering, medicine,
and management with each other, their faculty,
and industry
commitment to socially responsible licensing
guidelines and high involvement with AUTM
a project at the Center for Global Health and
Development to reduce neonatal mortality.
The Bayh-Dole Act: Accomplishing Its Goals
The Bayh-Dole Act, Public Law 96-517, enacted in
1980, was created
to promote the utilization of inventions arising
from federally-supported research or develop-
ment; . . . to promote collaboration between
commercial concerns and nonprofit organiza-
tions, including universities; . . . to promote the
commercialization and public availability of
inventions made in the United States by United
States industry and labor; [and] to ensure that
the Government obtains sufficient rights in feder-
ally-supported inventions to meet the needs of
the Government and protect the public against
nonuse or unreasonable use of inventions . . .
The Government Accountability Office (GAO) has
conducted studies on how well the Bayh-Dole Act has
met these expressed goals, and the results are that it
has done so extremely well. According to a 2009 report,
The Bayh-Dole Act: Selected Issues in Patent Policy
and the Commercialization of Technology, con-
ducted by the Congressional Research Service, the
academic sector has become a major source of
technological innovation for economic development
since the passage of the law, and the report sums up
expert surveys with Yale President Richard Levins quote
from The Economist that touted the law as probably
the most inspired piece of legislation to be enacted in
Americaover the past half-century. Thereport goes on
to say that the field of biotechnology in particular has
benefited from the Bayh-Dole Act, growing from a tiny
industry in the mid-1980s to an industry with annual sales
of morethan US$30 billionby 2005.
____
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Health-Care
Technology
D
emands on industrial research differ from those in
academic institutions. Although they are comple-
mentary, industrial research focuses more on practi-
cal issues, shorter timelines, and end-user needs,
whereas academic research tends to focus more on theoretical
issues and have longer term timelines. They both have novel
breakthroughs and attend to health-care needs.
This special issue of IEEE Engineering in Medicine and Biol-
ogy Magazine is aimed at highlighting health-care technology
from the aspect of industrial and clinical applications. Industrial
research work is driven by user needs and clinical insights. The
user can be the physician, nurse, patient, health-care system, hos-
pital, manufacturer of the medical system, or device, etc. Hence,
collaboration with clinical institutions is a necessity. Research
work resulting from such collaboration focuses on the problem in
hand without necessarily having to push the theoretical scientific
or engineering envelope. By focusing on patient problems,
applied research brings aspects to health-care solutions that theo-
retical research does not.
The IEEE Engineering in Medicine and Biology Society
(EMBS) is proposing the launch of a new journal in 2011 in col-
laboration with the American Medical Association and the Ameri-
can College of Clinical Engineers. This new journal will focus on
the development and use of applications for advanced technology
by industry engineers, clinical practitioners, and engineers. It will
also provide a forum in which practitioners, engineers, and devel-
opers may have open dialogue for improving health-care technol-
ogy for the future. This special issue of IEEE Engineering in
Medicine and Biology Magazine is a preamble to such a journal.
Nicolas Chbat received his Ph.D. degree
from Columbia University in 1995, at which
time he joined General Electric Global Re-
search Center (GE GRC) for seven years. He
is currently a senior researcher and quality
officer at Philips Research North America.
He leads efforts in clinical decision support
for cardiopulmonary medicine. In 2000, he
won the Dushman Award, GE GRCs Highest Technical Team
Achievement Award. He then spent four years at the Mayo Clinic,
Division of Engineering. In 2005, he won the Best Teacher of the
Year Award from the Mayo Graduate School. He holds 12 issued
and four filed patents. He coauthored the book Discrete-Time
Control Problems Using MATLAB and a chapter in the book
Advances in Healthcare TechnologyShaping the Future of
Medical Care, and 20 publications.
Emilio Sacristan received his Ph.D. degree
in biomedical engineering from Worcester
Polytechnic Institute (WPI) in 1993 and a
diploma in product design, development,
and management from Massachusetts Insti-
tute of Technology in 2003. He is a profes-
sor of electrical and biomedical engineering
at the UAM-Iztapalapa, Mexico, and direc-
tor of the Mexican National Center for Medical Instrumenta-
tion and Imaging Research, founder and chief science officer
of Innovamedica, Critical Perfusion, and Abdeo Medical. He
is an active member of the IEEE EMBS and CORAL, and has
served AdCom and other advisory boards in several capaci-
ties. His research interests focus on instrumentation for anes-
thesia and critical care.
Andrew Laine received his D.Sc. degree
in computer science from Washington
University School of Engineering and
Applied Science in 1989 and B.S. degree
from Cornell University, Ithaca. He was a
professor in the Department of Computer
and Information Sciences and Engineering
at the University of Florida, Gainesville,
from 1990 to 1997. He is now the director of the Heffner Bio-
medical Imaging Laboratory in the Department of Biomedical
Engineering at Columbia University and professor of biomedi-
cal engineering and radiology. He was a program chair for the
IEEE EMBS annual conference in 2006 held in New York City
and served as program cochair for IEEE ISBI in 2008, Paris,
France. He has served as the IEEE EMBS vice president of
publications since 2008. His research interests include quantita-
tive image analysis, cardiac functional imaging, ultrasound and
MRI, retinal imaging, intravascular imaging, and biosignal
processing. He is a Fellow of the IEEE and AIMBE.
BY NICOLAS CHBAT, EMILIO SACRISTAN,
AND ANDREW LAINE
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Industrial and Clinical Applications
IEEE ENGINEERING IN MEDICINE AND BIOLOGY MAGAZINE 0739-5175/10/$26.002010IEEE MARCH/APRIL 2010 17
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
The eICU Research
Institute
A
s the volume of data that is electronically available pro-
liferates, the health-care industry is identifying better
ways to use this data for patient care. Ideally, these
data are collected in real time, can support point-of-
care clinical decisions, and, by providing instantaneous quality
metrics, can create the opportunities to improve clinical practice
as the patient is being cared for. The business-world technology
supporting these activities is referred to as business intelligence,
which offers competitive advantage, increased quality, and
operational efficiencies. The health-care industry is plagued by
many challenges that have made it a latecomer to business intel-
ligence and data-mining technology, including delayed adoption
of electronic medical records, poor integration between informa-
tion systems, a lack of uniform technical standards, poor in-
teroperability between complex devices, and the mandate to
rigorously protect patient privacy. Efforts at developing a health
care equivalent of business intelligence (which we will refer to
as clinical intelligence) remains in its infancy. Until basic
technology infrastructure and mature clinical applications are
developed and implemented throughout the health-care system,
data aggregation and interpretation cannot effectively progress.
The need for this approach in health care is undisputed. As
regional and national health information networks emerge, we
need to develop cost-effective systems that reduce time and
effort spent documenting health-care data while increasing the
application of knowledge derived fromthat data [1].
As the availability of robust electronic clinical data grows, the
opportunities to translate that data into innovative health-care
practices are expanding. The eICUResearch Institute is an exam-
ple of a program delivering these results today. This article
chronicles the challenges in building such an infrastructure, pro-
vides a framework for collaboration between health-care pro-
viders and administrators, industry, and academic researchers,
and illustrates early lessons learned. While the focus is on a
large-scale program involving more than 40 health-care systems,
this model can also be applied to a variety of health-care projects.
Genesis of the eICU Research Institute
The electronic intensive-care unit (eICU) programthat started in
1998 provides technology that transforms care delivery via
remote monitoring of intensive-care unit (ICU) patients. Several
studies confirm improved clinical outcomes and a reduced cost
of care through shorter length of stay using this program[2], [3].
The systemevolved froma recognition that continuous monitor-
ing of critically ill patients can improve outcomes and reduce
complications [4]. Novel components such as Smart Alerts and
other population management tools as well as reporting best
practice compliance and severity-adjusted outcomes are impor-
tant factors in the success of the eICUProgram.
The technology behind these tools requires data collection
and integration. Standards-based interfaces allow different
hospital systems [laboratory, pharmacy, admit/discharge/
transfer (ADT) and enterprise clinical information systems]
and medical devices (vital sign monitors, ventilators, etc.) to
provide remote eICU clinicians with a comprehensive and
current view of the patient and enable powerful CDS tools.
On a quarterly basis, data collected from all health systems in
the eICU community are combined and analyzed to assess ICU-
specific compliance with evidence-based best practices and to
severity-adjusted outcomes such as mortality and length of stay.
Participating health systems also receive benchmark reports
comparing their compliance rates with those of peer institutions.
The highly consistent data schema used at all sites and the focus
on important clinical and financial outcomes allows health sys-
tems to integrate the information gained from these reports into
their clinical operations and the daily practice of patient care.
Although initially focused on quality improvement, the Phi-
lips VISICU eICU business (http://www.visicu.com) was
encouraged by several customers in 2006 to aggregate data
and allow observational studies of critical-care practices. In
2007, the eICU Research Institute (eRI) was formed as a pro-
gram within Philips VISICU, with a charter to create a perma-
nent, anonymized data repository and to use the collected data
for research purposes. The program was formally launched in
2008, and the first research proposals from eICU participants
were submitted in March 2009. The eRI database includes
more than 1,000,000 ICU admissions and is increasing by
nearly 275,000 newICUadmissions per year (see Table 1).
The eRI program represents a unique opportunity to gener-
ate evidence-based innovations for critical-care medicine, but
there were challenges to success that had to be overcome.
These obstacles reside in two main categories: organizational
BY MICHAEL MCSHEA, RANDY HOLL,
OMAR BADAWI, RICHARD R. RIKER,
AND ERIC SILFEN
EYEWIRE
A Collaboration Between Industry,
Health-Care Providers, and Academia
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
and technical. The organizational challenges include estab-
lishing a governance structure that allows customers control
over use of their data, and establishing mechanisms to insure
that academic independence is preserved without improper
influence from commercial interests. Given the lack of a
coherent national policy regarding data use for such projects
[5], the collaborative eRI framework is an excellent model for
vendor-sponsored programs.
Foremost among the technical challenges is ensuring that
patient privacy is maintained in the permanent data repository.
This requires a careful balance between reidentification risk
and preservation of the important information needed for
clinical research. Once the data are aggregated, getting value
from the data requires a good deal of hard work due to a vari-
ety of data anomaliesfrom the mundane to the peculiar.
These challenges are described in the sections that follow,
along with lessons learned fromearly research projects.
Organizational Framework
Even within the well-established community of eICU pro-
grams, each institution has its own set of legal policy and busi-
ness priorities that needs to be addressed up front in the
formation of the eICU Research Institute. Ultimately, through
collaboration with participating eICU customers, an organiza-
tional and governance structure was created that enabled the
value of the aggregate data to be unlocked in a way that met
all objectives of both Philips VISICU and its customers. The
operational model for the eRI is illustrated in Figure 1.
eRI Governance
Two committees are established to manage the eRI operations:
an Executive Steering Committee and a Publications Committee.
The Executive Steering Committee is charged with two tasks:
approving a charter for the eRI that defines the governance and
operating responsibilities for managing the eRI, and oversight of
ongoing research projects and activities. The Publications Com-
mittee is responsible for developing the tools and processes for
submitting, evaluating, selecting, and managing the overall
research agenda as well as selecting specific research proposals
based on clinical relevance and available budget and resource
capabilities. Both committees
include 78 appointees from
eICU customer organizations,
one representative from Philips
VISICU, and one representa-
tive fromour academic partner.
Charter and Legal
Framework
Created by the Executive Steer-
ing Committee, the eRI charter
is modeled after several existing
medical data repository proj-
ects, including the American
College of Cardiology-National
Cardiovascular Data Registry
(ACC-NCDR) (www.ncdr.
com/webncdr/common/) and
the Society of Thoracic Sur-
geons (STS) National Database
(www.sts.org/sections/stsnational
database/). The charter provides
authorship guidelines that follow the standards set by the Interna-
tional Committee of Medical Journal Editors (www.icmje.org) and
recommendations for editorial freedom by the World Association
of Medical Editors (www.wame.org/). The eRI charter also all
ows for future research proposals fromoutside the eICUcommu-
nity and for grant-based and sponsored research (within specified
bounds) to provide funding to create a self-sustaining business
model. This design maintains the independence and integrity of
academic research while providing for a sustainable program.
Partnership with Academia
Philips VISICUpartners with Academic Research Organizations
(AROs) to provide the independent analytic support for research
Table 1. eICU Research Institute database characteristics.
Descriptor Quantification
Health-system demographics
Health systems 32
Hospitals 188
ICUs and step-down units 356
ICU stays/patients
Total ICU admissions in database 1,013,706
Total ICU admissions in 2008 273,270
Total ICU admissions added per quarter 70,083
Stay characteristics
Average ICU length of stay 2 days
Average hospital length of stay 7 days
Percentage patients with a single stay 95
Basic outcome characteristics
Percentage stays resulting in death in ICU 5
Percentage stays resulting in death in
hospital
14
Data characteristics
Total laboratories >200,000,000
Total medications >20,000,000
Total diagnoses >8,000,000
Total vital sign measurements >600,000,000
Customer Participant
Research Proposals
Research Partner
Proposals
Third-Party Proposals
(Future)
Publications Committee
Proposals
Publications
Committee
Philips VISICU
Program Staff
Academic
Research
Partner(s)
Health System
Investigators
Governance
Executive
Steering
Committee
Analysis Outcomes
and Input to Manuscripts
Prioritized
Projects
Philips VISICU
Internal
Projects
Project-Specific
Deidentified
Data Sets
eICU Research Institute
Abstracts and Publications
Fig. 1. eICU Research Institute operational model.
______
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
projects. The organizational structure ensures the integrity of the
research process from an academic perspective, shortens the
learning curve on analytic techniques for operational data, mini-
mizes the upfront investment in staff and analytic tools, and pro-
vides a flexible approach for supporting research projects
conducted over time. Potential research partners are evaluated
based on experience working with large data sets, nationally rec-
ognized analytic expertise, programmatic and technical infra-
structure for conducting research, and proficiency with study
design and clinical governance. The University of Maryland
School of Pharmacy Pharmaceutical Research Computing Cen-
ter (PRC) is selected as the initial partner and is allocated one seat
on the Executive Steering and Publications Committees. Formed
in 1998, the PRCprovides data management and analytic support
for epidemiological and health services research studies (http://
www2.pharmacy.umaryland.edu/centers/prc/). The relationship
with the University of Maryland PRC offers access to institu-
tional review board evaluation and surveillance of clinical stud-
ies. Philips VISICU is responsible for extracting project-specific
data due to their familiarity with the database management sys-
tem but delegated the creation of the analytic data sets and
performance of biostatistical analyses to the AROpartner.
Research Selection Process
The Publications Committee convenes several times each year,
either via teleconferencing or in person to carry out the process
for reviewing the submitted proposals, screening and ranking
those proposals, selecting projects for assigning resources, and
monitoring ongoing research activity. The selection criterion are
based on the existing National Institutes of Health grading sys-
tem, assigning a primary and secondary reviewer for each
project, allowing comment by all committee members and rank-
ing each project from 1 (highest) to 5. Any committee members
who might have conflicts of interest due to geographic, profes-
sional, or social relationships to any submitted project or princi-
pal investigator must recue themselves from evaluation of that
proposal. Applications for support of clinical research projects
are submitted twice a year in March and September. Final deci-
sions regarding the proposed projects are made at face-to-face
meetings of the committee in May and November.
Research Project Life Cycle
The final dimension of the operational model (Table 1) is the
formation of a team to perform the clinical research. Once a
proposal is selected, a research teamis formed that includes data-
base engineers, software test engineers, and a clinical project
manager fromPhilips VISICU; a database engineer and lead bio-
statistician from the ARO; and the principal and coinvestigators.
A significant lesson learned from early projects is the need to
implement research project life cycle management, adopting the
discipline of the software development life cycle (SDLC).
Formal specifications describe the data-extraction phase and spe-
cifics of the data analysis protocol, capturing decisions made by
the research teamduring joint study design sessions.
The software development paradigm translates well into the
research domain with the following life cycle steps: study
proposal, study design, data requirements definition, data
extraction and verification, analytic data set preparation, data
analysis, and study outcome. Similar to software develop-
ment, the process is rarely linear, responding to newly recog-
nized problems as the process evolves. Clinical governance
throughout the project is critical for guiding decisions that
affect the integrity of the research results as is defining a clear
division of responsibility between principle investigators, the
ARO clinical researchers, and the industry clinical product
experts (see Clinical Work Flow Artifacts section).
Technical Challenges
The size and complexity of the source databases and the opera-
tional nature of the source systems pose multiple technical
challenges for the eRI that are relevant to other endeavors of
similar magnitude. The emphasis on patient privacy and confi-
dentiality in the governance structure carries through to the
technical implementation, where the tension between retaining
useful data for research and added complexity due to privacy
and confidentiality ultimately has to be resolved. The most
vexing challenges come from the interpretation of data early
on during research projectswhen artifacts are discovered in
the data that cannot be resolved without intense work flow
analysis and interviews with end users. Finally, for the program
to deliver value for academic and applied research objectives,
some internal capacity in the fields of machine learning and
pattern analysis is essential for the industry participant.
Patient Privacy and Health System Confidentiality
The HIPAA regulations allow for use of personal health infor-
mation (PHI) for research purposes if the data set is classified
as meeting safe harbor standards or meeting the scientific
standard for deidentification. The safe harbor approach
requires that 18 categories of fields be removed from the data
set. The scientific standard requires that the risk is very small
that the information could be used, alone or in combination
with other reasonably available information, by an anticipated
recipient to identify an individual who is a subject of the infor-
mation [6]. The scientific standard requires certification by a
qualified statistician, but there is no regulatory guidance as to
what constitutes a very small risk.
For the majority of parameters, the safe harbor approach is
straightforward. All patient identifiers are replaced with randomly
generated substitutes and event-based time references (e.g., admit
time, discharge time, treatment times, ventilator start/stop, and
time of death) are converted to relative time, referenced to ICU
As the availability of robust electronic clinical
data grows, the opportunities to translate that
data into innovative health-care practices
are expanding.
____________________________
____
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
admit time. Birth dates are converted to age, including a single
designator of >89 years of age for all patients aged 90 or older, as
required by HIPAA. All free text fields are removed (e.g., free
text fields in physician notes or comment fields in nursing flow
sheets), as these data fields can contain references to the patients
family members or other identifying information about the
patient. Additional confidentiality measures beyond those
required by HIPAA include replacing references to the source
health system, facility names, and caregivers with deidentified
randomly generated parameters. This masking insures health sys-
temand care provider confidentiality and further reduces reidenti-
fication risk. Relationships between patient-stays and units, and
unit-stays within a health systemare maintained to allowfor anal-
ysis across sequenced unit stays for the same patient.
The primary data of interest that violates safe harbor
standards are dates, e.g., date of admit, discharge, and death.
Classifying the database as a limited data set was considered
to preserve more data, as this type of data set is allowed to con-
tain dates more granular than year. The majority of eICU cus-
tomers did not consider eRI activities as falling within the
existing business associate agreement (BAA); however, this
approach would require institutional review board approval of
each and every study from each institution. As an alternative,
Philips VISICU constructed two parameters of interest for
future studies in place of actual dates, weekday/weekend flag,
and season of year. Inclusion of these parameters necessitated
an initial attempt at certification using the scientific standard.
To analyze reidentification risk, Philips VISICU contracted
with Privacert (www.privacert.com/). Privacert has developed
unique technology that allows measurement of reidentification
risk using an inference engine that analyzes the research data-
base schema against known publically and semipublically
available data, including state-based claims information, which
is not subject to HIPAA regulation and varies in every state.
The eRI deidentification approach failed because of health sys-
tem demographic information that indirectly revealed the state
in which the patient was treated, which then allowed easier cor-
relation between eRI data, state claims databases, and other
generally available databases, such as voter registration. Fortu-
nately, data that enabled identification of the state was easily
removed, and the eRI was tested against all 34 states that cur-
rently have monitored ICUs. Ultimately, the weekend/weekday
and seasonal information was also removed. The availability of
data in the public domain is changing too rapidly to insure that
the statistical reidentification risk will remain acceptable to
meet the scientific standard. For this reason, the final certifica-
tion is based on HIPAAsafe harbor standards.
Database Architecture and Processing
The span of potential investigations using the eRI data is both
broad, including longitudinal summarization for investigations
into general ICU practices, and deep, with narrow focus ques-
tions such as expected clinical response to a particular treat-
ment for a specific diagnosis. Support for such a wide range of
potential investigations requires the amalgamation of clinical
data from multiple sources into a single, unified corpus from
which subset data can be easily and quickly retrieved. The
architectural design of the eRI system, which is rapidly grow-
ing into a very large data set (>10 Terabytes), of true, online,
nonsegmented relational data, poses several distinctive chal-
lenges. Figure 2 illustrates the technical infrastructure. The eRI
is unique when compared with other data-driven research sys-
tems [7] due to the creation of high value-derived data via com-
plex clinical based algorithms that operate on the patient
record before the data warehouse processing. As shown, there
are four source databases, including raw delivered data, proc-
essed clinical benchmark data, processed Acute Physiology
and Chronic Health Evaluation (APACHE) input variables and
results, and a separate ICU and health system demographic
database that is updated through annual surveys.
Management of the eRI database management system is seg-
mented into two modes of operation: data load and data
retrieval. Data load operations include extraction, transforma-
tion, and load (ETL) from the source data sets, and data scrub-
bing and cleansing operations to ensure data quality, and
deidentification to meet HIPAA privacy standards. The ETL
process is constructed as sets of repeatable packages with con-
trol software for loading, unloading, and reloading data from
source databases with full audit trail to ensure data integrity. The
ETL process must recognize various source system versions, as
multiple schemas will always exist across the source databases.
Source data comes fromonline transaction processing (OLTP)
clinical systems. Transformation steps convert the highly normal-
ized data schemas of the OLTP sources into a typical star-schema
relational online analytic processing (ROLAP) model. To support
high-performance data retrieval, the hierarchical source data are
flattened (denormalized) and the star-schema is partitioned into
21 partition tables appropriately randomized for even distribution
of data across physical hard drives. This partitioning scheme
allows effective use of asymmetric parallel query processing by
the database engine for high-performance retrieval.
A key challenge to the incremental integration of multiple
data sources into the eRI data repository is reconciling the new
data inputs with the additional requirement of deidentifying
the data during the ETL run. This requires the creation of a dei-
dentification blinded surrogate key for the extraction process
to use as the relational linkage between data records. The dei-
dentification process for each ETL run uses the appropriate
identifying data to be extracted (e.g., patient IDand health-care
service provider ID), to create a new, unique nonidentifying
surrogate key [i.e., a pseudorandom globally unique identifier
(GUID)], in the extraction step. This surrogate key is registered
Efforts at developing a health-care equivalent
of business intelligence remains
in its infancy.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
onto a master-link list. The master-link list is used for recon-
struction mapping and is updated with the new surrogate keys
with each ETL run.
Research data subsets are extracted from the corpus to meet
specific study needs and are translated into suitable form for
the intended analytical tools to be used in the investigation. As
research studies are proposed and executed, database tuning
creates additional retrieval indices for high-performance data
extraction. These tuning efforts are cumulative as the range of
actual queries evolves to allow high-performance retrieval for
subsequent extractions with reduced effort. Based on the
underlying ROLAP star-schema design of the data corpus,
extracts can be created in a variety of formats including the
direct ability to create high performance, multidimensional
online analytic processing (OLAP) cubes that allow highly
flexible, interactive data exploration along with advanced ana-
lytic techniques.
Clinical Work Flow Artifacts
Real-world operational data sources present challenges with
data quality. Missing data, erroneous entries, test data, orphans,
and other faulty records must be identified and marked for
exclusion from the eRI data set. Data cleansing is an iterative
operation based on examination of exceptions identified in the
ETL process and basic exploratory analysis of the resultant
data for descriptive statistics and identification of outliers. To
eRI High-Level ETL Processing
VISICU Secured Server 1
eICU
Research
Database
H
Investigation
Queries
Investigation
DB
eCore
Manager
Client
Databases
VISICU
Processing
APACHE
Processing
eRI ETL
Process
eRI ETL
Index
Database
eICU/ICU
Survey Data
G
F
C C
D
B
B
A
E
Benchmark
Reporting
Data
APACHE
Reporting
Data
Legend
A = Quarterly Reporting Feed
B = Processed Data
C = Extracted Outcomes
D = Subset of Raw Archives
E = Blinded Surrogate Key
F = Demographic Survey
G = eRI Data Load
H = Research Data Extract
VISICU Secured Server 2
Fig. 2. eICU Research Institute technical infrastructure.
Clinical governance throughout the project is
critical for guiding decisions that affect the
integrity of research results.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
maintain the integrity of the data set, exceptions are marked for
exclusion rather than deleted. This ensures that exclusion rules
are independently verified, modified, and/or corrected after the
fact to validate or repeat any downstreaminquiry.
The context of these data anomalies can affect the ability to
extract useful conclusions fromthe data. Missing and erroneous
data are important to anticipate and methodically address dur-
ing the course of a research project. Artifact or data cleansing
rules are created and maintained over time. Ideally, they are fed
back into the design of the source software systems as input
data-checking rules. The action taken when data are missing or
erroneous has clinical significance, and proper action may vary
depending on the specific study. In some cases, missing data
may relate to inconsistent feature use between users or institu-
tions. Missing data may be expected, e.g., laboratory values not
obtained on a specific day may be imputed if absent.
Artifacts are even harder to interpret when underlying clini-
cal work flow is not understood. For example, in the course of
a study on readmissions, several such anomalies were discov-
ered, which serve to illustrate the issue. In one case, the analyst
reported that many patients had a unit stay extending after they
were reported to have died, with actual clinical data being
charted, making it unclear whether the second stay was a dif-
ferent patient or the death was invalid. Several weeks of
research were needed with actual users of the system before it
was determined that these ghost stays were in fact legitimate:
the patients were organ donors that had been pronounced brain
dead but continued to require ICU care before recovery of
organs. Figure 3 illustrates the multiple layers of data artifacts
that must be explored to insure a high-quality analytic data set.
Analysis of the data requires a high degree of system exper-
tise and knowledge of how the data are being recorded and
used, and in the case of eRI, this includes all the systems feed-
ing information to the eICUthrough interfaces. This knowledge
is not likely to exist with the principle investigator or the ARO
analysts. A clinical product expert is needed who often must
research issues during the analysis. Knowledge of how the sys-
temis used in real life is important in the both the research study
design and data-extraction phases, and continues throughout
the analytic phase of the project. Before accepting the results of
artifact cleansing, interpretive conclusions should be reviewed
by the research principal investigator and the ARO to provide a
degree of independent validation and clinical governance.
Analytic Techniques
As progress is made in refining best clinical practices in ICU
care and developing advanced CDS tools to assist these best
practices, advances in complex pattern analysis of finely
grained real-world clinical data is necessary. Traditional
regression analysis and statistical methods have proven to be
rigorous and efficient in finding linear relationships. However,
nonlinear patterns, which make up the bulk of clinical data
relationships, have until recently been handled with more ad
hoc, trial-and-error approaches. Machine learning, data min-
ing, kernel methods represent a class of theoretically well-
founded, robust, and powerful pattern analysis techniques that
offer exciting analytical possibilities. These techniques are
not without cost, posing new challenges to computational effi-
ciency and data manipulation in very large data spaces [8].
To derive the highest value from the data set for all of the
eICU Research Institute participants, Philips VISICU invested
substantial resources in developing in house capacity in the lat-
est analytic techniques. New kernel techniques (kernel meth-
ods) used with machine learning approaches, such as support
Missing Data
Exclude Record
Ignore or
Impute
Exclude Record
Ignore or
Impute
Definition,
Mapping,
Derivation
Filtering
Analytic
Data Set
Increasing Quality/Decreasing Number of Records
Bad Data Interpretation Study Cohort
Feature Version
and Evolution
Feature Usage
Patterns
Patient Needs
Missed Data Entry
Interface Failure
Sending System
Failure
Bad Data Entry
Bad Data Entry
in Sending System
Unexpected
Feature Usage
Clinical Work
Flow Artifacts
Terminology
Coding
Semantic
Meaning
Contextual
Meaning
System Based
Relationships
Time
Sequencing
Inclusion
Criteria
Exclusion
Criteria
Event
Windowing
Stabilization
Period
Refractory
Period
Fig. 3. Data artifacts and anomalies.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
vector machine analysis, are more adept at extracting meaning-
ful results from raw data, with different methods applicable
based on specific data quality considerations. With the large
eRI data set, even with highly specific patient exclusion/inclu-
sion criteria, large training data sets and test data sets can read-
ily be created. To accelerate learning and results, the Philips
VISICU staff is well-versed in these methods. Armed with
commercial mathematics software tools, analysts can proac-
tively contribute to discussions early in the study design phase.
Early Successes
As the eRI nears completion of its first year of operation, the
lessons learned have been applied to both internal product
R&D projects and external academic research projects along
with the expectation of multiple future publications. Several
projects described below illustrate the potential of the eRI
database for both product based and academic research.
ICU Readmission Analysis
Early discharge from the ICU is desirable (shorter length of
stay results in lower total cost) but can also lead to complica-
tions if the patient is transferred or discharged too soon (e.g.,
longer hospital stay or unanticipated death). Despite interest
in identifying predictors of readmission, an effective clinical
decision support (CDS) tool for helping physicians make this
important decision has not been identified [9][13]. The
hypothesis of the initial eRI study was that a discharge readi-
ness index could effectively predict the risk of ICU readmis-
sion or death within 48 h of discharge from an ICU. A
retrospective, exploratory cohort study was performed using
patient stays in the eRI database from 2005 through 2007 with
a calculated APACHE III score available [14]. Exclusion
criteria included do not resuscitate orders or similar care limi-
tations on ICU discharge. Regression tree analysis was per-
formed on a broad range of variables, including patient
demographics, ICU admission diagnosis, APACHE III score,
laboratory values, and physiologic variables present during
the last 24 h of the ICU stay. A logistic regression model was
created from the variables identified in the tree analysis.
A total of 123,848 ICU stays met criteria with 3,888 of these
resulting in death or readmission within 48 h of ICU discharge.
The analysis identified the following predictors of death or
readmission: age, admission diagnosis, admission APACHE
score, and the highest FiO
2
, average heart rate, lowest heart
rate, lowest oxygen saturation, and lowest systolic blood pres-
sure in the final 24 h of the ICU stay. The logistic regression
model produced a c-index of 0.71 (95% CI: 0.700.72) with a
bias corrected c-index of 0.71 and a calibration slope of 0.99.
When analyzed separately, the c-index for predicting death and
readmission within 48 h was 0.89 (0.870.90) and 0.61 (0.60
0.62), respectively. The model for the combined endpoint of
death or readmission produced moderate discrimination but
excellent generalizability, although there was a very good dis-
crimination when identifying death within 48 h of ICU. Further
research is currently being performed to refine the model and
prospectively validate its clinical utility as a decision support
tool. The analysis is now focused on support vector machine
learning approaches as described in the preceding sections.
Ongoing Academic Research
A sampling of current academic research projects highlights the
potential for the eRI to significantly contribute to ICU patient-
care improvements. The first is a cross-sectional study evaluat-
ing all ICU patients in the entire eRI database for the year 2008.
With more than 275,000 ICU admissions, this cohort will
provide a detailed viewof ICUdemographics, patient character-
istics, current practices, and clinical outcomes unavailable in
other registries and much larger than alternative databases.
Additionally, this data will serve as a baseline for comparison as
health-care standards and practices evolve. The second example
is a cohort study designed to evaluate the impact of conditions
that develop during the ICU stay but not present on admission
(such as hypotension, hypernatremia, tachycardia, and poor
glycemic control) on risk-adjusted outcomes for length of stay
and mortality. With the level of detail in this patient data and the
large sample size available in the eRI database, we hope to
identify relationships previously difficult to define.
Conclusions and Future Directions
While the historical and organizational characteristics of the
eRI are unique, there are several principles that contributed to
our success. Understanding these can help others successfully
standing up a research programwhether within a single
health system or across several systems such as in the case of a
consortium or Regional Health Information Organization
(RHIO). First, measures must be taken up front to insure
patient privacy, health system confidentiality, and academic
freedom, and to put in place proper governance that will man-
age the commercial interests sufficiently to legitimize the
endeavor. Second, to motivate the participants, there must be
some additional value that is achieved by providing the data.
The prospect of research alone may not be sufficient. Quar-
terly reporting of high-value clinical benchmarks and the
opportunity for additional research were both important in
the case of the eRI. Third, teaming with an ARO provides the
checks and balances that are needed for data providers and
prospective researchers to be satisfied that integrity is main-
tained throughout the project and provides core competencies
that are not likely to exist within the staff of the industry
participant. Finally, it is very important to treat the research
project with the same formality as a software development
project and to make sure that there is strong clinical involve-
ment throughout the project. As tempting as it is to believe the
research process can be automated, there always needs to be
an extensive evaluation by clinicians, researchers, and techni-
cal experts for every research project.
Additional technical challenges must be addressed for
large-scale public/private collaborative aggregation projects
for which the data will inevitably come from more diverse
sources and different data supplier organizations. We believe
that these technologies will be driven by developments such
as increased health-care information exchange, as data
standards take root in the entire health care systemwhether
this is years or decades ahead. Enabling technologies include
distributed deidentification mechanisms, ontological database
methodologies, automated data-collection techniques over a
robust communications infrastructure, and master patient
index technology to provide linkage across deidentified data
sources. The opportunity exists now for industry driven col-
laborative efforts using a framework similar to the eRI to
produce meaningful evidence-based improvements in care
practices and innovations in health-care information technol-
ogy that provides clinicians with data-driven clinical intelli-
gence in the form of CDS.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Michael McShea received his M.B.A.
degree from the Carey Business School of
Johns Hopkins University, an undergradu-
ate B.Sc. degree in electrical engineering
from Bucknell University, and a M.Sc.
degree in systems engineering from the
University of Pennsylvania. He is the pro-
gram manager for the eICU Research Insti-
tute and is director of product development for Philips Patient
Monitoring Informatics, which includes the Philips VISICU
eICU program, and other clinical information system prod-
ucts serving high-acuity health-care environments globally.
Before joining VISICU, which was acquired by Philips
Healthcare in 2008, he held numerous information technology
leadership positions in mission critical industries, including
health care, financial services, telecommunications, aviation,
and aerospace.
Randy Holl received his B.Sc. degree in
physics from the Massachusetts Institute
of Technology and M.Sc. degree in com-
puter science from Rensselaer Polytechnic
Institute. He has led the development of
high technology systems for more than 20
years. His career includes an impressive
track record with noteworthy technology-
driven companies across a wide span of technology sectors.
He has received numerous commendations including the
Chairmans Award for excellence in innovation from Har-
nischfeger Industries. Before joining the VISICU team in
2003, he held senior technical and management positions at
Hyperion Software, IBM Consulting, Harnischfeger Indus-
tries, and Notara. He also served as a qualified submarine offi-
cer in the U.S. Navy.
Omar Badawi earned his B.Sc. degree in
biology from the University of California,
Irvine, followed by a Pharm.D. degree
from the University of the Pacific School
of Pharmacy as a member of the Rho Chi
Honor Society. After earning his Pharm.D.
degree, he completed a residency in phar-
macy practice at the University of Califor-
nia, Irvine Medical Center, followed by a specialty residency
in critical care from the Mayo Clinic in Rochester, Minnesota.
He is also a board-certified pharmacotherapy specialist. He is
currently a clinical product manager with Philips VISICU and
an adjunct assistant professor at the University of Maryland
School of Pharmacy. Before joining Philips VISICU, he was
an assistant professor at the University of Maryland School of
Pharmacy, where he practiced in the cardiac intensive care
unit at the University of Maryland Hospital.
Richard R. Riker is a graduate from the
University of Vermont College of Medi-
cine with additional training and board
certification in internal medicine, emer-
gency medicine, pulmonary and critical
care medicine at Northwestern University
in Chicago and Maine Medical Center in
Portland, Maine. He is also a certified
physician investigator from the Association of Clinical
Research Professionals and has published numerous articles
dealing with care of critically ill patients. In addition to his
clinical work in pulmonary medicine, critical care, and neuro-
critical care, he is the vice chair of the Institutional Review
Board at Maine Medical Center and the associate director of
MaineHealth VitalNetwork in Portland, Maine.
Eric Silfen serves as senior vice president
and chief medical officer for Philips
Healthcare. Currently, he works closely
with academics and researchers in North
America and Europe to focus on health-
care research in the areas of molecular
medicine, computer-aided imaging, clini-
cal systems, and diagnostic evaluations.
He is responsible for crafting and disseminating a medical
consciousness that reflects Philips Healthcares commitment
to personal health and well being. Specifically, he is develop-
ing and managing clinical and health policy medical advice
networks; the clinical trials, comparative effectiveness, and
clinical evaluations of products, services and solutions for the
business groups; the reimbursement policy function and activ-
ities; and the health information management system and
exchange for clinical research and medical knowledge.
Address for Correspondence: Michael McShea, Philips
VISICU, 217 East Redwood Street, Suite 1900, Baltimore,
MD 21202. E-mail: michael.mcshea@philips.com.
References
[1] J. Harrison and A. Aller, Regional and national health are data repositories,
Clin. Lab. Med., vol. 28, no. 1, pp. 101117, 2008.
[2] C. M. Lilly and E. J. Thomas, Tele-ICU: Experience to date, J. Intens.
Care Med., vol. 25, no. 1, pp. 1622, 2010.
[3] E. T. Zawada, P. Herr, D. Larson, R. Fromm, D. Kapaska, and D. Erickson,
Impact of an intensive care unit telemedicine program on a rural health care sys-
tem, Post. Med., vol. 121, no. 3, pp. 160170, 2009.
[4] P. J. Provonost, T. Young, T. Dorman, K. Robinson, and D. C. Angus, Asso-
ciation between ICU physician staffing and outcomes: A systematic review,
Crit. Care Med., vol. 27, no. 12, p. A43, 1999.
[5] M. Bloomrosen and D. Detmer, Advancing the framework: Use of
health dataA report of a working conference of the American Medical
Informatics Association, J. Amer. Med. Inform. Assoc., vol. 15, no. 5,
pp. 715722, 2008.
[6] US Department of Health and Human Services, Standards for privacy of
individually identifiable health information: Final rule, 45 CFR pars 160 and
164 Fed. Reg., vol. 67, no. 157, 2002.
[7] J. A. Lyman, K. Scully, and J. H. Harrison, Jr., The development of health
care data warehouses to support data mining, Clin. Lab. Med., vol. 20, no. 1,
pp. 5571, 2008.
[8] J. Shawe-Taylor and N. Christianni, Kernel Methods for Pattern Analysis, 1st
ed. Cambridge: Cambridge Univ. Press, 2003.
[9] A. Kaben, F. Correa, K. Reinhart, U. Settmacher, J. Gummert, R. Kalff, and
Y. Sakr, Readmission to a surgical intensive care unit: Incidence, outcome and
risk factors, Crit. Care, vol. 12, no. 5, p. R123, 2008.
[10] T. Hanane, M. T. Keegan, E. G. Seferian, O. Gajic, and B. Afessa, The
association between nighttime transfer from the intensive care unit and patient
outcome, Crit. Care Med., vol. 36, no. 8, pp. 22322237, 2008.
[11] O. Gajic, M. Malinchoc, T. B. Comfere, M. R. Harris, A. Achouiti,
M. Yilmaz, M. J. Schultz, R. D. Hubmayr, B. Afessa, and J. C. Farmer, The
stability and workload index for transfer score predicts unplanned intensive care
unit patient readmission: Initial development and validation, Crit. Care Med.,
vol. 36, no. 3, pp. 676682, 2008.
[12] G. S. Cooper, C. A. Sirio, A. J. Rotonodi, L. B. Shepardson, and G. E. Rosenthal,
Are readmissions to the intensive care unit a useful measure of hospital perform-
ance? Med. Care, vol. 37, no. 4, pp. 399408.
[13] C. A. Chrusch, K. P. Olafson, P. M. Mcmillan, D. E. Roberts, and P.
R. Gray, High occupancy increases the risk of early death or readmission after
transfer from intensive care, Crit. Care Med., vol. 37, no. 10, pp. 27532758,
2009.
[14] O. Badawi, M. Breslow, P. Kolm, et al., Predicting death or readmission
within 48 hours of ICU discharge, Crit. Care Med., vol. 37 (12 Suppl.),
p. A287, 2009.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
___________________
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Nerve Conduction
Studies
N
erve conduction studies (NCSs) have played an
important role in the evaluation of neuromuscular
disease for the past 50 years. When patients present
with complaints of pain, numbness, tingling, or
weakness, NCS is often one of the earliest tests obtained by
physicians, because it enables the quantitative assessment of
peripheral nerve and muscle function [1] and, therefore, aid the
physician in identifying the physiological source of the
patients symptoms. NCSs are commonly employed in the
assessment of entrapment neuropathies such as carpal tunnel
syndrome and cubital tunnel syndrome, cervical and lumbar
radiculopathies caused by herniated disks, and polyneuropa-
thies such as those caused by diabetes. These are also used to
track disease progression over time and monitor nerve function
in situations where the possibility for nerve injury exists, such
as during the administration of neurotoxic drugs. Although
NCSs have been in clinical use since the 1950s, their utility has
been constrained by factors such as equipment designs that
limit instrument operation to subspecialist physicians, disagree-
ment among subject-matter experts about optimum study tech-
niques, and the subjective nature of NCS data interpretation.
So what, exactly, is an NCS? In essence, NCSs involve the
delivery of electric stimuli to peripheral nerves at accessible
locations on the human body and the recording of electrophysi-
ological responses. Responses are generally categorized as
motor or sensory. The former always refers to a study in which
the response is recorded from muscle. The short-latency-
evoked motor response is referred to as a compound muscle
action potential (CMAP). Longer latency motor responses such
as F-waves, A-waves, and the H-reflex (collectively termed
late responses) may be identified when appropriate stimulation
and recording techniques are employed. While some sensory
responses are the result of stimulation of a pure sensory nerve
(e.g., the sural nerve in the lower leg), other sensory responses
(such as those recorded from the fingers following ulnar,
median, or superficial radial nerve stimulation) are recorded
following stimulation of a mixed population of sensory and
motor axons. The responses in either case are commonly
referred to as sensory-nerve action potential (SNAP). For natu-
rally generated nerve impulses, the conduction pathway is
strictly one way: motor impulses originate from anterior horn
cells in the spinal cord and travel toward muscles in the periph-
ery; sensory impulses, on the other hand, originate at sensory
receptors in the periphery and travel toward the spinal cord.
Fortunately (for electrophysiologists), nerve-action potentials
evoked with electrical stimulation propagate bidirectionally
away from the site of stimulation. This behavior can be
exploited to investigate physiological functioning of proximal
nerve segments, as will be discussed later.
Peripheral nerve disease affects the recorded signal in
predictable ways that usually permit diagnostic classification.
However, the specific effect depends on whether the nerve axon
itself or its enveloping myelin sheath is the locus of pathology.
In demyelinating neuropathies, diffuse or focal damage to the
myelin sheath occurs. In the setting of diffuse injury, nerve con-
duction velocities fall, and the SNAP or CMAP become longer
in duration and smaller in amplitude. These changes in wave-
form morphology are largely the result of differential slowing
of axon conduction velocities among the collection of axons
that jointly contribute to the CMAP or SNAP recorded at the
detection electrodes. The temporal broadening of the waveform
is usually referred to as temporal dispersion. When demyelin-
ation is severe, the axon itself may become damaged. This is
referred to as secondary axonal injury and typically results in
further waveformamplitude reduction due to loss of signal.
Neuropathies that target the axon rather than the myelin
sheath represent a separate category of nerve pathology and most
commonly present with paresthesia and anesthesia of the feet
and distal leg, sometimes associated with weakness of foot
muscles or ankle dorsiflexors. They are often referred to as axo-
nal polyneuropathies, underscoring the involvement of multiple
nerves. A majority of them are associated with progressive
degeneration of the most distal portion of the axon. NCS changes
typically seen in this group of neuropathies are SNAP and
CMAP amplitude reduction caused by inability of degenerating
axons to convey an action potential to the detection electrodes.
The first clinically useful NCS instrument was engineered
through collaborations between Jasper and Johnston and Gol-
seth and Fizzell in the 1940s [2], [3]. The instrument consisted
of an amplifier, an oscilloscope, a loudspeaker, and a camera
(for capturing traces displayed). Subsequent NCS instruments
combine essentially the same elements: an electrical stimulator,
BY XUAN KONG, EUGENE A. LESSER,
AND SHAI N. GOZANI
EYEWIRE
Clinical Challenges and Engineering Solutions
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
an input terminal to convey bioelectric signals from either a
nerve or muscle to the instrument, a differential amplifier, a dis-
play screen, and a recording device.
In the 1980s, NCS instruments were integrated with desktop
PCs, thereby enabling digitization of the motor and sensory
responses, as well as their manipulation and storage. The
majority of NCS devices are physically large and complex to
operate, placing constraints on both the operator and environ-
ment in which NCS can be performed. Most instruments
employ computer algorithms to assign waveform cursors, per-
form elementary computations, and generate reports. Because
of their complexity and expense, as well as the need for special
expertise to interpret data, performing NCS historically has
been restricted to physiatrists and neurologists with a special
interest in electrodiagnostic testing and diseases affecting the
peripheral nervous system. Access to these specialists and thus
to these procedures may be limited because of physician man-
power shortages combined with geographic clustering of spe-
cialist physicians in urban areas. Consequently, wait times for
patients referred for these procedures may be long.
Improving access to care while advancing the state of the art
of NCS motivated our efforts to refine nerve conduction instru-
mentation and measurements [4]. This work originated at the
Massachusetts Institute of Technology (MIT) in the mid-1990s
and then migrated to a commercial enterprise later in that
decade. Historically, NCS equipment was primarily focused on
signal-acquisition hardware, and relatively little attention was
given to other aspects of the procedure. Our approach differed
from the past and current technology in that we considered the
entire diagnostic chain from electrodes, to instrumentation, to
computer support for report generation in a coherent and inte-
grated fashion. Furthermore, we applied modern disposable
electrode-construction techniques, signal processing, and com-
puter-based decision support to the challenge of both facilitat-
ing and improving nerve conduction measurements.
In this article, we first review how NCS is traditionally per-
formed. We then examine the technical challenges associated
with each step of performing an NCS and describe howengineer-
ing solutions could be realized to meet these challenges. The engi-
neering goals were several: improvement in NCS workflow, use
of prefabricated electrode arrays to standardize NCS technique
and reduce the errors associated with electrode placement, and
improvement of the overall accuracy and reliability of NCS.
NCS Background
Major steps for performing an NCS are summarized in the flow-
chart shown in Figure 1. Most NCS instruments offer testers the
ability to perform NCS manually. We will refer to these studies
as traditional and use median nerve testing to illustrate the work-
flow of a traditionally performed NCS as outlined in the left
column of the flowchart. The median nerve is one of the five main
nerves originating fromthe brachial plexus in the upper limb. It is
the only nerve passing through the carpal tunnel at the wrist. Its
compression at the carpal tunnel causes carpal tunnel syndrome.
To perform a median-motor NCS, an active surface-record-
ing electrode is affixed to the skin overlying the midpoint of
the abductor pollicis brevis (APB) muscle in the fleshy part of
the thumb (see Figure 2). The position of the active electrode
corresponds to the motor pointthe nervemuscle interface
where the majority of neuromuscular junctions are concen-
trated. An inactive electrode is placed distally on the thumb. A
reference electrode is affixed to either the palm or back of the
Nerve
Conduction
Study (NCS)
Recent
Engineering
Innovations in NCS
Individual
Electrode
Placement
Integrated
Electrode
Array
Acquisition Setup
(Gain, Stimulus,
Temperature, etc.)
Computer-
Assisted
Acquisition Setup
Manual
Stimulator
Control
Dynamic
Test
Control
Waveform
Display
and Analysis
Advanced Signal-
Processing
Algorithms
NCS
Parameter
Determination
Enhanced
NCS Parameter
Determination
Manual NR
Lookup and
Comparison
Rigorous
Statistical
Framework
NCS
Outcome
Expanded
NR
Database
Limited
Normal
Range (NR)
Patient Management
Fig. 1. Flowchart of an NCS test procedural steps and recent
engineering innovations that enhance the efficiency, accu-
racy, and reliability of NCS measurements.
Arigorous statistical framework has been
proposed and implemented for deriving NCS
parameter reference ranges as a part of our
development effort to refine NCS measurement.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
hand. Stimulator electrodes (cathode and anode) are posi-
tioned against the volar wrist in the midline, directly over the
median nerve, with the cathode closest to the active recording
electrode. The cathode is placed at a measured distance from
the active electrode or is located in relationship to an anatomi-
cal landmark such as the proximal wrist crease.
Once the electrodes are properly placed and connected to
the NCS instrument, the tester needs to adjust the gain and
sweep speed so that waveforms to be acquired can be properly
displayed and recorded. The tester uses the displayed wave-
forms to visually assess data quality and interference levels.
Most NCS waveforms are acquired and analyzed when all or
most of the axons in the nerve lying beneath the cathode are
activated (commonly referred to as supramaximal stimula-
tion). To achieve supramaximal stimulation, short-duration
electrical currents of increasing intensity are delivered to the
nerve while the tester monitors the size and morphology of the
motor response. As the stimulation current is increased,
greater numbers of axons lying beneath the cathode are depo-
larized (i.e., generate action potentials), resulting in larger
portions of the muscle being depolarized. Muscle depolariza-
tion, occurring in concert with muscle contraction, generates
the CMAP (see Figure 3). The growing CMAP is observed on
the display screen. When it becomes visually apparent to the
tester that the CMAP has stopped increasing in size, the
response is deemed maximal, and a final stimulus, usually
25% larger than the last, is delivered to ensure that all axons
are depolarized. Stimulation is then stopped and the waveform
saved for analysis. The tester then moves the stimulator to the
next median nerve motor-stimulation site, typically at the
elbow crease (see Figure 2, elbow-stimulation site), and the
process is repeated. The distance between the two stimulation
sites is measured and input into the NCS instrument so that a
segmental conduction velocity can be calculated. This con-
cludes the motor study, as traditionally performed.
F-waves are late motor responses that are recordable fromthe
same detection electrodes approximately 2030 ms following
the CMAP. Proper F-wave display for manual analysis requires
different gain and time-scale settings than that for CMAP since
F-waves are typically less than one tenth the size of CMAP (see
insert in Figure 3). When the nerve action potential invades the
axon hillock of the motor neurons, small groups of spinal motor
neurons discharge, generating new axon potentials. These prop-
agate distally and are recorded from APB as F-waves. Because
median nerve F-wave latencies reflect the time needed for the
peripheral stimulus to ascend through the arm and neck, pro-
voke a neuronal discharge, and have that discharge propagate
from the spinal cord to the APB muscle, they occur at long
latencies relative to the CMAP. Consequently, F-waves are
classified as late responses. While the time scale can be manipu-
lated such that both the CMAP and F-wave appear within the
display window, doing so results in horizontal compression of
the CMAP, which means that its waveform characteristics can-
not be accurately assessed. Therefore, F-wave acquisition and
analysis is traditionally performed as a separate study using
much higher gain and larger time-scale settings compared with
those used for CMAP. To obtain F-wave parameters, variable
Median Nerve
Elbow
Stimulation
Wrist
Stimulation
Carpal
Tunnel Syndrome
Reference
Electrode
Motor
Recording
Site
Sensory
Recording
Site
Fig. 2. Electrode placements for a typical median NCS. The
black disk is the active recording electrode, the red disk is the
inactive recording electrode, and the green disk is the refer-
ence electrode. Ring electrodes (black active, red inactive)
record the SNAP.The stimulation cathode (black) is interposed
between the anode and the active recording electrode.
Stimulation
Nerve Fibers
Spinal Neuron
Distal Muscle
CMAP
F-Wave
Fig. 3. An illustration of CMAP and F-wave recorded from a
muscle group distal to the nerve-stimulation site. The CMAP
is produced by the motor-nerve action potential propagat-
ing away from the stimulation site directly toward the
muscle recording site (purple segment). The F-wave is gen-
erated by the motor-nerve action potential propagating
toward the spinal cord (green segment).
Use of prefabricated electrode arrays designed
to conformto the standard paradigms would
simplify the electrode placement process and
ensure consistent NCS results between testers.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
numbers of F-waves need to be recorded, ranging fromless than
five to more than 20, according to the testers preference and
need [5]. Most NCS instruments offer computer-assisted cursor
placement functions for F-wave analysis.
At the conclusion of F-wave acquisition, a median nerve
sensory study is performed. The active and inactive electrodes
used for CMAP acquisition are removed and replaced with
ring electrodes positioned on the index or middle finger (see
Figure 2). The stimulator is again applied to the wrist, and
graded electrical stimuli are again delivered while the operator
monitors the development of a sensory response (SNAP).
Once SNAP amplitude appears stable, averaging is often
employed to improve the signal-to-noise ratio before SNAP
parameters are extracted fromthe averaged waveform.
Testing of a single nervethe median nerve in this caseis
now complete. If testing of additional nerves is required, as is
usually the case, the tester will iteratively repeat the process
described earlier until data collection is deemed complete. Fol-
lowing the acquisition of all motor and sensory responses, wave-
form parameters are analyzed by the physician interpreting the
study. The NCS data are usually compared with reference-range
data derived from a healthy population [6], although the exact
way in which this is performed is at the discretion of the inter-
preting physician. She may develop the reference range in her
laboratory or adopt values fromother laboratories. The compari-
son generally entails a straightforward determination of whether
a measured NCS parameter is above or below a normal limit.
This step can be performed by most modern NCS instruments as
long as the reference-range data has been input into the device.
Engineering Opportunities
Significant opportunities exist for improvement in NCS work-
flow as well as error reduction in the traditional test scenario
described earlier (see Figure 1). From a quality-control stand-
point, manual placements of individual stimulation and detection
electrodes may differ between testers or between multiple NCS
tests by the same tester, affecting both the reliability and reprodu-
cibility of results [7]. Connections fromthe instrument to individ-
ual electrodes are also error prone. A prefabricated integrated
electrode array with a keyed connector to the NCS instrument is
effective in preventing such errors and maintains consistent inter-
electrode distances to achieve maximum reliability and reprodu-
cibility. Simple measurement error involving the distance
between stimulation sites (motor studies) or between stimulation
and recording sites (sensory studies) will result in calculated
nerve conduction velocities that vary significantly, especially
over short nerve segments, with the potential for either obscuring
pathology or misclassifying normal nerves. Again, an integrated
electrode array with preconfigured electrode locations can
address this issue. Since traditionally performed NCSs are carried
out sequentially, a patient receives more electric stimulation than
would be required if motor, F-wave, and sensory studies could be
performed concurrently. Through multichannel recordings and
real-time waveformcursor assignment, total stimulation count as
well as test time could be significantly reduced. Failure of the
tester to achieve supramaximal stimulation during motor or
sensory NCS will result in inaccuracies of the measured wave-
form parameters. Real-time waveform monitoring and closed-
loop stimulator control can mitigate such errors. Finally, using
incorrect filter setting and the misconstruing of artifacts for the
waveform of interest are all potential problems that can benefit
from dynamic gain control, modern biosignal analysis, and pat-
tern-match algorithms.
Figure 4 illustrates waveforms commonly encountered dur-
ing an NCS. Under supramaximal stimulation, a CMAP is
expected to maintain its morphology and waveform features.
Commonly measured CMAP parameters include onset latency
in milliseconds, amplitude in millivolt, duration in millisec-
onds, and area (shaded region) in millisecond by millivolt. The
SNAP is generally small in amplitude and measured in micro-
volt; its recording places greater demands on the data-acquisi-
tion circuitry because it often has far lower signal-to-noise
ratios than are encountered in motor recordings. Unlike CMAP
parameters, SNAP parameters are often extracted froman aver-
aged waveformto improve signal-to-noise ratio. The number of
SNAPs required for adequate averaging varies depending on
the relative strength of the SNAP signal to the background
noise. Peak latency (ms) and peak-to-peak amplitude (lV) are
typically reported. However, with adequate signal-to-noise
ratio, SNAP onset latencies can easily be determined.
Late responses (Figure 4) consist of several waveform compo-
nents of clinical interest: A-waves [8] and F-waves [5], [9]. Late
responses often have a low-frequency baseline common to all
traces. The A-wave is a waveform with constant morphology that
occurs at the same time in a large percentage (40% or more) of
late response traces. In contrast, the F-wave is a waveformcompo-
nent with varying morphology from trace to trace. The frequency
at which F-wave components are present in late response traces is
referred to as F-wave persistence and can vary between 0 and
100%. Other commonly studied F-wave latency parameters
include minimum, mean, median, and maximum F-wave latency.
Latency parameter calculation is based on individual F-wave
latency, as marked by the v cursors. It is worth noting that more
than one F-wave component may appear in each trace. Whether a
waveform component belongs to an A-wave or F-wave can only
be determined by examining the entire set of late-response traces.
Engineering Solutions to NCS Problems
In this section, we describe in detail the NCS process aligned with
each procedural step as outlined in the flowchart in Figure 1. The
clinical relevance of technical innovations and engineering solu-
tions will then be discussed.
F-waves vary in their latency and morphology
in successive late wave traces even
though they are acquired under
identical stimulation conditions.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Electrodes and Electrode Placement
The first critical technical step in an NCS is to properly affix
recording electrodes to the skin. Since standard recording (i.e.,
electrode placement) paradigms have been used for years for
commonly tested nerves, use of prefabricated electrode arrays
designed to conform to the standard paradigms would simplify
the electrode placement process and ensure consistent NCS
results between testers. The electrode array designs are nerve spe-
cific, as is the case with traditional electrode placement. The
nerve-specific electrode (NSE) arrays incorporate stimulation,
detection, and reference electrodes; a temperature-monitoring IC
chip encoded with a unique identifier for electronic medical
record applications; and printed circuits affixed to a mylar
backing and adhesive coating. There are several advantages of a
prefabricated electrode array. First, the process of positioning
electrodes becomes standardized and, therefore, less error-prone
since the NSE array can now be positioned using standard ana-
tomic landmarks. This in turn increases the likelihood that the
active detection electrode will be correctly positioned over the
motor point. Second, since the distance between the cathode and
active electrode may be fixed in a preconfigured electrode array,
latency variations among patients or changes in measured latency
in a single patient over time can be confidently ascribed to patient
nerve physiology rather than to inconsistencies in electrode place-
ment. Since the diagnosis of median-nerve entrapment at the wrist
requires demonstrating that motor or sensory latencies are
CMAP Parameters Late-Wave Parameter
A-Waves F-Waves
Trace with F-Wave Only
Trace with A-Wave Only
Trace with Baseline Only
Traces with A-Wave and F-Wave
F-Wave Persistence: 4/6 = 0.67
SNAP Parameter
Peak Latency
P
e
a
k
-
t
o
-
P
e
a
k
A
m
p
l
i
t
u
d
e
Duration
Area
1 ms
1 ms
5 ms
Stimulus
Stimulus
Onset
Latency
Onset
Latency
2

m
V
2

V
0
.
2

m
V
A
m
p
l
i
t
u
d
e
(a)
(b) (c)
Fig. 4. Examples of NCS waveforms. (a) The CMAP is recorded using an active electrode affixed to the skin over a muscle
innervated by the nerve under study. (b) SNAP is a differential recording of the summated nerve-action potentials generated
by axons of sensory or mixed nerves. The SNAP waveform shown is an average of multiple raw SNAPs due to low signal-to-
noise ratio of individual SNAP waveform. (c) A- and F-waves are recorded from the same detection electrodes as CMAP but
at a later time window. F-waves occur when motor neurons fire in response to inbound nerve-action potentials generated in
the periphery. A-waves result from cross-talk between axons: an inbound action potential depolarizes an adjoining axon, pro-
ducing an outbound action potential. The v on late-wave panel marks individual F-wave latency.
Whether a waveformcomponent belongs to an
A-wave or F-wave can only be determined by
examining the entire set of late-response traces.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
prolonged relative to a normal population, a small error in the
measured latency may result in misclassification of nerve function
as normal or abnormal. Third, the spatial positions of active and
inactive detection electrodes influence the morphology of the dif-
ferentially recorded response, which in turn affects the measured
amplitude and duration of the CMAP or SNAP. Using a precon-
figured electrode array eliminates variability in these parameters
caused by inconsistent electrode placement. Further, using a
keyed connector on the NSEprevents incorrect connections to the
terminals of the NCS instrument. Figure 5 provides an example of
the preconfigured NSEfor median-nerve testing.
Since NSEs are designed to work on at least 95% of the adult
population, anthropometric variation was considered in determin-
ing the size and spacing of electrodes. Our initial approach was to
design NSEof different sizes, the choice of which would be based
on the gender and weight of the test subject. Physician feedback
on the difficulty of managing electrode inventories prompted us
to examine alternative design principles. A new set of single-size
NSEwas designed based on statistical analyses of anthropometric
databases and detailed studies of experimental data acquired from
a range of subjects with different anthropometric characteristics
and hand sizes. Rather than using a small detection electrode
placed directly over the motor point that would yield higher
amplitude signals, we used a larger size electrode to ensure that it
would overlie the motor point for subjects with a wide range of
hand sizes. For certain motor nerves, such as the deep branch of
the peroneal, with a motor point less readily identified based on
anatomic landmarks, we developed an interleaved detection
electrode array. Instead of requiring a tester to place the active
electrode over the motor point, the array concept only requires
the tester to place the electrode array near the motor point of the
extensor digitorum brevis (EDB) muscle. A pattern recognition
algorithm is tuned to recognize which of the collected CMAPs
represents the on-motor-point recording of EDB, and the instru-
ment then uses this CMAP for subsequent analysis.
Data-Acquisition Setup
Time Base and Gain Settings
Different NCS responses require different data-acquisition setups
depending on the waveform features of clinical interest. SNAPs
generally occur within 10 ms of stimulation and have amplitudes
less than 50100 lV, measured peak to peak. CMAPs are of simi-
lar time scale but can be 100 times larger in amplitude. Late
waves can appear as early as 20 ms in the upper extremity and as
late as 100 ms in the lower extremity [1], [9]. Because CMAPs
and late waves are recorded from the same detection electrodes
but with up to a 100-fold difference in dynamic ranges, different
gain-setting strategies have been implemented in NCS instru-
ments. The simplest approach is to set the acquisition gain to be
optimal for either CMAP or late-wave type. The disadvantage of
this approach is that only one waveformis usable with each stimu-
lation. In designing our NCS instruments, we first adopted an
approach to set the initial gain for proper CMAP acquisition and
then digitally switch to late-wave gain once the region of interest
for CMAP is passed. The advantage of this approach is that both
CMAPs and late waves are available for analysis for each stimula-
tion. The disadvantage is that a gain-switching artifact may be
introduced in the early segment of late-wave traces. More
recently, we have implemented a solution to allowuniformampli-
fication gain for both CMAP and late wave in conjunction with a
high-resolution (24 b) analog-to-digital converter circuit. This
arrangement allows adequate precision for analysis of low-ampli-
tude late waves while avoiding saturation of larger amplitude
CMAP. To improve the efficiency of waveformstorage and trans-
mission, the least significant bits of CMAP may be discarded
without losing any clinically meaningful information.
Temperature
Temperature has been shown to affect NCS parameters in
healthy subjects [10]. Lowering the temperature of a limb de-
creases motor and sensory conduction velocities [11] and
increases SNAP and CMAP amplitude [12]. One way to address
temperature variations is to warm or cool a limb if its surface
temperature falls outside a target range. However, this process is
time-consuming and may disrupt workflow. An alternative to
changing the temperature of the limb is to mathematically com-
pensate for the temperature effect on nerve conduction relative to
a standard temperature, since the relationship between tempera-
ture and NCS parameters, at least in healthy subjects, has been
Built-In
Temperature Sensor
Motor Recording Site
Sensory Recording Site
Fig. 5. A preconfigured NSE array for median nerve testing.
Locations of the prefabricated electrodes are shown, and
actual electrodes are on the underside of the electrode array.
Computer-assisted NCS is an enabling
innovation that took an existing diagnostic
technology, streamlined its operation, and
expanded its user base beyond the reach
of prior equipment.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
established [10], [13]. NSEs have a built-in temperature sensor
that records the skin surface temperature (see Figure 5). Compat-
ible NCS instruments can read the temperature and normalize the
measured NCS parameters.
Test Control
Once data acquisition is properly set up, the number of stimuli to
be delivered to a test subject depends on the minimum number
of response waveforms required to yield reliable NCS parame-
ters. Real-time test-control algorithms have been developed for
three distinct types of waveforms: CMAP, SNAP, and F-wave.
CMAP
The test control for CMAP acquisition focuses on the stability
of the CMAP parameters to ensure the quality and reliability of
the data. For example, the CMAP latency parameter is extracted
from each acquired CMAP waveform, and it is expected that
the latency values remain consistent from stimulus to stimulus
(e.g., variations limited to within one sample interval). If the
latency consistency cannot be achieved within a reasonable
number of stimuli, the test control will issue a warning, indicat-
ing that data-acquisition setup may be deficient. This feature
ensures that only the highest quality data are used to determine
NCS parameters by waveform-analysis algorithms.
SNAP
SNAP NCS parameters are often calculated based on an average
of multiple SNAPs because of the relatively low signal-to-noise
ratio of individual SNAPs. Motion artifact and measurement
noise are two common sources of contamination that degrade
individual SNAP quality. The primary objective of SNAP acqui-
sition test control is to collect sufficient number of responses such
that a stable average response waveform can be obtained. After a
minimumnumber of rawSNAPs are acquired, a tentative average
is constructed. A similarity measure between individual SNAP
and the tentative average is calculated. The similarity measure
can be in the form of correlation, normalized mean squared error,
or similar metrics. If the value of the similarity measure falls
below a threshold, the corresponding SNAP is tagged as having
excessive contamination and removed from future waveform
averaging. With each additional rawSNAP acquired, the tentative
average is updated, and the similarity measure is recalculated for
all SNAPs contributing to the tentative average. SNAP acquisi-
tion is complete when signal-to-noise ratio of the tentative aver-
age exceeds a threshold. The noise level is estimated based on the
differences between the individual SNAP and the average.
F-Wave
F- and A-waves are two distinct waveform components that can
be seen in late motor-response traces (see Figure 4). Both waves
are of clinical importance. Late-wave response may vary from
stimulus to stimulus because of the probabilistic nature of the
F-wave generation mechanism [9]. F-wave latency is one of the
most important NCS parameters for assessing proximal-nerve-
conduction properties. Individual late-wave responses may not
contain a measurable F-wave component. Latency statistics such
as mean F-wave latency are only reliable when a sufficient num-
ber of individual F-wave latencies are available. Late-wave
acquisition test control is designed to strike a balance between
minimizing the total number of stimuli and maximizing the accu-
racy and reliability of F-wave parameters. To illustrate, let us
consider a case where the F-wave component is not measurable.
Prior clinical research has shown that certain diseases suppress
generation of the F-wave [14]. Our initial test-control design was
rather straightforward: we stimulated the nerve the maximum
allowable number of times (say 40) and reported mean F-wave
latency if available. However, such a design led to NCS with the
maximum number of stimuli, but no reportable F-wave latency
results for certain patients. The F-wave test-control design was
revised to reduce the total number of electric stimuli delivered to
the patients nerve while maintaining the reliability and accuracy
of F-wave results.
The problem statement for the test-control redesign was, If
the first X (to be determined) late-wave traces do not contain a
single F-wave, can the late-wave acquisition be stopped, and can
mean F-wave latency be reported with high confidence as not
available? Given an observation that the first 14 late-wave
traces have no F-wave, we can reject the null hypothesis
(P = 0:05) that the probability for each trace to have a measura-
ble F-wave is not less than 0.19. At the maximum probability of
0.19, the expected number of measurable F-waves would be
slightly lower than five if we were to continue to stimulate an
additional 26 times, to obtain a total of 40 late-wave traces. If
reporting mean F-wave latency requires five or more individual
F-wave latencies, then we can conclude with high confidence
(Type I error < 0.05) that insufficient F-wave latencies will be
collected to produce a valid mean F-wave latency, even when the
maximum number of stimuli is applied. The proposed test-con-
trol scheme was prospectively validated with NCS data from764
nerves. Forty late-wave traces were acquired for each of these
764 nerves, and none of them had any F-wave in the first 14 late-
wave traces. The percentage of the nerves with five or more
F-waves out of 40 traces was only 0.005 (much smaller than the
targeted Type I error). The aforementioned late-wave acquisition
test control has been implemented in our NCS instruments,
thereby minimizing the number of stimuli delivered to the nerve.
Waveform Cursor Assignment and Analysis
Waveform cursors are placed on NCS waveforms to define fea-
tures of the NCS responses. For example, the maximum and
minimum of the SNAP are marked, and their magnitude differ-
ence defines the peak-to-peak amplitude (see Figure 4). Com-
puter-assisted cursor assignment can be found in almost all
modern NCS instruments. However, many were not designed to
handle the complexities of stimulus artifact, baseline variation,
noise, and complex waveform morphology. Since NCS instru-
ment manufacturers rarely publish the methods or the validation
results of their proprietary waveform analysis algorithms, it is
very difficult to conduct any meaningful comparison study.
Using late-wave analysis as an example, we highlight some key
engineering challenges related to waveform analysis and dis-
cuss the signal-processing solutions that have been imple-
mented in our instruments to address these challenges [15].
Two distinct waveform components in late motor responses
are of clinical importance: F- and A-waves (see Figure 4).
A-waves are commonly identified during acquisition of F-waves,
especially in the setting of nerve pathology. An A-wave is defined
as a waveform of constant latency and morphology that appears
in a significant percentage of the late-wave traces (i.e., approxi-
mately 40% or more). Unlike CMAPs or SNAPs, the F-waves
vary in their latency and morphology in successive late-wave
traces, even though they are acquired under identical stimulation
conditions. Individual F-wave latency refers to the onset of initial
deflection from baseline attributed to F-wave activity in each
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
trace. The highly variable morphology and low amplitude of
F-waves make latency assignment difficult because of complicat-
ing factors such as noise, power-line frequency interference (PFI),
and baseline disturbance. An algorithmis generally prone to error
if it searches an entire late-wave trace and assigns F-wave latency
to the first instance at which a deflection frombaseline is evident.
The algorithm may erroneously identify A-waves, noise, PFI,
and/or baseline disturbances as F-wave activity.
Any algorithm that could narrow the search for F-wave
latency in a trace to a limited time segment would significantly
improve the accuracy of F-wave latency cursor placement.
Diseases may create physiological conditions favorable to gen-
erating more than one F-wave in a late-response trace (see an
example of two F-wave regions in Figure 6). Searching for the
initial onset would identify only the first group of F-waves while
ignoring the second. However, the latency information from the
second group carries more clinically significant information, as
it indicates that certain axons contributing to F-wave generation
are conducting action potentials extremely slowly relative to the
axons producing the earlier F-waves. Correct identification of
the appropriate F-wave segments is therefore very important.
Late-Wave Segmentation
An F-wave segment is a temporal window across all late-wave
traces that mark the presence of a group of F-waves. Two
F-wave segments are evident in the sample late-wave traces
shown in Figure 6. To identify F-wave segments and differenti-
ate them from A-wave segments, two key feature spaces are cre-
ated: mean feature (MF) and differential feature (DF). The MF
captures constancy across the ensemble of traces at each time
sample. The MF illustrated in Figure 6 is based on arithmetic
averaging of samples across all traces. Samples corresponding to
A-waves will have high MF, because A-waves maintain latency
and morphology across different traces. On the other hand, sam-
ples related to F-waves will have lowMF, because F-waves from
different traces have little in common. The DF represents varia-
bility across the ensemble of traces at each time instance. The
DF shown in Figure 6 is based on standard deviations of the trace
values. The DF is expected to behave in a way opposite from the
MF: regions with A-wave will have low DF, because A-waves
share the same latency and morphology across different traces.
However, high-DF regions imply the presence of F-waves since
F-wave activities are not synchronized or repeatable. Often both
MF and DF need to be smoothed to create regions with clinically
reasonably duration (see heavier lines for MF and DF smoothing
results). Morphological filtering operations such as openings
and closings are used for such purposes [16]. The filter length is
selected so that two high-activity regions (or two low-activity
regions) will be merged into one if they are separated by a low-
activity region (or a high-activity region) with duration too short
to be clinically meaningful. Based on smoothed MF and DF, the
ensemble of late-wave traces is classified into A-wave (high
MF, lowDF), F-wave (lowMF, high DF), and noise regions.
Although the aforementioned method works very well for a
majority of the late-wave sets, it presents some ambiguity for seg-
ments where A-waves appear in late-wave traces with frequency
lower than 100%. Modifications were made to improve the
segmentation performance. When A-waves do not appear with
100% frequency, the DF becomes high for the A-wave region as
well since DF is defined by the differences between A-wave and
flat line. To differentiate A-wave from F-wave within a high-DF
region, waveform segments in the high-DF region are clustered
into groups based on a similarity metric such as correlation coef-
ficients. If a segment of 20 late-wave traces has only two clusters,
it is more likely to belong to an A-wave region, with one cluster
being A-waves and the other being flat lines. Therefore, a seg-
ment with high DF but a minimumnumber of clusters would also
be classified as an A-wave segment, to accommodate A-wave
with lower than 100%frequency.
F-wave Cursor Assignment
Once the F-wave activity regions are identified, individual trace
F-wave latency can be determined by assigning a cursor to the
earliest instance of deflection from the baseline for each late-
wave trace within the F-wave segment. An F-wave latency
estimation algorithm was developed, and its performance was
compared against manual latency assignments by an expert
neurologist [17]. The correlation coefficient between computer-
ized and manual assignments was 0.972 for individual F-wave
latencies and 0.995 for median F-wave latencies (see Figure 7
for scatter plots). Most published studies have used minimum
F-wave latency as the primary parameter (or only parameter
reported) for an ensemble of F-waves; nonetheless, numerous
studies have demonstrated that other F-wave parameters have
higher diagnostic accuracy and sensitivity [18][20]. Parame-
ters such as the mean and maximum F-wave latencies also have
the advantage of increased statistical robustness over the mini-
mal latency parameter [19]. Although the use of the minimum
F-wave latency persists in clinical practice, advances in
computer-assisted F-wave latency measurements such as those
described earlier should encourage broader adoption of more
robust F-wave latency parameters.
Normal-Range Comparison
Physicians use NCS data to support or refute clinical hypotheses
concerning neuromuscular diseases. The clinical interpretation
of NCS data relies on comparing an individual patients
High MF,
Low DF
A-Wave
Segment
Mean Feature (MF)
Differential
Feature (DF)
Low MF,
High DF
F-Wave
Segment
Low MF,
High DF
F-Wave
Segment
Fig. 6. Late-wave segmentation. An F-wave segment is a
temporal window across all late-wave traces where a group of
F-waves is identified. Two distinct F-wave segments are present
in this example, and they are identified through low MF activity
and high DF activity regions. In contrast, the early A-wave seg-
ment is characterized by high MF and low DF region.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
measurements with a reference range obtained from a healthy
population. NCS reference ranges determine a normal limit
above or below which a given NCS parameter is considered
abnormal. Traditionally, NCS reference ranges are developed in
individual laboratories, as each specialist has his own preferen-
ces in data-acquisition setups, such as filter settings and
electrode placements that impact the NCS parameter values. As
a result, some reference-range results are published based on
fewer than 40 subjects. Interlaboratory variability makes it diffi-
cult or impractical to compare NCS results with those from dif-
ferent test laboratories. Concerns about potential dissimilarity of
patient population characteristics also hinder adoption of refer-
ence ranges developed by different laboratories.
Some clinicians use a single normal limit threshold for all
patients without considering the influence of demographic
and physiologic covariates [1]. Although simple, such an
approach degrades both the sensitivity and specificity of NCS
testing. For example, specifying that an F-wave latency
greater than a stated threshold is abnormal may be useful in
the diagnosis of lumbosacral radiculopathy. However, F-wave
latency is a function of the length of the nerve segment
between the stimulation site and the spinal cord. Normal val-
ues vary with limb length. Failure to adjust F-wave normal
values for height likely will result in decreased sensitivity for
nerve pathology in short individuals and may lead to high
false-positive rates in tall individuals (decreased specificity).
Since nerve conduction velocity slows with normal aging
[21], and since both demyelinative and axonal neuropathies
produce F-wave latency prolongation owing to lower conduc-
tion velocities or loss of fastest conducting axons, age must be
taken into account when setting up reference ranges to avoid
misclassifying healthy individuals as diseased.
A rigorous statistical framework has been proposed and
implemented for deriving NCS parameter reference ranges as a
part of our development effort to refine NCS measurement
[13]. The statistical framework addresses the need to create
reference ranges for a large number of NCS parameters in an
objective and computationally efficient manner (see Figure 8).
The statistical framework has been applied to all NCS parame-
ters of a large reference-range data collection (1791,549
nerves depending upon the nerve type) pooled from five dis-
tinct IRB-approved studies. All studies used the same NCS
instrument (NC-stat) and preconfigured NSEs for data consis-
tency. Each NCS parameter is transformed by an appropriate
function since many NCS parameters are known to have non-
Gaussian distributions in their native domains (S1 in Figure 8).
For example, the SNAP amplitude histogramin a sample popu-
lation typically has an asymmetric tail extending toward larger
amplitudes (positive skew). Nonlinear mappings such as
square root and logarithmmay correct for this positive skew so
that the transformed measurements exhibit a Gaussian distribu-
tion. Bootstrap techniques are used to identify demographic
and physiologic covariates (S2 in Figure 8) that affect each
NCS parameter by creating multiple data sets through sam-
pling with replacement from the original data set [22]. Each
bootstrap data set simulates a different population sample (i.e.,
experiment), and the variabil-
ity of relevant covariate mem-
bership is observed in these
data sets. The observed varia-
bility explains wide variations
of covariate membership for
NCS parameters reported in
prior studies. To maintain the
highest possible consistency,
we retain only covariates that
are present in a majority of
bootstrapped data sets for each
NCS parameter. Linear re-
gression models are then used
to describe the dependence of
the NCS parameter mean on
the identified covariates (S3 in
Figure 8).
Because the standard mini-
mum mean square error ap-
proach is unduly influenced by
the presence of outliers when it
is used to estimate the regres-
sion model coefficients, we
used robust linear regression
[23] as an objective means to
manage the effects of outliers
on the regression model coeffi-
cientsan advantage over the
alternative method of relying
on the subjective assessment of
the researchers to include or
exclude data samples. The var-
iance of the regression model
Last 50 F-Wave Sets
All 80 F-Wave Sets
N
e
u
r
o
l
o
g
i
s
t

I
n
d
i
v
i
d
u
a
l

(
m
s
)
Yield = 95% (541)
Corr Coeff = 0.992
Yield = 89% (872)
Corr Coeff = 0.972
Yield = 90% (541)
Corr Coeff = 0.977
70
60
50
40
40
Neurologist Individual (ms)
C
o
m
p
u
t
e
r
i
z
e
d

I
n
d
i
v
i
d
u
a
l

(
m
s
)
Neurologist Individual (ms)
50 60 70 40 50 60 70
Fig. 7. Comparison of individual F-wave latency assignments between a neurologist and
computerized F-wave cursor assignment algorithm. The repeatability of the manual assign-
ments for the F-wave waveforms in the last 50 F-wave sets is given. A high correlation was
found between the manual and computerized latency assignment results on individual
F-wave waveform basis.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
residual (difference between the NCS parameter and the model
predicted value) is reduced in comparison with the original
parameter variance. The variance reduction is as high as 5969%
for mean F-wave latency of various nerves. Reducing the F-wave
latency variance enhances its clinical utility, as F-wave latency
variance was shown to be the determinant diagnostic factor in
lumbosacral root lesion detection through a detection theory-
based model [20]. Both mean and variance of the NCS parameters
are important for determining their normal limits.
In prior studies, variances of the NCS parameters were
generally assumed to be consistent, even though many nerve-
conduction parameter scatter plots clearly demonstrated hetero-
scedasticity (nonuniform variability for different parameter
means). For example, the average SNAP amplitude decreases
with age and is bounded below by zero with less scatter among
elderly persons [21]. We account for heteroscedasticity by mod-
eling the residual variance as a linear function of the parameter
mean (S4 in Figure 8). As a result, our normal limits for CMAP
and SNAP amplitudes remain positive throughout the whole
age range under consideration. In contrast, some reference-
range models would result in the SNAP amplitude normal limit
becoming negative for older subjects, rendering the limit mean-
ingless, as amplitude parameters can never be negative.
Advances in NCS instrumentation as well as the availability of
preconfigured electrodes offer a unique opportunity to standard-
ize NCS parameters and their reference ranges for commonly
tested nerves. We have implemented the reference-range models
described earlier. When an electronically coded NSEis connected
to a compatible NCS instrument, the instrument will determine
what data-acquisition conditions and setups to use for carrying
out the NCS. With the available demographic (e.g., gender, age,
height, and weight) and physiologic (i.e., temperature measured
with temperature sensor embedded in the NSE) covariate values
of a test subject, normal limits of the NCS parameters become
readily available. For each NCS parameter measured, comparison
with its normal limits is done instantaneously by the instrument
without further operator intervention. In addition to descriptive
flags to indicate normal or out of normal range status of individual
NCS parameters, percentile values are also calculated for the
parameters for which reference ranges have Gaussian distribu-
tions. Percentile values are especially useful when NCS parame-
ters are used to monitor neuropathic disease progression [24] and
the efficacy of interventions in clinical trials [25].
Discussion
Progressive improvements in NCS technology are enabled by
engineering and technological advances and guided by the
needs of timely delivery of quality patient care in office and
clinic settingswherever patients present with complaints
referable to the neuromuscular system. Traditional NCS instru-
ments are designed for use by specialists to support a wide vari-
ety of testing modalities, techniques, and approaches. New
NCS instruments have been designed to allowless highly speci-
alized yet still qualified medical personnel to carry out com-
monly performed NCS tests at the point of service. By focusing
on the commonly tested nerves and locations, the NSE solution
offers physician users a simpler and more convenient way of
performing NCS. The sequential nature of the NCS process
flow is especially suited for programmable computer assis-
tance. Because the NCS test protocol is widely accepted and
understood, guiding the test process flow with computer-
assisted control does not alter the standard of care.
Computer-assisted NCS is an enabling innovation that took an
existing diagnostic technology, streamlined its operation, and
expanded its user base beyond the reach of prior equipment.
These NCS instruments allow clinicians such as primary-care
physicians to make timely and objective decisions about patients
with neuromuscular symptoms. Deployment of NCS in the
primary-care setting streamlines patient management by reducing
the inconvenience, expense, and possible treatment delays
incurred by patients who would otherwise have to travel to alter-
native locations to obtain specialty services at a later date.
Advanced functionality of NCS also aids specialists with an accu-
rate and reliable complement to their traditional manual analyses
of NCS waveforms. Manual waveform analysis can be very
time-consuming and error-prone due to waveform complexity
Nerve
Conduction
Parameter
S1: Nonlinear Transformations
Transformed
Parameter
S2: Identify Relevant
Demographic and
Physiologic Covariates
S3: Determine
Regression
Coefficients
Regression
Residuals
S4: Establish
Variance Model
Normalized
Residuals
S5: Test for
Gaussian
Distribution
S5: Select Best Transformation
S6: Reference-
Range
Specification
Fig. 8. Flowchart of a rigorous statistical framework for deriving
NCS parameter reference ranges as a part of computer-
assisted NCS development effort.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
and variability. These factors contribute to the current restricted
use of certain NCS parameters. For this reason, well-designed
waveform processing methods that assist users to easily and accu-
rately analyze NCSparameters can be important. Prospective stud-
ies have reported high accuracy and reliability of the NCS
parameter results obtained with these waveform analysis methods
[17], [26][30]. Advanced signal analysis tools can also reveal
waveform features that are not readily detectable with manual
cursor assignment. One such example is the segmentation of the
late-wave traces to identify multiple F-wave latencies in each late-
response trace. Availability of the multisegment F-wave latencies
could lead to more sensitive diagnostic tools for peripheral nerve
diseases. NCS advances enhance repeatability and reliability of
NCS measures by removing subjectivity in manual waveform
analysis and by incorporating state-of-the-art engineering solutions
in signal processing and system control. The high repeatability of
NCS measurements makes them ideal candidates for longitudinal
tracking of peripheral nerve function in monitoring neuropathic
disease progression, assessing the efficacy of interventions in clini-
cal trials, and detecting peripheral neurotoxicity.
Xuan Kong received his Ph.D. degree in
electrical engineering from Johns Hopkins
University and M.B.A. degree from Boston
University. Currently he is the vice president
of research at Neurometrix Inc. Before join-
ing the company, he was a tenured associate
professor in the Department of Electrical
Engineering, Northern Illinois University,
Dekalb, Illinois. As a principal investigator, he had led many
federally and industrially funded projects in signal-processing
research and bioelectric signal-analysis applications. He has
published in leading signal processing, biomedical engineering,
and clinical neurophysiology journals. He is a Senior Member of
the IEEE and a member of Eta Kappa Nu and Beta Gamma
Sigma. He is certified as a professional risk manager by PRMIA.
His current research interests include signal-analysis theory and
its biomedical applications, risk management in medical device
and financial service industries, and complex-system modeling.
Eugene A. Lesser received his D.O.
degree from Michigan State University
College of Osteopathic Medicine, com-
pleted his neurology residency at Boston
University and the Boston VA Medical
Center, and obtained fellowship training in
clinical neurophysiology at Brigham and
Womens Hospital, Boston. He is a senior
medical director of NeuroMetrix, Inc., in Waltham, Massachu-
setts, and practices neurology and clinical neurophysiology at
Foundation Medical Partners in Nashua, New Hampshire.
Shai N. Gozani received his Ph.D. degree in neurobiology from
the University of California, Berkeley, in 1993. He also received
his M.D. degree from Harvard Medical School and the Harvard-
MIT Division of Health Sciences and Technology at MIT in 1994.
He founded NeuroMetrix, Inc., in 1996 and currently serves as
chair of the board of directors and as president and chief executive
officer. Before forming the company, he completed a neurophysi-
ology research fellowship in the laboratory of Dr. Gerald Fisch-
bach at Harvard Medical School. His research interests include
basic and clinical neurophysiology and biomedical engineering.
Address for Correspondence: Xuan Kong, NeuroMetrix,
Inc., 62 Fourth Ave., Waltham, MA 02451, USA. E-mail:
Xuan_Kong@neurometrix.com.
References
[1] J. Kimura, Electrodiagnosis in Diseases of Nerve and Muscle: Principles and
Practice, 3rd ed. Philadelphia, PA: Davis, 2001.
[2] R. L. Braddom, AAEM: The first five years. 1991 Presidential Address,
American Association of Electrodiagnostic Medicine, Muscle Nerve, vol. 15,
no. 1, pp. 118123, 1992.
[3] Y. Oester and J. Fudema, Historical notes on electromyography, AAEE
Bull., no. 1516, pp. 7980, 1969.
[4] E. A. Lesser, J. Starr, X. Kong, J. T. Megerian, and S. N. Gozani, Point-of-
service nerve conduction studies: An example of industry driven disruptive inno-
vation in health care, Perspect. Biol. Med., vol. 50, no. 1, pp. 4053, 2007.
[5] C. P. Panayiotopoulos and E. Chroni, F-waves in clinical neurophysiology:
A review, methodological issues and overall value in peripheral neuropathies,
Electroencephalogr. Clin. Neurophysiol., vol. 101, no. 5, pp. 365374, 1996.
[6] W. W. Campbell and L. R. Robinson, Deriving reference values in electro-
diagnostic medicine, Muscle Nerve, vol. 16, no. 4, pp. 424428, 1993.
[7] J. G. van Dijk, W. Kamp, B. J. van Hilten, and P. van Someren, Influence of
recording site on CMAP amplitude and on its variation over a length of nerve,
Muscle Nerve, vol. 17, no. 11, pp. 12861292, 1994.
[8] C. Bischoff, E. Stalberg, B. Falck, and L. Puksa, Significance of A-waves
recorded in routine motor nerve conduction studies, Electroencephalogr. Clin.
Neurophysiol., vol. 101, no. 6, pp. 528533, 1996.
[9] M. A. Fisher, F-wavesphysiology and clinical uses, Scientific World J.,
vol. 7, pp. 144160, Feb. 2007.
[10] K. Todnem, G. Knudsen, T. Riise, H. Nyland, and J. A. Aarli, The non-lin-
ear relationship between nerve conduction velocity and skin temperature, J. Neu-
rol. Neurosurg. Psychiatry, vol. 52, no. 4, pp. 497501, 1989.
[11] E. M. Halar, J. A. DeLisa, and T. L. Soine, Nerve conduction studies in
upper extremities: Skin temperature corrections, Arch. Phys. Med. Rehabil.,
vol. 64, no. 9, pp. 412416, 1983.
[12] R. Nelson, J. Agro, J. Lugo, E. Gasiewska, H. Kaur, E. Muniz, A. Nelson,
and J. Rothman, The relationship between temperature and neuronal characteris-
tics, Electromyogr. Clin. Neurophysiol., vol. 44, no. 4, pp. 209216, 2004.
[13] X. Kong, D. A. Schoenfeld, E. A. Lesser, and S. N. Gozani, Implementation
and evaluation of a statistical framework for nerve conduction study reference range
calculation, Comput. Methods Programs Biomed., vol. 97, no. 1, pp. 110, 2010.
[14] R. K. Olney and M. J. Aminoff, Electrodiagnostic features of the Guillain-
Barre syndrome: The relative sensitivity of different techniques, Neurology,
vol. 40, no. 3, pp. 471475, 1990.
[15] X. Kong and S. Gozani, F-Wave Wiley, in Encyclopedia of Biomedical
Engineering, vol. 3, M. Akay, Ed. New York: Wiley, 2006, pp. 17141725.
[16] J. Astola and P. Kuosmanen, Fundamentals of Nonlinear Digital Filtering.
Boca Raton, FL: CRC, 1997.
[17] M. A. Fisher, Comparison of automated and manual F-wave latency
measurements, Clin. Neurophysiol., vol. 116, no. 2, pp. 264269, 2005.
[18] M. D. Wells, A. P. Meyer, M. Emley, X. Kong, R. Sanchez, and S. N. Gozani,
Detection of lumbosacral nerve root compression with a novel composite nerve
conduction measurement, Spine, vol. 27, no. 24, pp. 28112819, 2002.
[19] J. A. Nobrega, D. S. Pinheiro, G. M. Manzano, and J. Kimura, Various
aspects of F-wave values in a healthy population, Clin. Neurophysiol., vol. 115,
no. 10, pp. 23362342, 2004.
[20] S. N. Gozani, X. Kong, and M. A. Fisher, Factors influencing F-wave
latency detection of lumbosacral root lesions using a detection theory based
model, Clin. Neurophysiol., vol. 117, no. 7, pp. 14491457, 2006.
[21] M. H. Rivner, T. R. Swift, and K. Malik, Influence of age and height on
nerve conduction, Muscle Nerve, vol. 24, no. 9, pp. 11341141, 2001.
[22] B. Efron, Bootstrap methods: Another look at the jackknife, Ann. Stat.,
vol. 7, no. 1, pp. 126, 1979.
[23] P. W. Holland and R. E. Welsch, Robust regression using iteratively reweighted
least-squares, Commun. Stat. Theory Methods, vol. A6, no. 9, pp. 813827, 1977.
[24] P. J. Dyck, P. C. OBrien, W. J. Litchy, C. M. Harper, C. J. Klein, and P. J. Dyck,
Monotonicity of nerve tests in diabetes: Subclinical nerve dysfunction precedes diag-
nosis of polyneuropathy, Diabetes Care, vol. 28, no. 9, pp. 21922200, 2005.
[25] V. Bril, R. Ellison, M. Ngo, B. Bergstrom, D. Raynard, and H. Gin, Electrophysio-
logical monitoring in clinical trials, Muscle Nerve, vol. 21, no. 11, pp. 13681373, 1998.
[26] M. B. Rotman, B. V. Enkvetchakul, J. T. Megerian, and S. N. Gozani,
Time course and predictors of median nerve conduction after carpal tunnel
release, J. Hand Surg. Am., vol. 29, no. 3, pp. 367372, 2004.
[27] X. Kong, E. A. Lesser, J. T. Megerian, and S. N. Gozani, Repeatability of
nerve conduction measurements using automation, J. Clin. Monit. Comput.,
vol. 20, no. 6, pp. 405410, 2006.
[28] S. J. Elkowitz, N. H. Dubin, B. E. Richards, and E. F. Wilgis, Clinical utility
of portable versus traditional electrodiagnostic testing for diagnosing, evaluating, and
treating carpal tunnel syndrome, Am. J. Orthop., vol. 34, no. 8, pp. 362364, 2005.
[29] J. F. Jabre, B. T. Salzsider, and K. E. Gnemi, Criterion validity of the NC-
stat automated nerve conduction measurement instrument, Physiol. Meas.,
vol. 28, no. 1, pp. 95104, 2007.
[30] X. Kong, E. A. Lesser, and S. N. Gozani, Repeatability of nerve conduction
measurements derived entirely by computer methods, BioMed. Eng. Online.,
vol. 8, p. 33, 2009.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
____________________
EYEWIRE
A Novel Phototherapy
Device
N
eonatal jaundice, also known as hyperbilirubinemia,
occurs when the blood has excessive levels of biliru-
bin, sometimes causing the skin and eyes to turn yel-
low. Hyperbilirubinemia is extraordinarily common
(Table 1), particularly in the developing world. In most cases,
infants with jaundice improve in a few days without interven-
tion. However, if not treated, severe cases can cause encephal-
opathy and kernicterus (a type of brain damage), leading to
deafness and other forms of disability.
Phototherapy is the standard treatment for severe cases of
hyperbilirubinemia in newborns [1]. Phototherapy exposes the
infants to light in the range of 400500 nmto isomerize uncon-
jugated bilirubin in the skin. Any light source that produces this
wavelength can be used, including the sun (though there is a
risk of sunburn), fluorescent or halogen lamps [2], or, more
recently, light-emitting diodes (LEDs) [3][5].
Despite the well-established efficacy of phototherapy devi-
ces and their relative simplicitybeing not much more than a
floor lampphototherapy devices are too expensive for
developing world hospitals to purchase, with typical hospital
models ranging fromUS$3,000 to US$5,000.
In addition, the resource-poor setting presents a more chal-
lenging engineering problem than most. Phototherapy devices
are frequently donated to developing world hospitals. How-
ever, donated phototherapy devices typically run for no more
than a few months once donated and, even then, offer little
value to some hospitals.
The most common donations are the older models that use
fluorescent bulbs. However, they do not use standard fluorescent
bulbs. The bulbs must be specially made to emit enough energy
of the correct wavelength to be effective. This raises the price of
the bulbs to several hundred dollars for one set. In addition, these
special fluorescent bulbs have an unusually short lifespan of
1,0002,000 h or about two to three months of continuous use.
Many developing world hospitals cannot afford to spend hun-
dreds of dollars on light bulbs every few months. And even if
they could, the special fluorescent bulbs are not available in
most developing world countries. Table 2 compares some of the
most commonly donated phototherapy devices, their sale price,
and the recommended time until the bulbs need to be replaced.
Developing world hospitals face another problem when
attempting to use phototherapy devices designed for the devel-
oped world. All of the commercially available devices required
electricity supplied by a wall outlet. However, in the develop-
ing world, it is rare that the electricity is on all day and night.
And, unlike many surgeries and procedures, births cannot be
scheduled to occur only when the electricity is on. So, even if
the hospital could continue to use the device after the bulbs
expired, it would only be useful for babies born when the
electricity was on, and then only those that required treatment
lasting less than the time until the electricity was off again.
Given this background, we set out to design a phototherapy
device specifically for use in the developing world. As a mini-
mum, we knew that it had to have a light source with an
extraordinarily long life span and have battery backup so that
it could run during frequent power outages. Our design uses
LEDs and is powered by a car or motorcycle battery.
The market size for our device, measured in dollars, is very
small; too small for traditional market-based engineering
design to be effective. We used an alternative volunteer-based
engineering approach. In this article, we present the full engi-
neering design cycle, staring with needs identification and con-
tinuing through several cycles of engineering field trials and
results with comments on the differences between the engineer-
ing design cycle executed in, and for, the developing world.
Methods
Needs Identification
The first step in our design process was needs identification
leading to specifications. Each year, DukeEngineering World
Health (EWH) sends 50 or more engineers to the developing
world. These engineers live and work in a community for up to
one month. Among the many tasks they complete during their
assignments is a needs-finding interview in a resource-poor hos-
pital. Each participant interviews four or more doctors, nurses,
administrators, and staff, probing to find opportunities where a
technology could solve health-care problems found on the
ground in a developing world village, clinic, or hospital. The
techniques used have been briefly described elsewhere and gen-
erally followed the principles suggested by McCracken [6] with
an emphasis on the staged model recommended by Rubin and
BY ROBERT MALKIN
AND VIJAY ANAND
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
The Design Community Approach
for the Developing World
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Rubin [7], stopping at the point of providing students main ques-
tions [7, pp. 152164]. We found that students were able to
develop their own probes, despite the language barriers [7, pp.
164172]. Guidance is given to the interviewers on follow-up
questions [7, pp. 173200]; however, only about 20% of inter-
viewers completed the follow-up questionnaires fully.
In a typical year, more than 200 interviews are conducted in
hospitals in Central America and Africa. From the interviews,
meetings and discussions are held with physicians, engineers,
and scientists in the United States to produce a list of projects
that are both needed and technically practical. This list is pub-
lished each year on the EWH Web site (www.ewh.org).
In 2005, the needs-identification process led to the selection
of the delivery of phototherapy in the developing world as a
problem worth solving. Nearly every surveyed hospital pos-
sessed a phototherapy unit. Almost none of them worked. An
analysis of 974 pieces of broken medical equipment showed
that, by far, the most common cause of phototherapy failure
was the lack of replacement light bulbs and of reliable power.
From these interviews and subsequent refinement inter-
views, specifications and requirements were derived (Table
3). Besides low cost, the most important unique specifications
were the need for a rechargeable, readily available battery,
and very long-lasting light source.
Design
Students taking a class at Duke University completed the first
design and later placed their design into the public domain. In their
design, an array of LEDs, designed to operate at 410490 nm
(MCD model MCDL-5013), was selected as the light source [4].
This light source is well characterized, produces light nearly exclu-
sively in the therapeutic range, and is inexpensive and efficient.
The initial testing of the students design demonstrated that
the principals of operation were sound and that it was
technically feasible to meet the required performance specifi-
cations. The initial design also served to identify several
weaknesses in reliability and manufacturability. Still, having
shown that it could be done and having removed the barriers
to intellectual property transfer, the initial design could be
leveraged by the product champion to recruit another design
teamto move the project forward.
In this case, a design team from a Durham, North Carolina,
design firm, Tackle Design, then took the design out of the pub-
lic domain and refined it. The second design focused on using a
simple approach to reliably manufacture the device using a vol-
unteer workforce. The specification of a volunteer workforce
was a consideration born out of the intended, nonprofit business
model and anticipated small manufacturing runs, typical of
many medical devices. The design team also considered the
need for field repairs with a minimal tool kit, another considera-
tion not in the original specifications. Neither the use of a volun-
teer workforce nor the need for repairs with minimal tools are
design challenges typical considered for medical devices. Again,
the design firmplaced their product in the public domain.
Figure 1 shows a photograph and a sketch of the second
design. The design consists of two parts, the head and the base.
The head mechanically consists of a flexible elastomer com-
pressed between two relatively rigid pieces of acrylic by a
dozen nuts and bolts. Electrically, foil strips run on both sides
of the elastomer connecting 55 blue LEDs in parallel. Each
LED is wired in series with a small resistor to prevent the
shorting of one LED from causing the entire device to fail.
The illuminating device head was dubbed the BluLine.
The base served as the power supply. As a result of engineer-
ing field trials, described below, two approaches were at-
tempted to power the device. The first used a commercially
available power supply. The particular commercial power sup-
ply was selected because the manufacturers specifications and
technical support staff suggested that this particular supply had
the greatest ability to withstand power-line fluctuations,
including spikes, surges, and brownouts. The second approach
was a simplified custom-developed design for this project,
shown in Figure 2. For the second approach, a commercially
available, low-cost wall transformer was used as a current
source to either charge a 6-V motorcycle battery or to power
the BluLine head. If the power failed, the battery immediately
began powering the device. The wall transformers are readily
available in the developing world.
Throughout the process, the American physicians were con-
sulted as to the design elements and their applicability to a
developing world setting. Most phototherapy devices are dis-
tributed with a stand. The most fre-
quent complaint received from the
American physicians about the design
was the lack of a stand. However, dis-
cussions with the staff of more than 20
developing world hospitals convinced
us that a stand was not necessary and
would not be used in many cases, leav-
ing the increased cost unjustified.
Instead, the device is distributed with
the equipment required to mount from
the ceiling, on an intravenous drip
stand, or from the wall. In many cases,
the local staff constructed a stand
unique to their environment.
Table 2. The commonly donated phototherapy devices, their sale price, and the
recommended replacement time.
Company Bulb Technology
Intensity of Light
(lW/cm
2
/nm)
Recommended
Life Span of
Bulbs (h)
Price of
Device (US$)
GE Medical Quartz-halogen 830 1,500 3,500
Drager Fluorescent tubes 422 1,000 3,000
Medela Fluorescent tubes 40 1,500 5,000
Most phototherapy devices are too expensive for developing world hospitals,
and the bulbs have a lifespan that is very short. The replacement bulbs are not
often available in the developing world.
Table 1. Neonatal jaundice: geography, mortality, and
morbidity.
Geography Mortality Morbidity
South East Asia 102,000 1,204,956
South America 9,800 115,770
Africa 60,250 711,751
Neonatal jaundice or hyperbilirubinemia kills more than
100,000 children each year, despite the availability of an
effective treatment, phototherapy.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Bench Testing
The devices were bench tested to determine that they met the
specifications. Figure 3 summarizes the most critical data from
bench testing, showing the intensity of light for each square
inch of the illuminated crib, following the methods suggested
by Vreman [8]. All of the designs tested in each of the field tri-
als passed all bench testing and met all design specifications.
Field Engineering Trials
Despite data from the bench that showed the device operated
in the manner anticipated, we wanted to undertake extensive
field testing. As discussed earlier, a device that is adequate for
a well-developed hospital may be inadequate for a developing
world setting.
A total of 114 units were constructed using volunteers coor-
dinated by a nonprofit corporation formed for the purpose of
manufacturing and distributing the BluLine: PhotoGenesis
Medical (www.photogenesismedical.com). These were di-
vided into three groups and shipped to hospitals in four countries.
The hospitals were chosen to range from small (approximately
20 beds) to very large (more than 500 beds) and from specialty
(maternity) to general and referral hospitals.
Every device was accompanied by two engineers. The engi-
neers attempted to install the device, train the staff to use the
device, and then interviewed the trained staff about the device.
For the first two groups, simulated clinical use was undertaken for
one month. Where simulated, clinical use includes at least occa-
sionally turning on the device and letting it run for several days.
Results
First Engineering-Field Trials
Fromthe first build, 56 units were completed by 6575 volunteers
(approximately 200 volunteers were turned away for this build
session). The cost for each unit for the first build was US$65.62.
This can be compared with a commercially sold unit whose retail
price can be several thousand dollars (Table 2). Each unit required
approximately five volunteer man-hours to complete. Despite
practice builds by the engineering team, eight procedures devel-
oped for the build required modification during the build.
Thirty-nine of the 56 units were shipped to hospitals in Tan-
zania, Nicaragua, and Honduras. Of those shipped, 30 worked
upon arrival and 38 worked after minor engineering correc-
tions (Table 4).
No developing world installation complained about the lack
of a stand. Most hospitals hung the device from the ceiling, as
the cribs were typically not moved, and in any case completely
filled the available space to overflowing, including several
infants in each crib. A few hospitals preferred a drip stand to be
used for the device, and others constructed a stand for the
device, the available manual labor being relatively inexpensive.
Table 3. The most important design criteria and require-
ments for phototherapy identified from interviews conducted
in developing world hospitals.
Power requirements
Rechargeable >8 h on battery alone after full
charge
Charge while light is running
and line power is on
Battery lifespan Two years with daily outages
Battery availability in
country
Comparable to a 12-V car
battery or 6-V motorcycle
battery
Light source
Intensity 830 lW/cm
2
/nm
Wavelength 425475 nm
Life time Bulb > 5 years normal use
Spot size 8 in
2
User interface On/off button only
The needs for a rechargeable battery and a very long-
lasting bulb would not have appeared on a similar speci-
fication list for a U.S. hospital.
(a)
(b)
Fig. 1. (a) and (b) The design of the phototherapy devices
illuminating head consists of three plastic layers (acrylic in
black, elastomer in gray) compressing and connecting 55
LEDs in parallel. The bolts, symbolized by the dashed lines,
hold the pieces together and also allow for easy field repairs
with minimal tools.
2 A Fuse D1 D2
D4 D3
R1
Branch 1
Branch 2
6 V dc
+
Branch 3
BlueRay
Unit
Fig. 2. Schematic showing the final power supply. The origi-
nal power supply used a 12-V source and commercially
available power handling but turned out to be excessively
sensitive to power-line fluctuation. The diodes allow for auto-
matic switching between the wall transformer (shown as a
current source) and the battery, when the power line fluctu-
ates or fails. The diodes also serve to drop the voltage to
the level required by the BlueRay unit.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Every installation experienced numerous power outages
(often several per day), with some hospitals having electricity
only a few hours per day. Power outages were accompanied
by power surges, brownouts, and spikes in most cases. At the
end of one month of simulated use, four units were identified
as malfunctioning. All of the malfunctioning units were
related to the power supply, either the regulator or the charger.
The staff of the hospitals noted several concerns. The first
was related to the wiring and switch of the device. The staff
complained that the device looked homemade. The staff com-
mented that they would be reluctant to use the device on
humans because of this homemade look. Even when specifi-
cally contrasted with the alternative, which in most cases was
not treating the patient, as the most of these hospitals had no
alternative treatment available, some staff members were
reluctant to use the device. The concern about the look and feel
of the device surprised the design team. The American physi-
cians had not indicated that this aspect of the design would be a
problem, assuming that, given the alternative of no treatment,
the look and feel of the device was considered insignificant.
After some discussion with the local nonmedical staff, the
engineering team hypothesizes that the use of devices that
look less than professional conveys a sense of inferiority to the
medical team. In other words, the medical team was proud of
their facilities and their work in general. Using equipment that
did not look professional, particularly if that equipment came
from a western source, contradicted their self-view.
Further discussion and experimentation revealed two issues
with the look of the device. The first was the use of black wire
and a white switch. The mismatched switch and wire color gave
the device a pieced together look. The engineers involved spec-
ulate that the staffs frequent exposure to repaired equipment
has taught them to quickly identify equipment that has been
refurbished or reworked. The mis-
matched colors for the switch and
wiring are one sign of reworked
equipment. Since reworked
equipment often malfunctions in
the developing world, the staff
were immediately suspicious.
Second, the staff had not used
a rotary switch before and did not
like the fact that they could not
easily tell on fromoff, despite the
fact that the light is only on when
the switch is in the on position.
Again, this may result froma lack
of trust in the equipment.
Only one hospital staff asked
about Food and Drug Adminis-
tration (FDA) or CEmark (Euro-
pean product approval). All of
the staff members were excited
to begin using the device as soon
as it became available. None
requested clinical efficacy data.
This may be because they
deemed the safety factor high
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
0 0.9 1.6 2.4 4.6 5.2 4.7 3.1 1.9 1
1 0.8 1.7 4.1 7.7 10 12 11 7.7 4.4 1.9 0.7
2 0.6 1.2 3.4 8 14 14 21 18 14 7.9 3.4 1.3
3 0.6 1.6 5.1 11 19 26 29 28 21 12 5.1 2 0.6
4
0.5 0.7 1.9 5.5 11 19 26 29 28 28 11 5.7 2.3 0.8 0.4
5 0.4 0.7 1.6 4.8 9.9 16 22 25 23 17 10 5.3 1.9 0.7 0.4
6 0.4 0.5 1 2.8 6.4 10 15 16 16 12 7.4 3.8 1.3 0.6 0.4
7 0.3 0.4 0.7 1.4 3.8 6 10 10 10 7.6 4.7 2.1 0.9 0.5 0.3
8 0.5 0.8 1.8 3.3 4.8 5.2 4.8 3.5 2.2 1
9 0.4 0.5 0.7 1.2 1.7 1.8 1.7 1.3 0.8 0.6
Fig. 3. Luminosity data. The intensity of light (lW/cm
2
/nm)
measured every inch for the BluLine mounted 18 in above
the crib. There are more than 90 in
2
of surface area above
1 lW/cm
2
/nm. This shows that the majority of the baby,
when properly placed under the light, will receive adequate
therapeutic radiation.
Table 4. Units shipped for field-engineering trials and the results of shipment.
Round 1 Round 2 Round 3 Results of Shipment
Units shipped for field engineering trials
Tanzania 14 15 0
Nicaragua 12 8 0
Honduras 13 12 0
Kenya 0 0 28
Total units 39 35 28
30 33 28 Units initially working
postshipping
38 1 0 Units requiring field repair dur-
ing month of testing in
hospital
4 0 0 Units that could not be fixed in
the field
10 0 0 Staff concerns and problems
noted
Despite extensive bench testing, the first round of engineering field trials resulted in
nearly 100% of the units needing to be repaired in the field. However, the third and final
round of units showed that successive redesign had fixed the first-round problems.
Phototherapy is the standard treatment for
severe cases of hyperbilirubinemia
in newborns.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
and the alternative unavailable. Or, this may be because of the
frequent availability of untested devices. It should also be
noted that there is little or no enforced medical device regula-
tions in the countries studied here.
Second Engineering-Field Trials
The devices were redesigned to use a custom power supply,
improve the look of the device, and replace the rotary switch
with a rocker switch.
A second volunteer build resulted in 45 working units. Each
unit required approximately 6 man-hours of labor to complete.
The cost for each unit for the second build was approximately
US$45, the drop in price coming from the use of a custom-
designed power supply.
Thirty-five units were shipped to mostly the same hospitals in
Tanzania, Nicaragua, and Honduras. Of those shipped, 33 worked
upon arrival and one required minor engineering corrections. One
device was too damaged during shipping to be repaired in the field.
No units were found to have failed after one month of use,
and none of the staff expressed concerns about the device.
Third Field-Engineering Trials
The third and final engineering trials were conducted in Kenya.
Twenty-eight units were shipped to several hospitals. All of the
lights worked upon arrival and were installed without incident.
The hospital staff expressed no concerns about using the device.
All of the staff members were excited to begin using the device
as soon as it became clinically available. Again, most of the Ken-
yan hospitals had no other devices for treating infantile jaundice.
Discussion
This article describes a phototherapy device that runs on a
motorcycle or car battery and uses very long-lasting bulbs.
The need was clear for this device in the developing world,
and the design was relatively simple. So, why has the product
not been introduced before?
Typical medical device development can consume several
years and cost US$10US$20 million [9]. Even if the device
can be developed, it must continue on to engineering field
trials, perhaps animal and clinical trials. The costs of clinical tri-
als can vary tremendously, even between developed nations. A
Class II device going through the 510K process (a relatively
simple approval process in the United States) can require clinical
testing that can range from a short study for European approval
to an expensive US$10 million study in the United States [9].
Even if the regulatory costs are ignored (perhaps arguing that
it is irrelevant because many developing world nations do not
have medical device regulations), the traditional development
costs may exceed the entire market size for the developing
world. Many poorer nations are completely dependent on out-
side donations, the ministry of health providing only a few per-
cent of the health-care expenditures within their countries [10].
With little or no chance of recovering the development costs,
and the possible expense of trials, no large, traditional medical
device manufacturer, e.g., Phillips, Siemens, or Medtronic, will
design a device that has exclusively a developing world market.
This leaves only nontraditional medical device design and devel-
opment and small companies. The small medical device companies
can use grant funding to design and develop devices [11]. How-
ever, grant funding may be sufficient to generate a fewdesigns, but
certainly not the thousands of products that an U.S. market enjoys
or the hundreds that the developing world markets need.
Since engineering labor is by far the largest cost of any
medical device design, the alternative must consider ways to
reduce engineering costs. In this work, we explored the use of
volunteer designers, builders, and field engineers as part of a
development cycle. Collectively, we refer to the designers,
builders, and field engineers as a community. While dramati-
cally reducing the development costs and therefore opening
the door for medical device design, the community design
approach reveals several significant differences between
design for the developing world and the developed world.
A Design Community Across
Development Boundaries
A medical device designed for the developing world by a
designer in the developed world must pass through essentially
the same stages as a device designed for the developed world.
However, because the designer is separated from the user by
language, culture, and thousands of miles, and because the
designer is a volunteer in our case, the process can be compli-
cated at points that are not typical of a commercially, domesti-
cally designed and used product.
Development consists of needs finding, abstraction, design,
and testing in what is called the engineering design cycle.
Needs finding across a cultural and linguistic barrier is a unique
challenge, and our approach is described earlier. Abstraction is
not significantly different between target environments. This
leaves design and testing.
As with any design, the engineer must design and test their
device to meet developed specifications. There is a relatively
well-developed infrastructure for electronic, collaborative
design in certain domains. Projects such as Wikipedia (www.
wikipedia.org) demonstrate that volunteer efforts can, in fact,
create tremendous work when properly organized. Engineering
accomplished by the volunteer communities have been equally
impressive, such as Linux (www.linux.org) and GNU (www.
gnu.org). Even in the medical field, the projects such as the
open prosthetics project (www.openprosthetics.org), the open
ECG project (www.open-ecg-project.org), and EWHs projects
that matter (www.ewh.org) demonstrate that an engineering
design can be accomplished by a volunteer community.
However, design is only effectively accomplished in conjunc-
tion with testing. One of the dramatic differences between design
for GNU and design of a medical device is that every member of
the GNU development community can fully test their designs,
often immediately. The entire engineering design cycle can be
executed in a matter of minutes, in some cases. However, very
fewmembers of the medical device design community are quali-
fied to fully test their products. Even bench testing of a medical
device for the developing world can be challenging.
In this project, thorough bench testing was attempted. However,
the bench testing proved insufficient to highlight the problems with
the power supply. Despite having selected the most robust power
supply available on the U.S. market, many of our devices failed in
the field because they could not cope with the frequent power out-
ages, fluctuations, and surges. We conclude that, when designing
for the developing world, bench testing must be augmented with
engineering field trials before continuing the engineering design
cycle. Partnership with developing world engineering teams, if
they can be identified, might provide better bench-testing data.
One could argue that designers often specify the environmen-
tal usage of their devices. The developing world is not unique.
For example, the device must be stored above 0 C or the line
____
____
____
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_____
power must be between 110 and 220 V. However, this approach
presents two problems when designing medical devices for the
developing world. First, environmental factors for designs used
in the United States or Europe are typically based on well-
studied ergonomics and well-characterized infrastructure. The
ergonomics and infrastructure are not well characterized for
developing world hospitals, including things such as storage and
use temperatures and line voltages. Second, defensively specify-
ing usability criteriathe practice of specifying that the device
be used only under ideal conditions such as in climate-controlled
operating roomsmay reduce the applicability of the device to
nearly zero in a developing world hospital. Given these con-
straints, only engineering field trials can be considered sufficient
to continue a given design, in our opinion.
Although it is not unusual for an American design firm to
trial initial, perhaps nonfunctional prototypes with the selected
physicians in the United States before the product is more
broadly introduced, these engineering field trials are often of a
very limited nature and can be considered a marketing practice
more than an engineering practice. Indeed, field engineering
trials typically consume nearly none of the cost of a typical
medical device development.
However, when designing for the developing world, our
experience suggests that extensive engineering field trials with
many hospitals are required. If a volunteer community is taking
on design and considering the cost of flying engineers to many
hospitals across the world, engineering field trials can be the
largest cost of medical equipment when community designed.
In summary, we feel that our project demonstrates that the
engineering design cycle can be executed by a community of
volunteers. However, costs shift significantly from design and
bench testing to the engineering field trials while the impor-
tance of engineering field trials is increased. Thus, engineering
field trials become the main obstacle to successful community
design of medical equipment for the developing world.
In this work, we solved the engineering field trials problem
by using volunteers participating in the Duke-EWH Summer
Institute. Wherever a hospital had technical staff, our instal-
ling engineers coordinated with the staff and attempted to train
themto the greatest extent possible. However, most of the hos-
pitals had no technical staff. Indeed, lack of human resources
is an enduring problem in the developing world.
Another approach would be a partnership with a single
developing world engineering school, where the engineers at
that specific school could, presumably, prototype and field test
designs. Perhaps an expansion and combination of approaches
would be a community of engineering schools in the develop-
ing world where any member could access the design materi-
als and participate. The latter begins to approach the power of
the design communities of GNUand Linux.
Alternatives to the Design Community
As mentioned earlier, an alternative and successful model to a
design community is the protected design house promoted by
PATH [11]. In this scheme, grants fund the design and develop-
ment of a medical device, typically within the confines of intellec-
tual property protection. Atraditional, perhaps large, manufacturer
then obtains the intellectual property and manufactures the device.
For example, the Uniject by PATH has been manufactured in the
millions if not billions of units.
However, the grant-funded approach also has limitations.
The most obvious is that it is limited by the amount of grants
available. At most, grants can be expected to create no more
than a handful of designs in a given year. Also, the grant-
funded approach limits the use of developing world compa-
nies to manufacture the devices, as they cannot afford to
obtain the intellectual property. Perhaps most importantly, the
grant approach is unlikely to create incremental improvements
in existing designs. Still, incremental introduction dominates
the medical device innovation pathway.
Another way to look at the lack of incremental improvements
using grant funding is to examine the nature of the problemthat it
addresses. Most developing world hospitals do not need the latest
technology. They are lacking the current generation of technol-
ogy. And, much of the latest technology will not work once it gets
there. Still, it is very difficult to obtain a grant to adapt the already
existing equipment for developing world applications.
Regulatory, Clinical Trials, and Manufacture
The design community addresses the cost of design. However,
there are other costs associated with bringing a medical device
to market. After a design is successfully completed, it must be
clinically tested, manufactured, and distributed. Most medical
devices are manufactured and distributed by companies. The
ability of a company to legally distribute a medical device
depends on the regulations of the countries involved, and those
regulations typically involve clinical trials. Thus, manufacture,
clinical trials, and regulatory affairs are inseparable issues.
We believe that very few, if any, medical-device manufacturer
in the developed world would build a device for distribution in the
developing world without FDA approval or a CE mark. Indeed,
under section 802 of the FDA rules, manufacture of a medical
device exclusively for export from the United States must still be
able to obtain approval, although actually obtaining such approval
is not required. Even if these rules are not enforced, the U.S. medi-
cal-device manufacturers would be unwilling to violate them.
If manufacturing in the developed world is not likely, then we
must consider manufacturing in the developing world for the
developing world. Since many developing world nations do not
have medical-device regulations, local manufacturing bypasses
the concern for regulatory costs [12]. For example, the Jaipur foot
makes its manufacturing plans freely available (www.jaipurfoot.
org) and has been produced for many thousands of patients
around the world in their home country. The projects for devel-
oping world production of foreign-designed wheelchairs abound
(www.doitfoundation.org and www.whirlwindwheelchair.org)
to the point where World Health Organization produced a sup-
porting guideline [13]. Published plans for baby incubators, clini-
cal water stills, and centrifuges that can be produced and used in
Africa are also available [14]. All of these devices went from
design to engineering field trials to manufacture/distribution in
the developing world without passing through formal clinical tri-
als or regulatory affairs.
If we assume that the cost of regulatory compliance is too high
to justify design for a developing world country, then we must
identify an alternative or there cannot be medical device innova-
tion for the developing world. As mentioned earlier, the FDAprob-
ably will not enforce regulation on the American-made products
exclusively for export to the developing world, and many develop-
ing world nations have no medical device regulations to enforce.
Therefore, one clear alternative to the cost of FDA or CE regula-
tory compliance is to ignore these regulatory authorities. This
should not be taken as a suggestion that the safety and efficacy of
the device should be compromised. However, design and testing a
___
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
device to International Standards Organization (ISO) standards,
such as operation at 0 C (as some ISO standards require) for use
in countries where the temperature never dips below25 C, can be
safely avoided with considerable savings in cost.
Another point of view is that the cost of regulatory affairs is not
excessive; it is only excessive when compared with the market
size. With this point of view, an alternative to avoiding American
or European regulatory authorities would be to aggregate sub-
Saharan African markets. If multiple government purchasers could
specify a newproduct and guarantee purchase of this product, man-
ufacturers would be more willing to innovate for the now larger
market, including regulatory approval. A disadvantage of this
approach is that it requires a high degree of organization and manu-
facturer trust that the orders will be placed at the end of the devel-
opment cycle. It may also be difficult for governments to specify
the purchase of a piece of equipment that does not, as yet, exist.
Finally, another alternative would be for the public sector (such
as ministries of health) or independent funding agencies (such as
the African Development Bank) to create funding opportunities
to obtain FDAapproval for devices that have demonstrated safety
and efficacy. One weakness of this proposal is that compliance
with American and European regulatory authorities is an ongoing
expense. Thus, the cost of all units for the life cycle of the device
would be increased, even if the initial costs were covered. Also,
considering that the average life cycle of a device is only about
two years before an incremental improvement requires new
approval, this approach might be excessively expensive.
The Role of the Design Champion
One thing that a Wikipedia page, the OpenECG project, and
this project have in common is a champion. Each Wikipedia
page has someone who is responsible for the integrity of the
page. In this project, Vijay Anand has served the role of the
project champion, recruiting volunteers, financial resources,
and cheerleading, when needed.
The role of the design champion has been most important at
a few critical points in the design cycle: when moving the
device from bench-tested prototype to redesign for manufac-
turability, when moving from manufacturable prototype to
engineering field trials, and, now, when moving from engi-
neering field trials to clinical trials.
The design champion has to have unfettered access to the
intellectual property and a working knowledge of medical
device design. However, the champion does not need to be
dedicated full time to the project. In this project, no member of
the core teamconsiders this work their full-time occupation.
Conclusions
We have presented the results of a community-designed and
developed phototherapy device. Several hundred volunteers and
a design community participated in the design, manufacture,
and engineering field trials of the device. Unlike traditional
manufacturing, community-designed devices for the developing
world need to focus a larger fraction of available resources on
engineering field trials. Regulatory affairs, clinical trials, and the
approach to manufacturing present the intertwined, next greatest
challenge to the introduction of medical devices community-
designed specifically for the developing world.
Acknowledgments
The authors thank Chuck Messer for his invaluable technical
assistance throughout the development of the BluLine.
Robert Malkin received his M.S. and
Ph.D. degrees in electrical engineering
from Duke University in 1991 and 1993,
respectively. He is a professor of biomedi-
cal engineering at Duke University in Dur-
ham, North Carolina. He was the Herbert
Herff Professor of biomedical engineering
at the University of Memphis and the
University of Tennessee. He was an assistant professor of
electrical engineering at the City College of New York, a
member of the graduate faculty at the City University of New
York, and a research associate at Columbia University. He is
a fellow of the American Institute for Medical and Biological
Engineering and is the director of Duke University-EWH and
The Global Public Service Academies. He is a member of the
World Health Organizations Technical Advisory Group on
Health-Care Technology.
Vijay Anand received his masters degree in
engineering management from Duke Univer-
sity in 2006. He founded PhotoGenesis Medi-
cal, which won the Annual Competition for
Underserved and Resource-Poor Economies
(CUREs) Business Plan held by the Engineer-
ing World Health. He is experienced traveling
to and working in developing countries,
including Kenya and India, to better understand the related issues
with jaundice in infants. With a degree in biomedical engineering,
and experience with GE Medical Systems, he works passionately
to make medical resources more easily available.
Address for Correspondence: Robert Malkin, Duke Univer-
sity, Biomedical Engineering, Hudson 136, Durham, NC
27708. E-mail: robert.malkin@duke.edu.
References
[1] American Academy of Pediatrics, Management of hyperbilirubinemia in the
newborn infant 35 or more weeks of gestation, Pediatrics, vol. 114, no. 1,
pp. 297316, 2004.
[2] M. J. Maisels, PhototherapyTraditional and nontraditional, J. Perinatol.,
vol. 21 (Suppl. 1), pp. S93S97, 2001.
[3] S. Nakamura and G. Fasol, InGaN single-quantum-well LEDs, in Q5 The
Blue Laser Diode, S. Nakamura, G. Fasol, and S. J. Pearton, Eds. Berlin:
Springer-Verlag, 1997, pp. 201221.
[4] D. S. Seidman, J. Moise, Z. Ergaz, A. Laor, H. J. Vreman, D. K. Stevenson,
and R. Gale, A new blue light-emitting phototherapy device: A prospective
randomized controlled study, J. Pediatr., vol. 136, no. 6, pp. 771774, June 2000.
[5] D. S. Seidman, J. Moise, Z. Ergaz, A. Laor, H. J. Vreman, D. K. Stevenson,
and R. Gale, A prospective randomized controlled study of phototherapy using
blue and blue-green light-emitting devices, and conventional halogen-quartz pho-
totherapy, J. Perinatol., vol. 23, no. 2, pp. 123127, 2003.
[6] G. McCracken, The Long Interview. Newbury Park, CA: Sage, 1988.
[7] H. Rubin and I. Rubin, Qualitative Interviewing. Thousand Oaks, CA: Sage,
2005, pp. 114122.
[8] H. J. Vreman, R. J. Wong, J. R. Murdock, and D. K. Stevenson, Standar-
dized bench method for evaluating the efficacy of phototherapy devices, Acta
Pediatr., vol. 97, no. 3, pp. 308316, 2008.
[9] A. V. Kaplan, D. S. Baim, J. J. Smith, D. A. Feigal, M. Simons, D. Jefferys,
T. J. Fogarty, R. E. Kuntz, and M. B. Leon, Medical device development from proto-
type to regulatory approval, Circulation, vol. 109, no. 25, pp. 30683072, 2004.
[10] WHO Group, An assessment of interactions between global health initia-
tives and country health systems, Lancet, vol. 373, pp. 21372169, June 2009.
[11] M. J. Free, Achieving appropriate design and widespread use of health care
technologies in the developing world. Overcoming obstacles that impede the
adaptation and diffusion of priority technologies for primary health care, Int. J.
Gynaecol. Obstet., vol. 85 (Suppl. 1), pp. S3S13, 2004.
[12] World Health Organization, Medical device regulations: Global overview
and guiding principles, WHO Report, 2003.
[13] WHO, Guidelines on the provision of manual wheelchairs in less resourced
settings, WHO, 2008.
[14] A. Platt and N. Carter, Making Health Care Equipment. Warwickshire, Eng-
land: Intermediate Technology Publications, 1990.
________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Continuous Monitoring of
Respiratory Flow and CO
2
I
n this article, the challenges of simultaneous respiratory gas
concentration and flow measurements in a breathing circuit
are reviewed. The tradeoffs that were considered in the
development of a clinically useful on-airway combination
CO
2
/flow sensor are discussed as well as the applications
enabled by this on-airway combination CO
2
/flow sensor.
Continuous respiratory gas measurements during the admin-
istration of anesthesia, intensive care, and other clinical envi-
ronments provide valuable information for assessment of
cardiorespiratory function as well as breathing circuit integrity.
In particular, measurements of gas concentrations, flows, vol-
umes, and pressures proximal to a patients airway can often
better reflect the underlying pathophysiology of the patient than
measurements made more distally. Continuous monitoring of
gas composition and volume in the complex breathing circuit
environment, where changes in humidity and temperature can
distort the measurement and instrumentation is further contami-
nated by sputum, pathogens, blood etc., presents a number of
difficulties/complications that have led to compromises and
less than optimal clinical and technical solutions. This environ-
ment demands simplicity, reliability, ease of use, and ability to
continue working in wet, often mucous-filled circuits for long
periods of time without operator intervention. Because these in-
circuit devices are typically disposed of after each patient use to
avoid cross-contamination, they should be relatively inexpen-
sive, lightweight, and disposable. Furthermore, these devices
need to have minimal volume so as not to interfere with therapy
and should work over wide flow ranges, while contributing
minimal resistance to air flow, and require minimal or no cali-
bration. Devices typically designed for the pulmonary function
laboratory employ sensors designed for brief procedures lasting
for a fewminutes rather than for continuous monitoring [1].
With respect to flow, volume, and pressure measurements, it
has become increasingly clear that to accurately deliver tidal
volumes to a critically ill patient, it is preferable to performthese
measurements as close to the patient as possible [2]. Tradition-
ally, the measurements of the delivered and expiratory flows
and volume have been, for technical reasons, located in the
ventilator. Proximal flow (i.e., measured at the patients airway)
can be substantially different from flow measured inside or at
the ventilator due to the wasted gas compression volume, breath-
ing hose expansion, and differences due to humidification. This
wasted portion of the tidal volume, i.e., compression and hose
expansion volume, does not ventilate the patient, but rather
remains within the breathing circuit tubing. A correction for this
effect which is proportional to the inspiratory peak pressure is
applied by some ventilator manufacturers given that the breath-
ing circuit volume and compliance is known. However, the com-
pliance may also depend on the inspiratory flow rate [3]. Even
with this correction applied, the precise estimation of the com-
pression volume is difficult because of variations between indi-
vidual breathing circuits, use of humidifiers, heat-moisture
exchange (HME) devices, and other circuit components. In a
typical breathing circuit, gas conditions such as temperature
may vary from room. Air-to-body temperature and humidity
may vary from dry air to fully saturated air. Water vapor phase
changes within the breathing circuit hoses also contribute to dif-
ferences in volume measurements between the distal (within the
ventilator) and proximal (patient airway) measurements. Thus,
for more accurate monitoring of delivered volumes and of the
patients expired volume, the flow sensor should be placed
between the breathing circuit wye and the endotracheal tube
(ET). For neonatal and pediatric patients, this problem has been
widely recognized by ventilator manufacturers who generally
offer proximal flowmeasurements for these patient populations.
The clinical benefits of on-airway measurements were under-
stood early in the development of clinical respiratory instru-
ments [4]. These clinical benefits include faster signal response
and precise time alignment between the flowand gas concentra-
tion signal. Early respiratory gas benches used in anesthesia
were nondiverting on-airway (i.e., mainstream) devices. Bulky
and heavy, the size of these mainstream benches obscured the
patients head from the surgeon [5]. For these reasons, gas con-
centration measurements have traditionally been made in
breathing circuits using diverting (or sidestream) gas benches.
However, with advances in technology, the size and weight
concerns of these early devices have been overcome. The clini-
cian need no longer accept the limitations of sidestream
approaches [6], [7], such as delayed and dampened waveforms
and susceptibility to clogging of the sampling line.
We have developed a family of integrated, disposable air-
way adapters (Figure 1) that are capable of monitoring CO
2
BY MICHAEL B. JAFFE
AND JOSEPH A. ORR
EYEWIRE
Challenges of On-Airway Measurements
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
concentration in real time, breath by
breath, using infrared (IR) absorption
techniques in combination with moni-
toring respiratory flowwith differential
pressure flowmeters under diverse inlet
conditions. These airway adapters
interface to an on-airway capnometer
(Figure 2) and spirometry measure-
ment module (Figure 3). Our solution
to the challenges of on-airway flow
CO
2
and the combination of these two
measurements are reviewed with re-
spect to the challenges highlighted in
this article.
This solution has been nearly 20 years
in the making, requiring a number of
key developments that leveraged on
internal technical developments, such as
a novel thick-film IR source enabling a
solid-state on-airway capnometer [5];
external alliances that resulted in a
robust, fixed orifice flow sensor [8]; a
single piece-combined CO
2
/flow sen-
sor [9]; and technological develop-
ments including cost-effective and
robust, yet extremely sensitive differ-
ential pressure sensors; and in-
creasingly integrated digital signal
processing chips. The Philips-Respir-
onics devices for on-airway volumetric
capnography have evolved since the
early 1990s from separate flow and
CO
2
sensors placed together in the
breathing circuit, with each connected
to a separate device (e.g., Ventrak 1500 and Ventrak 1550 sys-
tems and Capnoguard or CO
2
SMO monitors from Novametrix
Medical Systems) [10] to integrated CO
2
/flow-airway adapters
interfaced to the same host system.
On-Airway Flow Measurement
Various technologies have been used to measure airway flow.
Many of these techniques were developed strictly for precise
short-term laboratory measurements. These applications
require meticulous attention to detail, including calibration,
placement of a disposable filter element, and operator attend-
ance at all times. Laboratory airflow measurement systems
assume relatively dry gas and well-characterized gas inlet
conditions. The operators of these systems are expected to
pay great attention to accuracy, calibration, repeatability, and
precision when taking a laboratory measurement. Overall, in
respiratory research, the most widely used flow measurement
device is the Fleisch or Lilly-type differential pressure pneu-
motach with a heated microtube or screen orifice. A continu-
ous, bidirectional-airway flow measurement device that can
be placed proximal to the patient and used in critical care
environments has been of great interest. Because of the
requirements of continuous long-term monitoring in the criti-
cal care environment, a different type of flow device was
required. To address the requirements of the critical care
environment, fixed or variable orifice (or aperture) differen-
tial pressure pneumotachometers have been developed for the
last 20 years.
The Philips-Respironics family of
fixed orifice flowand CO
2
/flowsensors
grew from an alliance with a research
group in the Department of Anesthesi-
ology at the University of Utah. Early
prototypes consisted of adding pres-
sure-sensing ports on both the proximal
and distal sides of the optical window
in a mainstream CO
2
cuvette [11].
Later designs followed an iterative
design process including modeling and
testing. This testing often resulted in
features being added for improved
performance under varying inlet condi-
tions and flow-path contamination test-
ing. Scaled-up models of the adult flow
design were constructed, and airflow
patterns were visualized by introducing
filaments of smoke into the airstream.
Measurement of low flows using a
fixed orifice design requires extreme
sensitivity in the differential pressure
measurement signal. Commercially
available low-cost, low-pressure differ-
ential pressure sensors of the early
1990s were drifty, noisy, and sensitive
to orientation. Early analog-to-digital
(A/D) converters were limited in
resolution. Early designs of the flow-
measurement systems used clever de-
signs to compensate for these limita-
tions. New generations of differential
pressure sensors due to improvements
in fabrication and sensor design have
effectively overcome these limitations.
Fixed orifice type of flow-measurement sensors are simpler in
construction than alternative types of flowsensors, lowcost due to
the ability to manufacture as a single piece of molded plastic, and
thus can be disposable. They can be designed to operate in wet,
mucus-filled patient airway circuits and as such not sensitive to
moisture or subject to drift when used for many hours on a patient.
Most importantly, fixed orifice devices do not require individual
calibrations using a syringe as do variable orifice devices.
The Philips-Respironics fixed orifice flow sensors were first
developed as flow-only devices. Shortly after introduction of
Fig. 2. Capnostat 5 CO
2
sensor with electron-
ics (image courtesy of Philips-Respironics).
Fig. 3. FloTrak Elite spirometry module (image courtesy of
Philips-Respironics).
Fig. 1. Adult, pediatric, and neonatal com-
bined CO
2
/flow sensors (image courtesy of
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
the flow-only devices, sensors combining the differential pres-
sure flow-sensor geometry with an integrated carbon dioxide
(CO
2
) measurement cuvette with different size airway adapters
optimized for neonatal/infant, pediatric, and adult patients
were released. In the adult-combined sensor, the CO
2
cuvette
was placed on the patient side of the flow sensor so as to
provide known inlet geometry during expired flow. Test data
showed that the variability of the flow signal caused by differ-
ences in inlet geometry was more pronounced for the breathing
circuit components likely to be placed between the patient and
the sensor. Kofoed et al. [12] indicated variations in inlet con-
ditions (i.e., variations in the velocity profile of the flow gas
incident to the flow sensor) can lead to significant (e.g., up to
15%) error in the flow and volume measurement of some flow-
meters. For example, changing the proximal connections from
a direct ET connection to a connection via an elbow has a
dramatic effect on the cross-sectional profile of the flow enter-
ing the flow sensor. A robust flow-sensor design should toler-
ate such variations with minimal effect on the measurement of
gas flow(Figure 4).
The Philips-Respironics flow-sensor designs have sought to
minimize the effects of changing inlet conditions. These adult
flow and combined CO
2
/flow sensors (Figure 5) feature a tar-
get geometry composed of a center strut and side-mounted
flow restrictions mounted on the side of the center strut (hav-
ing a notch in the strut) designed to minimize localized
streamline effects about the
pressure sensor aperture. Ini-
tial tests of potential flow-sen-
sor designs showed that the
worst inlet condition error
occurs when flow at low
velocity exiting from a small
diameter (6 mm) endotracheal
tube forms a highly variable
flow profile. If the center of
this profile strikes the differ-
ential pressure generator, then
an erroneous flow measure-
ment was observed. The flow-
sensor designs incorporate a
strut intended to break up a
laminar flow profile ahead of
the differential pressure mea-
surement ports with the side-mounted flow restriction cen-
tered between the differential pressure-sensing ports. This
strut-port-restriction design significantly improves its per-
formance compared to variable orifice flowmeters, which
have varying signal strength with changes in upstream geome-
try. This design allows for immunity to unpredictable flow-
velocity profiles, without the need to add excessive length to
the flow-sensor adapter (minimal dead space). Additionally,
placing the pressure taps and flowobstruction on a strut, rather
than on the wall of the adapter, reduces the potential effects of
moisture and sputumon the measurement.
The Philips-Respironics neonatal flow and combined CO
2
/
flow sensors feature a target geometry composed of a center
strut without the side-mounted flow restrictions to maintain an
acceptable level of flow resistance. With the neonatal and
pediatric-combined CO
2
/flow sensors (Figure 6), the pressure
ports are located on opposite sides of the CO
2
measurement
cell section of the sensor, thereby using the pressure perturba-
tion and pressure loss of the CO
2
measurement cell section of
the sensor as part of the flow signal. A first port is placed on
the proximal (nearer to the patient) side, and the second port is
placed on the distal (farther from the patient) side of this sec-
tion. This design is particularly advantageous for use in situa-
tions where the respiratory tidal volumes are extremely small,
because it reduces the volume of rebreathed expired gases. In
both the neonatal and pediatric combined sensor design, a
flow-conditioning strut is placed at the inlet of the flow sensor
to break up jetting in the flow profile of the gas.
Tubing Ports
Strut
Pressure-
Sampling Ports
Flow
Restriction
(a) (b)
Fig. 5. Combination (a) adult and (b) neonatal CO
2
/flow sensors: side, top, and end sections
(image courtesy of Philips-Respironics).
Fig. 6. Perspective view of a combined pediatric CO
2
/flow
sensor (image courtesy of Philips-Respironics).
Flow
Profile
Elbow Wye Combination Adult CO
2
/
Flow Sensor
Fig. 4. Combination adult CO
2
/flow sensor in-circuit with
example expiratory flow velocity profiles (image courtesy of
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
As the equipment (i.e., apparatus) dead space in the ventila-
tor circuit (Figure 7) can be substantial relative to the total
tidal volume, especially when large HME filters are used, a
design goal for each of the flow sensors was to keep the added
dead space to a minimum. As such, the combination pediatric/
adult and neonatal CO
2
/flow sensors described are single
piece designs. Two pressure-sensing lines, one on each side of
the target geometry, are used to measure differential pressure.
The proximal pressure sensing is tapped to a gauge pressure
sensor and used to measure airway pressure.
Even small changes in the geometry of the breathing circuit
tubing relative to the flow sensor can potentially have a signifi-
cant effect on the measured flow. The adult and neonatal flow
sensor are only slightly affected by such changes, whereas
other devices can be significantly affected. Figure 8 illustrates
the various upstream breathing circuit components on the adult
flow sensor during exhalation. Typical design rules for an
annular fixed orifice gas flow sensor recommend a length of
510 diameters of straight tubing upstream of the differential
flow ports as an inlet flow conditioner [13]. Such a large-
volume flow conditioner is not viable in the critical-care appli-
cation. Therefore, a design that is less sensitive to variations in
inlet geometry was needed. For example, it has been demon-
strated that Fleisch pneumotachographs connected between
the wye and ETs exhibit a flow rate-dependent error in
measured flow up to 10% [14]. If sufficient entrance length is
provided in the flow sensor, then laminar flow and a consistent
flow velocity profile can be achieved. However, this is usually
not practical, so entrance length to a flowsensor must be traded
off against the design requirement of minimal dead space.
The addition of a flow sensor to a breathing circuit should
have a minimal impact on the measured quantityflow. One
of the design goals was to minimize the resistance as measured
by the pressure loss across the sensor while maintaining as
large as possible recoverable differential pressure drop
between the ports. It is important to remember that the pres-
sure loss associated with the flow sensor is different than the
measured differential pressure (Figure 9).
With a fixed orifice device, the differential pressure measured
between two pressure taps varies roughly as the square of the flow
[8]. The measured flow should be corrected by use of empirically
determined coefficients because
of variations from this relation-
ship and the assumptions made
in developing the flow equa-
tions that include incompressi-
bility, given the very low
differential pressures, and invis-
cid flow, given that the cross-
sectional area of the device is
large relative to the boundary
layer. The relationship between
the measured differential pres-
sure to flow (L/min) can be
derived by applying Bernoullis
equation to the upstream and
the downstream pressure taps,
equating each equation, rewrit-
ing the pressures at the two taps
as a differential pressure, and
applying the continuity equa-
tion for mass flowyielding
Flow =
P
m
T
std
P
std
T
m
K
DP
_
, (1)
where P
m
, P
std
, T
m
, and T
std
are the measured and standard
pressures (in mmHg) and temperatures (in K), respectively;
K is a correction factor that is flow dependent and includes gas
composition, flow-discharge coefficients, and other factors,
and DP is the differential pressure (in mmHg). The P
m
T
std
/
P
std
T
m
is the ideal gas law correction of calculated flow to
standard conditions. Inspiratory and expiratory phases are
treated separately with regard to temperature and gas compo-
sition. For example, in an unheated breathing circuit supplied
with room air, inspiratory air may be considered to be at near
room temperature and consisting of nominally 21% oxygen
and balance nitrogen and expiratory air at body temperature
(or near body temperature less than 24 C for temperature
drop from lungs to sensor) and consisting of nominally 16%
oxygen, 5% carbon dioxide, and balance nitrogen. While in a
heated circuit with elevated oxygen (such as 60% FiO
2
), the
values used for both temperature and gas composition would
be quite different.
7
Inlet Adaptor
6
5
4
3
2
1
0
1
2
3
4
5
P
e
r
c
e
n
t

E
r
r
o
r
E
l
b
o
w

0
E
l
b
o
w

9
0
E
l
b
o
w

1
8
0
E
l
b
o
w

2
7
0
6

m
m

E
T
T
7

m
m

E
T
T
8

m
m

E
T
T
H
M
E
2
2
1
5

m
m
A
d
a
p
t
o
r
C
O
2

A
d
a
p
t
o
r
Fig. 8. Effect of various inlet conditions for the adult flow sensor at a flow rate of 20 L/min
(with average percent error 2 standard deviation limits shown). Note that the effect of the
elbow adapter was tested at four angles (image courtesy of Philips-Respironics).
Rebreathed
Volume
(Shaded Area)
Patient Interface
ET Tube with
Inflated Cuff
Lost
Dead Space
Trachea
Carina
Fig. 7. Anatomic and apparatus dead space in an intuba-
ted patient (image courtesy of Philips-Respironics).
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Additionally, the design of the front end was simplified so
that each flow-sensor type (e.g., adult, pediatric, neonatal) gen-
erated the same maximal recoverable differential pressure drop
for its maximum flow rate (e.g., 10 in H
2
O). One of the chal-
lenges in choosing a fixed orifice design was the extreme range
of the differential pressures that must be accurately measured.
Dynamic range of flow for each sensor type was approximately
200:1, which required a measurable pressure range of more than
10,000:1. However, because of the relationship between differ-
ential pressure and flow, this resulted for most respiratory flows
a resulting differential pressure of less than 1 in H
2
O. For exam-
ple, adult flow sensor with a full-scale output of 10 in H
2
O at
180 L/min has an output of 3 3 10
4
in H
2
Oat 1 L/min. There-
fore, a low-noise front end with a high-resolution A/Dconverter
(_20 b) needs to utilized.
On-Airway Gas Concentration Measurements
IR gas measurement has been the preferred technology for on-
airway CO
2
gas sensors, since the filter wheel based Hewlett-
Packard CO
2
sensor was introduced in the 1970s [5]. Many of
the disadvantages of on-airway (i.e., mainstream) gas sensors
presented by some authors in the past are primarily technolog-
ical in nature and often relate to prior generations of that
technology. These disadvantages are often listed in older
reviews [15], [16] of the technology while more recent
reviews note otherwise [17]. This includes possible damage
during handling, increased mechanical dead space, issues of
weight on airway, and use limited to only intubated patients.
For example, mainstream IR benches in the past have been
termed vulnerable to costly damage [16]. While earlier IR
benches were vulnerable primarily due to the use of moving
parts such as chopper or filter wheels, some newer mainstream
IR benches utilize all solid-state designs that have been shown
to be robust enough to survive repeated drops onto hard floors
and have been in use for decades nowin high-impact areas such
as the emergency room, ambulances, and transport. Histori-
cally, the primary concerns of mainstream-based systems are
related to size and weight. However, the reductions in both size
and weight have alleviated these concerns. Current-generation
mainstream devices, besides being relatively light and low in
dead space, have generally demonstrated better performance
than conventional sidestream system in terms of signal fidelity
and end-tidal measurements, particularly at higher respiratory
rates in small children [18]. Airway adapter design and advan-
ces in microelectronics and optical technology incorporated in
the latest generation of on-airway IR sensors have reduced the
concerns for dead space and weight, even as the functionality
and complexity of these sensors has increased.
The beginnings of the Philips-Respironics mainstream CO
2
sensor can be found in a company started in 1981 as a division of
hybrid circuit manufacturer (including thermal print heads), which
was looking for the most direct method to measure breathing
and . . . came up with was an inexpensive way to measure exhaled
CO
2
[5]. This device, introduced as an infant respiration monitor,
was a qualitative CO
2
detector based on nondispersive infrared
(NDIR) technology, which was nonratiometric, nonchopped, and
somewhat drifty. After three years, the parent company sold to the
assets to a Midwest-based company (nowknown as BCI), and the
founders of this division found themselves unemployed. Still
believing in the technology, they put together a team at a new
company called Cascadia Technology Corporation to build a
quantitative system that was a true ratiometric, chopped, and sta-
ble NDIR system. This required a number of innovations, includ-
ing building an IR source with sufficient IR energy and very low
thermal mass that could also be pulsed on/off at rates of up to
100 Hz [5]. The original prototype Capnostat CO
2
sensor has
undergone significant changes over the past 20 years resulting in
the current highly integrated Capnostat 5 CO
2
sensor.
Airway Adapter
The measurement cell, referred to as the cuvette, also serves as
the airway adapter that defines the flow path for the gases being
monitored obviating the need for gas sampling, drying, water
trapping, and scavenging. The optical path, shown in Figure 10,
crosses the flow path of the gases through windows in the side-
walls of the cuvette aligned along opposite sides of the flow pas-
sage, allowing the beam of IR radiation to pass through the
cuvette. The fraction of IR radiation absorbed is dependent upon
the path length of the radiation through the gases, the wavelengths
One of the challenges in choosing a fixed
orifice design was the extreme range of the
differential pressures that must be
accurately measured.
Upstream Tap
Downstream Tap
Primary Element
Pipe Loss
Distance
P
f
Pressure Loss
Differential Pressure
Stagnation Pressure
Fig. 9. Theoretical changes in pressure across an orifice plate
along the flow axis The overall pressure loss may be less than
the measured differential pressure signal. Downstream pres-
sure tap may have a significant signal (modified from [8]).
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
of radiation being measured, and the gas molecule itself. This path
is constrained by dead space, resistance, and inner bore (i.e., flow
path) requirements. Also, the optical path length must be suffi-
cient to have adequate sensitivity to measure the gas. One study
using a simplified formula for cuvette path length optimized for
the received radiation intensity with CO
2
of 46% resulted in a
path length of approximately 7 mm[19], which is close to the dis-
tance between windows for existing designs. Additionally, the
end openings of the cuvette need to conform to International
Organization for Standardization (ISO) 5356-1 [20], which speci-
fies connections for 15 and 22 mm sizes intended for general use
in breathing systems. Condensed water or waterlike mixtures can
affect the windows of the cuvette for some optical designs. If
droplets appear within the cuvette optical path, severe scattering
and absorption can occur. Furthermore, the capnometer head may
incorporate a heater to heat the gas in the cuvette or hydrophobi-
cally treat the cuvette, thereby discouraging condensation of drop-
lets in the area of CO
2
measurement.
With the combined neonatal and pediatric CO
2
/flowsensors
from Philips-Respironics, the measurement cell with small
restrictions located on each side of cell chamber serves a dual
function by adding a differential pressure flow signal to CO
2
measurement. With the adult CO
2
/flow sensor, dead space is
less critical and as such the CO
2
measurement cell and flow
measurement portions are separate.
In reusable cuvettes, the windows are formed from materials
such as sapphire with a broad transmission range (up to 5 lm)
and high optical transmission (>80%).The cost of these cuvettes
has been reduced by replacing the relatively expensive sapphire
windows with windows fabricated from a polymer with consist-
ent IR-transmission characteristics and robust mechanical prop-
erties. The major problems encountered in replacing sapphire
cuvette windows with polymer windows is twofold: 1) establish-
ing and maintaining a precise optical path length through the
sample being analyzed affected by such factors as dimensional
stability in the polymeric material, the potential of wrinkles in
the windows, and the need for a system for retaining the win-
dows at precise locations along the optical path and 2) choosing
a polymer with sufficiently broad transmission range and high
optical transmission. Using elastically compressible snap-in
retainer rings in conjunction malleable homopolymer windows
allows a precise optical path through the sample to be main-
tained. Selecting an appropriate polymer allows a significant
portion of the IRradiation impinging upon it to be transmitted as
well as appropriate spectral transmission characteristics (i.e.,
similar absorbance at the data and reference wavelengths).
On-Airway IR Bench
The measurement cell interfaces directly to the IR bench, which
consists of a source and detector assembly located on opposite
sides of the CO
2
cuvette. The source emits IR radiation that
includes the absorption band for carbon dioxide. Carbon dioxide
has a very strong absorption band at 4.26 lm, due to the funda-
mental asymmetric O=C=O stretch, which lies between the two
very-intense water vapor bands. The IR radiation, after passing
through the gas sample flowing in the cuvette, is often filtered
using narrow-band optical filters and measured by appropriate
detectors configured as a single-beam ratiometric design. With
such a design, both the reference (or non-CO
2
sensitive) and data
(or CO
2
sensitive) channels see IR radiation after passing
through the measurement cell. The reference channel uses a fil-
ter with a center wavelength at which little or no CO
2
is
absorbed, and data channel uses a filter with center wavelength
at which CO
2
is strongly absorbed. Using the ratio of these two
measurements helps to allow for cancellation of the effects of
changes in the spectrally independent optical properties of the
system such as signal strength, sample chamber contamination,
and thermal drift. Thus, in principle, only the presence of CO
2
will affect the data channel while leaving the reference channel
unaffected, allowing CO
2
gas concentration to be distinguished
fromthe often larger effects mentioned above.
The filters used, usually of the narrowband type, are typically
manufactured using multiple thin-film vapor depositions on a
silicon substrate. The data channel filter with a center bandpass
wavelength of 4.26 lm and a half power bandwidth typically
less than 0.2 lm can effectively eliminate any interference from
water vapor or even closer bands of N
2
O. One of the most com-
mon detectors in use today, a photoconductive semiconductor
material, is lead selenide (PbSe). PbSe works well for on-airway
It is important to remember that the pressure
loss associated with the flowsensor is different
than the measured differential pressure.
Y
X
Z
Fig. 10. Capnostat 5 sensor with an airway adapter illustrat-
ing single beam ratiometric optical configuration showing
the IR source, beam splitter, mirror, and detectors (with a
reference filter at 3.7 lm and data filter at 4.26 lm) (image
courtesy of Philips-Respironics).
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
CO
2
measurement, in part due to the adequate spectral response
and fast response time constant. The bulk resistance of this mate-
rial drops proportionately to the amount of incident IR energy.
As this material is extremely temperature sensitive, a feedback
temperature control system is used to assure a stable detector
temperature, enabling accurate measurements.
The monitor has traditionally contained the electronics asso-
ciated with control and measurement functions of the IRbench.
Newer designs including the Capnostat 5 sensor from Philips-
Respironics, incorporating advancements in electronics pack-
aging and components, have moved these electronics into the
measurement head, turning the sensor into a complete headless
(i.e., lacking only display and controls) CO
2
analysis system.
This results in greater complexity in the sensor but greater
simplicity to the original equipment manufacturer (OEM)
customer and the end user by enabling concepts such as plug
and play capnography; that is, the user can choose between the
most appropriate type of CO
2
measurement technology for the
patient at hand. An on-airway CO
2
sensor can be selected for
monitoring patients on mechanical ventilation, and a side-
streamsolution for monitoring patients using nasal cannulas.
The Philips-Respironics Capnostat 3 and Capnostat 5 CO
2
sensors also employ a coaxial light path design, wherein a
dichroic beam splitter oriented at 45 to the IR light path is
employed to direct the appropriate band of wavelengths to each
of the two detectors such that each detector receives light
through precisely the same path. This prevents contaminants in
the cuvette from influencing the light reaching only one of the
detectors, causing a ratiometric shift that would otherwise cause
CO
2
measurement errors.
On-Airway Flow and Gas Measurement
Combining the Signal
Challenges of combining flow and CO
2
signals in volumetric
capnography include the following:
compensating for the rebreathing of gas in the tubing
that will cause CO
2
excretion rate (
_
VCO
2
) to be overes-
timated if not corrected
correcting for the delay of CO
2
signal relative to the flow
signals (much more of a consideration with sidestream sys-
tems) requiring proper frequency matching of the signals
consideration of the variations in temperature and vapor
content of expired gas affecting volume correction and
mixed expired CO
2
values (FECO
2
) (more of an issue
with sidestream systems that measure CO
2
and flow at
different points).
Issues of dead space and resistance as well as robustness to
the challenging environment (discussed earlier) at the airway
must also be kept in mind. Devices such as the combined CO
2
/
flow airway adapters allow for minimizing the added dead
space between the wye and elbow and permit accurate and
continuous measurement of carbon dioxide elimination and
volumetric capnogram-derived parameters such as dead space
to tidal volume ratios (Vd/Vt) [21].
During normal conditions, the lung will excrete CO
2
at the
same rate as the total body production rate and there will be no net
change in body CO
2
stores. CO
2
elimination (
_
VCO
2
), which is
often incorrectly referred to as CO
2
production, is the net volume
of CO
2
measured at the mouth or airway, and calculated as the
difference between expired and inspired CO
2
normalized to a
minute.
_
VCO
2
is computed by taking integral of the dot-product
of the flow and CO
2
waveforms over the entire breath cycle and
usuallyreportedat standardtemperature andpressure dry(STPD)
conditions. For breath-by-breath measurements it is calculatedas
_
VCO
2
=
X
FCO
2
(t) 3V(t) 3Dt 3RR; (2)
where FCO
2
(t) and V(t) are the sampled individual values of the
CO
2
and flow waveforms summed over the entire breath, RR is
the respiratory rate, and Dt is the sampling interval. When
present, inspired CO
2
, if not accounted for could result in an
error in the calculation of
_
VCO
2
of several percent [22]. In anes-
thesia and intensive care, components such as filters, HMEs,
connecting tubes, elbows, airway adapters, and suction adapters
are placed between the ET connector and wye, causing partial
rebreathing and therefore inspired CO
2
. Placing the sampling
site more proximal (e.g., on-airway) will potentially allow the
end-tidal CO
2
value to better reflect the alveolar concentration.
If the inspiratory carbon dioxide volume is ignored, the overes-
timation of
_
VCO
2
will increase with decrease in tidal volume
and/or increase in apparatus dead space.
Figure 11 illustrates the cross-multiplication process with the
plot of actual flowand CO
2
waveforms versus time of a mechani-
cal breath delivered in a volume control mode. Because of appa-
ratus dead space from the mainstream sensor, wye, and other
circuit components, a small volume of end-expiratory CO
2
(from
the previous breath) is rebreathed upon the initiation of inspira-
tion. Note that, in this patient, the expiratory CO
2
waveformrises
rapidly to a plateau and the CO
2
volume curve follows that of the
expiratory portion of the flow waveform.
_
VCO
2
would then be
the difference between the expiratory and inspiratory areas of the
dot products. If we plot PCO
2
and volume instead, carbon diox-
ide elimination (
_
VCO
2
), the net volume of CO
2
eliminated can
be viewed as the area between the expiratory and inspiratory
curves (Figure 12). With no rebreathing, the volume of CO
2
eliminated during a breath is the area under the volumetric cap-
nogram. However, the presence of inspiratory CO
2
must be
accounted for when reporting and interpreting
_
VCO
2
[22].
_
VCO
2
does not accurately reflect the underlying physiology
when there are leaks in the collecting system or where
Newer mainstreamIR benches utilize all
solid-state designs that have been shown to be
robust enough to survive repeated drops
onto hard floors.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
conditions exist such that all the gas that is considered part of
the alveolar ventilation volume cannot be measured i.e., pneu-
mothorax with leak or ET cuff leaking on exhalation. Because
of the complex interaction between tidal volume, physiologi-
cal dead space, and alveolar ventilation, the volume of CO
2
excreted by each breath is variable. The results of several
breaths are often averaged in an attempt to decrease the effect
of normal breath-to-breath changes in volume. Depending on
how
_
VCO
2
is used (metabolic measurements versus ventilator
adjustments) [23], different averaging intervals may be
required including a range of averaging intervals such as one
breath, eight breaths, 1 min, 3 min, and longer.
During steady-state conditions, the lung will excrete CO
2
at
the same rate as the total body production rate, and there will
be no net change in body CO
2
stores. In this case, measured
_
VCO
2
is representative of total body production and is propor-
tional to metabolic rate. Because the body retains a large
amount of CO
2
relative to the rate at which CO
2
is produced,
eliminated CO
2
can be different from metabolically produced
CO
2
for a long time (up to 1 h) following a change in ventila-
tion. However, changes in
_
VCO
2
can provide an instantaneous
indication of the change in effective alveolar ventilation [23].
With robust mainstream sensors for flow and carbon dioxide
located at or very close to the same point in the gas stream
the principal issue is one matching the frequency response of
the flow and gas analyzers [24]. Note that, since integration
of the dot product of the flow and CO
2
signals over an entire
breath generates a significant low pass filter, there is little need
to provide low-pass filtering of the input signals prior to integra-
tion. The frequency response characteristic of the flow measure-
ment is primarily limited by the pneumatic pathways between
the fixed orifice sensor and the differential pressure transducer
as well as the sampling rate and front-end hardware and software
low-pass filtering. The frequency response of the carbon dioxide
measurement is limited by the size of the measurement cell (i.e.,
volume interrogated by the IR beam) as well as the thermal
characteristics of the IRsource and time constant of the detector
(if certain technologies are employed). By properly delaying
and filtering the respective signals, a reasonable degree of
frequency matching can be achieved.
Applications Enabled
Thecombinationof flow, pressure, volume, andgas measurements
at the airway with a single integrated sensor provides a powerful
array of measurements to the clinician for the management of both
the spontaneously breathing and mechanically ventilated patient.
These measurements include bedside spirometry, lung mechanics,
and volumetric capnographic measurements such as carbon diox-
ide elimination and dead space. Applications enabled by volumet-
ric capnography include prediction of outcome with ARDS[25], a
screening test for pulmonary embolism [26] and determining dis-
ease staging of chronic obstructive pulmonary disease (COPD)
[27]. Additional measurements requiring perturbations to the sys-
tems also include noninvasive cardiac output via partial CO
2
rebreathingandfunctional residual capacitymeasurement [28].
The measurement cell also serves as the airway
adapter that defines the flowpath for the gases
being monitored obviating the need for gas
sampling, drying, water trapping,
and scavenging.
L
/
m
i
n
m
m
H
g
40
20
Flow
Inspired
CO
2
Volume
Expired
CO
2
Volume
Inspiration Expiration
PCO
2
0
20
40
20
0
Fig. 11. Plot of flow and CO
2
waveforms for an individual venti-
lator delivered breath with dot product showing inspired and
expired CO
2
volumes (image courtesy of Philips-Respironics).
5
0
0 80
Volume (mL)
C
O
2

(
%
)
Expiration
Inspiration
Fig. 12. Plot of PCO
2
versus volume illustrating both the expir-
atory and inspiratory portions. While often the inspiratory por-
tion is negligible, the net CO
2
volume per breath is the
difference between the area under the expired and inspired
portions of the curve or similarly the area within the loop.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Future Challenges
With the competitive pressures for increasing miniaturization,
greater regulation, and the future inclusion of new on-airway
measurements (e.g., oxygen) for enhanced capabilities (e.g.,
metabolic measurements), additional technological, and user
interface/ease-of-use hurdles will need to be overcome. However,
advances in microelectronics and other technologies suggest that
these hurdles will be surmounted in the not-to-distant future.
Michael B. Jaffe received his B.S. degree
from the Cooper Union for the Advance-
ment of Science, an M.S. degree from Dart-
mouth College, and a Ph.D. degree in
biomedical engineering from the University
of Southern California in 1994. He currently
works in advanced technology at Philips-
Respironics in Wallingford, Connecticut.
He has been employed since 1981 in the respiratory field, work-
ing for such companies as Beckman Instruments and Sensor-
medics. He has served as a coeditor of Capnography: Clinical
Aspects, Cambridge University Press, 2004. He is currently the
secretary of IEC/ISO Joint Working Group for the upcoming
revision of ISO 21647, respiratory gas monitors. He is also a
member of several international standards committees relating
to anesthesia and respiratory equipment and a member of the
Anesthesia Patient Safety Foundation Committee on Technol-
ogy. He is an author of 17 peer-reviewed publications and holds
more than 25 U.S. patents in patient monitoring. He is a Mem-
ber of the IEEE.
Joseph A. Orr received his B.S. degree in
electrical engineering from Arizona State
University in 1985 and a Ph.D. degree in bio-
engineering from the University of Utah in
1991. He is now a research associate profes-
sor of anesthesiology and is a codirector of
the Bioengineering Laboratory in the Anes-
thesiology Department at the University of
Utah School of Medicine. He is an author of 26 peer-reviewed
publications, an inventor of more than 40 issued U.S. patents,
and is the cofounder of two medical device companies.
Address for Correspondence: Michael B. Jaffe, Advanced
Technology, Philips-Respironics, LLC, 5 Technology Drive,
Wallingford, CT06492 USA. E-mail: mike.jaffe@philips.com.
References
[1] M. Saklad, M. Sullivan, J. Paliotta, and M. Lipsky, PneumotachographyA
new, low-dead-space, humidity-independent device, Anesthesiology, vol. 51,
no. 2, pp. 149153, 1979.
[2] R. A. Castle, C. J. Dunne, Q. Mok, A. M. Wade, and J. Stocks, Accuracy of
displayed values of tidal volume in the pediatric intensive care unit, Crit. Care
Med., vol. 30, no. 11, pp. 256674, 2002.
[3] S. Silvestri, The influence of flow rate on breathing circuit compliance and
tidal volume delivered to patients in mechanical ventilation, Physiol. Meas.,
vol. 27, no. 1, pp. 2333, 2006.
[4] J. O. Elam, E. S. Brown, and R. H. Ten Pas, Carbon dioxide homeostasis dur-
ing anesthesia. I. Instrumentation, Anesthesiology, vol. 16, no. 6, pp. 87685, 1955.
[5] M. B. Jaffe, Infrared measurement of carbon dioxide in the human breath:
Breathe-through devices from Tyndall to the present day, Anesth. Analg.,
vol. 107, no. 3, pp. 890904, 2008.
[6] R. A. Epstein, A. M. Reznik, and M. A. F. Epstein, Determinants of distor-
tions in CO
2
catheter sampling systems: A mathematical model, Respir. Phys-
iol., vol. 41, no. 1, pp. 12736, 1980.
[7] J. Schena, J. Thompson, and R. K. Crone, Mechanical influences on the cap-
nogram, Crit. Care Med., vol. 12, no. 8, pp. 6724, 1984.
[8] S. A. Kofoed and J. A. Orr, Differential pressure sensor for respiratory moni-
toring, U.S. Patent 5 535 633, July 16, 1996.
[9] S. A. Kofoed, J. A. Orr, and D. R. Rich, Multiple function airway adapter,
U.S. Patent 6 312 389, Nov. 6, 2001.
[10] J. Orr, D. Westenskow, and S. Kofoed, Evaluation of a system to measure
CO
2
production, J. Clin. Monit., vol. 12, no. 6, pp. 471472, 1996.
[11] J. Orr, S. Kofoed, and D. Westenskow, A respiratory flowmeter based on a
modified mainstream CO
2
cuvette, J. Clin. Monit., vol. 9, no. 3, p. 215, 1993.
[12] S. Kofoed, J. Orr, and D. Westenskow, Inlet conditions: Unexpected source of
error in respiratory flow measurements, J. Clin. Monit., vol. 11, no. 4, p. 283, 1995.
[13] R. Miller, Flow Measurement Engineering Handbook, 3rd ed. New York:
McGraw-Hill, 1996.
[14] J. W. Kreit and F. C. Sciurba, The accuracy of pneumotachograh measure-
ments during mechanical ventilation, Am. J. Respir. Crit. Care Med., vol. 154,
no. 4, pp. 91317, 1996.
[15] F. E. Block and J. S. Mcdonald, Sidestream versus mainstream carbon
dioxide analyzers, J. Clin. Monit., vol. 8, no. 2, pp. 13941, 1992.
[16] D. Hess, Capnometry and capnography: Technical aspects, physiologic
aspects, and clinical applications, Respir. Care, vol. 35, no. 6, pp. 55773, 1990.
[17] K. R. Ward and K. R. Yealy, End-tidal carbon dioxide monitoring in emergency
medicine, Part 1: Basic principles, Acad. Emerg. Med., vol. 5, no. 6, pp. 62836, 1998.
[18] R. C. Pascucci, J. A. Schena, and J. E. Thompson, Comparison of a side-
stream and mainstream capnometer in infants, Crit. Care Med., vol. 17, no. 6,
pp. 5602, 1989.
[19] J. S. Putyatina, Account of length of cuvette for capnometer, in Proc. 4th
Annu. 2003 Siberian Russian Workshop on Electron Devices and Materials, July
1-4, 2003, p. 259.
[20] Anaesthetic and Respiratory EquipmentConical ConnectorsPart 1:
Cones and Sockets, 3rd ed., ISO 53561-12004.
[21] R. H. Kallet, B. M. Daniel, O. Garcia, and M. A. Matthay, Accuracy of
physiologic dead space measurements in patients with acute respiratory distress
syndrome using volumetric capnography: Comparison with the metabolic monitor
method, Respir. Care, vol. 50, no. 4, pp. 4627, 2005.
[22] P. H. Breen, E. R. Serina, and S. J. Barker, Measurement of pulmonary
CO
2
elimination must exclude inspired CO
2
measured at the capnometer sampling
site, J. Clin. Monit., vol. 12, no. 3, pp. 2316, 1996.
[23] V. Taskar, J. John, A. Larsson, T. Wetterberg, and B. Jonson, Dynamics of
carbon dioxide elimination following ventilator resetting, Chest, vol. 108, no. 1,
pp. 196202, 1995.
[24] A. C. Jackson and A. Vinegar, A technique for measuring frequency response of
pressure, volume, and flow transducers, J. Appl. Physiol., vol. 47, no. 2, pp. 4627, 1979.
[25] T. J. Nuckton, J. A. Alonso, R. H. Kallet, D. M. Daniel, J. F. Pittet, J. F. Eisner,
and M. A. Matthay, Pulmonary dead-space fraction as a risk factor for death in the acute
respiratory distress syndrome, N. Engl. J. Med., vol. 346, no. 17, pp. 12816, 2002.
[26] F. Verschuren, G. Liistro, R. Coffeng, F. Thys, J. Roeseler, F. Zech, and
M. Reynaert, Volumetric capnography as a screening test for pulmonary embo-
lism in the emergency department, Chest, vol. 125, no. 3, pp. 84150, 2004.
[27] P. V. Romero, B. Rodriguez, D. De Oliveira, L. Blanch, and F. Manresa,
Volumetric capnography and chronic obstructive pulmonary disease staging,
Int. J. Chron. Obstruct. Pulmon. Dis., vol. 2, no. 3, pp. 38191, 2007.
[28] J. S.Gravenstein, M. B. Jaffe, and D. A. Paulus, Eds., Capnography: Clinical
Aspects. Cambridge, U.K.: Cambridge Univ. Press, 2004, p. 441.
Airway adapter design and advances in
microelectronics and optical technology
incorporated in the latest generation of on-
airway IR sensors have reduced the concerns
for dead space and weight.
____________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Quest for the Artificial
Pancreas
T
ype 1 diabetes mellitus (T1DM) is characterized by the
destruction of the insulin-producing b-cells in the pancreas.
Exogenous insulin administration is therefore required to
regulate the blood glucose concentration. The goal of dia-
betes management is to maintain homeostasis and blood glucose
near normal levels (80140 mg/dL) and thus to avoid immediate
life-threatening situations, such as severe hypoglycemia and ketoa-
cidosis, and long-term complications, such as cardiovascular dis-
ease, nephropathy, neuropathy, and retinopathy.
Treatment of T1DM requires either multiple daily insulin
injections or continuous subcutaneous (SC) insulin infusion
(CSII) delivered via an insulin infusion pump. Both treatment
modes necessitate frequent blood glucose measurements (eight to
ten times/day, including fasting, pre- and postprandial, before
bedtime, and in the middle of the night) to determine the daily
insulin requirements for maintaining near-normal blood glucose
levels [1][3]. With the advent of continuous glucose sensing,
which reports interstitial glucose concentrations (that reflect the
blood glucose) approximately every minute, and the development
of hardware and algorithms to communicate with and control
insulin pumps, the vision of closed-loop control of blood glucose
is approaching a reality. In individuals without diabetes, blood
glucose is controlled by various neural and hormonal inputs from
the brain, gut, liver, and pancreas that respond to various situa-
tions such as meals, stress, and exercise. Aclosed-loop systemfor
T1DM must be responsive to all daily challenges in life and be
able to accurately predict blood glucose levels in advance. This
closed-loop system, or artificial pancreas, would include an amal-
gamof features necessary to bring a person with diabetes as close
as possible to normoglycemia using SCinsulin therapy.
The artificial pancreas can be seen as a puzzle, represented in
Figure 1, in which several integral pieces formthe core, including
communication, modeling, control algorithms, learning, meal
detection, and safety algorithms. Optional features such as teleme-
dicine can further improve the lives of people with T1DM. With
current technology and development, the pieces, as described in
detail in the following sections, are coming together, though many
still need further refinement. The remaining challenges include
detecting and overcoming irregular or variable features in the life
of a person with T1DM, such as stress, illness, and exercise.
Communication: Artificial Pancreas System
The primary piece of the puzzle involves reliable communication
and secure data transfer between the glucose sensor, a control
algorithm, and the insulin delivery pump. This research team has
developed a novel platform for use in clinical trials and made
progress toward a portable unit. This prototype encapsulates
communication between the control algorithm and the pump and
sensors, as illustrated in Figure 2. The artificial pancreas system
(APS) facilitates communication and provides a simple and clear
human interface that presents all the information to the physician.
Additionally, the system allows all events to be logged electroni-
cally [4]. It also ensures safety by integrating interlocks, check-
lists, and alarms. Furthermore, the plug-and-play concept allows
the APS to serve as a test bed for various control algorithms with-
out regard to the specific sensors or pumps involved.
The APS has been validated and evaluated using a testing plat-
form known as hardware in the loop [5]. A simulated subject with
T1DM, instead of a human volunteer, is connected to the continu-
ous glucose monitor (CGM) and a CSII pump with the APS and
control algorithm in the same fashion as during a clinical trial.
Such a test allows researchers to conduct a full system validation
and verification including extreme scenarios. The current version
of the APS (v. 2.6) has received the approval of the Food and Drug
Administration (FDA) through a Master File (MAF-1625) to be
used in human clinical trials [6]. The APS allows investigators to
conduct clinical trials using the Insulet OmniPod pump and the
two CGMs, the DexComSeven systemand the Abbott FreeStyle
Navigator system, with their control algorithms. Additional
pumps and sensors will be added as they become available.
Control Algorithms
There is a rich control literature with numerous examples of
algorithms applied to diabetes [7]. However, regulation of blood
glucose is different from conventional process control, such as
temperature regulation by a thermostat or cruise control in a car.
Because of the complexity of human physiology, traditional
control algorithms need to be customized to meet the challenges
of controlling glucose concentrations in people with T1DM.
The effects of exogenous insulin and carbohydrate on blood
glucose concentrations are affected by diverse factors. For con-
trol purposes, these variations can be addressed at two levels.
During 24-h periods, the process is, effectively, continuous. A
BY REBECCA A. HARVEY, YOUQING WANG,
BENYAMIN GROSMAN, MATTHEW W. PERCIVAL,
WENDY BEVIER, DANIEL A. FINAN, HOWARD ZISSER,
DALE E. SEBORG, LOIS JOVANOVI
C, FRANCIS J. DOYLE, III,

AND EYAL DASSAU
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Combining Technology with Treatment
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
controller can be tuned to the individual
subject based on their insulin needs, thus
accounting for inter subject variations. In
contrast, when considering variations on a
day-to-day basis, the process can be viewed
as a batch operation, and the control algo-
rithms should be updated as such to
account for intrasubject variations. For this
reason, we have focused on model predic-
tive control for the continuous process and
iterative learning control (ILC) for the
batch process.
Model Predictive Control
The control of blood glucose concentra-
tions using SC insulin delivery, without
administering counterregulatory hormone
or hormones, is at a distinct disadvantage
compared with the non diabetic state, because SC insulin
administration is not able to be absorbed fast enough to balance
the rapid gut absorption of recently ingested carbohydrates [8].
Although a change in the basal infusion rate may improve post-
meal glucose excursion if food is not ingested, an automatic
change in basal rate can be dangerous.
The intersubject variation in insulin pharmacokinetics is rela-
tively large, requiring a personalized control algorithm [9]. These
factors motivate the use of MPC using a
subject-specific model of SC insulin and
carbohydrate effects on blood glucose con-
centrations. Basic MPC formulations can
be solved analytically and in some cases are
equivalent to classical proportional-inte-
gral-derivative (PID) controllers [10].
In addition to physical limitations on
the SC insulin pump, further constraints
should be considered, such as safety con-
straints that explicitly consider active
insulin, and soft constraints designed to
achieve a specific set of hierarchical
objectives [11]. With the addition of
constraints, there is typically no analyti-
cal solution, and a constrained quadratic
program (QP) must be solved. The solu-
tion of such a QP involves an iterative
optimization procedure with no guarantee of convergence.
Additionally, such problems are computationally expensive,
and in the context of a portable system, this could be an unde-
sirable drain on battery power.
The application of multiparametric programming techni-
ques to MPC (mpMPC) offers a means by which to retain opti-
mal control while minimizing online computation [12]. This
technique involves the reformulation of the MPC problem via a
multiparametric variable; all the optimal solutions are then
obtained offline for the expected operating conditions. Within
this expected operating space, there are a finite number of critical
regions within which an affine function of the state vector defines
the optimal control law. The online problem therefore reduces to
the evaluation of an affine function contained within a lookup
table. This technique has been shown to be feasible for intrave-
nous (IV) insulin delivery [13] and SCinsulin delivery [14].
Learning Patterns for Delivery
It is evident that there exist repetitive cycles in glucose-insulin
dynamics. For example, people generally consume meals at
similar times from day to day, and the corresponding meal
sizes are similar from one day to the next. In addition, insulin
sensitivity changes in a predictable manner throughout the
day because of circadian variation of hormone levels.
To exploit the repetitive nature of glucose-insulin dynam-
ics, run-to-run (R2R) control has been used to update the meal
bolus for outpatient subjects [15][17]. ILC can be considered
an enhanced version of R2R in a two-dimensional sense;
hence, ILC has been successfully implemented in silico for
glucose regulation. The general structure of ILC can be repre-
sented as follows:
u(t, k) = u(t, k 1) r(t, k), (1)
T1DM Subject
Glucose Sensor
and Transmitter
Glucose
Sensor
Interface
Device
R
F

S
i
g
n
a
l
R
F

S
i
g
n
a
l
RS-232/USB
RS-232/USB
Infusion Pump
Interface Device
Notebook Computer
(Interfaces and Controller)
CSII Pump
Fig. 2. The components of a closed-loop system with radio
frequency (RF) communication links. RS-232 is a standard for
serial binary data transfer. USB: universal serial bus.
Fig. 1. The essential pieces of the artifi-
cial pancreas puzzle.
The primary piece of the puzzle involves reliable
communication and secure data transfer
between the glucose sensor, a control
algorithm, and the insulin delivery pump.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
where u is the control signal designed by ILC, r is termed as
the updating law for ILC, t denotes the time step, and k is the
batch index. The equation indicates that the control signal for
the current batch is composed of the control signal from the
previous batch and the updating law.
In general, there are two ways to adjust the insulin delivery
rate using ILC, as shown in Figure 3 [18]: first, ILC can be uti-
lized to determine the insulin delivery rate directly, a method
that is called direct ILC; second, the delivery rate can be deter-
mined by a local controller and ILC is used to adjust the local
controller, a combination that is known as indirect ILC.
A key concern for direct ILC is the design of the updating
law. In our previous studies, e.g., [19], the updating law of ILC
was designed by MPC. On the other hand, for indirect ILC, two
essential issues need to be established: what algorithms are used
to design the local control and which parameters of the local
controller are updated by ILC. In our previous studies [20], [21],
the local control was designed by MPC, and the set point for
MPCwas updated by ILC. This novel combination is referred to
as learning-type MPC (L-MPC). L-MPC has been tested on
11 adult in silico subjects fromthe UVa/Padova diabetes simula-
tor, which has been accepted by the FDA as a surrogate for ani-
mal studies in developing a future artificial pancreas [22], [23].
The tracking performance under L-MPC is superior to that with
MPC, with an average reduction of the tracking error of 21.1%.
Specifically, the closed-loop control results for the adult average
subject under L-MPC are shown in Figure 4. After 20 days, the
blood glucose concentrations can be kept within 68145 mg/dL.
y(t, k)
y(t, k)
r (t, k)
r (t, k)
u(t, k)
u(t, k)
y(t, k)
y(t, k)
u(t, k1)
q(t, k1)
q(t, k)
q(t, k)
Memory
Memory
Local
Controller
Updating
Law
Updating
Law
+
+
+ +
Set Point
Yr
T1DM
T1DM
Glucose
Glucose
Insulin
ILC
ILC
(a)
(b)
u(t, k)
Insulin
Fig. 3. Block diagrams for the direct and indirect ILC, where t is
the time index and k denotes different days: (a) direct ILC; (b)
indirect ILC. The solid lines denote the real-time information;
the dotted lines denote the information in the previous batch;
the components in the dashed frames comprise the ILC.
The artificial pancreas systemfacilitates
communication and provides a simple and
clear human interface that presents all the
information to the physician.
0 2 4 6 8 10
(a)
12 14 16 18 20
0
100
200
300
400
G
l
u
c
o
s
e

(
m
g
/
d
L
)
Glucose 60 mg/dL
180 mg/dL Set Point
0 2 4 6 8 10 12 14 16 18 20
10
1
10
0
10
1
10
2
I
n
s
u
l
i
n

(
U
/
h
)
Time (day)
(b)
(c)
0 5 10 15 20
50
100
150
200
250
G
l
u
c
o
s
e

(
m
g
/
d
L
)
0 5 10 15 20
0
5
10
I
n
s
u
l
i
n

(
U
/
h
)
Time (h)
(d)
Fig. 4. Control performance under L-MPC in 20 days for adult
average, where the subject consumes three meals a day at
7:00 a.m., 1:00 p.m., and 6:00 p.m., with fixed amounts of car-
bohydrate, 60, 100, and 70 g, respectively. (a) Glucose con-
centration, (b) insulin delivery rate where logarithmic scale
was used for the Y axis, (c) last days glucose, and (d) last
days insulin [20].
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Patient Modeling
Different mathematical models can serve as candidates for the
model in an MPC-based artificial pancreas; however, MPC will
work best with a personalized model, since the glucose-insulin
dynamics differ greatly between subjects. The mathematical
model for MPC, for example, could be based on a simplified
two differential equation system such as the Bergman minimal
model [24]. More detailed models use several subsystems
including a SC phase, resulting in higher-order models that also
take into account exogenous delivery of meals and insulin
[25][27]. However, complicated models often contain varia-
bles difficult to measure physiologically and parameters that
need to be reevaluated in vivo. Moreover, methodological mod-
els that include physiological insights include nonlinear terms
that will transform the linear-MPC problem into a much more
difficult nonlinear-MPC problem that is less robust and more
difficult to adjust than linear-MPC. Alternatively, input/output
black-box models, such as autoregressive, exogenous input
(ARX)-models, have demonstrated the ability to capture the
blood glucose concentration dynamics of living subjects with-
out taking into account physiological insights [28]. The ARX
models are a more reasonable choice for control, being easy to
personalize, computationally efficient to estimate, and readily
updated with well-known recursive identification methods
[29]. The ARXmodels are based on a standard linear regression
that provides a straightforward way to personalize them to dif-
ferent subjects.
Meal Detection Algorithm
A fully functioning artificial pancreas must regulate glucose
levels during fasting periods as well as compensate for meals.
In control terminology, meals can be considered as disturbances
for a feedback control algorithm. An additional challenge is that
an automated systemthat is responsible for regulating the blood
glucose concentration of people with T1DM needs to perform
under different compliance levels, without relying on the ability
of the user to announce the meal or to correctly estimate its con-
tent. The concept of reliable and robust meal detection was
presented by Dassau et al. as a way to overcome compliance
issues and to allowa secondary line of defense [30].
The meal detection algorithm (MDA) combines four ways
to analyze glucose values and its rate of change to flag a meal
event. A set of binary inputs from the algorithm is evaluated
by a voting algorithm that issues an alarm to the controller in
case a predetermined majority is reached. Evaluation of the
algorithm on historical clinical data has indicated that a meal
can be detected at a mean time of 30 min from the onset of the
meal, and, importantly, the mean serum glucose was on an
average only 21 mg/dL higher at detection than at the meal
time. This rapid detection allows an automated control algo-
rithm to intervene and compensate for the meal disturbance.
The MDA can be used by any type of control algorithm to
inform the controller of a meal disturbance and then issue the
appropriate meal response, as demonstrated by simulation
with MPC controller [31]. The use of MDA can improve over-
all blood glucose control by minimizing the likelihood of
hyperglycemia due to no control action and, later, hypoglyce-
mia due to aggressive control action.
Safety Considerations
A functional artificial pancreas will be required to demon-
strate both efficacy and safety in controlling blood glucose
concentrations. As part of the development of such a system,
various alarms, interlocks, and algorithms are used to provide
a safe device for the end user. The main objective is to prevent
the overdosing of insulin that can result in immediate life-
threatening consequences that are associated with severe
hypoglycemia.
300
250
200
150
100
50
0
7 14 21 4
Time of Day (h)
G
l
u
c
o
s
e

(
m
g
/
d
L
)
300
250
200
150
100
50
0
7 14 21 4
Time of Day (h)
(a)
(b)
G
l
u
c
o
s
e

(
m
g
/
d
L
)
Glucose (mg/dL)
180 (mg/dL)
60 (mg/dL)
Fig. 5. Example of a glucose tracing with an ARX-based
MPC (b) with and (a) without the IOB constraint [33]. Note
that with the IOB constraint, glucose remains well above the
hypoglycemia threshold of 60 mg/dL.
The intersubject variation in insulin
pharmacokinetics is relatively large, requiring
a personalized control algorithm.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Insulin on Board
Understanding the dynamics and kinetics of exogenous insulin
in the human body can be utilized to ensure that the correct
dosage of insulin is administered without erroneous overdos-
ing. Insulin on board (IOB) is an empirical estimation
technique that predicts the amount of available insulin over a
period of time following exogenous insulin admissions.
The IOB concept was introduced in modern CSII pumps
with different decay insulin curves, both linear and curvilin-
ear, with a duration of action of two to eight hours, as part of
the smart pump bolus wizard [32]. The use of the IOB concept
was demonstrated as a practical dynamic constraint in pre-
venting overdelivery of insulin by the artificial pancreas; see
Figure 5 for an in silico example [33]. The synergistic effect
of IOB with MPCwill result in an aggressive but safe therapy.
The motivation for including IOB constraints was demon-
strated in a fully automated closed-loop clinical study at San-
sum Diabetes Research Institute in 2007. Initially, no
constraint was included, and trials had to be interrupted by
the physician to correct for excessive insulin delivery [Figure
6(a)]. The IOB constraint was tested and found to be
effective in the avoidance of excessive insulin delivery, as
shown in Figure 6 (b).
Hypoglycemia Alarming
Amajor goal of diabetes therapy is to ameliorate the long-term
microvascular and macrovascular complications associated
with the disease, i.e., retinopathy, nephropathy, neuropathy,
cardio-, and cerebrovascular disease [34]. The Diabetes Con-
trol and Complications Trial Research Group demonstrated
that intensive insulin therapy slows the onset and progression
of these complications, although intensive insulin therapy car-
ries an elevated risk of severe hypoglycemia or low blood
glucose concentrations [34]. To best improve the life of people
with T1DM, therapy to relieve long-term complications must
be supplemented with features to resolve more immediate
problems such as hypoglycemia.
One of the most feared risks for a person with T1DM is the
occurrence of hypoglycemia. Hypoglycemia causes symp-
toms that range from sweating, confusion, and dizziness in the
short term, to seizure, coma, and death if severe and prolonged
[35]. Nocturnal hypoglycemia is especially critical, as meals
180
160
140
120
G
l
u
c
o
s
e

(
m
g
/
d
L
)
G
l
u
c
o
s
e

(
m
g
/
d
L
)
100
80
60
9:00 a.m. 12:00 p.m. 3:00 p.m. 6:00 p.m.
Time
Time
180
160
140
120
100
80
60
9:00 a.m. 12:00 p.m. 3:00 p.m. 6:00 p.m.
Time
9:00 p.m.
60
50
I
n
s
u
l
i
n

I
n
f
u
s
i
o
n

R
a
t
e

(
U
/
h
)
40
30
20
10
0
60
50
I
n
s
u
l
i
n

I
n
f
u
s
i
o
n

R
a
t
e

(
U
/
h
)
40
30
20
10
0
9:00 a.m. 12:00 p.m. 3:00 p.m. 6:00 p.m.
Time
9:00 a.m. 12:00 p.m. 3:00 p.m. 6:00 p.m. 9:00 p.m.
Control Session
CGM
YSI
Meal
Control
Session
CGM
YSI
Meal
Control Session
Automatic Delivery
Physician Override
Control Session
Automatic Delivery
Physician Override
(a) (b)
Fig. 6. Clinical results of a fully automated closed-loop control from a clinical study conducted at Sansum Diabetes Research
Institute using MPC with and without IOB. (a) A trial without IOB, where excessive insulin was delivered and resulted in physician
override of the controller. (b) The results of MPC with IOB, conducted on the same subject. Note that there were no physician
overrides in (b). In both trials, the control session started at the hyperglycemia state to challenge the control algorithm.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
are less likely to be ingested. Additionally, evidence shows
that the counterregulatory response to hypoglycemia is
blunted during sleep, along with a marked suppression of
auditory response by the sleeping person [36].
Arobust alarm, which can awaken the person or another house-
hold member, is essential to minimize hypoglycemia and avoid
its complications. According to Buckingham et al. [36], seizure
occurs two to four hours after the onset of severe hypoglycemia,
providing a windowof opportunity in which corrective action can
be taken to avoid seizures. Corrective actions may include manual
correction or, after sufficient time with no response, pump suspen-
sion [36]. Ideally, an alarm will be sounded sufficiently far in
advance of the onset of severe hypoglycemia to avoid it altogether
or at least soften the landing, or blunt the rate of change and ulti-
mate nadir of blood glucose. Such an alarmis essential to provide
a layer of safety to the artificial pancreas.
An alarm suite that can easily be implemented into current
CGM systems has been developed by the authors along with
several collaborators. The hypoglycemia prediction algorithm
(HPA) includes five predictive algorithms incorporated into a
scheme in which a voting threshold must be reached to set off an
alarm[37]. Validation followed using a historical set of 22 Clinical
Research Center admissions of T1DM subjects with induced
hypoglycemia. With a voting threshold of four out of five algo-
rithms, 82% of events were predicted from 35 to 55 min before a
glucose threshold of 80 mg/dL [38]. The HPA suite appears to be
an effective systemto enable alarming and eventual pump suspen-
sion prior to a hypoglycemia event. This safety feature is a crucial
fail-safe for automatic glucose control in the artificial pancreas.
Clinical Results
The authors conducted a series of clinical trials at Schneider Child-
rens Medical Hospital, Israel, consisting of open-loop model
identification and closed-loop insulin delivery in people with
T1DM. ARX models were identified from ambulatory data com-
prising CGMmeasurements, CSII pump records, and subject-esti-
mated carbohydrate content (CHO). The best ARX model was
chosen based on model validation performance. The mpMPCcon-
trol lawwas formulated with this model and included the personal-
ized IOBsafety constraint, resulting in a lookup table representing
optimal control. This lookup table included ten to 100 entries and
required approximately 100 kBof RAM[39], [40].
Hyperglycemia was induced prior to the initiation of closed-
loop control. During the closed-loop trial, the subject consumed
unannounced meals of up to 30 g CHO.
The control algorithm restored euglycemia
in less than three hours. No hypoglycemic
events occurred. In no case did the physi-
cian override the controller. Exemplary
results are given in Figure 7[41].
Telemedicine
The other piece of the puzzle that can both
improve the safety of the system and assist
with system maintenance is telemedicine.
Telemedicine, the use of information and
communication technology to support
medical care and decision-making, has
been broadly used to enhance medical
care and to provide up-to-date medical
information [42]. Although telemedicine
is not a required piece in the evolution of
the artificial pancreas, it should and will
play an integral role in this emerging
technological treatment. Telemedicine
will allow both data logging and monitor-
ing of artificial pancreas users by their
medical providers and even remote tuning
of their control algorithm.
From a safety point of view, the teleme-
dicine feature of the artificial pancreas
will not only bring a means to remotely
monitor glucose, as is the case for parents
Basal
Delivered
IOB Constraint
Forecast
Set Point
Sensor
Prediction
0
10
20
30
I
n
s
u
l
i
n

D
e
l
i
v
e
r
y

(
U
/
h
)
9 a.m. 11 a.m. 1 p.m. 3 p.m. 5 p.m. 7 p.m.
9 a.m. 11 a.m. 1 p.m. 3 p.m. 5 p.m. 7 p.m.
50
100
150
200
250
G
l
u
c
o
s
e

S
e
n
s
o
r

(
m
g
/
d
L
)
Time
30 g CHO Meal
(Unannounced)
1.9 U Insulin
2.8 U Insulin
Initializing
Fig. 7. In this clinical trial, insulin was delivered in quantities to commensurate with the
subjects insulin requirements, thus normalizing blood glucose concentrations after clin-
ically induced hyperglycemia and an unannounced meal. The controller delivered
an appropriate amount of insulin without overdosing due to the IOB constraint [41].
Afully functioning artificial pancreas must
regulate glucose levels during fasting periods as
well as compensate for meals.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
of children with T1DM, but
also include an active layer of
defense. This safeguard can be
embedded such that the artifi-
cial pancreas will switch from
closed-loop to open-loop mode
or suspend insulin delivery in
cases of a hypoglycemia event
or system failure [43]. The use
of global positioning system
(GPS) technology with a moni-
toring system within a teleme-
dicine application can add
essential security to the artifi-
cial pancreas. Such an applica-
tion, termed E911, has been
suggested by the authors as a
way to monitor, alert, and
locate people with T1DM that
are experiencing a hypoglyce-
mia event. The proposed sys-
tem would issue a short message service (SMS), or text
message, to a predefined list in case of a pending event. An
example of a blood glucose tracing including a hypoglycemia
event with the visualization of an automatic SMS with GPS
coordinates is shown in Figure 8.
Present Challenges
The successful operation of an artificial pancreas will require
careful monitoring to detect underlying changes in the sub-
jects glucose-insulin dynamics because of factors such
as stress, illness, or unusual exercise. If such abnormal condi-
tions can be detected and diagnosed in a timely manner, the AP
can make appropriate adjustments.
Automatic Detection of Stress States
A recent study by the coauthors has demonstrated that a statis-
tically based monitoring technique, principal component analy-
sis (PCA), can detect stress states in Type 1 diabetes subjects for
ambulatory conditions [44]. Nine adults (six men and three
women) participated in the study. The data consisted of continu-
ous (5-min) glucose measurements, insulin pump records of basal
rates and bolus amounts and times, and subject-recorded estimates
of the times and CHOcontent of meals.
An empirical PCA model was calculated for each subject
from several days of normal data (i.e., the training data). The
subjects were then administered prednisone for three consecu-
tive days, causing decreased insulin sensitivity and thereby
simulating a physiological stress state. Then, newdata for both
stress and normal days (i.e., the test data) were evaluated. A
total of 37 test days for the nine subjects were available.
Of the 37test days, 33(89%) were classifiedcorrectly. Thus, the
proposed monitoring technique was able to differentiate between
normal and stress days with a high degree of accuracy. Typical
monitoring results for a single subject are shown in Figure 9 [44].
250
200
150
Alert User
SMS
100
50
6
:
0
0

p
.
m
.
6
:
3
0

p
.
m
.
7
:
0
0

p
.
m
.
7
:
3
0

p
.
m
.
8
:
0
0

p
.
m
.
8
:
3
0

p
.
m
.
Time
G
l
u
c
o
s
e

(
m
g
/
d
L
)
CGM Data
60 mg/dL
70 mg/dL
Alarm
(a) (b)
Fig. 8. (a) Blood glucose tracing in which a hypoglycemia event is forecasted. (b) The user is
alerted first, and after not responding for 15 min (evident by the unchanged trajectory), an
SMS is sent, including the GPS coordinates of the subject [43].
Insulin on board (IOB) is an empirical estimation
technique that predicts the amount of available
insulin over a period of time following
exogenous insulin admissions.
120
100
80
60
40
Q
20
0
N
1
N
2
N
3
N
4
Day
95% Confidence Limit
N
5
S
1
S
2
S
3
P
1
P
2
P
3
Fig. 9. PCA results for Subject 2 using the Q statistic. The PCA
model was developed from the normal training days N
1

N
4
. All subsequent days are test data. The stress days are
denoted by S
1
S
3
. The poststress days are denoted by P
1

P
3
. Poststress days are not included in the results because of
the unknown residual effects of the prednisone during these
days. Day N
5
is a normal test day [44].
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
The Q statistic is a standard metric of PCA model accuracy. A Q
value above the 95% confidence limit indicates an abnormal (i.e.,
stress) day.
Exercise
Another major impediment to closed-loop glucose control is
the physiological change related to glucose control induced by
exercise. To achieve automatic control, exercise must be
detected without the user informing the controller of impend-
ing exercise. This is especially important for young patients,
as they are generally very active and would be unable to
inform a controller of every exercise session.
The challenge in detecting exercise is that glucose monitor-
ing alone may not give an accurate picture of the influence of
exercise. The effect of exercise on glucose-insulin dynamics
changes over time, as opposed to the transient perturbation of a
meal or insulin bolus [45]. Adding to the complexity of detect-
ing exercise is the subjects fitness and the duration and inten-
sity (mild, moderate, or strenuous) of exercise, all of which
influence the bodys response differently. For example, mild-
to-moderate exercise can cause hypoglycemia in people with
T1DM, with the added complication of a blunted autonomic
nervous system response [46]; on the other hand, strenuous
activity may cause hyperglycemia after exercise without
recovery even when insulin is delivered [47]. As a result, the
duration and intensity of exercise must be detected with a cer-
tain degree of confidence to maintain glycemia within a safe
range. To do this accurately, a second metric must be used to
supplement blood glucose monitoring. The chosen measure-
ment must be both convenient for the patient and representa-
tive of the effect of exercise on glucose-insulin dynamics.
Challenges to Commercializing the Artificial Pancreas
It is envisioned that the artificial pancreas will be produced in
stages, each one a step closer toward full automation. Initial devi-
ces, such as those currently available in Europe, consist of over-
night pump suspension to prevent hypoglycemia. Intermediate
models may require user-defined inputs for frequent occurrences
such as meals and exercise. As physiological sensing technology
improves, user-defined inputs may be eliminated entirely, thus
effectively delivering automated glucose control of similar
performance to that found in nature. Market availability of any of
these products will depend upon regulatory body approval, which
will require extensive clinical trials. Product use will ultimately
rely upon a cost-benefit analysis by health-insurance companies.
Summary
The various components of the artificial pancreas puzzle are
being put into place. Features such as communication, control,
modeling, and learning are being realized presently. Steps have
been set in motion to carry the conceptual design through
simulation to clinical implementation. The challenging pieces
still to be addressed include stress and exercise; as integral
parts of the ultimate goal, effort has begun to shift toward over-
coming the remaining hurdles to the full artificial pancreas.
The artificial pancreas is close to becoming a reality, driven by
technology, and the expectation that lives will be improved.
Acknowledgments
This work was supported by the Otis Williams Fund at the
Santa Barbara Foundation, the National Institute of Health
(grants R21-DK69833 and R01-DK068683), and the Juvenile
Diabetes Research Foundation, which provided funding from
2006 to the present for the Sansum Diabetes Research Insti-
tute and the University of California, Santa Barbara (UCSB).
Dr. Cesar Palerm is acknowledged for his role in some of the
clinical studies discussed in this article.
Rebecca A. Harvey received her B.S.
degree in chemical engineering from North-
eastern University in Boston, Massachusetts,
in May 2008, graduating summa cum laude.
Along with her undergraduate course work,
she completed 18 months of cooperative
education at three major industrial firms in
the Boston area. Currently, she is in the
Ph.D. program in the Chemical Engineering Department at
UCSB. She has been a guest investigator at the Sansum Diabetes
Research Institute since January 2009. Her research focus is on
developing algorithms for monitoring and control of T1DM.
Youqing Wang received his B.Sc. degree
in mathematics from Shandong University,
Jinan, Shandong, China, in 2003, and his
Ph.D. degree in control science and engi-
neering from Tsinghua University, Beijing,
China, in 2008. He worked as a research
assistant in the Department of Chemical
Engineering, Hong Kong University of Sci-
ence and Technology from February 2006 to August 2007. He
is currently a senior investigator in the Department of Chemi-
cal Engineering, UCSB. From March 2008 to present, he has
been a guest investigator at Sansum Diabetes Research Insti-
tute. His research interests include the artificial pancreatic
b-cell, iterative learning control, model-predictive control, and
fault-tolerant control. He is a member of the American Diabe-
tes Association (ADA) and American Institute of Chemical
Engineering (AIChE) and is a Member of the IEEE.
Benyamin Grosman received his B.Sc. degree in 1997, his
M.Sc. degree in 2002, and his Ph.D. degree in 2008, all at the
Technion Institute of Technology, Haifa, Israel. He currently
Amajor goal of diabetes therapy is to
ameliorate the long-termmicrovascular and
macrovascular complications associated
with the disease.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
participates in the postdoctoral program in
the Department of Chemical Engineering at
UCSB, working in parallel on T1DMcontrol
and modeling, and on a heat-shock project
that involves modeling of both protein-sig-
naling pathways and heat-transfer phenom-
ena. His research interests are in process
systems engineering, with a particular em-
phasis on control and evolutionary computational methods.
Matthew W. Percival received his
M.Eng. degree in chemical engineering
with first-class honors from Imperial Col-
lege London in 2005 and is currently
pursuing the Ph.D. degree in chemical
engineering at the UCSB. He has been a
guest investigator at the Sansum Diabetes
Research Institute since 2006. His research
interests include the application of control theory to enhance
drug delivery in next-generation biomedical devices.
Wendy Bevier received her undergraduate
degree at UC Davis and earned a masters
degree at the University of Oregon. She
received her Ph.D. degree in biology/physiol-
ogy at UCSB and was a postdoctoral fellow
in the Stanford University School of Medi-
cine. She is an associate investigator with
Sansum Diabetes Research Institute and is
coordinating studies for the artificial pancreas project. She is
responsible for patient recruitment, conducting experiments, and
collecting and organizing data. The studies are a collaboration
with the UCSB, Department of Chemical Engineering, and the
goal of the project is to develop mathematical formulas (algo-
rithms) that will work with external continuous blood glucose
sensors and insulin pumps to control blood glucose levels. She
has a certificate from Stanford University Medical Center as a
Fellow in medicine/gerontologyexercise physiology and is a
member of the American Physiological Society and ADA.
Daniel A. Finan received his B.S. degree
from the University of Colorado in 2003 and
his Ph.D. degree fromUCSBin 2008, both in
chemical engineering. His graduate and post-
doctoral research has focused on modeling,
monitoring, and control strategies for type 1
diabetes, and he also has a strong interest in
clinical diabetes research. He is currently a
postdoctoral fellow at the Technical University of Denmark in
the Copenhagen area. In his spare time, he writes crossword puz-
zles for The NewYork Times and other publications.
Howard Zisser received his B.S. degree
from the University of Florida, Gainesville,
and his M.D. degree from the Johns Hop-
kins University School of Medicine, Balti-
more, Maryland. He is currently the director
of Clinical Research at Sansum Diabetes
Research Institute, Santa Barbara, Califor-
nia. He also serves as an adjunct professor
in the Department of Chemical Engineering at the University of
Santa Barbara. His current research interests include the testing
of new drugs and devices for the treatment of diabetes. He is
working with a team of researchers to combine these new tech-
nologies into an artificial pancreas.
Dale E. Seborg received his B.S. degree from
the University of Wisconsin and his Ph.D.
degree from Princeton University. He is a
professor and vice chair of chemical engineer-
ing at the University of California. He has
published more than 200 articles on process
control and related topics and is the coauthor
of a widely used textbook, Process Dynamics
and Control, 2nd ed., in 2004, which has been translated into Chi-
nese. He is the recipient or corecipient of several national awards
that include the American Statistical Associations Statistics in
Chemistry Award in 1994 and the American Automatic Control
Councils Education Award in 1993.
Lois Jovanovic received her M.D.,
F.A.C.E., F.A.C.P., F.A.C.N. degrees. Her
area of specialization is endocrinology. She
is the chief executive officer and chief scien-
tific officer of Sansum Diabetes Research
Institute, Santa Barbara, California. She is a
clinical professor of medicine at the Univer-
sity of Southern California-Keck School of
Medicine in Los Angeles, a member of the internal medicine
faculty at Santa Barbara Cottage Hospital, an attending
physician at the Santa Barbara County Diabetes/Endocrine
Clinic, and an adjunct professor in the Biomolecular Sci-
ence and Engineering Program at the UCSB.
Francis J. Doyle, III, received his B.S.E.
degree from Princeton in 1985, C.P.G.S.
degree from Cambridge in 1986, and Ph.D.
degree from Caltech in 1991, all in chemical
engineering. He is the associate dean for
research in the College of Engineering at
UCSB, and he is the associate director of the
Army Institute for Collaborative Biotechnol-
ogies. He holds the Duncan and Suzanne Mellichamp Chair in
process control in the Department of Chemical Engineering as
well as appointments in the Electrical Engineering Department
and the Biomolecular Science and Engineering Program. He is a
Fellow of the IEEE and a Fellow of the International Federation
of Automatic Control (FAC). He served as the editor-in-chief of
IEEE Transactions on Control Systems Technology from 2004
to 2009 and currently holds the associate editor positions with
Journal of Process Control, SIAM Journal on Applied Dynami-
cal Systems, and Royal Societys Interface. His research interests
are in systems biology, network science, modeling and analysis
of circadian rhythms, drug delivery for diabetes, model-based
control, and control of particulate processes.
Eyal Dassau received his B.Sc., M.Sc.,
and Ph.D. degrees in chemical engineering
from the Technion Israel Institute of
Technology, Haifa, Israel, in 1999, 2002,
and 2006, respectively. He is currently a
senior investigator and diabetes research
manager at the UCSB, Santa Barbara, Cal-
ifornia. His research interests include the
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
design and control of the artificial pancreatic beta cell for
T1DM, application of systems theory, modeling, simulation,
optimization and control to medical, biological, and other
complex systems such as process and product design, with
emphasis on medical and biomedical applications. He is a
Member of the IEEE. He is a senior member of the AIChE
and a member of the ADA.
Address for Correspondence: Francis J. Doyle, III, Depart-
ment of Chemical Engineering, University of California,
Santa Barbara, CA 93106-5080 USA. E-mail: frank.doyle@
icb.ucsb.edu.
References
[1] L. Jovanovic, Rationale for prevention and treatment of postprandial
glucose-mediated toxicity, Endocrinologist, vol. 9, no. 2, pp. 8792, 1999.
[2] L. Jovanovic and C. M. Peterson, Insulin and glucose requirements during
the first stage of labor in insulin-dependent diabetic women, Amer. J. Med.,
vol. 75, pp. 607612, Oct. 1983.
[3] C. M. Peterson, L. Jovanovic, and L. H. Chanoch, Randomized trial of
computer-assisted insulin delivery in patients with type I diabetes beginning pump
therapy, Amer. J. Med., vol. 81, pp. 6972, July 1986.
[4] E. Dassau, H. Zisser, C. C. Palerm, B. A. Buckingham, L. Jovanovic, and F.
J. Doyle, III, Modular artificial b-cell system: A prototype for clinical research,
J. Diab. Sci. Technol., vol. 2, no. 5, pp. 863872, 2008.
[5] E. Dassau, C. C. Palerm, H. Zisser, B. A. Buckingham, L. Jovanovic, and F.
J. Doyle, III, In silico evaluation platform for artificial pancreatic beta-cell
developmentA dynamic simulator for closed-loop control with hardware-in-the-
loop, Diab. Technol. Ther., vol. 11, pp. 187194, Mar. 2009.
[6] E. Dassau, H. Zisser, F. J. Doyle, III, and L. Jovanovic, Sansum/UCSB arti-
ficial pancreas software (APS), Food and Drug Administration, Silver Spring,
MA, Master File MAF-1625, 2009.
[7] R. Hovorka, The future of continuous glucose monitoring: Closed loop,
Curr. Diab. Rev., vol. 4, pp. 269279, Aug. 2008.
[8] R. Hovorka, Continuous glucose monitoring and closed-loop systems,
Diab. Med., vol. 23, pp. 112, Jan. 2006.
[9] U. Fischer, W. Schenk, E. Salzsieder, G. Albrecht, P. Abel, and E. J. Freyse,
Does physiological blood glucose control require an adaptive control strategy?
IEEE Trans. Biomed. Eng., vol. BME-34, no. 8, pp. 575582, 1987.
[10] M. W. Percival, H. Zisser, L. Jovanovic, and F. J. Doyle, III, Closed-loop
control and advisory mode evaluation of an artificial pancreatic b-cell: Use of
proportional-integral-derivative equivalent model-based controllers, J. Diab. Sci.
Technol., vol. 2, no. 4, pp. 636644, 2008.
[11] P. Dua, F. J. Doyle, III, and E. N. Pistikopoulos, Multi-objective blood
glucose control for type 1 diabetes, Med. Biol. Eng. Comput., vol. 47, no. 3,
pp. 343352, 2009.
[12] E. N. Pistikopoulos, V. Dua, N. A. Bozinis, A. Bemporad, and M. Morari,
On-line optimization via off-line parametric optimization tools, Comput. Chem.
Eng., vol. 24, no. 27, p. 188, 2000.
[13] P. Dua, F. J. Doyle, III, and E. N. Pistikopoulos, Model-based blood
glucose control for type 1 diabetes via parametric programming, IEEE Trans.
Biomed. Eng., vol. 53, no. 8, pp. 14781491, 2006.
[14] M. W. Percival, E. Dassau, H. Zisser, L. Jovanovic, and F. J. Doyle, III,
Closed-loop control of an artificial pancreatic beta cell using multi-parametric
model predictive control, in Proc. 2008 AIChE Annu. Meeting, Philadelphia,
PA, 2008, session 160, paper a.
[15] H. Zisser, L. Jovanovic, F. J. Doyle, III, P. Ospina, and C. Owens, Run-to-run con-
trol of meal-related insulin dosing, Diab. Technol. Ther., vol. 7, no. 1, pp. 4857, 2005.
[16] H. Zisser, C. C. Palerm, W. C. Bevier, F. J. Doyle, III, and L. Jovanovic,
Clinical update on optimal prandial insulin dosing using a refined run-to-run
control algorithm, J. Diab. Sci. Technol., vol. 3, no. 3, pp. 487491, 2009.
[17] C. C. Palerm, H. Zisser, W. C. Bevier, L. Jovanovic, and F. J. Doyle, III,
Prandial insulin dosing using run-to-run control: Application of clinical data and
medical expertise to define a suitable performance metric, Diab. Care, vol. 30,
pp. 11311136, May 2007.
[18] Y. Wang, F. Gao, and F. J. Doyle, III, Survey on iterative learning control,
repetitive control, and run-to-run control, J. Process Control, vol. 19, no. 10,
pp. 15891600, 2009.
[19] Y. Wang, E. Dassau, and F. J. Doyle, III, Closed-loop control of artificial
pancreatic b-cell in type 1 diabetes mellitus using model predictive iterative learn-
ing control, IEEE Trans. Biomed. Eng., to be published.
[20] Y. Wang, H. Zisser, E. Dassau, L. Jovanovic, and F. J. Doyle, III, Model
predictive control with learning-type reference: Application in artificial pancreatic
b-cell, AIChE J., to be published.
[21] Y. Wang and F. J. Doyle, III, Indirect iterative learning control: Application
on artificial pancreatic b-cell, in Proc. 21st Chinese Control and Decision Conf.,
Guilin, China, 2009, pp. 17391744.
[22] B. P. Kovatchev, C. Breton, C. Dalla Man, and C. Cobelli, In silico preclin-
ical trials: A proof of concept in closed-loop control of type 1 diabetes, J. Diab.
Sci. Technol., vol. 3, no. 1, pp. 4455, 2009.
[23] B. P. Kovatchev, D. M. Breton, C. Dalla Man, and C. Cobelli, In silico model
and computer simulation environment approximating the human glucose/insulin utili-
zation, Food Drug Administration, Silver Spring, MA, Master File MAF-1521, 2008.
[24] R. N. Bergman, L. S. Phillips, and C. Cobelli, Physiologic evaluation of
factors controlling glucose tolerance in man: Measurement of insulin sensitivity
and beta-cell glucose sensitivity from the response to intravenous glucose, J.
Clin. Invest., vol. 68, pp. 14561467, Dec. 1981.
[25] R. Hovorka, V. Canonico, L. J. Chassin, U. Haueter, M. Massi-Benedetti, M.
O. Federici, T. R. Pieber, H. C. Schaller, L. Schaupp, T. Vering, and M.
E. Wilinska, Nonlinear model predictive control of glucose concentration in sub-
jects with type 1 diabetes, Physiol. Meas., vol. 25, no. 4, pp. 905920, 2004.
[26] M. E. Wilinska, L. J. Chassin, H. C. Schaller, L. Schaupp, T. R. Pieber, and
R. Hovorka, Insulin kinetics in type-I diabetes: Continuous and bolus delivery
of rapid acting insulin, IEEE Trans. Biomed. Eng., vol. 52, pp. 312, Jan. 2005.
[27] C. Dalla Man, R. A. Rizza, and C. Cobelli, Meal simulation model of the
glucose-insulin system, IEEE Trans. Biomed. Eng., vol. 54, pp. 17401749, Oct. 2007.
[28] D. A. Finan, H. Zisser, L. Jovanovic, W. C. Bevier, and D. E. Seborg, Prac-
tical issues in the identification of empirical models from simulated type 1 diabe-
tes data, Diab. Technol. Ther., vol. 9, pp. 438450, Oct. 2007.
[29] D. A. Finan, F. J. Doyle, III, C. C. Palerm, W. C. Bevier, H. C. Zisser,
L. Jovanovic, and D. E. Seborg, Experimental evaluation of a recursive model
identification technique for type 1 diabetes, J. Diab. Sci. Technol., vol. 3, no. 5,
pp. 11921202, 2009.
[30] E. Dassau, B. W. Bequette, B. A. Buckingham, and F. J. Doyle, III, Detec-
tion of a meal using continuous glucose monitoring (CGM): Implications for an
artificial b-cell, Diab. Care, vol. 31, no. 2, pp. 295300, 2008.
[31] E. Dassau, P. Herrero, H. Zisser, B. A. Buckingham, L. Jovanovic, C. Dalla
Man, C. Cobelli, J. Veh , and F. J. Doyle, III, Implications of meal library &
meal detection to glycemic control of type 1 diabetes mellitus through MPC con-
trol, in Proc. 17th IFAC World Congr., Seoul, Korea, 2008, pp. 42284233.
[32] H. Zisser, L. Robinson, W. Bevier, E. Dassau, C. Ellingsen, F. J. Doyle, III,
and L. Jovanovic, Bolus calculator: A review of four smart insulin pumps,
Diab. Technol. Ther., vol. 10, no. 6, pp. 441444, 2008.
[33] C. Ellingsen, E. Dassau, H. Zisser, B. Grosman, M. W. Percival,
L. Jovanovic, and F. J. Doyle, III, Safety constraints in an artificial pancreatic
b-cell: An implementation of model predictive control with insulin-on-board,
J. Diab. Sci. Technol., vol. 3, no. 3, pp. 536544, 2009.
[34] The Diabetes Control and Complications Trial Research Group, The effect
of intensive treatment of diabetes on the development and progression of long-
term complications in insulin-dependent diabetes mellitus, N. Engl. J. Med.,
vol. 329, no. 14, pp. 977986, 1993.
[35] E. P. Widmaier, H. Raff, and K. T. Strang, Vanders Human Physiology:
The Mechanisms of Body Function, 11th ed. Boston: McGraw-Hill Higher Educa-
tion, 2008.
[36] B. Buckingham, D. M. Wilson, T. Lecher, R. Hanas, K. Kaiserman, and
F. Cameron, Duration of nocturnal hypoglycemia before seizures, Diab. Care,
vol. 31, pp. 21102112, Nov. 2008.
[37] E. Dassau, F. M. Cameron, H. Lee, B. W. Bequette, F. J. Doyle, III,
G. Niemeyer, P. Chase, and B. Buckingham, Real-time hypoglycemia prediction
using continuous glucose monitoring (CGM), a safety net to the artificial
pancreas, in Proc. 68th American Diabetes Association Meeting, San Francisco,
CA, 2008, vol. 57 (Suppl. 1), p. A13.
[38] E. Dassau, F. Cameron, H. Lee, B. W. Bequette, H. Zisser, L. Jovanovic, H.
P. Chase, D. M. Wilson, B. A. Buckingham, and F. J. Doyle, III, Real-time
hypoglycemia prediction suite using continuous glucose monitoring (CGM), a
safety net for the artificial pancreas, Diab. Care, accepted for publication.
[39] E. Dassau, H. Zisser, M. W. Percival, B. Grosman, L. Jovanovic, and F.
J. Doyle, III, Design, validation and clinical evaluation of a fully automated artificial
pancreatic b-cell with unannounced meal using mpMPC and IOB, in Proc. 3rd
Conf. Advanced Technologies and Treatments for Diabetes, Basel, Switzerland, 2010.
[40] E. Dassau, H. Zisser, M. W. Percival, B. Grosman, L. Jovanovic, and F.
J. Doyle, III, Clinical evaluation of fully automated artificial pancreatic b-cell
with unannounced meal using mpMPC and insulin-on-board, Diabetes Care,
submitted for publication.
[41] M. W. Percival, B. Grosman, E. Dassau, H. Zisser, L. Jovanovic, and F.
J. Doyle, III, Automated insulin delivery system demonstrates safe and effica-
cious control of glycemia, in Proc. 69th American Diabetes Association Meet-
ing, New Orleans, LA, 2009, abstract 3LB.
[42] R. Bellazzi, Telemedicine and diabetes management: Current challenges
and future research directions, J. Diab. Sci. Technol., vol. 2, no. 1, pp. 98104,
2008.
[43] E. Dassau, L. Jovanovic, F. J. Doyle, III, and H. Zisser, Enhanced 911/GPS
wizard: A telemedicine application for the prevention of severe hypoglycemia
Monitor, alert and locate, J. Diab. Sci. Technol., vol. 3, no. 6, pp. 15011506,
2009.
[44] D. A. Finan, Modeling and monitoring strategies for type 1 diabetes,
Ph.D. dissertation, Dept. Chem. Eng., UC-Santa Barbara, 2008.
[45] M. Breton, Physical activity-the major unaccounted impediment to closed
loop control, J. Diab. Sci. Technol., vol. 2, no. 1, pp. 169174, 2008.
[46] A. C. Ertl and S. N. Davis, Evidence for a vicious cycle of exercise and
hypoglycemia in type 1 diabetes mellitus, Diab. Metab. Res. Rev., vol. 20,
pp. 124130, Mar.-Apr. 2004.
[47] T. H. Mitchell, G. Abraham, A. Schiffrin, L. A. Leiter, and E. B. Marliss,
Hyperglycemia after intense exercise in IDDM subjects during continuous sub-
cutaneous insulin infusion, Diab. Care, vol. 11, pp. 311317, Apr. 1988.
________
________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Optical Coherence
Tomography
D
uring breast-conserving surgeries, axillary lymph
nodes draining from the primary tumor site are
removed for disease staging. Although a high number
of lymph nodes are often resected during sentinel and
lymph-node dissections, only a relatively small percentage of
nodes are found to be metastatic, a fact that must be weighed
against potential complications such as lymphedema. Without a
real-time in vivo or in situ intraoperative imaging tool to
provide a microscopic assessment of the nodes, postoperative
paraffin section histopathological analysis currently remains
the gold standard in assessing the status of lymph nodes. Optical
coherence tomography (OCT), a high-resolution real-time
microscopic optical-imaging technique previously used to
image breast cancer tumor margins intraoperatively in humans
and lymph-node microarchitecture in a rat animal model, is
being presented for the intraoperative ex vivo imaging and
assessment of axillary lymph nodes. OCT provides real-time
microscopic images up to 2 mm beneath the tissue surface in
axillary lymph nodes. Normal (13), reactive (1), and metastatic
(3) lymph nodes from 17 human patients with breast cancer
were imaged intraoperatively with OCT. These preliminary
clinical studies have identified scattering changes in the cortex,
relative to the capsule, which can be used to differentiate
normal from reactive and metastatic nodes. These optical scat-
tering changes are correlated with inflammatory and immuno-
logical changes observed in the follicles and germinal centers.
These results suggest that intraoperative OCT has the potential
to assess the real-time node status in situ, without having to
physically resect and histologically process specimens to visu-
alize microscopic features.
Breast Cancer
Breast cancer continues to affect a significant proportion of
women, as 192,370 new cases of invasive breast cancer and
62,280 cases of ductal carcinoma in situ (DCIS) are expected
in the United States during 2009, making it the most widely
diagnosed cancer (26% of new cases) in women [1]. The num-
ber of deaths attributed to breast cancer (40,610 expected in
2009second only to lung cancer) has started to decrease
over the last few years, largely attributed to the effectiveness
of breast-cancer screening [1]. These developments have led to
the increased detection of breast-cancer lesions at earlier
stages, resulting in smaller breast lesions and a decreased like-
lihood of lymph-node involvement and cancer metastasis. The
management and treatment of breast cancer has continued to
improve, as evidenced by the gradual increases in the five-year
survival rates for all stages of breast cancer [2]. Currently,
patients electing to undergo lumpectomies or mastectomies
will typically have a sentinel lymph-node dissection (SLND),
which may be accompanied by the removal of additional axil-
lary lymph nodes via axillary lymph-node dissection (ALND)
to help stage disease progression [2]. Staging is based on the
size of the primary tumor, the involvement of lymph nodes,
and the metastatic spread to secondary sites.
Mapping of Lymph Nodes
Lymph-node mapping is currently performed by injecting a
radioactive tracer (technetium-99) and/or a dye (isosulfan
blue) preoperatively near the site of the primary tumor. The
tracer and/or dye is allowed to circulate for several hours, after
which the sentinel node(s) are located and resected during
lumpectomy or mastectomy [3]. Other lymphatic mapping
methods include standard X-ray and computed tomography
(CT) in conjunction with contrast agents, or fluorescence tech-
niques using near-infrared (NIR) quantum dots [4], [5] or
indocyanine green (ICG) dye [6], which provide better local-
ization of the lymph node. Lymph nodes are typically only
removed to stage the disease progression via histopathological
analysis. A positive sentinel node status will prompt surgeons
to consider patients for axillary lymph-node dissections,
further node sampling via biopsies, or whole-body imaging to
determine the extent of the metastatic spread of cancer.
Current and Experimental Nodal
Assessment Techniques
The current standard of care for lymph-node assessment is
postoperative paraffin section histopathological analysis. No
real-time intraoperative microscopic technique is being cur-
rently widely used for the assessment of lymph nodes. The use
of frozen-section histopathology of lymph nodes is also an
accepted nodal assessment method, but often takes 2030 min
to perform, lengthening the time that the patient is under
BY FREDDY T. NGUYEN, ADAM M. ZYSK, ERIC J. CHANEY,
STEVEN G. ADIE, JAN G. KOTYNEK, URETZ J. OLIPHANT,
FRANK J. BELLAFIORE, KENDRITH M. ROWLAND,
PATRICIA A. JOHNSON, AND STEPHEN A. BOPPART
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
The Intraoperative Assessment of
Lymph Nodes in Breast Cancer
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
general anesthesia [7], [8]. Several other intraoperative techni-
ques are currently under investigation, including touch imprint
cytology [9], [10], molecular marker assays [11], frozen-sec-
tion analysis [7], [8], and Fourier transform infrared micro-
spectroscopy [12]. All these techniques require the physical
sectioning of the lymph node, leading to the physical destruc-
tion of the nodal structural integrity and cellular architecture
to perform the molecular assay, staining, or probing protocols.
The use of wide-field NIR fluorescence imaging has also been
developed to better visualize the dynamics of the lymphatic
system[13], [14].
Clinical Importance of Lymph-Node Assessment
Lymph nodes serve as the primary site for filtering, sequester-
ing, and degrading foreign particles that travel through the
lymphatic system. They are typically classified as normal,
reactive, or metastatic. Both a reactive node and a metastatic
node are relatively enlarged in size as lymphocytes and macro-
phages are recruited to the node in response to the presence of
foreign particles or abnormal cells. However, in a metastatic
node, there is an increased presence and involvement of cancer
cells in addition to the inflammatory response. When the
lymphatic system becomes impaired because of the disruption
of the lymphatic network, lymphatic obstruction develops,
leading to the accumulation of lymph fluid or lymphedema.
Currently, nearly 20% of the patients who undergo ALND suf-
fer from these cases of edema [15], [16]. In addition, a recent
study reported that the presence of micrometastases in the
axillary lymph nodes contributed to a decrease in the five-year
survival rates for women with early breast cancer [17].
OCT in Breast Cancer
OCT is a high-resolution microscopic optical imaging
technique that yields real-time multidimensional images of
subsurface tissue structure [18][24]. The use of NIR light
enables micron-scale resolution, providing structural images
on a resolution scale similar to histopathology. The penetra-
tion depth in tissue has been found to be 12 mm and is highly
dependent on the type of tissue imaged. OCT has been investi-
gated in a large number of clinical applications ranging from
ophthalmology, cardiovascular disease, Barretts esophagus,
and more recently in oncology [25]. OCT has also been used
to image tumor margins for breast cancer in an NMU carcino-
gen-induced rat mammary model and intraoperatively for the
assessment of tumor margins [21], [26], [27]. High-scattering
signals are often attributed to a combination of the increase in
nuclear-to-cytoplasm (N/C) ratio and the increase in cellular
density during the focal proliferation of cancer cells [28], [29].
OCT is a suitable imaging modality for assessing the lymph-
node architecture [30] but, more importantly, for imaging and
assessing the morphological changes observed in the cortex,
which can be used to differentiate between normal nodes and
reactive and metastatic nodes. By imaging through the capsule
of an intact lymph node, OCT can provide this assessment
without compromising the structural integrity of the lymph
node, making it a potential candidate as an in vivo nodal-
assessment technique.
Materials and Methods
Instrument
A clinical spectral domain OCT (SD-OCT) system [31] was
used to assess sentinel lymph nodes for this feasibility imaging
study (Figure 1). The OCT system employed a superlumines-
cent diode (B&W Tek Inc., SLD1C), centered at 1,310 nm
with a bandwidth of 92 nm. A 60-mm achromatic lens was
used to focus 4.75 mW of light to a 35.0-lm spot size (trans-
verse resolution) with an in-focus depth-of-field (confocal
parameter) of 1.47 mm. The broad bandwidth of the source
resulted in an axial resolution of 8.3 lm in free space or 5.9
lm in tissue. The penetration depth of OCT imaging was 12
mm in the axillary lymph nodes. With camera exposure times
ranging from 24.4 to 408.4 ls, the measured signal-to-noise
ratio (SNR) ranged from 96 to 116 dB, respectively. The
imaging system acquired OCT images at a rate of 5,000 axial
scans per second or approximately eight frames per second for
an image with 600 axial scans covering a lateral distance of 10
mm. A commercial OCT system with similar instrument spec-
ifications and performance parameters as the custom-built
Staging is based on the size of the primary
tumor, the involvement of lymph nodes, and the
metastatic spread to secondary sites.
SLD1,310 nm
OC
5/95 FC
Col
Col
Col
Reference
Arm
Sample Arm
Detector
(a) (b)
RM
PP
Fig. 1. (a) Photograph of the clinical system with (b) clinical
spectral domain optical coherence tomography system
schematic. Light from a superluminescent diode (SLD) cen-
tered at 1,310 nm is directed into the optical circulator (OC)
that passes the light into the fiber coupler (FC) that splits the
light into the reference arm (5%) and the sample arm (95%).
The light is collimated through a set of fiber collimators (Col).
The reflected light from each arm is coupled back together
through a set of polarization paddles (PP) and interfered
with each other through the 5/95 fiber coupler (FC) and
spectrally dispersed onto a line scan camera that serves as
the detector.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
system used for this study was also used. Both systems used
light sources centered at 1,310 nm with a bandwidth of 90
100 nm, used the same detection system, and had similar opti-
cal elements in the sample arm, yielding comparable axial and
transverse resolutions. The main difference between the two
systems was in the beam-scanning mechanism, which enabled
the custom-built system to scan over longer lateral distances
(10 mm) to produce two-dimensional (2-D) images, while the
commercial system was able to scan more than a 5 35 mm
2
region to produce three-dimensional (3-D) volumetric images.
The acquisition times were approximately 25 s for a single 2-
D image (10 mm) for the custom-built system and approxi-
mately 25 s for a full 3-D volumetric block of images (5 35
mm
2
) for the commercial OCT system.
Imaging Protocol
The patients identified and recruited for this study had primary
breast tumors (invasive and/or in situ carcinoma) in need of
surgical resection and/or sentinel lymph-node dissection, as
determined by their physicians at Carle Foundation Hospital
and Carle Clinic Association, Urbana, Illinois, based on previ-
ous radiological films, biopsy results, and other relevant diag-
nostic information. Patients were consented before each
surgery based on the approved protocols from the Institutional
Review Boards at the University of Illinois at Urbana-Cham-
paign and Carle Foundation Hospital. After resection of the
sentinel and/or axillary lymph node(s) by the surgeon and
before microscopic assessment by the pathologist, the lymph
node was imaged in the operating room using the clinical
OCT system. Of the lymph nodes resected during each proce-
dure, one sentinel lymph node, as determined by the surgeon,
was imaged per patient for this feasibility study. In cases with
multiple sentinel lymph nodes, only the first resected sentinel
node was chosen and imaged. No additional nodes were
imaged because of time constraints, and no other criteria were
used in determining which lymph node would be imaged. In
cases where the sentinel lymph node could not be clearly
distinguished from other resected lymph nodes, the first axil-
lary lymph node excised was imaged. The intact outer capsule
of the sentinel lymph node was exposed for imaging without
affecting the structural integrity of the tissue specimen. The
OCT beam was laterally scanned across the node over a
10-mm distance to produce a 2-D cross-sectional OCT image.
Subsequently, multiple parallel OCT images (spaced 1 mm
apart) were acquired orthogonal to the long axis of the node.
In cases where the commercial OCT system was used,
5 35 mm
2
regions of images were acquired covering the
entire node. Upon completion of the imaging session, the sen-
tinel node was returned to the surgical staff for standard speci-
men processing and transported to the pathology department
for sectioning, staining, and histopathological analysis.
OCT Image Processing and Evaluation Protocol
OCT images were standardized by subtracting the background
power spectrum from the raw data, resampling the data using
the cubic spline interpolation technique, and displaying the
images on the same intensity scale to account for day-to-day
system variations. The background power spectrum was
acquired by acquiring an image with the sample arm blocked,
and recording the power spectrum of the light reflected in the
reference arm. This process removes image artifacts that are
inherent to the OCT system at the time of imaging. Fifteen
patients were enrolled as part of this study to evaluate the
potential of OCT imaging for the real-time intraoperative
assessment of lymph nodes using the custom-built system.
Two patients were enrolled additionally to evaluate the
commercial OCT system and the advantages of 3-D volumet-
ric data acquisition. OCT images were evaluated by a single
interpreter intraoperatively, allowing for consistent identifica-
tion of suspicious areas based on the level and distribution of
scattering intensity in the outer layers and cortex of the nodes,
and the ability to distinguish the boundary between the
capsule and cortex of the nodes. Since nodal status is not cur-
rently determined intraoperatively by surgeons but rather
postoperatively by pathologists, no information about the
nodes was transmitted between the surgeons or surgical staff
and the researcher imaging and evaluating the OCT images.
Using these image features, a single researcher in a single set-
ting evaluated and classified OCT images into normal and
abnormal categories where both metastatic and reactive nodes
were considered abnormal.
Histopathological Image Evaluation Protocol
According to the standard of care, the sentinel lymph node was
bisected, paraffin embedded, and sectioned for histopathologi-
cal analysis in the pathology department. The tissue sections
were stained with hematoxylin and eosin (H&E) and in some
cases also stained immunohistochemically for CK7 to further
confirm the presence of and to characterize the cancer cells in
the sentinel node. The histology slides were digitized using a
light microscope at 43 magnification and autostitched togeth-
er (Adobe Photoshop CS3) to provide a single montage for
viewing and comparison purposes. The H&E-stained histology
slides were reviewed by board-certified pathologists and clas-
sified as normal, reactive, or metastatic. The histopathological
processing and analysis were performed as part of the standard
of care and reported by the pathology department.
Results
Specimen Information
A total of 30 sentinel lymph nodes and an additional 106 axil-
lary lymph nodes were surgically resected from 17 patients as
Lymph nodes serve as the primary site
for filtering, sequestering, and degrading
foreign particles that travel through the
lymphatic system.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
determined by the surgeons as part of the standard of care. A
subset (16 sentinel, one axillary) of the lymph nodes (one per
patient) chosen, based on the order they were resected, were
imaged with OCT intraoperatively. Of all the nodes resected,
11 (8%) were found to be metastatic and even fewer, 2 (1%)
were found to be reactive, resulting in 91% of all nodes
resected being classified as normal. When considering only
the sentinel lymph nodes as labeled by the surgeons, four of
the 30 (13%) sentinel lymph nodes and one of the 30 (3%)
were classified as metastatic and reactive, respectively, by
histopathological analysis, resulting in 84% of the sentinel
nodes being reported as normal. OCT images of 17 lymph
nodes from 17 patients were used to identify unique image
features that could potentially be used to classify lymph nodes
as normal, reactive, or metastatic. OCT images were acquired
from 16 of the 30 sentinel lymph nodes from 16 patients, and
one axillary lymph node from the 17th patient, where identifi-
cation of the sentinel node was not possible by the surgeon. Of
the 17 nodes imaged with OCT, histopathological analysis
classified three nodes as metastatic, one node as reactive, and
the remaining 13 nodes as normal. The OCT images from two
of the three metastatic nodes were excluded from the study, as
they could not be adequately interpreted as the node capsule
was not well exposed from the surrounding fatty tissue, plac-
ing the nodal architectural features beyond the imaging depth
of OCT. The third metastatic node was imaged using the
commercial OCT system.
Representative OCT Images
Representative intraoperative images from normal (Figures 2
and 4), reactive (Figure 3), and metastatic (Figure 5) axillary
lymph nodes are shown. The normal lymph nodes are all char-
acterized by a distinct capsule that is highly scattering, in
comparison to the lower scattering cortex, as seen in the repre-
sentative OCT images (Figures 2 and 4). Under OCT, the
transverse sinuses separating the lymphoid follicles of the
lymph node can be observed in Figure 4. The corresponding
H&E-stained histological slides show normal nodes with
expected architecture that are either largely lipid filled (Figure
2) or contain more eosinophilic (pink) structures indicative of
the presence of intracellular and extracellular proteins (Figure
4). The subcapsular sinus separates the capsule from the cor-
tex, which is organized into discrete lymphoid nodules with
germinal centers observable in normal lymph nodes. The dis-
tinct boundary between the capsule and the cortex is no longer
visible in reactive (Figure 3) and metastatic (Figure 5) nodes
where the entire node becomes more highly scattering, match-
ing the scattering intensity level fromthe capsule, and forming
a single homogeneous scattering layer.
The corresponding H&E histology slides presented in Fig-
ures 3 and 5 have more basophilic structures as exhibited by
the darker blue stain, which stains for the nucleic acids found
in the ribosomes, the chromatins in the nucleus, and the RNA
in the cytoplasm. The metastatic lymph node shows a loss of
1 mm
(a)
(b)
Fig. 3. (a) Reactive lymph node OCT image from a reac-
tive axillary lymph node with (b) corresponding H&E histol-
ogy. In this case, the increased cellular density in the cortex
when the lymph node becomes reactive contributes to the
increased scattering signal observed under OCT. The
change in scattering intensity begins to match the scatter-
ing intensity of the capsule, decreasing the ability to distin-
guish the capsule from the cortex under the OCT images
(arrows).
3-D OCT
500 m
(a) (b)
Fig. 4. Three-dimensional rendering of a normal lymph node.
(a) OCT image volume acquired from a normal sentinel
lymph node is shown with (b) the corresponding H&E histol-
ogy. The normal lymph node shows a clear capsule that is
easily differentiated from the low-scattering cortex. The OCT
images correlate to the regions in the histology that are
highlighted by the red boxes. The OCT image block
measures 5 35 31.7 mm
3
.
1 mm
1 mm
(a) (b)
Fig. 2. (a) Normal lymph node OCT images from a normal
sentinel lymph node with (b) corresponding H&E histology
demonstrate an intact capsule structure, which is easily dis-
tinguishable from the cortex of the lymph node. The OCT
images are highlighted (orange boxes) with the regions that
correlate with the histology images.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
normal tissue architecture as there is an obvious lymphoid
depletion characterized by a decrease in the number and size of
follicles with some or no germinal centers. At the same time,
the decrease of sinus spaces and increased density of cells
observed is due to the high level of cancer-cell proliferation.
Three-dimensional OCT reconstructions are presented in
Figures 4 and 5, demonstrating the added level of details avail-
able through the reconstructions. These renderings identify
the same features as those found in the 2-D images but further
allows for the arbitrary optical sectioning of the tissue along
planes that are not necessarily the same as the OCT imaging
acquisition planes. These intraoperative findings from human
lymph nodes extend our observations from previously
reported OCT image data of lymph nodes from a rat mammary
model [30]. The normal rat lymph node exhibited a highly
scattering layer corresponding to the capsule, compared with
the low-scattering cortex [30]. In the same report, OCT images
from a necrotic metastatic lymph node exhibited a high level
of scattering intensity and capsule disruption.
Discussion
The relatively high percentage of normal nodes resected dur-
ing lymph-node dissection procedures, as was exhibited for
the patients enrolled in this study, points to the significant
need for a real-time intraoperative microscopic nodal-assess-
ment technology. OCT has the potential to fill this crucial role
by providing real-time intraoperative information through the
identification of unique architectural features observed in the
capsule and outer cortex of the lymph node. These features
correspond to known morphological changes in reactive and
metastatic lymph nodes and are all accessible within the pene-
tration depth of OCT.
Identification of Normal Nodes from Reactive
and Metastatic Lymph Nodes
When cancer cells metastasize through the lymphatic system,
the nodes serve as the primary line of defense to sequester the
cancer cells and initiate a tumor-specific immunological
response. This process will initially cause an enlargement of
the lymph node as a part of the reactive inflammatory
response, but this enlargement can also be attributed to the
replication of cancer cells in the nodes. The presence of
inflammation in the node, however, is not sufficient to indi-
cate the presence of cancer cells. With hypercellularity, either
from the inflammatory response or from the replication and
expansion of cancer cells, the lymph node will not only
increase in relative size but also increase in cellular density
within the node. These changes will typically correspond to an
increase in the scattering observed in OCT, matching the scat-
tering levels observed in the capsule. As the scattering
intensity in the capsule and cortex areas become similar, the
differentiation of the capsule from the cortex becomes less
distinct. The increased scattering changes observed in the cor-
tex of reactive and metastatic nodes are more likely due to
increased cellular density but could also be attributed to the
increased number of cancer cells that have a higher N/C ratio.
These changes are more predominant in the cortex region
where the germinal centers and follicles, the primary sites of
lymphocyte recruitment and aggregation, are located.
Potential Clinical Impact
It is recognized that the majority of lymph nodes resected for
staging purposes are negative, and that the clinically relevant
data fromlymph node resection is primarily to knowwhether or
not the lymph node contains metastatic disease. In addition, the
primary surgical risk factor for lymphedema is nodal dissection
of the axillary lymph nodes, as well as the associated scarring
from surgical intervention. Therefore, reducing the number or
eliminating the need to resect normal nodes during surgery will
not only reduce the number of lymph nodes resected but also
decrease the risk of developing lymphedema. With a 12-mm
imaging penetration depth, OCT is a suitable modality to visu-
alize the microscopic lymph-node architecture, including the
Optical coherence tomography is a
high-resolution microscopic optical
imaging technique that yields real-time
multidimensional images of
subsurface tissue structure.
3-D OCT
(a) (b)
500 m
Fig. 5. Three-dimensional rendering of a metastatic lymph
node. (a) OCT image volume acquired from a metastatic
axillary lymph node in a single session is shown with (b) the
corresponding H&E histology. As with reactive lymph nodes,
increased scattering from the node is observed in the OCT
data, and the ability to differentiate distinct boundaries
between the node capsule and cortex has been lost. The
OCT images correlate to the regions in the histology that
are highlighted by the red boxes. The OCT image block
measures 5 3 5 31.7 mm
3
.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
capsule and the outer cortex containing the follicles and germi-
nal centers. In normal nodes, OCT is capable of clearly defining
the boundary between the capsule and the cortex based on the
scattering intensities. This boundary is lost in abnormal lymph
nodes. With no suitable real-time intraoperative method to
assess the lymph node before its surgical resection, OCT is able
to fill a niche by providing high-resolution microscopic images
in real time. Histopathological analysis from paraffin sections,
the current standard of care, is unable to provide surgeons with
this real-time information on the nodal status that could be used
to guide the surgical resection of lymph nodes. OCT is capable
of providing this information intraoperatively in real time, with
the potential for in vivo or in situ imaging, thereby potentially
eliminating the need for further lymph-node resections and
reducing the number of resulting complications.
Future Directions
The intraoperative OCT imaging data from the lymph nodes
in this feasibility study demonstrate promising results. From
the images presented, distinct morphological differences are
observed between the defined classifications of lymph nodes,
which are all well correlated to the corresponding histopatho-
logical findings and are all within the penetration depth of the
OCT imaging system. These initial intraoperative ex vivo
imaging studies demonstrate that the relevant diagnostic infor-
mation needed to classify lymph nodes is accessible by OCT
from outside of the lymph node capsule, leaving the lymph
node structurally intact. Transcapsule imaging by OCT, along
with the implementation of a fast scanning handheld probe
[32], will allowlymph nodes to be rapidly imaged in situ. Nee-
dle-based OCT imaging probes would allowfurther minimally
invasive probing of the node if it is not fully exposed or easily
accessible, such as in performing transcutaneous needle-
biopsy procedures of lymph nodes [33][35].
Future in vivo studies will further determine the clinical
impact of OCT on reducing the number of normal lymph nodes
removed and subsequently reducing the rate of morbid compli-
cations associated with lymph node dissections and breast can-
cer surgeries. The implementation of recent technological
advances to further increase the resolution of OCT [36] and
data acquisition speed [37] and to computationally classify tis-
sue types [27], [33] will improve the usability and effective-
ness of OCT in the intraoperative assessment of lymph nodes
by enhancing the capability to examine the underlying lymph
node microarchitecture, by rendering 3-D OCT data sets for
multisectional assessment and by rapidly classifying features
and tissue types using computer-aided detection.
Acknowledgments
This study was supported by the National Institutes of Health
NIBIB, R01 EB005221, Carle Foundational Hospital, and the
Grainger Foundation. Stephen A. Boppart is the principal
investigator on all of these grants and he also receives patent
royalties from the Massachusetts Institute of Technology for
technologies associated with OCT. Freddy T. Nguyen was
supported by the U.S. Department of Defense grant
BC073292. The authors acknowledge the technical contribu-
tions of Dr. Daniel Marks, the assistance with histological
analysis by Dr. Marina Marjanovic, and the contributions of
Ann Benefiel, Mary Collins, and Barbara Hall in the recruit-
ment and consenting of patients. They also thank the physi-
cians, staff, and administration at Carle Foundation Hospital
and Carle Clinic Association for their assistance in this
research. Additional information can be found at http://
biophotonics.illinois.edu.
Freddy T. Nguyen received the B.S. degree
in chemistry and the B.A. degree in mathe-
matics from Rice University in 2002. Since
2003, he has been at the University of Illi-
nois at Urbana-Champaign where he is an
M.D.Ph.D. candidate. He currently holds a
predoctoral fellowship from the Department
of Defenses Congressionally Directed
Medical Research ProgramsBreast Cancer Program. He was
a research engineer in the G.R. Harrison Spectroscopy Labora-
tory at the Massachusetts Institute of Technology from 2002 to
2003. He is a member of the American Medical Association,
the American Association for the Advancement of Science, the
American Association for Cancer Research, and the American
Chemical Society. His research interests have broadly been in
the development of spectroscopic and imaging techniques and
their associated contrast agents for direct biomedical applica-
tions such as cancer and atherosclerosis.
AdamM. Zysk received the B.S., M.S., and
Ph.D. degrees in electrical engineering from
Purdue University and the University of Illi-
nois at Urbana-Champaign in 2001, 2003,
and 2007, respectively. Since 2007, he has
been a postdoctoral researcher in the Electri-
cal and Computer Engineering Department,
Illinois Institute of Technology, Chicago,
Illinois. As a graduate student, his research focused on the appli-
cation of OCT to breast-cancer imaging. He currently researches
phase-contrast X-ray imaging techniques and their medical
applications. He is a Member of the IEEE, a member of the Opti-
cal Society of America, a member in training of the Radiological
Society of North America, and a member of the SPIE.
Eric J. Chaney received the B.S. degree in biology from the
University of Evansville, Evansville, Indiana, in 1992. From
The primary surgical risk factor for lymphedema
is dissection of lymph nodes, as well as the
associated scarring fromsurgical intervention.
________________
____
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
1993 to 1997, he worked as a research
assistant at the Indiana University School of
Medicine, Indiana State University, Terre
Haute. From 1997 to 2000, he worked as a
transmission electron microscope techni-
cian at the University of Illinois, Urbana-
Champaign, Urbana. Since 2000, he has
been a research specialist in molecular biol-
ogy at the Biophotonics Imaging Laboratory, Beckman Insti-
tute, University of Illinois, Urbana-Champaign.
Steven G. Adie received a B.Sc. (Hons.)
degree in chemical physics in 1997 and a
Ph.D. degree in electrical and electronic
engineering in 2007, both from the Univer-
sity of Western Australia. Between his
undergraduate and graduate studies, he
worked as a research engineer at a
company developing new technologies for
LASIK eye surgery. He has been a postdoctoral research asso-
ciate in the Biophotonics Imaging Laboratory at the Beckman
Institute, University of Illinois at Urbana-Champaign since
2007. His research interests include development of low-
coherence imaging techniques, namely interferometric syn-
thetic aperture microscopy and OCT, and their application to
breast cancer imaging. His Ph.D. research was on the devel-
opment of alternate contrast mechanisms for OCT.
Jan G. Kotynek received a B.A. degree in
psychology in 1971 followed by an M.D.
degree in 1975 from the University of Illi-
nois. He then completed a residency in
surgery at the University of Wisconsin in
1982. Since then, he has practiced as a gen-
eral surgeon, the last 18 years at Carle Clinic,
and has served on the faculty of the Univer-
sity of Illinois as clinical assistant professor of surgery. In 2008,
he retired from clinical practice. He is a member of the American
College of Surgeons, Illinois Surgical Society, American Medical
Association, and Alpha Omega Alpha Honor Medical Society.
Uretz J. Oliphant received the B.A.
degree from Boston University and the
M.D. degree from the University of Minne-
sota in 1983. He completed an internship in
general surgery at the University of Chi-
cago in 1985, residency in general surgery
at the University of Illinois in 1991, and a
fellowship in trauma/critical care at the Illi-
nois Masonic Medical Center. Since 1992, he has been a
surgeon at the Carle Clinic. Since 1994, he has been a profes-
sor in clinical surgery and internal medicine at the University
of Illinois, Urbana-Champaign. He is currently the head of the
Department of Surgery at the University of Illinois College of
Medicine, Urbana-Champaign. He is a member of the Ameri-
can Medical Association, the American Society of Breast Sur-
geons, president of the Illinois Surgical Society, and a fellow
of the International College of Surgeons.
Frank J. Bellafiore received the B.S. degree in mathematics
and chemistry from Creighton University, Omaha, Nebraska,
in 1986, and the M.D. degree from Washington University,
St. Louis, in 1990. He completed a resi-
dency in anatomic and clinical pathology,
including a fellowship in surgical pathology
and cytopathology in 1995 at the University
of Virginia Health Sciences Center. He is
currently a pathologist and director of
histology and flow cytometry in the Depart-
ment of Pathology at the Carle Clinic Asso-
ciation. He is a clinical assistant professor at the University of
Illinois College of Medicine, Urbana-Champaign.
Kendrith M. Rowland received the B.A.
degree from Lawrence University, Apple-
ton, Wisconsin, in 1976, and the M.D.
degree from the University of Illinois Col-
lege of Medicine, Chicago, in 1980. He
completed a residency in internal medicine
at the University of Illinois Hospital, West-
side V.A., and a fellowship in oncology at
Rush-Presbyterian, St. Lukes Medical Center. He is currently
a hematologist and oncologist in the Department of Medicine
at the Carle Clinic Association and Carle Foundation Hospital
since 1985. He is the principal investigator of the Carle Cancer
CenterCommunity Clinical Oncology Program. He is a
clinical assistant professor of medicine at the University of
Illinois College of Medicine, Urbana-Champaign. He is a
member of the American Society of Clinical Oncology, the
International Association of the Study of Lung Cancer, Phi
Beta Kappa, and the ASCO Clinical Trials Committee.
Patricia A. Johnson earned a B.S. degree
in microbiology from the University of
Maryland in 1969, and M.S. and Ph.D.
degrees in microbiology from the Univer-
sity of Illinois in 1970 and 1974, respec-
tively, and M.D. degree at Rush University
Medical School in 1977. She completed a
residency in internal medicine in 1980 and
a fellowship in medical oncology in 1982 at Rush Presbyte-
rian St. Lukes Medical Center, Chicago. She has been a
medical oncologist at Carle Clinic, Urbana Illinois since
1982, where she has served as president of the medical staff,
chair of the board of governors, medical director of medical
subspecialties, and medical director of the Mills Breast Can-
cer Institute. She is a member of American Society of Clinical
Oncology, American Medical Association, Illinois State
Medical Society, Champaign County Medical Society, Phi
Beta Kappa, Phi Sigma, Phi Kappa Phi, Sigma Alpha Omi-
cron, and Alpha Lambda Delta.
Stephen A. Boppart received the B.S.
degree in electrical and bioengineering
and the M.S. degree in electrical engineer-
ing from the University of Illinois,
Urbana-Champaign, in 1990 and 1991,
respectively, the Ph.D. degree in electrical
and medical engineering from the Massa-
chusetts Institute of Technology, in 1998,
and the M.D. degree from Harvard Medical School, in 2000.
He completed residency training in internal medicine at the
University of Illinois, Urbana-Champaign, in 2005. He is cur-
rently a professor of electrical and computer engineering,
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
bioengineering, and medicine. His research interests include
the development and translation of novel optical-imaging
technologies for biological and medical applications, with
particular emphasis in cancer detection and diagnosis. He is a
Senior Member of the IEEE. He is a fellow of the Optical
Society of America and the International Society for Optical
Engineering (SPIE). In 2003, he was named as one of the Top
100 Young Innovators in the world by MITs Technology
Review Magazine for advances in medical imaging technol-
ogy. He received the IEEE Engineering in Medicine and Biol-
ogy Society Early Career Achievement Award in 2005.
Address for Correspondence: Stephen A. Boppart, Beckman
Institute for Advanced Science and Technology, University of
Illinois at Urbana-Champaign, 405 North Mathews Avenue,
Urbana , IL 61801, USA. E-mail: boppart@illinois.edu.
References
[1] Cancer Facts & Figures 2009, Atlanta: American Cancer Society, 2009.
[2] Breast Cancer Facts & Figures 20072008, Atlanta: American Cancer Soci-
ety, 2008.
[3] D. Krag, D. Weaver, T. Ashikaga, F. Moffat, V. S. Klimberg, C. Shriver,
S. Feldman, R. Kusminsky, M. Gadd, J. Kuhn, S. Harlow, and P. Beitsch, The
sentinel node in breast cancera multicenter validation study, N. Engl. J. Med.,
vol. 339, pp. 941946, Oct. 1998.
[4] J. V. Frangioni, S. W. Kim, S. Ohnishi, S. Kim, and M. G. Bawendi, Senti-
nel lymph node mapping with Type-II quantum dots, Methods Mol. Biol.,
vol. 374, pp. 147160, 2007.
[5] S. Kim, Y. T. Lim, E. G. Soltesz, A. M. De Grand, J. Lee, A. Nakayama, J.
A. Parker, T. Mihaljevic, R. G. Laurence, D. M. Dor, L. H. Cohn, M.
G. Bawendi, and J. V. Frangioni, Near-infrared fluorescent type II quantum dots
for sentinel lymph node mapping, Nat. Biotechnol., vol. 22, pp. 9397, Jan.
2004.
[6] H. Nimura, N. Narimiya, N. Mitsumori, Y. Yamazaki, K. Yanaga, and
M. Urashima, Infrared ray electronic endoscopy combined with indocyanine
green injection for detection of sentinel nodes of patients with gastric cancer,
Br. J. Surg., vol. 91, pp. 575579, May 2004.
[7] S. A. Mclaughlin, L. M. Ochoa-Frongia, S. M. Patil, H. S. Cody III, and L.
M. Sclafani, Influence of frozen-section analysis of sentinel lymph node and
lumpectomy margin status on reoperation rates in patients undergoing breast-
conservation therapy, J. Am. Coll. Surg., vol. 206, pp. 7682, Jan. 2008.
[8] R. Ali, A. M. Hanly, P. Naughton, C. F. Castineira, R. Landers, R. A. Cahill,
and R. G. Watson, Intraoperative frozen section assessment of sentinel lymph
nodes in the operative management of women with symptomatic breast cancer,
World J. Surg. Oncol., vol. 6, p. 69, 2008.
[9] K. Motomura, S. Nagumo, Y. Komoike, H. Koyama, and H. Inaji, Accuracy
of imprint cytology for intraoperative diagnosis of sentinel node metastases in
breast cancer, Ann. Surg., vol. 247, pp. 839842, May 2008.
[10] A. J. Creager, K. R. Geisinger, S. A. Shiver, N. D. Perrier, P. Shen, J. Ann
Shaw, P. R. Young, and E. A. Levine, Intraoperative evaluation of sentinel
lymph nodes for metastatic breast carcinoma by imprint cytology, Mod. Pathol.,
vol. 15, pp. 11401147, Nov. 2002.
[11] T. B. Julian, P. Blumencranz, K. Deck, P. Whitworth, D. A. Berry, S.
M. Berry, A. Rosenberg, A. B. Chagpar, D. Reintgen, P. Beitsch, R. Simmons,
S. Saha, E. P. Mamounas, and A. Giuliano, Novel intraoperative molecular test
for sentinel lymph node metastases in patients with early-stage breast cancer, J.
Clin. Oncol., vol. 26, pp. 33383345, July 2008.
[12] B. Bird, M. Miljkovic, M. J. Romeo, J. Smith, N. Stone, M. W. George, and
M. Diem, Infrared micro-spectral imaging: Distinction of tissue types in axillary
lymph node histology, BMC Clin. Pathol., vol. 8, p. 8, 2008.
[13] E. M. Sevick-Muraca, R. Sharma, J. C. Rasmussen, M. V. Marshall, J.
A. Wendt, H. Q. Pham, E. Bonefas, J. P. Houston, L. Sampath, K. E. Adams, D.
K. Blanchard, R. E. Fisher, S. B. Chiang, R. Elledge, and M. E. Mawad, Imag-
ing of lymph flow in breast cancer patients after microdose administration of a
near-infrared fluorophore: Feasibility study, Radiology, vol. 246, pp. 734741,
Mar. 2008.
[14] R. Sharma, J. A. Wendt, J. C. Rasmussen, K. E. Adams, M. V. Marshall, and
E. M. Sevick-Muraca, New horizons for imaging lymphatic function, Ann. N.
Y. Acad. Sci., vol. 1131, pp. 1336, 2008.
[15] H. L. Bumpers, I. M. Best, D. Norman, and W. L. Weaver, Debilitating
lymphedema of the upper extremity after treatment of breast cancer, Amer. J.
Clin. Oncol., vol. 25, pp. 365367, Aug. 2002.
[16] I. Langer, U. Guller, G. Berclaz, O. R. Koechli, G. Schaer, M. K. Fehr,
T. Hess, D. Oertli, L. Bronz, B. Schnarwyler, E. Wight, U. Uehlinger,
E. Infanger, D. Burger, and M. Zuber, Morbidity of sentinel lymph node biopsy
(SLN) alone versus SLN and completion axillary lymph node dissection after
breast cancer surgery: a prospective Swiss multicenter study on 659 patients,
Ann. Surg., vol. 245, pp. 452461, Mar. 2007.
[17] M. de Boer, C. H. van Deurzen, J. A. van Dijck, G. F. Borm, P. J. van
Diest, E. M. Adang, J. W. Nortier, E. J. Rutgers, C. Seynaeve, M. B. Menke-Plu-
ymers, P. Bult, and V. C. Tjan-Heijnen, Micrometastases or isolated tumor cells
and the outcome of breast cancer, N. Engl. J. Med., vol. 361, pp. 653663,
Aug. 2009.
[18] S. A. Boppart, B. E. Bouma, C. Pitris, J. F. Southern, M. E. Brezinski, and J.
G. Fujimoto, In vivo cellular optical coherence tomography imaging, Nat.
Med., vol. 4, pp. 861865, July 1998.
[19] S. A. Boppart, B. E. Bouma, C. Pitris, G. J. Tearney, J. F. Southern, M.
E. Brezinski, and J. G. Fujimoto, Intraoperative assessment of microsurgery with
three-dimensional optical coherence tomography, Radiology, vol. 208, pp. 81
86, July 1998.
[20] S. A. Boppart, T. F. Deutsch, and D. W. Rattner, Optical imaging technol-
ogy in minimally invasive surgery. Current status and future directions, Surg.
Endosc., vol. 13, pp. 718722, July 1999.
[21] S. A. Boppart, W. Luo , D. L. Marks, and K. W. Singletary, Optical coher-
ence tomography: Feasibility for basic research and image-guided surgery of
breast cancer, Breast Cancer. Res. Treat., vol. 84, pp. 8597, Mar. 2004.
[22] B. E. Bouma and G. J. Tearney, Handbook of Optical Coherence Tomogra-
phy. New York: Marcel Dekker, 2002.
[23] J. G. Fujimoto, C. Pitris, S. A. Boppart, and M. E. Brezinski, Optical coher-
ence tomography: An emerging technology for biomedical imaging and optical
biopsy, Neoplasia, vol. 2, pp. 925, Jan.Apr. 2000.
[24] D. Huang, E. A. Swanson, C. P. Lin, J. S. Schuman, W. G. Stinson,
W. Chang, M. R. Hee, T. Flotte, K. Gregory, C. A. Puliafito, and J.
G. Fujimoto, Optical coherence tomography, Science, vol. 254, pp. 1178
1181, Nov. 1991.
[25] A. M. Zysk, F. T. Nguyen, A. L. Oldenburg, D. L. Marks, and S.
A. Boppart, Optical coherence tomography: A review of clinical development
from bench to bedside, J. Biomed. Opt., vol. 12, p. 051403, Sept.Oct. 2007.
[26] F. Nguyen, A. M. Zysk, E. J Chaney, J. G. Kotynek, U. Oliphant, F.
J. Bellafiore, K. M. Rowland, P. A. Johnson, and S. A. Boppart, Intraoperative
evaluation of breast tumor margins with optical coherence tomography, Cancer.
Res , vol. 69, pp. 87908796, Nov. 2009.
[27] A. M. Zysk and S. A. Boppart, Computational methods for analysis of
human breast tumor tissue in optical coherence tomography images, J. Biomed.
Opt., vol. 11, p. 054015, Sept.Oct. 2006.
[28] D. Arifler, et al., Light scattering from normal and dysplastic cervical cells
at different epithelial depths: Finite-difference time-domain modeling with a per-
fectly matched layer boundary condition, J. Biomed. Opt., vol. 8, pp. 484494,
July 2003.
[29] R. Drezek, et al., Light scattering from cervical cells throughout neoplastic
progression: Influence of nuclear morphology, DNA content, and chromatin tex-
ture, J. Biomed. Opt., vol. 8, pp. 716, Jan. 2003.
[30] W. Luo, et al., Optical biopsy of lymph node morphology using optical
coherence tomography, Technol. Cancer. Res. Treat., vol. 4, pp. 539548, Oct.
2005.
[31] F. T. Nguyen, A. M. Zysk, J. G. Kotynek, F. J. Bellafiore, K. M. Rowland,
P. A. Johnson, E. J. Chaney, and S. A. Boppart, Portable real-time optical coher-
ence tomography system for intraoperative imaging and staging of breast cancer,
in Proc. SPIEPhotonics West BiOSAdvanced Biomedical and Clinical Diag-
nostic Systems V, San Jose, CA, 2007.
[32] X. Liu, M. J. Cobb, Y. Chen, M. B. Kimmey, and X. Li, Rapid-scanning
forward-imaging miniature endoscope for real-time optical coherence tomogra-
phy, Opt. Lett., vol. 29, pp. 17631765, Aug. 2004.
[33] B. D. Goldberg, N. V. Iftimia, J. E. Bressner, M. B. Pitman, E. Halpern, B.
E. Bouma, and G. J. Tearney, Automated algorithm for differentiation of human
breast tissue using low coherence interferometry for fine needle aspiration biopsy
guidance, J. Biomed. Opt., vol. 13, p. 014014, Jan.Feb. 2008.
[34] A. M. Zysk, S. G. Adie, J. J. Armstrong, M. S. Leigh, A. Paduch, D.
D. Sampson, F. T. Nguyen, and S. A. Boppart, Needle-based refractive index
measurement using low-coherence interferometry, Opt. Lett., vol. 32, pp. 385
387, Feb. 2007.
[35] A. M. Zysk, E. J. Chaney, and S. A. Boppart, Refractive index of carcino-
gen-induced rat mammary tumours, Phys. Med. Biol., vol. 51, pp. 21652177,
May 2006.
[36] R. J. Zawadzki, B. Cense, Y. Zhang, S. S. Choi, D. T. Miller, and J.
S. Werner, Ultrahigh-resolution optical coherence tomography with monochro-
matic and chromatic aberration correction, Opt. Express., vol. 16, pp. 8126
8143, May 2008.
[37] R. Huber, D. C. Adler, and J. G. Fujimoto, Buffered Fourier domain mode
locking: Unidirectional swept laser sources for optical coherence tomography
imaging at 370,000 lines/s, Opt. Lett., vol. 31, pp. 29752977, Oct. 2006.
______________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Deep-Tissue Dynamic
Monitoring of Decubitus
Ulcers
T
he use of deep-tissue multiwavelength imaging is of
paramount importance in clinical settings as a noninvasive
solution to identify and monitor the progression of decubi-
tus ulcers. A point-of-care multiwavelength imager is
being developed, whose utility results from the provision of
important physiological characteristics and blood flow metrics
via analysis of deep-tissue response to light. The noninvasive
real-time monitoring and analysis of tissue focusing on wound
imaging is integral, because it allows for quantitative in situ
measurements that characterize tissue to assess the progress of
either tissue healing or necrosis.
Decubitus ulcers are a common problem in health care, cost-
ing US$13 billion annually [1], and are caused by the effects of
extended pressure applied over a particular area of the body on
tissue. Despite the widespread problem of decubitus ulcers,
there is still a relatively simple and iterative approach used in
hospitals to turn patients periodically, which still subjects
patients to ulcer formations. The solution is a deep-tissue mul-
tiwavelength imager that characterizes tissue health in a
continuous, real-time environment, thereby streamlining work-
flows and enhancing diagnostic capabilities.
Various solutions have been investigated with some limita-
tions in the context of wound care and assessment. Techniques
such as hyperspectral imaging (HSI) currently exist, whose tech-
nologies are being significantly advanced, but the acceptance of
this imaging model is limited in medical communities because of
its limited scope of near-surface analyses that adequately corre-
spond to wound tissue factors [2]. This limitation is significant
despite the importance of time-averaged physiological informa-
tionoxygen saturation mapsthat it provides, because the
near-surface sensitivity precludes its use for assessment of deeper
surface tissue ailments that affect patients with decubitus ulcers,
skin transfers, and burns [3]. Optical tomography (OCT) is
another near-surface sensitive technique that has met the needs in
ophthalmology meaningfully [4]. Pulse oximetry is based on dif-
fuse optics with displaced illuminationthe detector collects the
light exiting tissue at a location distinct from the location of light
injectionbut is not imaging in nature, although imaging solu-
tions have been investigated [5]. Time-resolved spectroscopy
(TRS) coupled with displaced illumination has resulted in a
modality with limited spatial imaging capability [6]. More
recently, optical imaging for noninvasive assessment of brain
activity and the development of physics-based models have been
proven useful to aid tissue visualization [7], [8].
The technology discussed herein combines coincident and
displaced light-injection patterns to generate a three-dimen-
sional (3-D) spatial map with a fourth dimension of time. This
combination of depth and temporal resolution is a unique
aspect lacking in other existing imaging techniques.
Procedure
Preliminary Studies
The initial experiments involved the use of a monochromatic
CMOS image sensor in tandem with light sources at two mono-
chromatic wavelengths of 660 and 860 nm. The following imaging
technique used reflectance ratios between red and infrared images
to provide information about perfusion based on oxygenation. A
finger was sampled under different stress environments. This first
study was meant to demonstrate comparable capability to existing
systems for near-surface perfusion analysis.
The experiment was conducted in an invariably dark room
in which images of a human finger were taken under different
source lighting conditions (red and infrared). Each image was
of 512 31,024 resolution in a gray scale. Peripheral source
lighting, four LEDs at each wavelength, and a central image
sensor were placed in a matte black mount to limit signal
reflections. The imaged area in the following measurements
was coincident with the illuminated area. For this coincident
light injection and imaged area setup, the light captured by the
imager is dominated by photons that traversed only shallow
regions directly below the imaged arealess than 3 mm for
human tissue [3]. Clear glass was mounted on the top surface
to support the finger sampled (Figure 1).
The first set of images was taken of a normally situated finger
with uninhibited circulation. The images captured at each wave-
length were then compared, resulting in a final composite of red
and infrared signals that depicted relative oxygenation (Figure 2).
Next, a rubber band temporarily occluded blood circulation in
the finger. The image was sampled a minute after the band had
been applied to deteriorate blood flowin that area. This is proven
in the image, seen by the lighter appearance of the finger. The
BY ROHIN MOZA, J. MICHAEL DIMAIO,
AND JOSE MELENDEZ
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Wound Care and Assessment
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
left-to-right gradient of greater-to-lesser perfusion is also
indicative of appropriate physiological responses (Figure 3).
The finger in a postoccluded state presents a much darker area
than the normal state of a finger, indicating higher levels of perfu-
sion. The bodys attempt to compensate for the previous lack of
blood flowis clearly visible by the increasing amount of blood sent
to that part of the finger. This phenomenon is known as hyperemia
and can be attributed to flowautoregulation (Figure 4).
The last case was a finger to which an unmeasured amount
of pressure was applied. The image shows that the area to
which force was applied had lower perfusion levels, as
expected. The medium of blood flow has been compressed,
thus resistance to blood flow is greater. The lighter region cor-
roborates hindered blood flow (Figure 5).
Time-Dependent Observations
The present system is not static and as such is not limited by
temporal resolution or the ability to penetrate deep-tissue sur-
faces, a restriction seen in other imaging techniques. By
avoiding this concern, the presented technology can capture
events at a frequency of multiple times per cardiac cycle and
critical tissue depths that are required to assess wound-related
physiologies [3]. It is of value to capture the pulsatile nature of
blood when making tissue assessments in situ.
Light-Tissue Contact Study
In this study, a finger was illuminated at the red and infrared
wavelengths at two separate time intervals, and the video was
captured at 30 Hz at each wavelength (Figure 6).
Imaging
Unit Computer
RJ45
(Image Capture)
FPGA
Process
Data
Fig. 1. Setup for preliminary studies.
Base of Finger Finger Tip
Fig. 2. Normal state of finger.
Left to Right Gradient of Finger
Point of Occlusion
Fig. 3. Finger occlusion.
Extreme Dark Areas Indicating
Hyperemic Condition
Fig. 4. Finger postocclusion.
Pressure Applied
Fig. 5. Finger pressure.
Imaging Unit
Computer
Signal Observed
PCI Framegrabber
(Video Capture)
DSP
Process
Data
Fig. 6. Setup for light-tissue contact observation.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
The LEDs were placed in direct contact with a portion of
the finger, and imaging was performed in an area adjacent to
the illumination target as indicated in Figure 7. An opaque
shield to reduce direct light from the source to the camera
headlimiting direct surface reflections that would present
near-surface tissue information and noise from light that had
not diffused the tissuemasked the LED.
This reduction of specular reflections and other point satura-
tions of light is a significant factor in being able to identify pulsa-
tile components of tissue fromarteries that generally span deeper
tissue regions and layers. A grid of pixels of the imaged finger,
indicated in Figure 7, was averaged for each frame of video cap-
tured, and the corresponding values of that grid were plotted over
time, as indicated in Figures 8 and 9. The benefit of minimizing
such reflections and saturations is evident in the signal, as it
clearly appears periodic with characteristic photoplethysmo-
graph features. The displaced illumination in this optical system
allowed the elucidation of this
signal, whereas coincident illu-
mination would not demon-
strate such a response due to
the confounding factors. The
heart rate from the photople-
thysmograph is consistent with
the pulse rate measured before
and after sampling at the radial
artery. One signal feature
worth noting is the small dip
and upstroke in signal demon-
strated in both waveforms. This
is an important information that
represents the difference be-
tween the systolic and diastolic
components of a pressure
wave, and the feature is termed
a dicrotic notch. Such features
would not be detected in a
static analysis of the blood that
inherently overlooks important
information about the cardio-
vascular processes contained
within the time-varying pulsa-
tile waveform. Additional fea-
tures in the dynamic responses
relate to blood pressure, car-
diac output, respiration, and
oxygen saturation [9].
It is important to note that
these results represent a single
illumination condition not
mapped to a specific tissue
depth, but the depth of pene-
tration is known to be higher
than the previously described
optical setup. This is due to
the increased spacing between
the detector and illuminator.
As the spacing increases, the
light passes deeper trajecto-
ries through tissue and experi-
ences more attenuation. The
path length of light through
the tissue assumes the shape of a banana curve [10]. The out-
put of the detector at any given point is a summation of signals
from multiple paths that exit the skin at that point. This rela-
tionship between the detector output and the sampled region is
X: 743 Y: 246
Index: 0.1765
RGB: 0.204, 0.204, 0.204
Fig. 7. Area of analysis: infrared capture.
I
n
t
e
n
s
i
t
y
0.021
0.0205
0.02
0.0195
0.019
0.0185
0.018
0.0175
0.017
0 1 2 3 4 5 6 7 8 9
Time (s)
Time-Dependent Behavior in Red With FPS = 30
Number of Frames = 250 Sample Frequency = 30 Hz
X: 4.133
Y: 0.01761
X: 4.867
Y: 0.01747
81 BPM
Fig. 8. Signal observed in imager for red illumination.
I
n
t
e
n
s
i
t
y
0.0112
0.0111
0.011
0.0109
0.0108
0.0107
0.0106
0.0105
0 1 2 3 4 5 6 7 8 9
Time (s)
81 BPM
.0112
.0111
0.011
.0109
.0108
.0107
.0106
.0105
0 1 2 3 4 5 6 7 8 9
81 BPM
X: 3.1
Y: 0.01074
Time-Dependent Behavior in IR with FPS = 30
Number of Frames = 250 Sample Frequency = 30 Hz
X: 3.833
Y: 0.01077
Fig. 9. Signal observed in imager for infrared illumination.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
further complicated in cases where the injections occur over
large areas or asymmetrically as has been the case in previous
investigations [5], [11], [12].
Light-Tissue Noncontact Study
Athird optical setup was devised to performnoncontact analy-
sis of perfusion. A single LED at each of the 650- and 875-nm
wavelengths was used with narrow beam angles of 4 and 10,
respectively. Each LED was encased by an opaque shield and
barrel to shape the beam to implement a displaced light injec-
tion scheme.
When performing this analysis at the wrist, periodicity
similar to the contact approach was observed with respect to
the pulse rate. Additionally, a lower frequency, respiratory
component presented as the major signal component in the
analysis. Breathing cycles were manually monitored and
coincident with the plotted data while also consistent with
physiologically expected respiration rates (Figures 10 and 11).
A separate image analysis of a patient after surgery with
multiple lesions and an open wound spanning the abdomen,
sized 5 3 1 cm, was performed while using the IR LED (850
nm) for illumination. A frame capture of the imaged area is
shown in Figure 12. A video lasting 60 s, or 1,800 frames was
captured. Given that the data collected spans four dimensions
time, pixel intensity, and two spatial dimensionsFourier trans-
forms were used to perform a time-based analysis at a pixel
resolution spanning the 512 31,024 31,800 image matrix.
The most common peak frequency found throughout the image
was selected, and the corresponding power was plotted for that
frequency at each pixel in a color map representative of pulse
rate activity. The pulse rate was manually measured prior to
sampling and in agreement with the selected frequency compo-
nent of 1.17 Hz or approximately 70 beats/min. The color map
corresponding to the spatial power values of the pulse frequency
component is illustrated in Figure 13. The regional gradient
fromblue to red represents the areas fromthe lowest to the high-
est pulse rate intensities, respectively.
During the healing process of a wound, the inflammation
response recruits more blood to the site, earlier noted as the
hyperemic effect. Because of this, affected sites, such as feed-
ing tube (G-Tube) insertion and chest tube insertion in this
particular patient, will demonstrate such intensely distributed
areas. Two subsurface conditions are worth noting. Region 1,
indicated in Figures 12 and 13, reflects a subsurface condition
where deeper tissue was punctured to allow for the passage of
the chest tube. This intensity
level provides that the area
may still be healing and per-
haps infected. Also, Region 2
coincides with the location of
a liver abscess, possibly dem-
onstrating another subsurface
condition. This topographic
layout of pulse intensity dis-
tribution provides a visualiza-
tion that represents a series of
frames for that capture.
Real-Time Analysis
A digital signal processor
(DSP) was used in conjunc-
tion with the CMOS camera to
process algorithms in real time
to define areas of interest. The
example demonstrates the live
detection of edges of a hand
using the Hough transform
X: 505 Y: 270
Index: 1.222e+04
RGB: 0.184, 0.184, 0.184
Fig. 10. Wrist captured with red illumination.
I
n
t
e
n
s
i
t
y
2,800
2,750
2,700
2,650
2,600
2,550
2,500
0 5 10 15
Time (s)
Time-Dependent Behavior in Red with FPS = 30
Number of Frames = 1,000 Sample Frequency = 30 Hz X = 505 Y = 270 ROI Area = 41 x 41
20 25 30 35
Fig. 11. Respiration cycles captured at wrist.
Decubitus ulcers are a common problem in
health care and are caused by the effects of
extended pressure applied over a particular
area of the body on tissue.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
(Figures 14 and 15). This demonstrates that the system can
provide real-time analysis output for complex transform analy-
sis algorithms. This real-time detection was enabled through
four layers of white bed sheets and a clear mattress. The ability
to discern objects through these layers demonstrates a unique
capability of the technology and the feasibility in monitoring
patients with an imager built into a bed surface or mattress.
This is readily applicable to patient monitoring to identify
errant needle caps, waste matter, and other risk factors for ulcer
formation that would not be visible to nurses unless the patient
was turned. Similar spatial transforms can be applied to define
regions of necrosis or susceptibility to tissue damage.
The frequency analysis of the time-dependent signals in
each wavelength will allow for dynamic in situ patient moni-
toring. Specific frequency bands contain characteristic physi-
ological information that can be used to determine a frequency
signature of stages of necrosis and ascertain progress of tissue
healing. Currently, the requirements to dynamically monitor
tissue are supported by definitions of digital illumination pat-
terns and transfer functions mapping tissue depths and
2
1
Fig. 13. Intensity map of pulse rate at wound site with out-
lined subsurface conditions.
Fig. 15. Upper left quadrant shows the unprocessed hand
image. Remaining quadrants show real-time transform
analyses and results utilizing differing algorithm parameters.
DSP
Imaging
Unit
Imaging Unit (Underneath Mattress)
Fig. 14. Setup for real-time display.
Patient inclined and
supported by blanket.
Wound site is in granulation phase
but also purulent.
G-Tube Insertion
1
2
1 cm
5 cm
Chest Tube
Insertion
Fig. 12. Abdominal area captured.
This effort aims to focus on time-dependent
and tissue depth-dependent behavior that
demonstrates progression or regression
of tissue health.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
locations to an image coordinate. The data acquired demon-
strates that more dense and patterned illumination boards
coupled with frequency-based algorithms further the physio-
logical information derived from these samples. Given that
the implemented algorithms demonstrate real-time detection
abilities, the application of such an analysis is consistent with
the demonstrated system and quantity of data.
Current and Future Work
Currently, work is being done to advance algorithms to classify
the periodic nature of tissue perfusion in tandem with methods
to automatically highlight the identification of those regions for
further clinical analysis. Frequency analysis continues to be
developed to further reveal physiological meaning of images
captured. These frequency signatures are the basis for classify-
ing tissue health. Infrared and red signals in the time and
frequency domains are correlated and synchronized to generate
conclusions about tissue oxygenation to drive such assessments.
The identification of specific sets of digital illumination condi-
tions based on characteristic light-tissue profiles and detector-
illuminator spacing is necessary to ensure systematic and com-
plete sampling of an entire tissue region. Figure 16 demonstrates
the patterns that are used to illuminate the sample. Each of the
diagrams represents the field of view of the detector while the
white regions are directly illuminated, and the gray regions are
not. Each example requires two sampling cycles per region to
ensure complete coverage. The gray squares provide deeper sur-
face information while the white squares primarily present near-
surface content. Establishing a diffuse sampling mathematical
model to convey the relationship between this lighting scheme
and the tissue response and regions sampled enables the novel
extraction of physiologically based optical and infrared cross-
sectional image maps as a function of both tissue depth and time.
Commercial Clinical Deployment
The applications of this technology address a pervasive health-
care need, and this is evident as 257,412 Stage 3 and 4 decubi-
tus ulcers cost Medicare US$43,180 each in fiscal year (FY)
2007 [13]. Recent Centers for Medicare and Medicaid Services
(CMS) rule changes, as of 2008, shift the burden of these costs
to the hospitals. The point-of-care capability in this device pro-
vides on-demand analysis by clinicians, and the diagnostic
information becomes incorporated into patient records. Hospi-
tals, clinics, and nursing homes will purchase instruments to
reduce loss of CMS/insurance reimbursements in the light of
rule changes as well as costs in the order of thousands of dollars
per incident. Additional revenues will be generated from dis-
covered wounds, reduced physician labor, and better wound
care. The value proposition intersects health-care priorities of
affordable universal access. Procedural reimbursement is de-
sirable but not required to justify return on investment. The use
of LED illumination in the device does not provide the regula-
tory obstacles a laser-based system might, and power levels of
the device are within a safely operable range.
There are 294,297 serviceable care providers [14], and an
independent market estimate of US$1.0 billion per year [15].
The first product planned is a portable solution that can be
transported to patient rooms in the ICU to assess local tissue
healing. Relationships have been established with component
and subsystem providers as well as with potential distributors
and commercialization partners. The challenges of clinical
validation are being met by engaging leading clinicians in
areas such as wound healing, plastic surgery, physical medi-
cine and rehabilitation, and burn trauma. Early direct sales will
be used to establish a market and create distributor demand.
Data will be gathered to demonstrate return on investment and
care benefits. Success stories will be promoted at industry
trade shows and publications, while distributors will also be
engaged to promote widespread product sales and support.
Conclusions
The use of a discrete set of wavelengths to image skin allows
the penetration of tissue to derive valuable perfusion and oxy-
gen-metric data. This association with specific tissue physiol-
ogies in tandem with analysis of the components that form the
tissue layers allow for a relevant and meaningful analysis of
blood perfusion. Variables such as wound age may be derived
from the frequency behaviors and used as a metric for deter-
mining wound severity.
When performing imaging based on a reflectance concept,
certain allowances have to be made with illumination schemes
(a)
(b)
Fig. 16. Alternating (a) checker board pattern and (b)
window pattern.
The use of a discrete set of wavelengths
to image skin allows the penetration of tissue
to derive valuable perfusion and
oxygen-metric data.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
so as to limit direct light interference at the optical head.
Examples include the reduction of specular reflections to limit
local image saturation and light environments that maximize
coherency. Optical design, digital illumination patterning,
real-time signal processing, depth of penetration, as well as
frequency-based algorithms are components of this system
that form the basis of the integrated imager capable of provid-
ing relevant and accurate monitoring of tissue health.
These studies have provided real-time data and physiologi-
cal visualization via Fourier transforms, tissue imaging and
analysis of noncontact photoplethysmograph images, and
optical injection schemes that can be used to analyze the tissue
state and oxygenation of a patient.
Existing point-of-care solutions are limited and unable to
monitor risks of decubitus ulcers or provide tissue data that
will allow for their early identification or prevention. Iterative
methods of turning patients have been hindered by shortages
of nursing staff as well as the inability to visually detect ische-
mic conditions at a subskin level, the origin of these ulcers.
Incorporating a dynamic monitoring solution to assess hemo-
dynamic behavior with optical imaging of tissue physiology
by variable digital illumination will contribute significantly to
the treatment, care, and prevention of decubitus ulcers. This
effort aims to focus on time-dependent and tissue depth-
dependent behavior that demonstrates progression or regres-
sion of tissue health. Clinical studies are being developed, as
the next phase of this project is integrating and deploying this
system as a product in a health-care setting.
Acknowledgment
This work was supported by the University of Texas South-
western Medical Center (UTSW) at Dallas.
Rohin Moza received his B.S. degree in
electrical engineering and B.S. degree in
mathematical biology from the University of
Michigan at Ann Arbor. He is currently a
Ph.D. student in biomedical engineering with
the Cardiovascular and Thoracic Surgery
Department at the University of Texas
Southwestern Medical Center (UTSW) at
Dallas. His research involves image processing, optical tissue
modeling, optical systems development, translational medicine,
and mathematical modeling of pathologic and healing tissue.
J. Michael DiMaio received his medical
degree from the University of Miami in
1987 and completed his internal medicine,
general surgery, and thoracic surgery res-
idencies at Duke University Medical
Center. While at Duke, he completed a
surgical research fellowship focused on
gene therapy, immunology, and transplan-
tation. He joined the faculty at UTSW in 1998. He currently
oversees clinical research projects focused on cardiac and
thoracic research and has an active laboratory studying the
effects of protein and pharmacologic agents involved with
myocardial and neurological recovery. He is the director of
Cardiovascular and Thoracic Surgery Research and holds the
Laurence and Susan Hirsch/Centex distinguished chair in Heart
Disease at UTSW. His clinical interests include transplantation,
cardiac surgery, ventricular aneurysm repair, aortic stenting,
minimally invasive cardiac and thoracic surgery, and usage of
holmium-yttrium-aluminum-garnet (YAG) laser therapy for
tracheal and thoracic obstructive processes.
Jose Melendez received his Ph.D. degree
in electrical engineering from Stanford
University and B.S./M.S. degree in electri-
cal engineering and computer science from
the Massachusetts Institute of Technology.
He is an accomplished leader in the devel-
opment and commercialization of optical
systems, wireless communication devices,
voice/data convergence and microelectromechanical system
(MEMS) through 14 years of contributions to Texas Instru-
ments (TI) combined with six years of entrepreneurial experi-
ence. He served as a general manager of TIs worldwide
wireless infrastructure business. He presently serves as the pres-
ident and chief scientist of Spectral MD Inc., a high-technology
image systems company founded to make fundamental contri-
butions to the improvement of health-care delivery, opera-
tions, and patient outcomes. He also founded and held senior
management positions at Commoca Inc., a supplier of con-
verged enterprise and residential client devices and services.
He has 26 U.S. patents.
Address for Correspondence: Rohin Moza, University of
Texas Southwestern Medical Center at Dallas, Dallas, TX
75390 USA. E-mail: rohin.moza@utsouthwestern.edu.
References
[1] Department of Health and Human Services, Proposed Rules, Fed. Register,
vol. 72, no. 85, p. 24720, May 3, 2007.
[2] L. Khaodhiar, T. Dinh, K. T. Schomacker, S. V. Panasyuk, J. E. Freeman,
R. Lew, T. Vo, A. A. Panasyuk, C. Lima, J. M. Giurini, T. E. Lyons, and
A. Veves, The use of medical hyperspectral technology to evaluate microcircula-
tory changes in diabetic foot ulcers and to predict clinical outcomes, Diab. Care,
vol. 30, no. 4, pp. 903910, 2007.
[3] T. Binzoni, A. Vogel, A. H. Gandjbakhche, and R. Marchesini, Detection
limits of multi-spectral optical imaging under the skin surface, Phys. Med. Biol.,
vol. 53, pp. 617636, 2008.
[4] N. A. Nassif, B. Cense, B. H. Park, M. C. Pierce, S. H. Yun, B. E. Bouma, G.
J. Tearney, T. C. Chen, and J. F. De Boer, In vivo high-resolution video-rate
spectral-domain optical coherence tomography of the human retina and optic
nerve, Opt. Express., vol. 12, no. 3, pp. 376376, 2004.
[5] K. Humphreys, T. Ward, and C. Markham, Noncontact simultaneous dual
wavelength photoplethysmography: A further step toward noncontact pulse oxi-
metry, Rev. Sci. Inst., vol. 78, no. 4, pp. 044304-1044304-6, 2007.
[6] M. W. Hemelt and K. A. Kang, Determination of a biological absorber depth
utilizing multiple source-detector separations and multiple frequency values of
near-infrared time-resolved spectroscopy, Biotechnol. Prog., vol. 15, no. 4,
pp. 622629, 1999.
[7] D. A. Boas, M. A. Franceschini, A. K. Dunn, and G. Strangman, In Vivo Opti-
cal Imaging of Brain Function. Boca Raton, FL: CRC, 2002, pp. 194216.
[8] R. Sharp, J. Adams, and R. Machiraju, Physics-based subsurface visualiza-
tion of human tissue, IEEE. Trans. Visual Comput. Graphics, vol. 13, no. 3,
pp. 620629, 2007.
[9] M. H. Chen, C. W. Kerechanin II, D. G. Greenspan, T. B. Criss, S. C. Franckowiak,
J. A. Vincent, and R. S. Pattay, Development of a thermal and hyperspectral imag-
ing, Johns Hopkins APL Tech. Digest, vol. 26, no. 1, pp. 6774, 2005.
[10] E. B. Aksel and A. Akin, Depth resolution by continuous-wave imaging,
in Proc. SPIE., vol. 6629, no. 66290G.
[11] K. Humphreys, T. Ward, and C. Markham, A CMOS camera-based pulse
oximetry imaging system, in Proc. 2005 IEEE Engineering in Medicine and
Biology 27th Annu. Conf., Shanghai, China, Sept. 2005, pp. 34943497.
[12] F. P. Wieringa, F. Mastik, and A. F. W. Van Der Steen, Contactless multi-
ple wavelength photoplethysmographic imaging: A first step toward SpO2
camera technology, Ann. Biomed. Eng., vol. 33, no. 8, pp. 10341041, 2005.
[13] CMS Report for FY 2007, CMS-1390-F, Washington, DC: Rep. U.S. Federal
Register, 2008.
[14] U.S. Census Bureau, 2002 economic censusAmbulatory health care serv-
ices, EC02-62I-01, Washington, DC: U.S. Census Bureau, Nov. 7, 2005.
[15] R. McBride. (2008, Mar. 17). New hypermed chief to drive commercializa-
tion. Mass. High. Tech. [Online]. Available: http://www.masshightech.com/stories/
2008/03/17/story12-New-HyperMed-chief-to-drive-commercialization.html.
___________________________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_____________________
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Robotic System
for Transrectal Biopsy
of the Prostate
T
he high incidence and mortality rate of prostate cancer
means that its accurate and early diagnosis is fundamental
for choosing the right type of therapy. Magnetic reso-
nance imaging (MRI) has shown considerable promise to
image suspected lesions within the prostate. To harness the
possibility of real-time guidance of MRI, a robotic system has
been developed to perform transrectal prostate biopsy inside a
1.5-T closed bore scanner. A specially developed MR pulse
sequence is capable of tracking the needle location in real time
while dynamically updating the scan planes to always include
the needle and target.
Clinical Application
Prostate cancer is a very common form of cancer and consti-
tutes a major global public health issue. In the United States,
about 190,000 new prostate cancer cases were estimated in
2008 alone with more than 28,000 deaths [1]. There is a need
for reliable and efficient diagnostic techniques for selecting
the right type of therapy to be used in each case.
If an abnormal result is detected in a patients serum pros-
tate-specific antigen (PSA) level or digital rectal examination, a
needle biopsy of the prostate is carried out for pathological
analysis to diagnose the presence of cancer. The current gold
standard for needle guidance is transrectal ultrasound (TRUS),
because of its simplicity, low cost, and real-time nature.
Systematic sampling of the prostate under U.S. guidance is gen-
erally performed with preferential sampling of the peripheral
zone, the site of origin of more than 70% of prostate cancers
[2]. However, the limited signal-to-noise ratio (SNR) of ultra-
sound combined with the low sensitivity of TRUS for detecting
lesions means that TRUS-guided biopsy leaves approximately
1531%of prostate cancers undetected [3], with many men fac-
ing repeat biopsies and rising PSAlevels.
These limitations have provoked interest for MRI as an
alternative method for tumor detection and image guidance for
biopsy of the prostate, especially when using an endorectal coil
capable of enhancing SNR in the vicinity of the gland. The
good soft tissue contrast of MRI and its high spatial resolution
allow the central gland and peripheral zone of the prostate to
be clearly distinguished in T2-weighted (T2w) images and the
detection of focal lesions in the prostate. Recent research
shows that MRI is more sensitive for tumor localization than
sextant prostate TRUS biopsy, although the study also recog-
nized that the specificity using MRI alone is lower than the
TRUS-guided biopsy counterpart [4]. Hence, although MRI of
the prostate alone cannot at present substitute the biopsy
process for diagnosing pathology, the use of MRIs for both
lesion localization and real-time guidance during the biopsy
procedure has the potential to increase its diagnostic accuracy,
overcoming some of the main drawbacks of TRUS guidance.
High-quality images of the prostate are best obtained using
closed-bore MRI scanners because of their higher field strengths.
However, these clinical scanners, which have a typical bore size
of around 500700 mm in diameter and 125200 cm in length,
impose severe spatial restrictions on the practitioner when trying
to access patient anatomy for biopsy and treatment. By locating a
manipulator alongside the patient inside the bore of an MRI scan-
ner, accurate positioning of a biopsy needle into the prostate
under real-time image guidance could be accomplished.
Engineering Challenges
The MR environment provides significant engineering chal-
lenges when developing MRI-compatible mechatronic devices.
The device must operate under intense static and switching
magnetic fields present, without posing a safety hazard to the
patient or the equipment and without degrading in excess the
quality of the images. This requires developing devices without
components based on electromagnetism (e.g., dc motors and
magnetic sensors) and materials presenting very low magnetic
susceptibilities (e.g., polymers and ceramics) that would other-
wise experience large torques and forces. Any electronic circuit
in the system must be shielded to avoid electromagnetic noise
from interfering with the receiver coils on the scanner. In addi-
tion, conductive materials such as metal components or cables
can induce heating and potentially burn the patient, and mea-
sures must be taken to avoid these.
Currently, the field of MRI-compatible manipulators is
experiencing a rapid development with an increasing number
of systems reported in the research literature [5][8] as well as
the presence of a few commercial devices such as Exablete, an
MR-guided focused ultrasound ablation device for uterine fib-
roids [9], and Innomotion, an MR/computerized tomography
BY HAYTHAM ELHAWARY, ZION TSZ-HO TSE,
MARC REA, ALEX ZIVANOVIC, BRIAN L. DAVIES,
COLIN BESANT, NANDITA M. DESOUZA,
DONALD MCROBBIE, IAN YOUNG,
AND MICHAEL U. LAMP
ERTH
EYEWIRE
Real-Time Guidance Under MRI
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
(CT)-compatible needle puncture device for percutaneous
interventions [10], [11], which was recently acquired by Syn-
thes Inc. Clinical applications that have been addressed by
these devices include breast biopsy [12], liver therapy [13],
neurosurgery [14], prostate biopsy [15], [16], and brachyther-
apy [17]. Technological solutions to the challenge of accurate
actuation within the MR scanner bore have included a custom-
designed pneumatic stepper motor [18], adapted pneumatic
cylinders [16], and the use of piezoceramic motors [14], [19].
This article presents the design of a robotic system able to
perform targeted transrectal biopsy of the prostate inside a
1.5-T closed-bore MRI scanner under real-time image guid-
ance. The main contributions of the system are twofold: 1) it
can be operated within the scanner field of view(FOV) without
degrading image quality while being actuated using specially
engineered piezoceramic motors and 2) a custom-developed
imaging sequence allows dynamic update of the scanner planes
to always include the needle trajectory and biopsy target pro-
viding real-time image guidance for the intervention.
System Design
Design Specifications
After consultation with radiologists from the Royal Marsden
Hospital, Surrey, United Kingdom, the following high-level sys-
tem specifications were set. 1) The device should be able to per-
form a transrectal prostate biopsy inside a 1.5-T closed-bore
MRI scanner froma predefined region in the prostate. Given that
lesions of 1 cm or larger were to be initially targeted [16], a
minimum target accuracy of 5mm was set. 2) The system
should be controlled at all times by the practitioner, that is, it will
not be autonomous. By including the practitioner in the control
loop, an additional safety measure is ensured. 3) The control
console of the device will be located inside the scanner room so
that the practitioner can be close to the patient during the inter-
vention in case of an emergency. 4) The system must not
provide any danger to patients, staff, and equipment, and it must
not reduce in excess the SNRor diagnostic quality of the MRIs.
System Overview
The prostate biopsy system consists of a robotic manipulator
that is situated by the patient inside the MRI scanner, as shown
in Figure 1. The patient would be positioned in the left lateral
decubitus position, and the robot is located beside the anus and
proximal to the scanner isocenter. The robotic device moves an
endorectal probe inside the rectum, with the probe containing a
biopsy needle inside a needle channel and a radio frequency
(RF) receiver coil to increase local SNR in the vicinity of the
prostate. The practitioner will be located at a control console
2 m away from the entrance to the scanner bore, which assures
that any ferromagnetic materials included in the console are
not affected by the magnetic fields. The practitioner then
provides commands to the robot manipulator while receiving
real-time updated MRIs of the prostate region and the needle
trajectory. Once the needle is correctly aligned with a target
position in the prostate, it is fired and a biopsy sample obtained.
1
3
4
2
1
2
3
4
MRI Scanner
Patient in Left
Decubitus Position
Robot Biopsy
Manipulator
Practitioner at
Control Console
(a) (b) (c)
Fig. 1. Setup of robotic prostate biopsy device inside MRI scanner bore. (a) Schematic representation of the system and
patient. (b) Touch-screen controller located close to scanner bore from which the physician controls the robotic device and
receives updated MRIs used for guidance. (c) Robotic device prepared for phantom tests.
Prostate cancer is a very common form
of cancer and constitutes a major global
public health issue.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Mechanical Design of the Robot
The robot manipulator presents 5-degrees of freedom (DoF)
and is presented in Figure 2(a). It consists of a 3-DoF Carte-
sian stage, which is connected to an endorectal probe by
means of a passive gimbal mechanism that allows rotation in
the two axes. The endorectal probe is inserted into the rectum,
and by moving the Cartesian stages, the probe pivots around a
rotation point at the anus by changing its pitch, yaw, and
insertion depth. Attached to the end of the endorectal probe
are the remaining 2-DoF composed of a biopsy needle inser-
tion and firing mechanism, which are actuated once the nee-
dle is correctly positioned. The kinematic diagram of the
robot is illustrated in Figure 2(b). In [20], we reported the
development of MRI-compatible 1-DoF linear stage. These
stages could act as individual linear modules or connect
together to form multi-DoF kinematic chains. Each stage con-
sists of two piezoceramic motors tightly coupled to a linear
slider, with a surface mount incremental optical encoder to
provide the position information. The stages were demon-
strated to operate inside a 1.5-T MRI scanner even when
located inside the imaging FOV, which was made possible by
filtering and shielding all the cables to the motors and
encoders, as described below. A maximum load capability of
approximately 17 N was reported for each axis, and the motor
drive electronics allowed speeds of up to 6.1 mm/s. To formthe
3-DoF Cartesian stage of the prostate biopsy robot, three 1-DoF
stages were connected together in perpendicular directions, as
shown in Figure 2(c). Each stage was manufactured using Del-
rin (acetal homopolymer) as its structural material and held
together with brass and nylon screws. The endorectal probe was
made using a rapid prototyping technique to design an ergo-
nomic shape. The probe is made from Duraform, a glass-filled
nylon powder that can be sterilized (autoclavable), and provides
relatively high rigidity. The probe [shown in Figure 2(e)]
includes a recess designed to carry a needle channel that guides
the biopsy needle into the prostate. In addition, two passive
fiducial markers (tuned microcoils) were embedded, which will
allowthe position of the probe to be tracked by the MRI scanner
for real-time guidance.
The needle insertion and firing mechanism can be seen in
Figure 2(d). For needle insertion, joint actuation is produced
by means of a single piezoceramic motor that provides transla-
tion of the carriage along the linear guides, with a surface-
mount optical encoder for position. The carriage has a
specially shaped groove designed to allow a MR-compatible
18-gauge biopsy needle to be attached. Between the carriage
and the gimbal, a pair of piezoresistive force sensors have
been included to provide force feedback. We previously eval-
uated the use of the force sensor in an MR environment in
[21], which showed only a marginal incidence on the MRIs.
By mounting this sensor to the needle insertion axis, a force
profile can be generated, which shows the force as the needle
3 DoF
Cartesian
Stage
Needle Insertion
Needle Insertion
Needle
Insertion
Gimbal
Biopsy Needle
Biopsy Needle
Endorectal
Probe
Needle Firing
Needle Firing
Passive Gimbal
Y Axis
X Axis
Z Axis
Needle Firing
Piston-Cylinder
Endorectal
Receive Coil
Endorectal
Probe
Needle
Channel
Gimbal
Axis 1
Gimbal
Axis 2
Three Linear
Stages
Fiducial
Markers
T
Y
7
Y
6
Y
4
Y
3
Y
2 Y
1
Y
0
Y
5
X
7
X
6
X
4
X
3
X
1
X
0
X
2
X
5
Z
7
Z
6
Z
5
Z
4
Z
3
Z
1
Z
0
Z
2
N
(a) (b) (c)
(d) (e) (f)
Fig. 2. (a) 5-DoF robotic manipulator for transrectal biopsy of the prostate, (b) kinematic diagram of the robot, (c) 3-DoF Car-
tesian stage formed by connecting three 1-DoF linear modules together with a horizontal axis attached to a plastic bearing
and placed on a support block to minimize deflection, (d) needle firing and translation stage, (e) endorectal probe with RF
receiver coil and fiducial markers embedded, and (f) detailed view of the passive two-axis gimbal mechanism.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
is inserted into the prostatic tissue, providing an extra real-
time indicator of when the needle pierces the prostate capsule.
Once the needle has been inserted, it must be fired to obtain
the biopsy. A custom-made pneumatic piston and cylinder
were developed that provides large forces at high speeds and
changes between the two binary positions: needle retracted
and needle fired. Pressurized air at 1.5 bar was provided by the
internal air supply of the MR room after regulation or by a
portable compressor located in the control room. A valve is
required to turn on or off the air flow to the cylinder, which
will control the firing of the needle.
Control Console and Hardware
To allow an intuitive controller for the robotic device, several
hardware components were required. This consisted of a motion
controller and a touch screen-based PC that hosted a user inter-
face. The motion controller Compact Rio 9004/9104 included a
field-programmable gate array (FPGA) chassis, a real-time
processor, and input and output modules to receive the electrical
pulses from the optical position sensors and to send high-
frequency signals to the motor drive electronics. Closed-loop
position control with a proportional-integral differential (PID)
controller was implemented using LabView 8.1 on each of the
robot axes that were actuated by means of the piezoceramic
motors. This controller provided an root mean square (RMS)
position error of 0.049 0.014 mm (P < 0.05) for each of the
linear axes. As the system requires cables going from the con-
trol box to the robot inside the scanner, measures are required
to avoid these cables acting as antennas bringing into the
scanner RF noise from inside the box. These measures
included shielding all cables, installing hardware low-pass
filters (TUSONIX 4209-053) at the wall of the shielded
enclosure to attenuate any high-frequency signals which
could be picked up by the sensitive scanner receiver coils,
and connecting the shield of the box to the ground of the scan-
ner. As the system specifications required the control console
of the robot to be located inside the scanner room, all these
hardware components were introduced inside a large shielded
enclosure that acts as a Faraday cage. The enclosure was
made from aluminum panels of 1.5 mm thickness and is
shown in Figure 1(a). The system software and graphical user
interface were run on an embedded PC (Intel Pentium M,
400 MHz, 1 GB RAM, Windows XP embedded) and inter-
faced to a 15-in touch screen for user input and output. The
touch screen was covered with a shielded film to reduce inter-
ference with the images. The embedded PC and the motion
controller were connected via a network connection.
MR Real-Time Tracking
Position sensors are included in the robot; however, relying
exclusively on these to calculate the position of the biopsy
needle can mean that organ deformation, mechanical backlash
and inaccuracies in the kinematic model, control algorithm or
registration process can be ignored, compromising the accu-
racy of the biopsy. Another method for end-effector position
calculation would be to use the MRIs. In [22] and [23], we
described the development of an imaging technique that com-
bines the use of passive fiducial markers and real-time
processing of scanner images to provide end-effector localiza-
tion and tracking inside the FOV at a real-time update rate with
accuracy. The method has a twofold objective: 1) to provide
end-effector localization and tracking updated in real-time
and 2) to provide automatic slice alignment of the scan
planes to produce images that include the end effector at all
times. This section provides a summary of the technique.
Passive Fiducial Markers
Tracking the endorectal probe in the rectum provides an ideal
reference from which to calculate the needle tip, as there is a
fixed geometric relationship between the needle and the probe.
In addition, this relationship does not rely on any of the system
kinematics or anus pivot point localization and is thus free
fromthe sources of error associated to these. Two passive fidu-
cial microcoils [see Figure 3(b) and (c)] were implanted in the
head of the probe, as shown in Figure 2(e), with a separation of
35 mm. Only two fiducials were necessary to track the probe in
three dimensions, as rotation around its own axis (roll rotation
of the probe) is constrained. These passive fiducial microcoils
are designed to resonate in a 1.5-T MRI scanner, locally ampli-
fying the signal produced by a vinyl plastisol gel material sold
under the name of Dermal Pads contained inside the coil. The
gel will thus appear with increased SNR against the back-
ground on a lowflip-angle MRI, leaving the surrounding tissue
relatively unperturbed and available for subsequent imaging.
Scan Plane Alignment, Fiducial Tracking,
and Needle Localization
Three imaging planes were defined with respect to the endor-
ectal probe to track its position, as shown in Figure 3(a). Two
planes were used for tracking while the third was used for
imaging the needle and biopsy target [22]. At the beginning of
the biopsy procedure, the user at the host computer manually
Intersection
Point
Needle Tip
Fiducials
Inductance Coil
(a)
(b) (c)
Vinyl Gel
Capacitor
3 mm
3

m
m
Needle Plane
Fiducial Plane
Sagittal Plane
y
x
z
Fig. 3. (a) Endorectal probe showing the relative orientation
of the three selected imaging planes. The fiducial and sagit-
tal planes are used as tracking slices, whereas the needle
plane includes the needle trajectory and biopsy target. The
fiducials form a straight line that crosses the needle at the
intersection point (cross-hair), from which the needle tip can
be easily calculated. (b,c) Fiducial microcoil markers tuned
to resonate at 1.5 T, which appear as bright spots on low
flip-angle MRIs; each fiducial measures approximately 3 3 3 3
5 mm
3
in size. This figure is slightly modified from [22].
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
aligns the previously defined planes on the probe, and a set of
images is taken. As the images taken have a low flip angle of
only 2, everything on the image appears dark, except for the
fiducial markers that appear as bright spots due to the local
magnification of the field in their proximity. These tracking
images are then processed by the scanner image reconstruc-
tion machine following a specially developed pulse sequence
and tracking algorithm [22], [23]. The three-dimensional
(3-D) coordinates of the two fiducials are then extracted from
the images. The accuracy of the tracking algorithm to detect
fiducial positions gave an average RMS error of 0.36 0.17
mm (P < 0.05) [22]. With the extracted 3-D coordinates of the
fiducials, two new tracking planes centered on the fiducials
were then calculated and fed back into the scanner via the
local network. The minimum update rate of the imaged track-
ing planes was 2.4 s. An image with a higher flip angle can
then be obtained to show the needle trajectory and target anat-
omy. The algorithm is capable of tracking the fiducial
positions as the robot moves the probe as long as the fiducial
markers do not move out of the image planes in the time
between each image update. By choosing a slice thickness of
8 mm and considering the slow speeds of the robot this is gen-
erally assured, so the coordinates can be tracked and the scan
plane updated in real time to always include the markers.
With the fiducial coordinates known, the needle end-effector
position can be easily calculated as there is a fixed geometric
relationship between the needle tip and the fiducial markers. A
graphical representation of the needle trajectory is then overlaid
onto each of the scan planes by our system software. An over-
viewof the full measurement protocol is given in Figure 4.
System Evaluation
MR Compatibility
It is paramount that any systemdesigned for use within an MRI
scanner should prove to be safe to the people and equipment
present in the scanner room.
Additionally, the device must
function properly under the
effects of the scanner fields.
We previously performed a
qualitative evaluation of the
forces, torques, and RF heat-
ing induced on piezoceramic
motors, optical encoders, and
structural materials used in the
robotic device [20]. Addi-
tional tests, as defined by
American Society for Testing
and Materials (ASTM) stan-
dard F-2119, were also per-
formed to quantify any image
artifacts produced by these
elements, with the maximum
image artifact measured being
39 mm [20]. When designing
the manipulator for use inside
the FOV of the scanner, care
was taken to ensure that the
location of these components
was such that they did not affect
the imaged region of interest.
In terms of evaluation of the
entire device when actuated
within the FOVof the scanner,
SNR tests were performed to
quantify any electromagnetic
noise introduced into the
MRIs due to the presence of
Scanner
Host
Computer
Sagittal Plane
Fid 2, X/Y/Z = 80.0/110.5/17.6(mm)
Fid 1, X/Y/Z = 82.4/111.8/19.60(mm)
Target, X/Y/Z = 62.7/111.5/10.2(mm)
Intercept, X/Y/Z = 84.6/113.0/52.2(mm)
StoT = 3.8
CtoT = 2.1
StoT = 27.5
CtoT = 2.1
Tracking Slice
Fid 2, X/Y/Z = 80.0/110.4/19.0(mm)
Fid 1, X/Y/Z = 82.3/111.8/18.0(mm)
Target, X/Y/Z = 62.7/111.5/9.6(mm)
Intercept, X/Y/Z = 84.4/113.1/51.7(mm)
StoT = 3.6
CtoT = 2.1
StoT = 27.3
CtoT = 2.1
Tracking Slice
Target, X/Y/Z = 62.7/111.5/10.2(mm)
Intercept, X/Y/Z = 84.6/113.0/52.2(mm)
StoT = 27.5
CtoT = 2.1
Tracking Slice
Fiducial Plane Needle Plane
Calculate Fiducial
in Plane Coords
Calculate Fiducial
3-D Coordinates
Calculate Slice
Orientations
Draw Graphical
Overlays
Image Processor
Images
1 2 6
5
7
4 3
Feedback Data
Fig. 4. The designed pulse sequence for imaging, fiducial coordinate calculation, and scan
plane orientation feedback. The two tracking images (1, 2) are obtained and the scanner
pauses while the images are processed (3) and the resultant 3-D fiducial locations are calcu-
lated (4). These positions are then used with knowledge of the probe geometry to calculate
the new slice orientations (5), which are fed back into the scanner permitting imaging of the
needle plane (6) and graphical overlay of end-effector position (7) before repeating the
loop. This figure is taken from [22] with some modifications.
High-quality images of the prostate are best
obtained using closed-bore MRI scanners
because of their higher field strengths.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
the system hardware inside the scanner bore. Scans were taken
under the following conditions: 1) the presence of a cylindrical
phantomfilled with a CuSO
4
-doped solution (1.25 g/L concen-
tration) located at the scanner isocenter, 2) the robot manipu-
lator located next to the phantom with one axis powered and
actuated, 3) manipulator with two axes actuated, 4) manipu-
lator with three axes actuated, and 5) manipulator with all
axes actuated. These experiments were done using the body
coil of a 1.5-T Avanto and with a True fast imaging with
steady state precession (FISP) gradient echo sequence time
to repeat (TR)/time to echo(TE) = 6.46/3.05 ms, nine slices,
5-mm slice spacing, matrix size 256 3256, flip angle=80,
voxel size = 0.9 30.9 35 mm
3
) and a turbo spin echo
sequence (TR/TE =2,000/85 ms, nine slices, 5-mm slice spac-
ing, matrix size 256 3256, voxel size = 1.2 31.2 35 mm
3
).
The average SNR of the image slices was computed using the
following definition [24]: SNR =P
center
/SD
corner,
where P
center
is the mean signal in the 40340 pixel region in the center of
the phantom and SD
corner
the standard deviation of the signal in
the 40340 pixel region at the bottom right-hand corner of the
image. The results of these tests are presented in Figure 5, and
they show that the maximum SNR degradation (measured as an
average over the imaged slices) produced due to the presence of
the actuated system being 1.1% with the gradient echo True
FISP sequence, which is small enough to be acceptable for this
application. Hence, the system is considered apt for use within
the MRenvironment of a 1.5-T Siemens Avanto MRI scanner.
Temperature Heating of the Endorectal Probe
Temperature tests were performed to make sure that the effect
of repeat scanning would not cause the circuitry inside the
endorectal probe to heat to dangerous levels. A Luxtron 812
fiber-optic thermometer was used to record any temperature
rise induced by coil heating. Saline bags were located in close
proximity to the coil while running a number of sequences
including the clinical sequences and the tracking sequence.
The two temperature-sensing fibers were attached to the head
of the probe, and the temperature was measured to an accuracy
of 0.1 C. These tests showed that there was no significant
temperature rise during the scans, with the temperature read-
ing from the probes being always between 23.5 and 24.5 C.
Phantom Tests in the Scanner
To test the functionality of the system inside an MRI scanner,
a phantomwas built to simulate the anus, rectum, and prostate.
The phantom [see Figure 1 (b)] contained a flexible anus pivot
point and a commercial mock prostate gland made from a
deformable silicone gel. A 10-mm thick layer of vinyl plasti-
sol gel material was placed over the mock prostate that
produces high signal acting like the peripheral zone of the
prostate gland. To perform the phantom biopsies, a target
point within the vinyl plastisol gel was selected. Figure 1
shows the hardware setup for the phantom tests, showing the
control console located close to the entrance of the scanner
bore and the probe inserted into the phantom.
Robot-to-image registration was done using a rigid trans-
form, obtained by implanting three passive fiducials in several
locations on the robot, which clearly showed up on the MRIs.
The biopsy target was then defined on the images, and the
practitioner used motion of the individual joints to finely align
the biopsy needle to the target using the MRIs as feedback.
Once the needle was satisfactorily positioned, it was then
inserted and fired, and the force profile of the needle was
obtained. The needle tip position was then measured on the
scanner image and the RMS error with respect to the desired
position was calculated.
In Table 1, the results for the phantom biopsy tests are
presented. The first column indicates the desired target position,
and the second column shows the coordinates of the needle tip
position in the MRI after the biopsy needle was fired. The errors
vary from a minimum of 1.97 to a maximum of 2.72 mm (aver-
age 2.3 mm), which is within the target accuracy stated in the sys-
temspecifications. MRIs showing the fiducial tracking algorithm
as the needle is aligned toward the target position are shown in
Figure 6, along with the images of the needle tip inside the vinyl
plastisol gel. It shows the low-flip angle tracking images with the
bright fiducials and the graphical overlay presented to the practi-
tioner indicating the end-effector position if the biopsy needle
was to be fired. Figure 7 presents the force profile as the needle
was inserted into the prostate phantom, showing the exact point
at which the needle breaks through the plastisol gel.
There are several sources of error that contribute to the error
measurement obtained during the phantom trials: 1) registra-
tion error: an average error of approximately 1.2 mm was
obtained in the registration process due to the accuracy of the
GE (TrueFISP)
SE (TSE)
1.10
1.05
1.00
0.95
0.90
0.85
0.80
0.75
0.70
P
h
a
n
t
o
m
O
n
e

A
x
i
s
T
w
o

A
x
i
s
T
h
r
e
e

A
x
i
s
F
o
u
r

A
x
i
s
Experiment Conditions
N
o
r
m
a
l
i
z
e
d

S
N
R
1.1% 0.7%
Fig. 5. SNR tests performed with the prostate biopsy manipu-
lator introduced inside the scanner bore while being actu-
ated, showing a maximum degradation of 1.1%. The image
sequence parameters are also displayed.
Table 1. Error measurements using robot kinematics as
gross positioning for the needle and the fiducial tracking
images for fine positioning.
Phantom Trial Kinematics
Target Position
(mm)
Measured Position
(mm)
Error
(mm) X Y Z X Y Z
80.2 72 21.8 82 72 21 1.97
75.3 72.9 11.4 77 74 10 2.46
77.4 85.4 10.7 80 85 10 2.72
83.7 85.5 28.2 83 85 26 2.36
84.8 88.9 25.2 87 89 25 2.21
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
fiducial coordinate extraction algorithm (0.36 mm). 2) Needle
deflection in tissue: as the needle is inserted into the tissue (in
this case the vinyl gel), it is subject to forces that can cause the
needle to deflect. This motion is not considered by the fiducial
tracking algorithm when calculating the end-effector position.
3) Tissue deformation: when performing the biopsy, a prede-
fined target was selected in the phantom. As the probe moves
against the prostate, the tissue may deflect causing our desired
target position to move. Unless the target can be seen on the
MR-tracking images, it may be necessary to take a high-reso-
lution image of the prostate before firing the needle to ensure
correct needle alignment to the anatomical target. 4) Needle
tip measurement: once the needle is fired into the vinyl gel,
the needle tip position was measured at the scanner console on
the host computer. However, the exact coordinates of the
needle tip are subject to error
as an artifact is produced at
the tip, as displayed in Figure
6(g) and (h).
Conclusions
and Discussion
We have developed a prototype
of a 5-DoF robotic system
designed to perform transrectal
biopsy of the prostate in a
closed bore 1.5-T MRI scanner.
The robot axes are powered by
piezoceramic motors, with opti-
cal encoders assuring accurate
position control, and piezoresi-
sitive force sensors providing
force feedback of the needle
insertion. This system provides
a considerable advantage over
several of the previous devel-
oped systems, as the device can
be located and actuated inside
the scanner FOV close to the
patient without distorting the
scanner images. By providing a
technology that is compatible with the MR environment and still
capable of accurate and reproducible motion, we have established
a base upon which many other devices can be developed for other
clinical applications related with MR-guided percutaneous inter-
ventions. In addition, this device can be extended for therapy,
such as prostate brachytherapy, with only minor modifications to
the design.
Another important characteristic of the system is the develop-
ment of an imaging pulse sequence, which could provide real-
time tracking of the end-effector location and updated scan
plane images to always include the needle trajectory and target
anatomy. At the moment, the fastest update time of the fiducial
tracking algorithm was 2.4 s using a FLASH 2-D gradient echo
sequence. However, that time could be considerably reduced if a
faster pulse sequence, such as a gradient echo True FISP, was to
be used. However, as the pulse sequence had to be developed
specifically for this application, we were obliged to make it
scanner specific, in this case for a 1.5-T Siemens Avanto. This
provides one of the limitations of the system, reducing the port-
ability of the device with scanners fromdifferent manufacturers.
The tracking sequence is used to provide guidance during the
positioning of the biopsy needle by the robotic system. The low
resolution of these tracking images requires that the planned
biopsy target on the prostate be selected from high-resolution
images, such as those acquired with a spin echo sequence,
obtained at the beginning of the intervention. Once the needle is
positioned and aligned with the target by the robot, before it is
fired, another high-resolution image should be taken to confirm
alignment with the suspected lesion.
From an engineering perspective, we have demonstrated cor-
rect functioning of the device within the MR environment, and
phantom trials show its potential efficacy and safety for clinical
trials in humans. Preliminary trials on patients are necessary to
demonstrate the robots robustness, accuracy, and adequacy to
perform an intervention within the constraints of a clinical set-
ting. Currently, trials with a small cohort of patients are being
6
5
4
3
2
1
0
0 10 20
Displacement (mm)
F
o
r
c
e

(
N
)
30 40
Fig. 7. Force profile of biopsy needle as it is inserted into the
prostate phantom. An increase in the force is registered at
an insertion depth of 18 mm, which is the point in which the
needle enters in contact with the plastisol gel.
Fid 1
Fid 2
Needle
Artifact
(a) (b) (c) (d)
(e) (f) (g) (h)
Needle
Artifact
Fig. 6. (a)-(f) MRIs showing the fiducial tracking algorithm as the needle tip approximates the
target point (FLASH 2-D, TR/TE = 9.2/4.8 ms, 5-mm slice thickness, 2 flip angle, matrix size
1283128, endorectal probe). The red and blue boxes indicate the detected fiducial markers
while the yellow box indicates the tip of the biopsy needle which needs to align with the
biopsy target, shown as a yellow crosshair. Needle tip inserted into the vinyl plastisol gel in (g)
sagittal plane and (h) needle plane (same flash 2-D sequence but with a 30 flip angle).
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
undergone with this device at our collaborative site at the Royal
Marsden Hospital, although these trials are more focused on
proving the systems capabilities than on cost effectiveness of
the procedures compared to other alternative biopsy methods
such as ultrasound guided biopsy. Longer-term validation with
cost effectiveness will require more time and further testing at a
wide range of sites and operators. The work presented in this
article is predominantly about classic medical technology devel-
opment encompassing the initial steps of device/component
design and development, system integration, and in the near
future, the completion of phase I/II clinical trials. Long-term
economic feasibility will only be possible if the procedures
become easier, cheaper, and faster than at present, with this
being an important goal for the technology we hope to develop.
This robotic system provides an alternative to TRUS-guided
biopsy for prostate cancer diagnosis. The main expected advant-
age is the capability to perform targeted biopsies from specific
suspicious regions in the prostate as highlighted by high-resolu-
tion MRIs, as opposed to the multiple blind sampling that occurs
with the TRUS-guided approach. However, it is also clear that
MR-guided biopsy will require access to the already scarce and
expensive MR suite, is expected to take a longer procedure time
when considering robot setup and biopsy, and adds a significant
level of complexity compared to its TRUS counterpart. These
issues may suggest that although MR-guided prostate biopsy
may not completely substitute routine TRUS-guided procedures
in the future, the approach may well be used for those cases
where repeated TRUS biopsies show negative diagnosis while
PSA levels still increasing. For such a subset of patients, MR-
guided biopsy may be of significant importance.
The multidisciplinary nature of this project meant that a
team with a broad range of expertise was required to fulfill the
objectives of this project. A team comprising engineers,
computer scientists, physicists, and clinicians was thus
assembled, involving collaboration between the Mechanical
and Electrical Engineering Departments of Imperial College
London for system design and development, the Radiological
Sciences Unit from Hammersmith Hospital for the MRI and
sequence development expertise, and the Royal Marsden Hos-
pital for their clinical input and subsequent clinical trials.
Informal collaborations have also been fundamental to the
success of this project, with Siemens, the scanner manufac-
turer, offering considerable technical support for the integra-
tion of our custom developed RF coils and sequences.
Collaborations and teams such as these are required if research
is to be applied to solve real clinical needs and developed in a
setting that will facilitate its acceptance into clinical trials and
ultimately future practice for patient care.
Acknowledgments
This project has been funded by the New and Emerging
Applications of Technology (NEAT grant D083) program
from the U.K. Department of Health.
Haytham Elhawary received his M.Eng.
degree in mechanical engineering from the
School of Engineering (TECNUN) of the
University of Navarra, Spain. He received
his Ph.D. degree in medical robotics at the
Mechatronics and Medicine Laboratory at
Imperial College London, United Kingdom.
He is currently undergoing postdoctoral
research at the Brigham and Womens Hospital of the Harvard
Medical School. His current research interests are in the field of
medical robotics, and more specifically, the design and develop-
ment of hardware and software for image-guided interventions.
Zion Tsz-ho Tse obtained his Ph.D.
degree at Imperial College London in the
area of medical robotics in 2009. He is
currently a radiology research fellow in
Brigham and Womens Hospital, Harvard
Medical School. His current research inter-
ests include MR-compatible surgical ro-
bot, haptic control and tactile sensing, and
MR-compatible 12-lead EKG system. His research also fo-
cuses on modeling and suppression of magnetohydrodynamic
(MHD) effect to improve gating and differentiation of MR-
cardiac imaging and interventions.
Marc Rea obtained an M.Phys. degree from
the School of Physics, Lancaster University,
in 1999. He then went on to achieve the
M.Sc. degree in Interactive Multimedia Sys-
tems (Liverpool James Madison University)
and Medical Physics (Leeds University) and
is currently working as a trainee clinical
scientist in MRI, while working toward a
Ph.D. degree at Imperial College London, United Kingdom. His
current research interests are in the field of MRI compatibility
of devices and systems, including the automatic tracking of
mechatronic devices within the MRI environment.
Alex Zivanovic gained his B.Sc. (Hons.)
degree in computer systems engineering
(informatics) and an M.Sc. degree in elec-
tronic engineering, both from the University
of Kent at Canterbury, United Kingdom.
He earned his Ph.D. degree in mechanical
engineering from Imperial College London,
United Kingdom, where he is currently a
research associate. His research interests are in the applications
of mechatronics in medicine, in particular, the development of
MRI-compatible manipulators and haptic systems for training
and intervention.
Brian L. Davies is the technical director of
the Acrobot Company, Ltd., London, United
Kingdom, and an emeritus professor of
medical robotics at Imperial College, Lon-
don, where he founded and leads the Mecha-
tronics in Medicine Research Group. He was
recently recognized by the award of a D.Sc.
for his international contribution to medical
robotics and was awarded the Maurice Muller Prize by the Inter-
national Society of Computer-Aided Orthopedic Surgery for his
contributions to medical research. He also is the recipient of the
IEE award for Outstanding International Contributions to
Computer-Aided Surgery.
Colin Besant earned an engineering degree at London
University and completed a Ph.D. degree in nuclear engineer-
ing in 1960. He then joined Imperial College London, United
Kingdom, where he built up a large research group working
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
on engineering design and manufacturing
systems with particular emphasis to indus-
trial needs. He was made a Fellow of the
Royal Academy of Engineering in 1987
and a professor of computer-aided manu-
facture in 1989. He is the chief executive
officer of the spin-off company Turbo
Genset.
Nandita M. deSouza received her B.Sc.
(Hons.) in physiology in 1978 from the
University of Newcastle Upon Tyne, Unit-
ed Kingdom, M.B.B.S. degree in 1983,
M.R.C.P. degree in 1986, F.R.C.R. degree
in 1991, M.D. degree in 1996, and F.R.C.P.
degree in 2001. She is a reader in imaging,
Imperial College, 20002002, and a consul-
tant in Imaging Hammersmith Hospital NHS Trust, 20022004.
She is a codirector of the MRI Unit at the Institute of Cancer
Research. Her main research interests are in gynecologyl, pros-
tate, and breast tumors, using MRI to understand biology,
improve staging, and monitor treatment response. She holds a
Cancer Research U.K. programme grant in MRI and several
project and studentship grants. She has more than 90 peer-
reviewed articles, several book chapters, and is the editor of a
multiauthor book on Endocavitary MRI of the Pelvis.
Donald McRobbie received his B.Sc.
degree in physics and M.Sc. degree in
medical physics from the University of
Aberdeen. He was awarded his Ph.D.
degree in magnetic resonance imaging
from the University of London in 1992. He
is currently the head of Radiological and
MR Physics, deputy director of the Radio-
logical Sciences Unit in the Hammersmith Hospitals NHS
Trust, and an honorary senior lecturer in imaging at Imperial
College, London. His research interests include functional
MRI and voxel-based morphometry, 3-D imaging of the air-
ways, MR device tracking and guidance, and dosimetry in
computed tomography. He is an author of the award-winning
textbook MRI from Picture to Proton.
Ian Young is a senior research fellow in the Electrical and
Electronic Engineering Department, Imperial College Lon-
don, United Kingdom. He has been working in the field of
MRI for more than 30 years.
Michael U. Lamperth obtained his
M.Sc. and Ph.D. degrees at Imperial Col-
lege London. He received his education
as a mechanical engineer from the MSW
Winterthur Professional School and Col-
lege of Technology, Switzerland. He is
currently a lecturer at the Mechanical
Engineering Department, Imperial Col-
lege London, where he leads a number of researchers in the
Mechatronics in Medicine Laboratory. His research interests
are mainly twofold. He works in the field of medical robotics
and implants, with a focus on using MRI for real-time guid-
ance and planning. He also conducts research in hybrid elec-
tric vehicles.
Address for Correspondence: Michael U. Lamperth, Mecha-
tronics in Medicine Laboratory, Mechanical Engineering
Department, Imperial College London, South Kensington
Campus, London SW7 2AZ, United Kingdom. E-mail:
m.lamperth@imperial.ac.uk.
References
[1] A. Jemal, R. Siegel, E. Ward, Y. Hao, J. Xu, T. Murray, and M. Thun,
Cancer statistics 2008, CA Cancer. J. Clin., vol. 58, no. 2, pp. 7196, 2008.
[2] C. Tempany, S. Straus, N. Hata, and S. Haker, MR-guided prostate interven-
tions, J. Magn. Reson. Imag., vol. 27, no. 2, pp. 356367, 2008.
[3] A. V. Damico, C. M. Tempany, R. Cormack, N. Hata, M. Jinzaki, K. Tun-
cali, M. Weinstein, and J. P. Richie, Transperineal magnetic resonance image
guided prostate biopsy, J. Urol., vol. 164, no. 2, pp. 385387, 2000.
[4] A. E. Wefer, H. Hricak, D. B. Vigneron, F. V. Coakley, Y. Lu, J. Wefer,
U. Mueller-Lisse, P. R. Carroll, and J. Kurhanewicz, Sextant localization of
prostate cancer: Comparison of sextant biopsy, magnetic resonance imaging and
magnetic resonance spectroscopic imaging with step section histology, J. Urol.,
vol. 164, no. 2, pp. 400404, 2000.
[5] H. Elhawary, Z. T. H. Tse, A. Hamed, M. Rea, B. L. Davies, and M. U.
Lamperth, The case for MRI compatible robotics: A review of the state of the
art, Int. J. Med. Robot. Comput. Assist. Surg., vol. 4, no. 2, pp. 105113, 2008.
[6] H. Elhawary, A. Zivanovic, B. Davies, and M. Lamperth, A review of mag-
netic resonance imaging compatible manipulators in surgery, J. Eng. Med.,
vol. 220, no. 3, pp. 413424, 2006.
[7] R. Gassert, A. Yamamoto, D. Chapuis, L. Dovat, H. Bleuler, and E. Burdet,
Actuation methods for applications in MR environments, Concepts Magn.
Reson. B Magn. Reson. Eng., vol. 29, no. 4, pp. 191209, 2006.
[8] N. V. Tsekos, A. Khanicheh, E. Christoforou, and C. Mavroidis, Magnetic
resonance compatible robotic and mechatronics systems for image-guided inter-
ventions and rehabilitation: A review study, Annu. Rev. Biomed. Eng., vol. 9,
no. 1, pp. 351387, 2007.
[9] InsightecUSA, ExAblate2000Focused Ultrasound Ablation device [Online].
Available: www.insightec.com.
[10] Innomotion, Innomedic, 2007 [Online]. Available: http://www.innomedic.de/
en/index.php
[11] E. Hempel, H. Fischer, L. Gumb, T. Hohn, H. Krause, U. Voges, H. Breit-
wieser, B. Gutmann, J. Durke, M. Bock, and A. Melzer, An MRI-compatible
surgical robot for precise radiological interventions, Comput Aid. Surg, vol. 8,
no. 4, pp. 180191, 2003.
[12] B. T. Larson, A. G. Erdman, N. V. Tsekos, E. Yacoub, P. V. Tsekos, and I.
G. Koutlas, Design of an MRI-compatible robotic stereotactic device for mini-
mally invasive interventions in the breast, Trans. ASME J. Biomech. Eng.,
vol. 126, no. 4, pp. 458465, 2004.
[13] N. Hata, F. Ohara, R. Hashimoto, M. Hashizume, and T. Dohi, Needle
guiding robot with five-bar linkage for MR-guided thermotherapy of liver tumor,
Lect. Notes Comput. Sci., vol. 3217, pp. 2161168, 2004.
[14] K. Masamune, E. Kobayashi, Y. Masutani, M. Suzuki, T. Dohi, H. Iseki, and
K. Takakura, Development of an MRI-compatible needle insertion manipulator for
stereotactic neurosurgery, J. Image Guid. Surg, vol. 1, no. 4, pp. 242248, 1995.
[15] A. Krieger, R. C. Susil, C. Menard, J. A. Coleman, G. Fichtinger, E. Ata-
lar, and L. L. Whitcomb, Design of a novel MRI compatible manipulator for
image guided prostate interventions, IEEE. Trans. Biomed. Eng., vol. 52, no. 2,
pp. 306313, 2005.
[16] G. S. Fischer, I. Iordachita, C. Csoma, J. Tokuda, S. P. DiMaio, C. M.
Tempany, N. Hata, and G. Fichtinger, MRI-compatible pneumatic robot for
transperineal prostate needle placement, IEEE/ASME Trans. Mechatron.,
vol. 13, no. 3, pp. 295305, 2008.
[17] D. Stoianovici, D. Song, D. Petrisor, D. Ursu, D. Mazilu, M. Muntener,
M. Schar, and A. Patriciu, MRI Stealth robot for prostate interventions, Min-
imal. Invas. Ther. Allied Technol., vol. 16, no. 4, pp. 241248, 2007.
[18] D. Stoianovici, A. Patriciu, D. Petrisor, D. Mazilu, and L. Kavoussi, A
new type of motor: Pneumatic step motor, IEEE/ASME Trans. Mechatron.,
vol. 12, no. 1, pp. 98106, 2007.
[19] K. Chinzei and K. Miller, Towards MRI guided surgical manipulator,
Med. Sci. Monit., vol. 7, no. 1, pp. 153163, 2001.
[20] H. Elhawary, A. Zivanovic, M. Rea, B. Davies, C. Besant, D. McRobbie,
N. de Souza, I. R. Young, and M. Lamperth, A modular approach to MRI
compatible robotics: Interconnectable one DOF stages, IEEE. Eng. Med. Biol.
Mag., vol. 27, no. 3, pp. 3541, 2008.
[21] Z. T. H. Tse, H. Elhawary, and M. U. Lamperth, Adding haptics in MR guided
interventions for soft tissue palpation and stiffness measurements, presented at 4th
Int., Workshop on Medical Imaging and Augmented Reality, Tokyo, Japan, 2008.
[22] M. Rea, D. Mcrobbie, H. Elhawary, Z. T. Tse, M. Lamperth, and I. Young,
System for 3D real-time tracking of MRI-compatible devices by image process-
ing, IEEE/ASME Trans. Mechatron., vol. 13, no. 3, pp. 379382, 2008.
[23] M. Rea, D. Mcrobbie, H. Elhawary, Z. T. H. Tse, M. Lamperth, and I.
Young, Sub-pixel localisation of passive micro-coil fiducial markers in interven-
tional MRI, MAGMA Magn. Reson. Mater. Phys. Biol. Med., vol. 13, no. 3,
pp. 379382, 2008.
[24] K. Chinzei, N. Hata, and F. A. Jolesz, MR compatible surgical assist robot: Sys-
tem Integration and preliminary feasibility study, presented at the 3rd Int. Conf.
Medical Image Computing and Computer-Assisted Intervention, Pittsburgh, PA, 2000.
__________________
______
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Image-Based Motion
Detection
T
he use of image guidance in medical applications is
constantly growing because of its tremendous impact
on the future of health care. Although image-based tis-
sue tracking has been thoroughly explored in the aca-
demic literature for years, it has not yet matured to become
widely accepted by clinicians. Undetected tissue movements
in image-based clinical procedures may cause safety and effi-
cacy difficulties. We introduce an image-based approach for
detecting tissue movements during clinical procedures. Our
method has been validated in more than 600 true clinical
cases. The results show that our algorithm agrees with an
expert analysis in 98% of the cases, showing zero events of
false alarms and zero events of undetected motion. The results
show that the approach provides a clinically ready motion-
detection algorithm. These robust results are achieved by
introducing the concept of weighted directional descriptors
(WDDs). The technique analyzes the directivity and confi-
dence level of each anatomical feature and uses it to weight
local inputs resulting in a robust motion vector. The robust-
ness is further increased by a novel preprocess that screens out
features that may be misleading or are repeated in the adjacent
search zone.
The technique meets the requirements, as defined by our
clinicians, and is now integrated in true medical systems. In
particular, our approach has been uniquely developed and
integrated into a clinical product. ExAblate is the first Food
and Drug Administration (FDA)-approved magnetic reso-
nance (MR)-guided noninvasive surgical device using focused
ultrasound therapy. It is used in commercial clinics and in
leading medical academic research institutions, attesting to
the success of our method and its practical clinical value.
Image-Guided Therapy
In recent years, medical imaging is moving from being an off-
line diagnostic tool to an intraoperative modality, which
together with noninvasive therapy procedures provide a revo-
lutionary concept of image-guided therapy systems [1], [2]. A
typical image-guided treatment starts by acquiring a set of pre-
operative images used to analyze the anatomy and plan the
treatment [3]. However, during the treatment, the anatomy
may change shape and shift from its original placement caus-
ing the planning images and the treatment plan to become
invalid. This is a concern to both safety and efficacy of a treat-
ment that depends heavily on image guidance. Therefore,
there is a need for a tool that will detect motion events and
issue a warning so that the proper actions can be taken by the
physician. A critical success factor of any motion detection
technique lies in its reliability, mainly having zero events of
undetected motion. In addition, it needs to be robust by mini-
mizing the events of false alarm where no motion actually
occurred. Finally, to be integrated in a clinical environment,
the method needs to be fast, requires very little interaction,
and be robust to image quality and user inputs. Our approach
is based on user input to define the area to be monitored. The
target region is analyzed to identify key features to be tracked
for displacements based on updated images of the same anat-
omy taken during treatment.
In this work, we develop a method that meets these require-
ments for adoption in a clinical setting. Our proposed method
was developed as part of a clinical product ExAblate. It is the
first and only FDA-approved medical device for using focused
ultrasound under MR guidance and control. It is commercially
available in the United States, Europe, and parts of Asia with
more than 5,000 treatments performed worldwide. It is
installed in more than 70 clinics and hospitals and in more than
30 top academic centers researching additional potential clini-
cal applications that can benefit fromthis technology.
Although our approach can be generalized to other clinical
applications, in this work, we focus on a medical device
using ultrasound to heat the target tissue to a point of irre-
versible coagulation. Since ultrasound energy can traverse
through soft tissue, this can be done completely noninva-
sively. However, to make it safe and controlled, the energy
needs to be guided to the target, and the thermal rise needs to
be measured and controlled in real time. To achieve this, the
system is integrated inside an MR scanner. Initially, MR
images are acquired to describe the anatomy and plan the
treatment. During the treatment, the target is continuously
scanned to measure the thermal effect and to confirm there
was no movement of the anatomy. This is a concern of clini-
cians that our approach was challenged with. Movement of
the target may result in a situation where the treatment is
BY EYAL ZADICARIO, SHLOMI RUDICH,
GHASSAN HAMARNEH,
AND DANIEL COHEN-OR
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Using the Concept of Weighted
Directional Descriptors
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
being delivered to nontargeted tissue or that some target tis-
sue remains untreated. Both are undesired and are the key
motivation for implementing a motion-detection mechanism
in the treatment procedure.
Our approach is not a generic image analysis technique but
rather a tailored approach to the clinical application we were
challenged with. The device is currently being evaluated and
used for treatment of the targets in uterus, bones, brain, breast,
and prostate. These are quasi-static organs in which motion is
not expected during the treatment.
Background and Related Work
Image-guided therapy of soft tissue organs has been widely
explored in recent years in an effort to use technological devel-
opment in computer-assisted interventions toward improving
patients health care [5], [7][9]. Vast research has evolved in
elastic deformable models for image-based organ deformation
measurements, e.g., [10][12]. Despite these efforts, very little
of this has matured to be accepted in routine clinical applica-
tions. The technological challenge is great since the clinical
environment demands a robust solution with little tolerance to
errors. Tracking of soft tissue organs presents a special chal-
lenge since they undergo nonrigid deformations. Various mod-
els for segmentation and tracking of deformable anatomy have
been proposed using parametric and geometric active contours
[13], [14]. However, the results were insufficiently robust in
situations of objects that have poor contrast [15].
Motion detection was also addressed by attempting registra-
tion techniques. The registration of medical images is in general
a difficult problem with numerous proposed methods [16]. Lin-
ear transformations do not have enough degrees of freedom to
track the deformable anatomy accurately, and dense deforma-
tion fields have too many degrees of freedom requiring long
computation times. Most techniques that estimate dense defor-
mation fields require the setting of the amount of spatial regula-
rization to produce smooth results and counteract image
degradation. However, setting the amount of regularization is a
nontrivial task. Further, the large degrees of freedom renders
the method more susceptible to entrapment into local minima
of the similarity criteria [22][25]. To address these difficulties,
various researchers suggested matching a sparse set of points as
the basis of image registration or shape matching, e.g., [17] and
[18]. The point feature is the simplest form of a feature that can
be extended to curves or surfaces. Variations of this concept
include elastic registrations based on thin-plate splines, in
which a set of corresponding anatomical-point landmarks drive
a minimum bending energy warp [17], [19]. However, point-
based matching has several drawbacks. The presence of noise
and the existences of outliers make it impractical for routine sit-
uations [26]. Finally, the geometric transformations may need
to incorporate high-dimensional nonrigid mappings to account
for deformations of the point sets [17].
A successful and highly adopted approach for extracting
distinctive invariant features that can be used for image
matching [using scale invariant features transform (SIFT)]
was suggested by Lowe (2004) [20], [27], [28]. It has been
shown that SIFT descriptors can be used to achieve a fairly
robust object detection in still 2-D scalar images. SIFT has
been used in medical imaging [4], extended to scalar images
of arbitrary dimensionality, and applied to determine point
correspondence between pairs of medical images in 3-D [28].
However, the approach has several drawbacks. It ignores key
attributes of anatomical structures and may be easily misled
by duplicated features within the area of interest. A compre-
hensive robustness comparison was performed between mo-
tion detection based on SIFT and our proposed approach.
Methods
Overview
The proposed method makes use of information available dur-
ing the existing treatment flow. Planning images are acquired
Local Motion Vectors
N
Y
P
r
e
o
p
e
r
a
t
i
v
e
I
n
t
r
a
o
p
e
r
a
t
i
v
e
Global Motion Test
Do Local Motion
Vectors Exceed Application
Thresholds?
Preprocess
Analysis of Each ADC to Identify Key
Feature Pixels (KFP)
Reference Data
1. Acquire Reference Image Set
2. Mark Anatomical Descriptor Cells (ADC)
Analysis
1. Acquire Intraoperative Image Set
2. Analyze Displacement for Each KFP
3. Analyze Directional Weighted
Average to Obtain Displacement for
Each ADC
Anatomical Descriptor Cells (ADC)
Key Feature Pixels (KFP)
Issue Motion Warning
Fig. 1. An overview of algorithm flow method.
Reference Edges
Passed Grade Filtration
(a) (b)
Fig. 2. Motion-detection components. (a) ADC marked by
the user (circles) and the patch in the image (square) to be
considered in the analysis of the anatomical descriptor. (b)
KFP: an example of one patch from (a). Pixels with high-
gradient value (white) have been screened to select pixels
having autocorrelation above threshold (red).
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
at the beginning of the treatment (e.g., MRT2-weighted image
sequences) covering the entire volume of the treatment target.
At planned intervals (e.g., every 3 s), updated images of some
portion of the target are being rescanned. These data are proc-
essed by the algorithm, and in case motion has been estimated
to exceed predefined thresholds, a motion warning is issued to
the physician.
Our algorithm is separated into four steps (Figure 1). Ini-
tially, a baseline reference image set is acquired to include the
entire volume of interest. In the second step, the volumetric
data are preprocessed to identify landmark features in the
anatomy, anatomical descriptor cells (ADCs), and to select
and classify the most robust key feature pixels (KFPs) within
each cell. The third step is applied during the treatment. The
anatomy is periodically rescanned. In the analysis phase,
the displacement for each
KFP is estimated, contribut-
ing to the local motion vector
of each cell. Finally, in step
four, the motion vectors are
compared with predefined cri-
teria, and a motion warning is
issued if needed.
Preprocessing the
Reference Image Set
A robust motion detection
technique needs to distinguish
between tissue displacements
that are inevitable or have no
effect on the planned proce-
dure (e.g., blood vessels) and
ones that need to raise a clini-
cal concern. This requires a
clinical understanding of the
anatomy and procedure. In
our approach, gestures are
given by the clinical user
marking the areas of clinical
interest. The automated analy-
sis is limited to cells surround-
ing each key feature identified
by the user, labeled as {Ci}.
Figure 2 shows sample cells
(green bounding boxes) that
have been placed surrounding
inputs given by the clinical
user. In each ADC, we need to
screen the KFPs that may
provide reliable evidence for
local displacement.
First, we apply an edge detector [6] to identify pixels that have
significant local gradient to become candidate KFPs. Figure 2(b)
shows the selected pixels for the top cell in Figure 2(a). The fol-
lowing screening steps are based on an autocorrelation matching
score (ACMS) to quantify the similarity between two image
patches (P
1
, P
2
) as defined in (1). Repeating this for each pixel
results in a 2-D autocorrelation map, as can be seen in Figure 3.
The goal of the second step is to distinguish between true features
and noisy pixels. We compute an integral autocorrelation value
for pixel P by normalized sum of pixels in its autocorrelation
map, as defined in (2). A significant anatomical feature has an
integral autocorrelation value that is above a threshold whereas a
noise pixel does not. Figure 3(a) shows an example of a feature
that has low-integral value and (c) a robust feature where the inte-
gral value is above the threshold. The third screening method
(a) (b) (c) (d)
S
W
SSSSSSSSSSSS
WWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWW
Fig. 3. Autocorrelation sample analysis cases. (a) Poor correlation: pixel with correlation value
below the required threshold (CCM = 0.21). (b) Duplicate correlation: Situation of duplicate
feature that may later become misleading (CCM = 0.44 and directional grade = 0.85). (c)
Nondirectional correlation: sample of pixel with high correlation value with a nondirectional
behavior (CCM = 0.62). (d) Directional correlation: directional feature where displacement
along S can be very reliable and displacement along W should be ignored (CCM = 0.51
and directional grade = 0.75).
ExAblate is the first FDA-approved MR-guided
noninvasive surgical device using focused
ultrasound therapy.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
eliminates candidate features that are repeated and can be mis-
leading in the motion analysis phase. The repeated features are
identified by applying a labeling algorithm[21] to the auto cor-
relation map and confirming there is not more than a single-
connected label. Finally, we characterize each KFP according
to its directivity. The motivation for this comes fromthe under-
standing that a key feature may have a directional nature and
this should be properly accounted for in the analysis phase. As
seen in Figure 3(d), some features have a directional nature.
We classify the KFPs to be omni [Figure 3(c)] or directional
[Figure 3(d)]. The directivity score of a KFP (P) is calculated
by a 2-D ray scan on the autocorrelation map. Equation (3)
describes howthe directional score is computed for each ray. If
a certain ray has a significantly higher value than the average,
the pixel is considered to have a preferred directional nature.
The segment with the high value represents its weak (W) direc-
tion and orthogonal to it is the strong direction (S). Figure 3(d)
shows an example of a directional KFP where the vectors of
Wand S are identified.
ACMS(P
R
1
, P
R
2
, ref) =
X
n
i=1
X
m
j=1
ref(P
R
1
)
ij
ref(P
R
1
)
ref(P
R
2
)
ij
ref(P
R
2
)
X
n
i=1
X
m
j=1
ref(P
R
1
)
ij
ref(P
R
1
)
2
8
>
>
>
>
:
9
>
>
>
>
;
X
n
i=1
X
m
j=1
ref(P
R
2
)
ij
ref(P
R
2
)
2
8
>
>
>
>
:
9
>
>
>
>
;
v
u
u
u
u
u
u
u
u
t
:
(1)
Integral Autocorrelation Value (P)
=
1
nm
X
n=2|
x=n=2|
X
m=2|
y=m=2|
ACM
P
(x, y): (2)
directional grade
h
=
1
2m 1
1
X
m
x=1
(ACM
P
(x, y(x, h)) ACM
P
(x, y(x, h)))
" #
: (3)
where
y(x, h) = tan (h)x 9(1 tan (h))|
and
directional grades vector = directional grade
h
h=0:15:165
:
Intraoperative Motion Vector Analysis
During the clinical treatment, the anatomy is periodically
scanned, and an updated input image set is acquired. On the
basis of this updated information, we analyze each ADC. The
cross-correlation computation is repeated for each KFP by
comparing the patch from the original reference image,
ref(P1), with patches from the input image (P2), as defined in
(4), resulting in a cross-correlation score (CCMS). The terms
are identical to the one described in (1). Only in (4), we com-
pare two different images: the reference image and input
image. A pixel with a high score determines a high corre-
spondence between the two patches. However, the location
with highest correlation value is not necessarily the best esti-
mate for the displacement. To enhance robustness, we as-
sume local rigidity within each cell. Equation (5) shows an
energy function that combines a measure for how well the
proposed location corresponds to the original image data
given by (4) and a second term (neighbors correlation) that
reflects the similarity to the displacements of neighboring
KFPs, as in (6). The coefficients for the weighted ratio deter-
mine how much we enforce local rigidity with respect to vari-
ability in local displacement.
CCMS(P
R
1
, P
I
2
, ref, input)
=
X
n
i=1
X
m
j=1
input(P
I
1
)
ij
input(P
I
1
)
ref(P
R
2
)
ij
ref(P
R
2
)
X
n
i=1
X
m
j=1
input(P
I
2
)
ij
input(P
I
1
)
2
8
>
>
>
>
:
9
>
>
>
>
;
X
n
i=1
X
m
j=1
ref(P
R
1
)
ij
ref(P
R
1
)
2
8
>
>
>
>
:
9
>
>
>
>
;
v
u
u
u
u
u
u
u
u
t
: (4)
Correlated Displacement Value
= a 3(Max Cross-Correlation Score)
b 3(Neighbors Corrolation): (5)
Neighbors Corrolation (P
I
j
)
= 1
X
N
n=1
P
I
j
P
I
n
P
R
j
P
R
n
2
: (6)
The estimated displacement of each cell Ci is a directional
weighted averaging of all its related pixels. The weighted
averaging is designed to ensure two things: 1) pixels with
higher cross-correlation value will have more contribution to
the average displacement and 2) directional pixels contribute
only in the appropriate (S) direction [Figure 3(d)] as seen in
(7) and (8).
DX
AVG
=
P
N
i=1
dX
i
3CCVx(Pi)
P
N
i=1
CCVx(Pi)
DY
AVG
=
P
N
i=1
dY
i
3CCVy(Pi)
P
N
i=1
CCVy(Pi)
: (7)
CCV
x
(Pi) =Max Cross-Correlation Score(Pi) 3cos h:
CCV
y
(Pi) =Max Cross-Correlation Score(Pi) 3sin h: (8)
Figure 5 shows an example of how the directional and non-
directional KFPs are combined to result in the local motion
vector. The original location of the cross mark is shown in a
dashed outline. The original KFPs are shown in green cross
marks (labeled 16). A diagonal displacement has been
applied (gray arrow). The contribution of nondirectional KFPs
is taken in both axes (green vectors), whereas in the directional
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
case (blue vectors), only its strong component (red vectors)
affects the final motion vector (gray vector).
The final part of the algorithm compares the displacements
with a predetermined threshold. This threshold is expected to
vary according to the clinical procedure, its safety require-
ments and depending on the treated organ. Since we do not
assume the overall displacement to be rigid, there is no reason
to assume that different anatomical descriptors, which repre-
sent different locations in the anatomy, will agree on the same
displacement. Therefore, the analysis is based on comparing
each descriptor cell with the motion threshold. Figure 4 shows
an example of a global motion test. Figure 4(a) shows the
reference image and Figure 4(b) reveals the input image (after
the motion occurred) taken during treatment. The green circles
show the original location of the anatomic descriptors on both
images. The red circles show the estimated location of the
descriptor on the input image. Figure 4(c) shows a 2-D plot of
the displacement for each descriptor in a sample situation.
When a single or a group of anatomical descriptors have a dis-
placement that exceeds the threshold defined for the applica-
tion, we issue a motion warning.
Data Sets
To test our approach, we used a clinical database that contains
imaging data of hundreds of treatments performed with a nonin-
vasive MR-guided focused ultrasound therapy system. The data
sets were anonymized and originated fromtreatments of uterine
fibroids (i.e., benign tumor in the uterus). We used this sample
database to test the algorithm with specific thresholds defined
by clinicians, setting a displacement of 4 mm (or higher) to be
large and an error below2 mmto be small and insignificant.
MR images are acquired and used to describe and monitor
the patients anatomy during the treatment. Typically, the
treatment volume is covered by 2030 sagittal images ac-
quired for reference at the beginning of the treatment. The
annotation on the reference images includes easily noticeable
landmarks in and adjacent to the target volume. The intraoper-
ative images are of the same modality and scanner and include
one sample slice acquired every 3 s.
The algorithm is implemented using MATLAB infrastruc-
ture running on Windows XP. Performance was typically less
Original Key Feature Pixels
Directional Displacements
Max Correlation Displacements
+
4
3
2
6
5
1
+
+
+
+
+
+
Fig. 5. Local motion vector. Input image with diagonal dis-
placement. Nondirectional contribution (green) and strong
component (red) of nondirectional (blue) are combined to
result in final local motion vector.
(a) (b)
15 10 5 0
X (mm)
5 10 15
15
10
5
0
5
10
15

Y

(
m
m
)
Global Motion Test
Descriptor Movements
(c)
Fig. 4. (a) User-defined anatomical descriptors on the refer-
ence image. (b) Analyzed descriptors (red) on the input
image. User-defined descriptors (green) are superimposed.
(c) Global motion test: the green and red areas mark the
allowed and critical displacements, respectively.
Our approach has been uniquely developed
and integrated into a clinical product.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
than 3 s to analyze ten features in a pair of images of 256 3
256 running on a 3-GHz PCmachine.
Results
Gold Standard Analysis
The most straightforward quantitative test is a comparison to
gold standard. Overall, 606 anatomical descriptors were
selected by the clinical expert users in 32 randomly selected
treatments from the clinical database.
The algorithm results are compared with the results of a
manually marked displacement done by expert clinicians (Fig-
ure 6). According to our clinical criteria, an error lower than
2 mm is considered a hit (acceptable). An error exceeding
2 mm is considered misplaced. An error exceeding 4 mm is
defined as a false alarm. Finally, the category associated with
safety aspects and with the highest concern is a case of motion
not detected by the algorithm(undetected motion or false nega-
tives). Our approach is robust and has zero events of undetected
motion. In Figure 7, it can be seen that 98%, vast majority of
the cases, is classified as hits, and only less than 2% of the
descriptors is misplaced by the algorithm. From safety aspects,
it is important to note that, throughout these tests, there were
zero events of undetected motion or false alarms. Both situa-
tions, even at low rate, would cause safety or reliability con-
cerns that would lead clinicians to reject the approach.
Localization Sensitivity
Our approach relies on user input for marking the anatomical
feature that requires tracking. This part cannot be automated
since it requires an understanding of the anatomy and the clini-
cal procedure performed. However, our approach does not
demand that this input be very precise but rather a mere
gesture. To test the sensitivity of the algorithm to the user
input, we repeated the same analysis after randomly altering
the location of the user input by 3 mm.
The results show that the average displacement error
increased by only 0.15 mm (from 4.08 to 4.23). The rest of the
results were the same as in the original test. This is a key
attribute in providing an approach that is clinically acceptable.
We do not have to impose stringent and stressful constraint on
our users by requiring very accurate labeling. The approach is
satisfied with a mere gesture of where the target KFPs are, not
requiring accurate user input.
Comparison to SIFT
The use of SIFT is very common in computer vision and can be
used for motion tracking and registration [20]. To compare the
performance of our proposed concept to SIFT, a typical set of
M
i
s
p
l
a
c
e
d
:

9
M
i
s
p
l
a
c
e
d

>

4

(
m
m
)
:

0
F
N

(
I
n

R
a
n
g
e
)
:

0
F
P
:

0
H
i
t
s
:

5
9
7
600
500
400
300
200
100
0
Total ADCs Count = 606
597
Hit
Err <
2 mm
9
Miss
Err > 2 mm
Zero
Err > 4 mm
Zero
False
Alarms
Zero
Undetected
Motion
0 0
Fig. 7. Results of the gold standard analysis. Histogram shows
the result of 606 tests comparing results from our approach
with expert user manually analyzed displacement.
The robust results are achieved due to the
weighted directional descriptors approach.
Reference Points
Reference Points
Algorithm
User Points +
(a)
(b)
Fig. 6. An example of a gold standard procedure. (a) Ana-
tomical descriptors on reference image set. (b) Input image
with original location descriptors (circles) and expert user
gold standard placement (cross marks). In red are the algo-
rithm results.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
brain MR images were used. A gold standard displacement
was manually marked by expert clinicians. This result was
compared with the estimated displacement of SIFT and of the
proposed approach. The mean error for SIFT was 17.1 mm
compared with 1.76 mmin our approach. SIFT, although a pop-
ular approach in academic research, is not robust for clinical
use. The error is ten times larger than the proposed approach,
which would lead to clinically unacceptable situations.
Discussion
The uniqueness of the approach is that it deals with true intrao-
perative combination of challenges making it practical for
immediate use in a clinical environment, thus translating a
huge academic research challenge into true accepted medical
practice. Ensuring safety via low rate of false alarms and mini-
mizing events of undetected motion to zero leads to a success-
ful implementation in the clinical arena.
Medical images are taken with a wide field of view to
provide the physician with the full view of the relevant anat-
omy. However, the scanned volume includes structures that
are of no interest to the procedure being applied. The notion of
excluding the nonrelevant anatomy requires medical under-
standing of the anatomy and the procedure performed. Ignor-
ing this would result in false alarms that would be not
accepted clinically. This is why our approach relies on user
input for the designation of the anatomical descriptors that are
to be tracked. Although user input is essential, it is important
to keep it from being demanding or time consuming. We mini-
mize the user interaction to mere gestures pointing to the fea-
tures of interest, and we show the approach is resilient to
errors and the localization does not have to be precise. The
detailed evaluation methods included also a sensitivity analy-
sis to image quality. As could be expected poor image quality
affects the results. It leads to more false alarms and an increase
in displacement errors. However, it is important to mention
that the key safety criterion of zero undetected motion events
is still maintained with our approach.
The analysis of anatomical features shows that a unique pre-
process is essential to identify the key pixels to be tracked. Fea-
tures that have a pattern that is repeated in several locations
within the search range must be ignored since they may be mis-
leading during the analysis phase. In many situations, a feature
has a directional nature (e.g., edges). Such features may
provide very reliable displacement in one direction (i.e., nor-
mal to the edge), but any other displacement must be ignored
[Figure 3(d)]. These realizations are essential to reliably esti-
mate the displacement as well as to avoid false alarms.
Medical images are challenging for many automated
approaches that are commonly used elsewhere due to the com-
plexity of the details and the demanding clinical requirements.
These factors became evident when comparing our WDDs
approach with the notable approach of SIFT, which is very
successful in other image processing applications. The SIFT
approach was misled by repeated features in the image [Figure
3(b)]. The SIFT approach ignores information given by dis-
placement of edges because of their directional nature. We
found that, without introducing these concepts, the results are
not robust enough to be accepted in clinical environment.
The approach has been implemented for quasi-static organs
where motion is not expected during the treatment (uterus,
breast, prostate, and brain). This was a fundamental phase in
validating the proof of concept of the image-based motion
detection technique. Further development is now being done
to adapt the approach to rapidly moving organs such as liver
where the treatment target area is constantly moving. The
main challenges are in high-temporal resolution and higher
performance algorithmic implementation. Since the organ is
not static, there are additional algorithmic challenges in esti-
mating the motion and comparing it with a predictive model
and updating the treatment plan accordingly.
The development of this novel technique for intraoperative
motion detection was challenged by more than just the techni-
cal algorithmic difficulties. There was a need to translate clini-
cal needs to technical requirements that would meet the
expectations of clinicians. The robustness and performance
are key factors in the acceptance of this tool and relying on it
in true clinical situations. InSightec is developing additional
products for treating prostate cancer, brain tumors, and breast
cancer, which will benefit from our experience that will be
implemented in these applications in the future. In coming
years, clinical studies will be conducted to develop statistical
data needed for reimbursement coverage and additional clini-
cal application will be investigated to reach wide use of this
technology for improving patient-care and health-care costs.
Conclusions
The use of image guidance in clinical procedures brings huge
benefits to patients and to health-care applications. When rely-
ing on preoperative imaging, tissue movement can create
safety concerns if not detected. Although studied thoroughly
in the academic arena, image-based motion detection is a chal-
lenge that has not yet met the goal of being widely accepted in
the true clinical environment.
The robustness of the proposed algorithm has been vali-
dated on true operative data. The encouraging robust results
have been shown to be resilient to inaccurate user input and
imperfect image quality. The results show accurate motion
analysis with very low rate of false alarms and zero events of
undetected motion under routine clinical conditions. The
approach has matched the clinical requirements and is inte-
grated in an image-based noninvasive therapeutic system
under MR imaging system. This technique is bound to have a
The use of scale invariant feature transformis
very common in computer vision and can be
used for motion tracking and registration.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
huge impact in procedures where image-based tissue motion
detection is needed.
The approach has been validated by clinicians and is
already in use in medical device products. In future, it will
be implemented in additional clinical procedures and will
have a significant impact by improving health care delivered
to patients.
Eyal Zadicario graduated in aeronautical
engineering (cum laude) from the Techn-
ion, Israel Institute of Technology and
completed his M.Sc. degree (summa cum
laude) in computer science from the Tel
Aviv University School of Mathematics.
He is currently the director of Neuro Pro-
grams in InSightec. Previously, he was a
part of the software development team responsible for
research development of SW applications and techniques for
medical device applications.
Shlomi Rudich received his B.Sc. degree
in biomedical engineering from Tel-Aviv
University in 2007 and a Software Devel-
oper Certification from the Technion in
2009. He is a clinical applications engineer
at InSightec.
Ghassan Hamarneh received his bache-
lors degree from Jordan University in 1995
and masters degree, with distinction, in
digital communications from Chalmers
University, Sweden, in 1997. He completed
his doctoral studies at Chalmers University
of Technology, Sweden, in 2001 and was a
predoctoral research fellow at the Univer-
sity of Toronto, Canada, from 2000 to 2001. He is an associate
professor in computing science at Simon Fraser University
(SFU), Canada. He is the codirector and cofounder (2003) of
the Medical Image Analysis Laboratory (MIAL) at SFU. His
research interest includes medical image analysis (segmenta-
tion, registration, and anatomical shape modeling and analy-
sis). Before joining SFU in 2003, he was a postdoctoral fellow
at the Hospital for Sick Children (Mouse Imaging Center) and
the University of Toronto, Canada from2001 to 2003.
Daniel Cohen-Or received his B.Sc. cum
laude in both mathematics and computer
science in 1985, an M.Sc. cum laude in
computer science in 1986 from Ben-Gurion
University, and his Ph.D. degree from the
Department of Computer Science, State
University of New York, Stony Brook, in
1991. He is a professor at the School of
Computer Science. He received the 2005 Eurographics Out-
standing Technical Contributions Award. His research inter-
ests include computer graphics, in particular, rendering and
modeling techniques; image synthesis, motion, and transfor-
mations; shapes and surfaces; and surface reconstruction.
Address for Correspondence: Eyal Zadicario, InSightec
LTD, 5 Nahum Heth, Tirat Carmel, Israel 39120. E-mail:
eyalz@insightec.com.
References
[1] S. Dimaio, T. Kapur, K. Cleary, S. Aylward, P. Kazanzides, K. Vosburgh,
R. Ellis, J. Duncan, K. Farahani, H. Lemke, T. Peters, W. Lorensen, D. Gobbi,
J. Haller, L. Clarke, S. Pizer, R. Taylor, R. Galloway, Jr., G. Fichtinger, N. Hata,
K. Lawson, C. Tempany, R. Kikinis, and F. Jolesz, Challenges in image-guided
therapy system design, NeuroImage, vol. 37, Suppl. 1, pp. S144S151, 2007.
[2] D. Verellen, M. Ridder, and G. Storm, A (short) history of image-guided
radiotherapy, Radiotherapy Oncol., vol. 86, no. 1, pp. 413, 2007.
[3] D. Jaffray, P. Kupelian, T. Djemil, and R. M. Macklis, Review of image-
guided radiation therapy, Expert Rev. Anticancer. Ther., vol. 7, no. 1,
pp. 89103, Jan. 2007.
[4] M. Moradi, P. Abolmaesoumi, and P. Mousavi, Deformable registration
using scale space keypoints, Proc. SPIE, vol. 6144, pp. 791798, 2006.
[5] E. R. Mcveigh, MRI of myocardial function: Motion tracking techniques,
Magn. Reson. Imag., vol. 14, no. 2, pp. 137150, 1996.
[6] Z. Fang, T. Moller, G. Hamarneh, and A. Celler, Visualization and explora-
tion of time-varying medical image data set, in Proc. Graphics Interface 2007
(ACM Int. Conf. Proc. Ser., 234), 2007, pp. 281288.
[7] M. J. Murphy, Tracking moving organs in real time, Semin. Radiat. Oncol.,
vol. 14, pp. 91100, 2004.
[8] G. C. Sharp, S. B. Jiang, and H. Shirato, Prediction of respiratory tumor
motion for real-time image-guided radiotherapy, Phys. Med. Biol., vol. 49,
pp. 40425, 2004.
[9] J. Tokuda, S. Morikawa, T. Dohi, and N. Hata, Motion tracking in MR-
guided liver therapy by using navigator echoes and projection profile matching,
Acad. Radiol., vol. 11, pp. 111120, 2004.
[10] B. J. Matuszewski, J. K. Shen, L. K. Shark, and C. J. Moore, Estimation of
internal body deformations using an elastic registration technique, in Proc. Int. Conf.
Medical Information VisualizationBioMedical Visualization, 2006, pp. 1520.
[11] J. Cho, Elastically deformable model-based motion-tracking of left ventricle,
in Proc. 26th Annual Int. Conf. IEEE EMBS 1, Sept. 2004, vol. 3. pp. 19251928.
[12] X. Papademetris, A. J. Sinusas, D. P. Dione, R. T. Constable, and J.
S. Duncan, Estimation of 3-d left ventricular deformation from medical images
using biomechanical models, IEEE Trans. Med. Imag., 21, vol. 7, pp. 786800,
July 2002.
[13] Y. Chenoune, E. Delechelle, E. Petit, T. Goissen, J. Garot, and
A. Rahmouni, Segmentation of cardiac cine-MR images and myocardial defor-
mation assessment using level set methods, Computer. Med. Imag. Graph. 29,
vol. 8, pp. 607616, Sept. 2005.
[14] T. Mcinerney and D. Terzopoulos, Deformable models in medical image
analysisA survey, Med. Image Anal. 1, vol. 2, pp. 91108, June 1996.
[15] F. Huang and J. Su, Face contour detection using geometric active con-
tours, in Proc. 4th World Congress on Intelligent Control and Automation, Nov.
2002, vol. 3. pp. 20902093.
[16] J. B. A. Maintz and M. A. Viergever, A survey of medical registration,
Med. Image Anal., vol. 2, pp. 136, Mar. 1998.
[17] H. Chui and A. Rangarajan, A new point matching algorithm for non-rigid
registration, Comput. Vis. Image Understand., pp. 114141, Feb.Mar. 2003.
[18] Z. Yefeng and D. David, Robust point matching for nonrigid shapes by pre-
serving local neighborhood structures, IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 28, pp. 643649, Apr. 2006.
[19] K. Rohr, H. S. Stiehl, R. Sprengel, W. Beil, T. M. Buzug, J. Weese, and M.
H. Kuhn, Point-based elastic registration of medical image data using approxi-
mating thin-plate splines, Visual. Biomed. Comput., pp. 297306, 1996.
[20] D. G. Lowe, Distinctive image features from scale-invariant keypoints,
Int. J. Comput. Vis., pp. 91110, Nov. 2004.
[21] R. M. Haralick and L. G. Shapiro, Computer and Robot Vision. Reading,
MA: Addison-Wesley, 1992, pp. 4048.
[22] M. I. Miller, A. Trouve, and L. Younes, On the metrics and euler-lagrange
equations of computational anatomy, Annu. Rev. Biomed. Eng., vol. 4, pp. 375
405, 2002.
[23] M. F. Beg, M. I. Miller, A. Trouve, and L. Younes, Computing large defor-
mation metric mappings via geodesic flows of diffeomorphisms, Int. J. Comput.
Vis., vol. 61, no. 2, pp. 139157, Feb. 2005.
[24] G. E. Christensen, P. Yin, M. W. Vannier, K. S. C. Chao, J. L. Dempsey,
and J. F. Williamson, Large-deformation image registration using fluid land-
marks, in Proc. 4th IEEE Southwest Symp. Image Analysis and Interpretation,
pp. 269273, 2000.
[25] J.-P. Thirion, Image matching as a diffusion process: An analogy with Max-
wells demons, Med. Image Anal., vol. 2, no. 3, pp. 243260, Sept. 1998.
[26] I. K. Sethi and R. Jain, Finding trajectories of feature points in a monocular
image sequence, IEEE. Trans. Pattern Anal. Mach. Intell., vol. 9, no. 1, Jan. 1987.
[27] M. Moradi and P. Abolmaesumi, Medical image registration based on
distinctive image features from scale- invariant (SIFT) keypoints, in Proc.
Computer-Assisted Radiology and Surgery (CARS) Conf., 2005, p. 1292.
[28] W. Cheung and G. Hamarneh, n-SIFT: n-Dimensional scale invariant feature
transform, IEEE. Trans. Image Process., vol. 18, no. 9, pp. 20122021, 2009.
______________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
EYEWIRE
Electronic Data-Capturing
Technology for Clinical
Trials
E
lectronic data-capturing (EDC) clinical database devel-
opment experience for a large global postmarketing study
was presented to illustrate how information technology
(IT), standards, and systems development methodology
were utilized to assist in the building of a clinical trial database,
overcome challenges, and monitor benefits gained via partner-
ing with both internal and external functional groups. Designing
an electronic case report form (eCRF) is an interdisciplinary
systems engineering task and requires not only technical skills
in utilizing information technology (IT) tools and systems
approach but also expertise and scientific reasoning in the sub-
ject-therapeutic areas. EDC technology may be utilized to
facilitate clinical trials and industry-wide study initiatives by
offering real-time data availability, automatic query manage-
ment functionality, and improved data quality.
The use of EDC and eCRF to collect data in clinical trials has
grown to gradually replace paper-based collection (PDC)
approach and affected the activities of clinical research opera-
tions for industry sponsors, clinical research organizations
(CROs), and clinical sites [1][3]. An eCRF design must ensure
superior data quality by following the clinical study protocols,
corporate therapeutic or standard consortium guidelines, and
good clinical practice rules with supportable technology. EDC
technology must comply with applicable regulatory require-
ments and offer flexible, configurable, scalable, and auditable
system features [4]. The design, development, and support of an
eCRF database or clinical data management system(CDMS) for
a clinical study must followthe systems development methodol-
ogy to ensure compliance. Additionally, postmarketing safety
data collection and clinical risk assessment based on observatio-
nal data are critical for evaluating and characterizing a products
risk profile and for making informed decisions on risk minimi-
zation [5]. Moreover, both payer groups and regulators require
more postmarketing data from sponsor companies. Therefore,
designing an eCRF to address postmarketing study safety and
pharmacovigilance seems to be specifically challenging.
A drug-eluting stent (DES) was approved by the U.S. Food
and Drug Administration (FDA) on 2 July 2008 and was
launched in Europe and other international markets in October
2006. Both the XIENCE USA and India studies were designed
to evaluate this everolimus-eluting coronary stent system
(EECSS) performance in the real world when used by a broad
group of physicians at a variety of health-care facilities. The
primary endpoint of these two studies is a measure of stent
thrombosis every year out of five years, as defined by the Dub-
lin/Academic Research Consortium(ARC). The ARCdefinition
of late-stent thrombosis was developed to eliminate variability
in the definitions across various DES trials. Adjunctive antipla-
telet therapy is a critical factor in optimizing long-term DES
safety. Despite established guidelines that recommend 612
months dual antiplatelet therapy, patients with DES implants
frequently stop taking their medication early. Consequently,
the postmarketing study followup will document patient
adherence and persistence with adjunctive antiplatelet drug
therapy at several time points throughout the study. Such
long-term postmarketing studies may help elucidate mecha-
nisms responsible for death, myocardial infarction, and late-
stent thrombosis risks not observed during controlled premar-
ket trials (see Table 1 about the trial characteristics).
The purpose of this article is to address three questions: What
were the EDC technologies employed in a typical industry-
sponsored clinical study? Howis the developed systemmeeting
the clinical research need? What would we want more fromthis
EDC technology? This article is prepared from industry per-
spectives to present and analyze the advantages, benefits, and
challenges in applying EDC technologies to address industrys
clinical trial operational needs based on a systematic overview.
Methods
Case Report Form Engineering
from Clinical Protocols
Designing eCRFs is an interdisciplinary systems engineering
task. The original materials for this critical design are the draft-
yet stable clinical protocol, the corporate therapeutic unit
standard forms, and clinical data acquisition standards harmo-
nization (CDASH) guidelines. Such systems engineering work
requires cross-functional team collaboration and input. We
used the Phase Forward InForm Architect tool to create the
initial set of forms based on existing standard library-form
XML files. Further modification was initiated and completed
based on the protocol and combined business requirements,
BY ZHENGWU LU
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Experience with a Global
Postmarketing Study
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
owing to the fact that a single consolidated database would be
developed for both U.S. and Indian geographies. The U.S.
and Indian clinical operation followed different standard
operation procedures (SOPs); and, this added a few more
challenges. It is mission critical that all functional teams
including science, safety, biostatistics, regulatory compliance,
and IT are represented in formreviewmeetings, and their feed-
back is incorporated into the revised and finalized forms [6].
Systems development methodology and controlled process
were followed for eCRF design and development to ensure that
regulatory requirements were met (see Figure 1).
Computer Software or Technology Tools
The system configured and adopted is based on Phase For-
wards InFormtechnologies to provide Web data entry by sites,
some basic reporting functionalities of the InForm reportable
database, query management features for both sponsor and site
users, and auditing or other regulatory needs. Captured data via
Web browser would pass through corporate firewall and come
to Clintrialan Oracle-based clinical data management sys-
tem. There were other multiple technology tools introduced:
CardioNow: A Web-based solution for clinical trial
image and information management. This digital imag-
ing and communications in medicine (DICOM)-compli-
ant imaging application was employed to assist
automation in handling, storing, transmitting information
in medical imaging, and allowing remote adjudication
committee to perform review and further analysis on
site-uploaded cardiovascular angiographies.
Interactive Voice or Web Response System (IVRS/IWRS):
To automate subject enrollment process, subject ID crea-
tion, and automatic casebook creation in InForm.
Table 1. XIENCE USA and India trial basic information retrieved from http://Clinicaltrials.gov.
Geography Study ID
ClinicalTrials.gov
Identifier FPI
Number
of Sites
Number
of Est.
Enrollment
Est. Primary
Completion
Date
Est. Study
Completion
Date Health Authority
USA 06-374 NCT00676520 July 2008 250 5,000 July 2009 July 2013 FDA
India 07-378 NCT00631228 June 2008 16 1,000 June 2009 June 2013 Central Drugs
Standard Control
Organization
FPI: first patient in.
Therapeutic
Standard Form
XML Files
Draft Yet Stable
Clinical Protocols
Release-Controlled
Change Management
Process
Any Ongoing Protocol
Revision or Operational
Impact on CRFs?
Yes
Clinical Data
Collected
Clinical Site Web
Entry Using
eCRFs
Create or Modify
Autoedit Checks and
Deploy eCRF in
Hosting Server
Approved
eCRFs
Yes
Cross-
Functional
Approval?
Finalize
Draft eCRF
Finalized and
Regulators
Approved
Protocols
Changes
Needed?
Postmarketing
Pharmacovigilance
Needs Such as Coding
Review with
Cross-Functional Teams
and Applicable CROs
Clinical Data Management Creates/
Modifies/Prepares Initial Set of
eCRFs Using InForm Architect Tool
No
No
Yes
CDASH
Guidelines
Fig. 1. The engineering workflow of designing, developing, and maintaining eCRFs in a typical industry-sponsored controlled
environment and regulated processes. eCRF: electronic case report form; CRO: contract research organization; XML: exten-
sible markup language; CDASH: clinical data acquisition standards harmonization.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Call Center Customer Relationship Management System:
The call center logistics and system were required for man-
aging distribution, collection, and data entry of the Seattle
Angina Questionnaire (SAQ) to assess patient- reported
outcomes and quality of life by CRO for U.S. sites only.
SAQ management was performed by sites in India.
Siebel Clinical Trial Management System (CTMS): This
system offers clinical document control and tracking,
site payment, and sponsors clinical research associate
(CRA) site monitoring support.
ePharma: The ePharma solutions were employed to
provide investigator, protocol, and safety trainings and
training document tracking for U.S. sites only.
Intranet Cognos-Based Reporting Application: A custom
application using Cognos technology to build multiple
complex reports based on data from Clintrial, system
application product (SAP; an enterprise resource plan-
ning system), and an enrollment projection spreadsheet
mainly for product-usage prediction.
Reconciliation (RECON): Another custom web applica-
tion to offer a user interface for clinical safety monitors
and product performance personnel to review, analyze,
and reconcile adverse events (AEs) reported through the
EDC user interface or health consumer hotline reports.
Secure File Transfer Protocol (sFTP): This is to establish
bridging for automatic casebook creation from IVRS/
IWRS to InForm and also provide a sharing mechanism
between CRO and sponsor for statistical analysis system
(SAS) data sets, listings, reports, or other documents.
Coding Application: World Health Organization drug
dictionary (WHOdd) and Medical Dictionary for Regula-
tory Activities (MedDRA) provide pragmatic, medically
valid terminology, with an emphasis on ease of use for
data entry, retrieval, analysis, and display, as well as a
suitable balance between sensitivity and specificity with-
in the regulatory environment [7].
Clinical Data Reporting Tool: A custom-built windows-
based application using Visual Studio.NET technology
and Oracle-structured query language (SQL) were
employed to create numerous intelligent clinical data
management reports to assist identifying issue data in
the trial. The scenarios covered by these reports are gen-
erally complex and cannot be automated through the
autoedit check specifications.
Citrix Technology: Software and services specialized in
virtualization and remote access software for delivering
applications over a network and Internet. This would
allow our CRO to access several reporting applications
such as the clinical data reporting tool to generate data
management reports for data cleaning.
Table 2 and Figure 2 summarize how potential clinical sys-
tems or tools are typically utilized to achieve clinical research
objective and quality data, the major players, and interconnec-
tivity or interactions between various systems or applications
for a typical industry-sponsored EDC study.
Results
The XIENCE global postmarketing study is currently in opera-
tion in United States and India, and data are being collected
through Web technology at nearly 300 sites in both countries.
Though there were two different study management units coordi-
nating and supporting the studies, there was just one single data-
base in the back end to maximize return on investment. Clinical
followups were created in InForm architect per protocol require-
ments, and approximately 30 forms were created in CDASH-
recommended domains such as demographics, AEs, medical
history, physical examination, laboratory test results, concomi-
tant medications, and ECG test results [8]. The deployed EDC
systempresents a Web browser-based user interface to offer easy
site data entry, query answering, investigators signoff on desig-
nated forms or casebook, sponsors current accessibility of data,
issuing or resolving queries, and reporting (see Figure 3).
The InFormEDCsystemis based on Visual Studio technology
and has addressed compliance, auditing, and security issues by
the vendor. Based on multiple years of InForm implementation
experience and gained expertise, we built a single database to
manage both XIENCE USA and India clinical data collection,
query resolution, AEadjudication process, and safety monitoring.
The interim reviews by safety, data management, or operational
committee were assisted by utilizing multiple InForm, Cognos-
based, or data management reports created through the clinical
data-reporting tool or Cognos-based technologies. One key bene-
fit the EDC brings to the sponsor is the autoedit checks that man-
age a large number of autoqueries based on subject knowledge
and functional teaminput. Aconsiderable effort was spent on col-
lecting, reviewing, refining, and finalizing hundreds of edit
checks with science and safety team for the proper range or
threshold setup to offer sites with immediate feedback to correct
the most common errors such as missing or out-of-range data.
Relatively complex edit checks based on cross-panel logic were
also programmed and deployed in place. To ensure timely safety
reviewof any potential serious AEs, several InFormreports based
on safety filtering needs were developed for automatic reporting.
One interesting issue we ran into for one AE InForm report
was that our safety monitors were getting a lot more records
for which they would not want to see some. It turned out that
InForm has a form-modified date field and it gets updated
whenever any data column is updated including any coding
fields on AE forms. To resolve this issue, we redeveloped this
AE update report via the clinical data-reporting tool using the
EDCtechnology may be utilized to facilitate
clinical trials and industry-wide study initiatives
by offering real-time data availability,
automatic query management functionality,
and improved data quality.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Clintrial as data source, where we are able to show AE records
based on certain specified key fields depending on the user
types. In the end, we were able to show records meeting the
specified criteria but modified by site users only. Dealing with
technology vendors could be challenging yet exciting: collab-
oration and patience are the key. Majority vendor organiza-
tions are process driven, and it takes days to deliver even
minor changes. In implementing IVRS for our Indian sites, we
Table 2. Computer software or technology tools utilized in a typical EDC-enabled clinical study and main functionalities.
Ref Software Tool Basic Characterization
[1] Clinical data report-
ing tool
A custom windows application using Visual Studio.NET technology. Oracle
SQL was employed to create numerous complex data management
reports to assist proactive data-cleaning strategy. These reports were
compiled into dynamic linking libraries (DLLs), deployed, and can be
accessed by designated personnel.
[2] CardioNow A commercial Web-based solution for clinical trial imaging and information
management.
[3] IVRS/IWRS Interactive voice or Web response system. Typically used for clinical trial
randomization, subject ID creation, and automatic casebook upload.
[4] InForm Architect An eCRF design tool to assist creating forms, rules, and testing with XML
based technology.
[5] OpenClinica A free, open-source, and Web-based clinical trial software platform for EDC
clinical data management in diverse research settings. It is developed
using the Java J2EE framework, with a database abstraction layer intero-
perable with PostgreSQL 8.x and Oracle 10g (custom ports to DB2, SQL
Server, or MySQL also possible) and based on leading independent
standards to achieve interoperability and regulatory compliance.
[6] Call center
customer relation-
ship management
A call center customer logistic system to support and manage distribution,
collection, and data entry of the clinical surveys to assess patient-reported
outcomes and quality of life.
[7] Siebel CTMS A commercial CTMS for clinical document control and tracking, CRA moni-
toring support, and site payment.
[8] ePharma A training tool to offer investigator, protocol, and safety training and training
document tracking.
[9] RECON A custom-Intranet application to assist sponsor personnel in AEs reconcilia-
tions between clinical data management system and clinical safety
system.
[10] Coding application Typically, this is to support MedDRA (safety) and WHOdd (medication) coding.
[11] Citrix A virtualization and remote access software for delivering applications over
a network and the Internet.
[12] User management
tool (UMT)
A commercial and Web-based account management tool.
[13] Clintrial or Oracle
clinical or other
vendor supplied
database
Clintrial or Oracle clinical are roles-based clinical database management
system based on the Oracle database technology. They offer a graphic
user interface to enable the users to create and maintain a clinical data-
base without needing to be an Oracle database expert. These software
products enable life sciences companies to unify their clinical data man-
agement needs, regardless of the source (electronic data collection or
paper based) into one system.
SQL: structured query language; eCRF: electronic case report form; EDC: electronic data-capturing; CRA: clinical
research associate; MedDRA: medical dictionary for regulatory activities; WHOdd: WHO drug dictionary; XML: extensible
markup language.
The market acceptance of EDCtechnology
has fueled newdemands for improvement,
configurability, and intelligent features.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
noticed that we were not able to get the local time in site enroll-
ment confirmation e-mails or faxes issued via IVRS/IWRS sys-
tem. It took days for the IVRS vendor to eventually have this
fixed; it turned out that the technical group had to add this special
value into one of their systemdrop down. It is also recognized that
multiple, interconnected clinical systems participate and support
an EDC clinical trial operation,
indicating the absolute necessity
of taking contextual systems
methodology in investigating
and resolving any potential
issues. Indeed, one must realize
that current clinical systems and
applications are interconnected
but not interoperable yet due to
regulation and standardization.
Table 3 summarizes some other
production experiences in deal-
ing with EDC technologies and
what we performed to address
the issues.
Discussion
The market acceptance of
EDC technology has fueled
new demands for improve-
ment, configurability, and in-
telligent features. The need to
improve clinical efficiencies and accelerate study times continues
to grow, driving industry sponsors to seek an eClinical environ-
ment that provides and promotes flexible, speedy eCRF design
and trial build; robust data management; real-time data visibility,
reporting and analysis; and global trial management and study
scalability [9]. We presented a Web-based EDCsystemto support
Web Server
InForm Service
InForm Server
Trial Trial
Clintrial
Integration
Solution
Clintrial
Sponsors InForm 4.5 Reportable DB
Sites
Trial
DB
Trial
DB
Fig. 3. InForm production environment. Site research coordinators enter data via Web browser
and data get submitted to and saved into each trial specific database. This database is hosted
by a third-party vendor and will proceed to Clintrial with a few hours delay, which is hosted in-
house through an intermediary process named clintrial integration solution. DB: database.
Regulatory
Submissions
Health Consumers
Report AEs Via
Phone, Mobile, or
Web Entry
Clinical Data
Reporting Tool
Cognos Web
Application
SAP
Global Corporate
Safety/Product
Performance System
Clinical Data
Management System
(Clintrial)
Generate SAS
Data Sets for
Stats Analysis
Regulatory Reports
(E-Mail, Fax, or
Snail Mail)
Regulator(s)
Regulatory
Submissions
CTMS
CTMS-EDC
Integration
InForm Reportable
Database
Coding
Application
and Standard
Dictionaries
Web Data Entry by Site
InForm
IVRS/IWRS
InForm Trial
Database
CardioNow
DICOM System
InForm Ad Hoc
Listings
Data
Management
Reports
Data
Management
Reports
Projection or
Other Reports
IVR
d
(C
sFTP
RECON
Fig. 2. The interconnected clinical systems and data flow from sites data entry to InForm reportable database and to sponsor
Clintrial or to SAS data sets. The diagram also shows how the different systems or applications interact with each other to
provide clinical data management, reporting, and service analysis and submission needs in the end. AE: adverse event; sFTP:
secure file transfer protocol; SAP: system application product (an enterprise resource planning system); RECON: reconcilia-
tion; SAS: statistical analysis system; CTMS: clinical trial management system; EDC: electronic data capturing.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
T
a
b
l
e
3
.
E
x
p
e
r
i
e
n
c
e
d
d
e
v
e
l
o
p
m
e
n
t
o
r
p
r
o
d
u
c
t
i
o
n
i
s
s
u
e
s
i
n
d
e
a
l
i
n
g
w
i
t
h
X
I
E
N
C
E
U
S
A
a
n
d
I
n
d
i
a
t
r
i
a
l
s
.
R
e
f
C
h
a
l
l
e
n
g
e
W
o
r
k
a
r
o
u
n
d
L
o
n
g
-
T
e
r
m
S
o
l
u
t
i
o
n
[
1
]
B
e
c
a
u
s
e
o
f
s
i
t
e
t
r
a
i
n
i
n
g
i
s
s
u
e
,
s
i
t
e
r
e
s
e
a
r
c
h
c
o
o
r
d
i
n
a
t
o
r
s
e
n
r
o
l
l
e
d
t
h
e
s
a
m
e
p
a
t
i
e
n
t
s
t
w
i
c
e
f
o
r
a
p
p
r
o
x
i
m
a
t
e
l
y
t
e
n
s
u
b
j
e
c
t
s
i
n
o
n
e
s
i
t
e
.
I
n
t
h
i
s
s
c
e
n
a
r
i
o
,
t
w
o
c
a
s
e
-
b
o
o
k
s
w
e
r
e
c
r
e
a
t
e
d
i
n
I
n
F
o
r
m
f
o
r
t
h
e
s
a
m
e
s
u
b
j
e
c
t
s
.
E
n
t
e
r
S
C
f
o
r
m
f
o
r
o
n
e
o
f
t
h
e
d
u
p
l
i
c
a
t
e
s
.
E
n
t
e
r
E
n
r
o
l
l
m
e
n
t
E
r
r
o
r
f
o
r
E
a
r
l
y
T
e
r
m
i
n
a
t
i
o
n
R
e
a
s
o
n
i
n
O
t
h
e
r
,
s
p
e
c
-
i
f
y
.
F
r
e
e
z
e
a
n
d
l
o
c
k
t
h
e
c
a
s
e
b
o
o
k
.
F
o
r
a
l
l
a
f
f
e
c
t
e
d
e
n
r
o
l
l
m
e
n
t
m
a
t
r
i
x
r
e
p
o
r
t
s
,
s
u
c
h
p
a
t
i
e
n
t
I
D
s
m
u
s
t
b
e
m
a
n
u
a
l
l
y
f
i
l
t
e
r
e
d
o
u
t
.
E
m
p
h
a
s
i
z
e
o
n
-
s
i
t
e
i
n
i
t
i
a
t
i
o
n
t
r
a
i
n
i
n
g
s
.
R
e
v
i
s
e
S
C
f
o
r
m
t
o
i
n
c
l
u
d
e
a
s
e
p
a
r
a
t
e
r
a
d
i
o
o
p
t
i
o
n
o
f
E
n
r
o
l
l
m
e
n
t
E
r
r
o
r
s
o
t
h
a
t
r
e
p
o
r
t
s
c
a
n
b
e
p
r
o
g
r
a
m
m
e
d
b
e
f
o
r
e
h
a
n
d
.
[
2
]
B
e
c
a
u
s
e
o
f
U
.
S
.
a
n
d
O
U
S
r
e
g
u
l
a
t
i
o
n
d
i
f
f
e
r
e
n
c
e
f
o
r
2
1
C
F
R
p
a
r
t
1
1
c
o
m
p
l
i
a
n
c
e
,
w
e
o
r
i
g
i
n
a
l
l
y
h
a
d
c
o
m
e
u
p
w
i
t
h
t
w
o
s
l
i
g
h
t
l
y
d
i
f
f
e
r
e
n
t
a
f
f
i
d
a
v
i
t
s
f
o
r
t
h
e
i
n
v
e
s
t
i
g
a
t
o
r
e
l
e
c
t
r
o
n
i
c
s
i
g
n
o
f
f
f
o
r
A
E
a
n
d
S
C
f
o
r
m
s
.
T
h
i
s
t
u
r
n
e
d
o
u
t
t
o
b
e
w
r
o
n
g
.
I
f
w
e
w
e
r
e
t
o
c
o
n
t
i
n
u
e
,
a
l
l
s
i
t
e
s
r
e
g
a
r
d
-
l
e
s
s
o
f
U
n
i
t
e
d
S
t
a
t
e
s
o
r
I
n
d
i
a
w
o
u
l
d
h
a
v
e
t
o
s
i
g
n
t
w
o
t
i
m
e
s
u
s
i
n
g
b
o
t
h
U
.
S
.
a
n
d
I
n
d
i
a
n
a
f
f
i
d
a
v
i
t
s
.
A
f
t
e
r
r
o
u
n
d
s
o
f
d
i
s
c
u
s
s
i
o
n
a
n
d
a
g
r
e
e
m
e
n
t
w
i
t
h
d
i
f
f
e
r
e
n
t
g
e
o
g
r
a
p
h
i
e
s
,
w
e
a
p
p
l
i
e
d
t
h
e
c
u
r
r
e
n
t
U
.
S
.
a
f
f
i
d
a
v
i
t
t
o
b
o
t
h
U
.
S
.
a
n
d
I
n
d
i
a
n
s
i
t
e
s
.
W
e
w
o
u
l
d
l
i
k
e
t
o
c
o
n
t
i
n
u
e
p
u
t
t
i
n
g
t
h
e
c
h
a
l
l
e
n
g
e
c
a
l
l
t
o
t
h
e
v
e
n
d
o
r
a
s
t
h
i
s
i
s
a
t
y
p
i
c
a
l
s
c
e
n
a
r
i
o
w
h
e
n
a
n
y
b
i
o
p
h
a
r
-
m
a
c
e
u
t
i
c
a
l
f
i
r
m
s
r
u
n
a
g
l
o
b
a
l
c
l
i
n
i
c
a
l
s
t
u
d
y
.
[
3
]
F
o
r
I
n
d
i
a
,
s
i
t
e
s
w
o
u
l
d
p
e
r
f
o
r
m
a
d
m
i
n
i
s
t
r
a
t
i
o
n
a
n
d
d
a
t
a
e
n
t
r
y
f
o
r
S
A
Q
.
H
o
w
e
v
e
r
,
s
i
t
e
s
s
h
o
u
l
d
h
a
v
e
r
e
a
d
-
o
n
l
y
a
c
c
e
s
s
t
o
S
A
Q
f
o
r
a
l
l
U
.
S
.
s
i
t
e
s
.
S
A
Q
d
a
t
a
e
n
t
r
y
i
s
p
e
r
-
f
o
r
m
e
d
b
y
a
C
R
O
f
o
r
a
l
l
U
.
S
.
s
i
t
e
s
.
C
r
e
a
t
e
a
s
e
p
a
r
a
t
e
S
A
Q
D
a
t
a
i
t
e
m
g
r
o
u
p
.
A
u
s
e
r
r
o
l
e
S
A
Q
E
n
t
r
y
w
a
s
c
r
e
a
t
e
d
f
o
r
t
h
e
C
R
O
d
a
t
a
e
n
t
r
y
.
W
e
k
e
p
t
t
h
e
d
e
f
a
u
l
t
S
i
t
e
R
C
a
n
d
I
n
v
e
s
t
i
g
a
t
o
r
r
o
l
e
s
f
o
r
I
n
d
i
a
s
i
t
e
s
b
u
t
c
r
e
a
t
e
d
t
w
o
m
o
r
e
u
s
e
r
r
o
l
e
s
S
i
t
e
R
C
S
A
Q
R
e
a
d
O
n
l
y
a
n
d
I
n
v
e
s
t
i
g
a
t
o
r
S
A
Q
R
e
a
d
O
n
l
y
f
o
r
U
.
S
.
s
i
t
e
s
.
W
e
b
e
l
i
e
v
e
t
h
a
t
t
h
e
w
a
l
k
a
r
o
u
n
d
m
a
y
b
e
t
h
e
l
o
n
g
-
t
e
r
m
s
o
l
u
t
i
o
n
t
o
a
d
d
r
e
s
s
a
n
y
p
o
t
e
n
t
i
a
l
o
p
e
r
a
t
i
o
n
a
l
d
i
f
f
e
r
e
n
c
e
w
h
e
n
r
u
n
n
i
n
g
g
l
o
b
a
l
s
t
u
d
i
e
s
.
[
4
]
P
r
o
d
u
c
t
-
l
a
u
n
c
h
t
i
m
i
n
g
d
i
f
f
e
r
e
n
c
e
i
n
U
n
i
t
e
d
S
t
a
t
e
s
a
n
d
I
n
d
i
a
c
a
u
s
e
d
u
s
t
o
r
e
l
e
a
s
e
t
h
e
i
n
i
t
i
a
l
E
D
C
w
i
t
h
o
u
t
t
h
i
n
k
i
n
g
o
f
t
h
e
f
u
l
l
p
e
r
s
p
e
c
t
i
v
e
a
n
d
i
m
p
a
c
t
o
n
d
a
t
a
q
u
a
l
i
t
y
.
S
o
o
n
,
w
e
h
a
d
t
o
d
o
a
s
e
r
i
e
s
o
f
r
e
l
e
a
s
e
s
j
u
s
t
t
o
f
i
x
u
p
m
u
l
t
i
p
l
e
c
h
a
n
g
e
s
.
C
o
r
p
o
r
a
t
i
o
n
m
a
n
a
g
e
m
e
n
t
m
u
s
t
r
e
a
l
i
z
e
t
h
a
t
t
h
e
t
e
c
h
n
o
l
-
o
g
y
w
i
l
l
n
o
t
s
o
l
v
e
a
n
y
p
r
o
b
l
e
m
i
f
t
h
e
i
s
s
u
e
i
s
n
o
t
d
e
f
i
n
e
d
y
e
t
.
A
l
s
o
,
i
m
p
l
e
m
e
n
t
a
t
i
o
n
o
f
E
D
C
t
e
c
h
n
o
l
o
-
g
i
e
s
t
a
k
e
s
t
i
m
e
.
F
o
l
l
o
w
s
y
s
t
e
m
s
d
e
v
e
l
o
p
m
e
n
t
m
e
t
h
o
d
o
l
-
o
g
y
w
i
t
h
a
d
e
f
i
n
e
d
a
g
r
e
e
a
b
l
e
o
b
j
e
c
-
t
i
v
e
a
n
d
r
e
g
u
l
a
t
o
r
s
a
p
p
r
o
v
e
d
c
l
i
n
i
c
a
l
p
r
o
t
o
c
o
l
.
A
v
o
i
d
m
u
l
t
i
p
l
e
E
D
C
r
e
l
e
a
s
e
s
.
[
5
]
T
h
e
c
l
i
n
t
r
i
a
l
i
n
t
e
g
r
a
t
i
o
n
s
o
l
u
t
i
o
n
(
C
I
S
)
i
s
t
h
e
m
i
d
d
l
e
c
o
m
p
o
n
e
n
t
o
f
b
r
i
n
g
i
n
g
c
l
i
n
i
c
a
l
d
a
t
a
f
r
o
m
I
n
F
o
r
m
t
r
i
a
l
s
p
e
c
i
f
i
c
d
a
t
a
b
a
s
e
h
o
s
t
e
d
i
n
t
h
i
r
d
v
e
n
d
o
r
p
a
r
t
y
t
o
i
n
-
h
o
u
s
e
C
l
i
n
t
r
i
a
l
.
W
e
o
n
c
e
i
n
a
w
h
i
l
e
g
o
t
t
h
e
s
y
n
c
b
r
o
k
e
n
w
a
r
n
i
n
g
s
b
e
c
a
u
s
e
s
i
t
e
u
s
e
r
s
e
n
t
e
r
s
p
e
c
i
a
l
c
h
a
r
-
a
c
t
e
r
i
n
t
o
t
h
e
c
o
m
m
e
n
t
f
i
e
l
d
s
i
n
I
n
F
o
r
m
.
W
e
c
o
m
m
u
n
i
c
a
t
e
d
t
h
e
m
e
s
s
a
g
e
t
o
t
h
e
s
i
t
e
s
t
h
r
o
u
g
h
o
u
r
C
R
A
s
(
I
n
d
i
a
)
o
r
C
R
O
C
R
A
s
(
U
S
A
)
s
o
t
h
a
t
s
i
t
e
u
s
e
r
s
w
i
l
l
n
o
t
e
n
t
e
r
t
h
e
s
e
s
p
e
c
i
a
l
c
h
a
r
a
c
t
e
r
s
a
g
a
i
n
i
n
t
h
e
f
u
t
u
r
e
i
n
I
n
F
o
r
m
c
o
m
m
e
n
t
.
W
h
y
i
s
n
t
t
h
e
r
e
a
p
r
o
g
r
a
m
m
a
t
i
c
w
a
y
p
r
e
v
e
n
t
i
n
g
t
h
e
i
s
s
u
e
f
r
o
m
h
a
p
p
e
n
i
n
g
i
n
t
h
e
b
e
g
i
n
n
i
n
g
e
v
e
n
i
f
s
u
c
h
s
p
e
c
i
a
l
l
e
t
t
e
r
s
a
r
e
e
n
t
e
r
e
d
t
h
r
o
u
g
h
I
n
F
o
r
m
c
o
m
m
e
n
t
?
I
p
u
t
t
h
i
s
a
s
a
n
o
t
h
e
r
c
h
a
l
-
l
e
n
g
e
c
a
l
l
t
o
t
h
e
v
e
n
d
o
r
.
[
6
]
W
e
s
t
a
r
t
e
d
u
s
i
n
g
t
h
e
U
M
T
f
o
r
a
c
c
o
u
n
t
m
a
n
a
g
e
m
e
n
t
o
f
t
h
e
s
e
t
w
o
t
r
i
a
l
s
.
S
o
m
e
h
o
w
,
w
h
e
n
a
n
e
w
s
i
t
e
w
a
s
c
r
e
-
a
t
e
d
i
n
I
n
F
o
r
m
,
t
h
e
I
n
F
o
r
m
v
e
r
s
i
o
n
f
o
r
t
h
e
n
e
w
s
i
t
e
s
t
a
r
t
e
d
w
i
t
h
a
n
o
l
d
o
n
e
.
O
n
e
U
.
S
.
s
i
t
e
h
a
p
p
e
n
e
d
t
o
u
s
e
t
h
e
o
l
d
r
e
l
e
a
s
e
m
e
d
i
c
a
t
i
o
n
(
M
E
D
S
)
f
o
r
m
f
o
r
n
i
n
e
s
u
b
j
e
c
t
s
.
B
e
c
a
u
s
e
o
f
t
h
e
f
o
r
m
n
a
t
u
r
e
c
h
a
n
g
e
s
(
f
r
o
m
f
l
a
t
t
o
r
e
p
e
a
t
i
n
g
)
,
t
h
e
M
E
D
S
d
a
t
a
f
o
r
t
h
e
s
e
s
u
b
j
e
c
t
s
i
s
n
o
t
i
n
C
l
i
n
t
r
i
a
l
b
u
t
o
n
l
y
a
c
c
e
s
s
i
b
l
e
v
i
a
I
n
F
o
r
m
r
e
p
o
r
t
i
n
g
o
r
d
i
r
e
c
t
b
r
o
w
s
i
n
g
.
F
o
r
t
h
e
s
e
n
i
n
e
s
u
b
j
e
c
t
s
,
a
n
y
n
e
w
m
e
d
i
c
a
t
i
o
n
r
e
c
o
r
d
s
h
a
v
e
t
o
b
e
e
n
t
e
r
e
d
v
i
a
t
h
e
A
d
d
E
n
t
r
y
s
u
b
f
o
r
m
i
n
t
h
e
o
l
d
M
E
D
S
f
o
r
m
.
W
e
f
i
x
e
d
t
h
i
s
n
a
g
g
i
n
g
i
s
s
u
e
b
y
a
d
d
i
n
g
a
d
y
n
a
m
i
c
r
u
l
e
t
o
t
h
e
o
l
d
M
E
D
S
f
o
r
m
-
o
p
e
n
i
n
g
q
u
e
s
-
t
i
o
n
,
i
n
d
i
c
a
t
i
n
g
w
h
e
t
h
e
r
t
h
e
r
e
i
s
a
m
e
d
i
c
a
t
i
o
n
s
o
t
h
a
t
t
h
e
n
e
w
M
E
D
S
f
o
r
m
w
i
l
l
a
p
p
e
a
r
.
W
e
t
r
a
i
n
e
d
s
i
t
e
s
t
o
r
e
e
n
t
e
r
t
h
e
a
f
f
e
c
t
e
d
m
e
d
i
c
a
t
i
o
n
r
e
c
o
r
d
s
i
n
t
o
n
e
w
M
E
D
S
f
o
r
m
.
B
e
c
a
u
s
e
o
f
a
g
r
e
a
t
d
i
f
f
e
r
e
n
c
e
i
n
t
h
e
v
e
r
-
s
i
o
n
c
h
a
n
g
e
s
,
s
i
t
e
s
n
e
e
d
t
o
b
e
i
n
v
o
l
v
e
d
f
o
r
t
h
e
r
e
e
n
-
t
r
y
.
I
n
t
h
e
e
n
d
,
w
e
w
e
r
e
a
b
l
e
t
o
g
e
t
a
l
l
M
E
D
S
d
a
t
a
i
n
C
l
i
n
t
r
i
a
l
.
H
o
n
e
s
t
l
y
,
t
h
e
r
e
i
s
a
l
a
c
k
o
f
a
u
t
o
m
a
t
i
o
n
f
r
o
m
U
M
T
t
o
t
h
e
I
n
F
o
r
m
d
a
t
a
b
a
s
e
b
y
t
h
e
v
e
n
d
o
r
.
T
h
e
r
e
w
e
r
e
s
i
m
i
l
a
r
o
c
c
u
r
-
r
e
n
c
e
s
l
i
k
e
t
h
i
s
p
r
o
b
a
b
l
y
d
u
e
t
o
t
h
e
f
a
c
t
t
h
a
t
m
a
n
u
a
l
e
r
r
o
r
s
a
r
e
i
n
t
r
o
d
u
c
e
d
i
f
s
o
m
e
b
o
d
y
h
a
s
t
o
m
a
n
u
a
l
l
y
c
h
a
n
g
e
r
i
g
h
t
s
o
r
a
s
s
o
c
i
a
t
i
o
n
s
i
n
I
n
F
o
r
m
b
e
c
a
u
s
e
o
f
s
o
m
e
n
e
w
u
p
d
a
t
e
s
i
n
U
M
T
.
I
n
d
e
e
d
,
w
e
u
r
g
e
t
h
e
v
e
n
d
o
r
t
o
o
f
f
e
r
t
h
e
m
u
c
h
-
n
e
e
d
e
d
a
u
t
o
m
a
t
o
n
t
o
s
t
o
p
p
r
o
b
l
e
m
s
l
i
k
e
t
h
i
s
f
r
o
m
h
a
p
p
e
n
i
n
g
.
S
C
:
s
t
u
d
y
c
o
m
p
l
e
t
i
o
n
;
O
U
S
:
o
u
t
s
i
d
e
o
f
U
S
;
C
F
R
:
c
o
d
e
o
f
f
e
d
e
r
a
l
r
e
g
u
l
a
t
i
o
n
s
;
R
C
:
r
e
s
e
a
r
c
h
c
o
o
r
d
i
n
a
t
o
r
;
S
A
Q
:
S
e
a
t
t
l
e
A
n
g
i
n
a
Q
u
e
s
t
i
o
n
n
a
i
r
e
;
C
R
O
:
c
l
i
n
i
c
a
l
r
e
s
e
a
r
c
h
o
r
g
a
n
i
z
a
-
t
i
o
n
;
E
D
C
:
e
l
e
c
t
r
o
n
i
c
d
a
t
a
c
a
p
t
u
r
i
n
g
;
C
R
A
:
c
l
i
n
i
c
a
l
r
e
s
e
a
r
c
h
a
s
s
o
c
i
a
t
e
;
U
M
T
:
u
s
e
r
m
a
n
a
g
e
m
e
n
t
t
o
o
l
.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
global postmarketing studies. We described what technologies
were involved in these studies, how they were utilized to serve
clinical trial effectiveness in a nutshell, the challenges overcome,
the benefits gained, and the lessons learned. On the other hand,
EDCvendors need to take on business input, partner with industry
sponsors, and offer service-oriented architecture to tackle evolv-
ing clinical research dynamics, address technology limitations,
and make technology improvement.
Our experience indicated that there were some misconcep-
tion about EDC and clinical data management: EDC is not just
an electronic data capture interface but a computerized system
designed for the collection of clinical data in electronic format
in supporting clinical trials [3]. Typically, EDC systems
provide: 1) a graphic user interface (GUI) component for data
entry, 2) a validation component to check user data, 3) a
reporting tool for analysis of the collected data. It seems that
sponsor or CRO data management activities revolve around
these three components in any typical EDC studies. Moreover,
other multiple systems played critical roles in ensuring all clini-
cal operations proceeded with effectiveness. In fact, not all sys-
tems and applications were interconnected and automated yet
in a sponsor clinical research environment. Some connections
may not be bidirectional, such as the interaction between
CDMS and CTMS. Clearly, there are many improvement areas.
Though there were multiple challenges, we started the develop-
ment of this large global trial with a science-focused proactive
eCRF design realizing the simple rule: garbage in, garbage out
based on our previous implementation experience using the
InForm technology. We worked with other functional teams
and business partners in following systems development
methodology with consistent and properly documented busi-
ness-driven process and in reaching a compromise between two
different clinical operational geographies to utilize a single
database. The consolidated single database approach provides
benefits of saving cost with one set of data safety monitoring
board (DSMB), clinical event committee (CEC), IVRS/IWRS
vendor, and easy pooling data for analysis. In addition, dedi-
cated resources were allocated to develop a solid thorough
eCRF completion guideline through collaboration with all func-
tional teams and operational units. To ensure data quality and
proactive data management practice in these trials due to the
fact that a small percentage (2030%) of patients data would
be source verified, data management initiated, developed,
tested, and deployed a series of approximately 30 data manage-
ment reports with intelligent and complex logics using a custom
data reporting tool and the Citrix technology to offer accessibil-
ity for CRO, India, and U.S. sponsor personnel to achieve
timely diagnosis of potential data issues across multiple forms.
We believe that data management should be the owner of
driving clinical data-cleaning process. However, cleaning data
need to involve other stakeholders: operations, safety,
biostatistics, science, quality assurance, and clinical sites [10]. It
is recognized that data management needs to be on top of tech-
nologies to tackle, drive, and provide better customer service in
using eCRF to capture protocol and compliance-necessitated
clinical data. Equally important, data management needs to
advise clinical stakeholders in using technology more intelli-
gently and more efficiently via partnering with EDC vendors
and clinical operations to customize and tailor technology to
conduct EDC studies. Though EDC technologies offer superior
advantages over traditional paper-based system, collecting,
monitoring, coding, reconciling, and analyzing postmarketing
safety data can be challenging as evidenced by XIENCE global
studies. To realize the full potential of technology advantage in
clinical research, both sponsor and site users will probably have
to change the way their offices and days are organized, howthey
enter and retrieve patient information, the process by which they
issue, answer, or close queries, the SOPs, and the ways in which
they relate to colleagues and CROs and interact with their
patients. To address the challenges of the eClinical environment,
biopharmaceutical firms need to move away from the legacy of
paper-based procedures. Technology should be tapped to add
process efficiencies and not to engender redundancy. For in-
stance, one interesting observation we noticed in interaction
with our CRO is how much information we should put on the
source-tracking sheets if site medical records or charts do not
have all the information required by the eCRF. Our perspective
is that source tracking sheets should be used only if medical
source does not have all the information needed and should con-
tain only those eCRF questions not covered by the source. When
used, these source-tracking sheets would be treated as source
documents. Additionally, safety monitors will have to find ways
to understand, analyze huge amounts of safety information
across different studies or systems, and coordinate with third-
party independent committee to enter adjudication results. It
seems reasonable to state that the effective use of technologies
depends as much on managing change as it does on information
management, and change has never been easy for sponsor e-
clinical systemimplementation and integration.
Our experience indicates that understanding the limitations
and opportunities offered by EDC vendor, configuring EDC
system to meet data-capturing needs based on sponsor IT or
data management profile, and collaborating with vendors to
offer flexible configurations are key to EDC implementation
success [11]. EDC technology vendors, however, think differ-
ently in that they would call upon sponsors to revise their busi-
ness processes to fit into the EDC features. Clearly, there is a
fine balance and flexibility from both ends to move forward
with issue resolutions. How to meet future needs, how to inte-
grate EDC clinical trial data with electronic health records
(eHR) system, how to enable interoperable standard-based
systems engineering, and how to persuade EDC vendors to
One key benefit the EDCbrings to the sponsor is
the autoedit checks that manage a large
number of autoqueries based on subject
knowledge and functional teaminput.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
invest in such innovative systems are something involving
collaborative efforts from many players. In fact, there are
some standardization initiatives in reaching interoperability
between differential clinical or e-health systems among
several standard consortiums such as CDISC, HL7, NCI, and
FDA. We anticipate that introduction of such revised imple-
mentation level standards based on XML will have an impact
on EDC technology and vendor product line [12].
Clearly, technology itself needs to be analyzed on a contin-
uously improving basis to meet the evolving clinical research
needs as evidenced by our experienced issues using EDC [13].
The modern EDC system should have more interoperability
and global standardization to enable interchangeable mecha-
nisms between EDC and eHR and will have the potential to
offer more service-oriented functionalities, increased scalabil-
ity, and more interoperable configurations to help address
clinical study objectives due to standards evolution [14], regu-
latory changes, medical science advances, health economical
implications, and technology innovations. Successful imple-
mentation of modern EDC technology requires reengineered
clinical trial operations and organizational changes to support
automated clinical and safety operations. Setting up the sys-
tem to function and ensuring its interoperability with multiple
other systems such as clinical data management system, cod-
ing application, CTMS, or clinical safety system will be a
daunting task yet achievable, but making sure the autoqueries
or safety alerts or signals it generates are scientifically rigor-
ous, reliable, and clinically valuable will be of paramount crit-
icality. Ultimately, modern EDC technology is a tool like all
other IT ventures, and one is still likely to be driven by humans
to engender long-term clinical efficiencies and cost benefits.
Acknowledgments
Thanks to Susanne Svensson and Garima Gaur, clinical data
managers from Santa Clara, California, and Delhi, India,
respectively, for their contribution in reviewing data manage-
ment reports specification and providing valuable feedback
and Wouter Van der borght, business application administra-
tor from Brussels, Belgium, for his assistance in compiling
and deploying data management reports.
The author has disclosed that opinions or views expressed
through this paper represent individual perspectives only.
Zhengwu Lu received his B.Sc. degree in
clinical medicine from Anhui Medical
University China in 1988 and his Ph.D.
degree in pharmacology from Beijing
University Health Sciences Center in
1994. He received a program certificate in
computer programming and database
development from the Praxis Training
Institute, Ottawa, Ontario, Canada, in 1998. He published
more than 20 scientific papers in peer-reviewed biomedical
journals and conference proceedings. His major contributions
have been in the research and development of animal models
to study psychneuroendocrine-immunity-disease susceptibility
and interventions involving immunopharmacology and tech-
nology applications in clinical research. In recent years, he has
performed clinical applications design and development in
various biomedically related industries. He has participated in
clinical study protocol preparation/review, developed clinical
applications, led clinical/laboratory systems development/
validation projects, and designed clinical trial case report
forms using electronic data-capturing technologies. He is cur-
rently a senior IEEE member, volunteers as a reviewer or edi-
tor for a number of peer-reviewed biomedical journals, and is
working as a clinical research consultant in the biopharma-
ceutical clinical trial industry.
Address for Correspondence: Zhengwu Lu, 1006 S De Anza
Blvd#K104, San Jose, 95129 CA USA. E-mail: zhengwu.
lu@ieee.org.
References
[1] I. Pavlovic, T. Kern, and D. Miklavcic, Comparison of paper-based and elec-
tronic data collection process in clinical trials: Costs simulation study, Contemp.
Clin. Trials, vol. 30, no. 4, pp. 300316, 2009.
[2] Z. W. Lu, Information technology in pharmacovigilance: Benefits, challenges,
and future directions from industry perspectives, Drug Healthcare Patient Saf.,
vol. 1, no. 1, pp. 3545, 2009.
[3] K. Ei Emam, E. Jonker, M. Sampson, K. Krleza-Jeric, and A. Neisa, The use
of electronic data capture tools in clinical trials: Web-survey of 259 Canadian tri-
als, J. Med. Internet Res., vol. 11, no. 1, p. e8, 2009.
[4] FDA, Guidance for industry: Computerized systems used in clinical investi-
gations, 2007.
[5] FDA Clinical Medical, Guidance for industry-good pharmacovigilance
practices and pharmacoepidemiologic assessment, Biotechnology Law Report,
vol. 24, no. 3, pp. 344356, 2005.
[6] W. E. Hammond, The power of working together, in Proc. Annual Meeting
of DIA Data Conf., Philadelphia, PA, Mar. 911, 2009.
[7] C. Toneatti, Y. Sa di, V. Meiffredy, P. Tangre, M. Harel, V. Eliette,
J. Dormont, and J. P. Aboulkera, Experience using MedDRA for global events
coding in HIV clinical trials, Contemp. Clin. Trials, vol. 27, no. 1, pp. 1322,
2006.
[8] CDISC CDASH Core and Domain Teams, Clinical data acquisition
standards harmonization (CDASH), Rep. No. CDASH_STD-1.0, Austin, TX,
Oct. 2008.
[9] Global Industry Analysts, Inc., Electronic data capture (EDC)-global strate-
gic business report, Res. Markets, Rep. No. 362140, Dublin, Ireland, Oct. 2006.
[10] E. Mcfadden, Management of Data in Clinical Trials, 2nd ed. New York:
Wiley, 2007.
[11] Z. W. Lu, Technical challenges in designing post-marketing eCRFs to
address clinical safety and pharmacovigilance needs, Contemp. Clin. Trials,
vol. 31, no. 1, pp. 108118, 2009.
[12] K. Buetow, Building a 21st century biomedical system: The cancer biomed-
ical informatics grid (caBIG), in Proc. Annual Meeting of DIA Data Conf., Phil-
adelphia, PA, Mar. 911, 2009.
[13] L. Hyveled, P. Karpur, and H. Nakskov, Knowledge innovation and clinical
operational excellence, Drug Inform. J., vol. 43, no. 2, pp. 159170, 2009.
[14] R. L. Richesson and J. Krischer, Data standards in clinical research: Gaps,
overlaps, challenges and future directions, J. Amer. Med. Inform. Assoc., vol. 14,
no. 6, pp. 687696, 2007.
EDCtechnology must comply with applicable
regulatory requirements and offer flexible,
configurable, scalable, and auditable
systemfeatures.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
________
______
EYEWIRE
Wireless Sensing Systems
in Clinical Environments
M
ultiple studies suggest that the level of patient care
may decline in the future because of a larger aging
population and medical staff shortages. Wireless
sensing systems that automate some of the patient
monitoring tasks can potentially improve the efficiency of patient
workflows, but their efficacy in clinical settings is an open ques-
tion. In this article, we introduce the challenges that such wireless
sensing systems must overcome and provide insights on the tech-
niques and features that systemdesigners should consider for suc-
cessful deployments in clinical settings. We do so through
MEDiSN, a wireless sensor network (WSN) designed to continu-
ously monitor the vital signs of ambulatory patients. We validate
the usefulness of MEDiSN with test bed experiments and results
from a pilot study performed at the Emergency Department,
Johns Hopkins Hospital. Promising results indicate that MEDiSN
can tolerate high degrees of human mobility, is well received by
patients and staff members, and performs well in real clinical
environments. We leverage our experience fromthis hospital pilot
study to outline outstanding issues and argue about the steps nec-
essary to bring wireless sensing applications to commercial use.
Wireless Sensor Networks
An increasingly aging population combined with nursing staff
shortages [1] and decreasing hospital capacities [2] suggest that
maintaining the current level of hospital care is becoming an
increasing challenge. Inefficient and labor-intensive procedures,
such as recording the patients vital signs periodically, included in
current hospital workflows are some of the underlying causes for
the mismatchbetween health-care capacity and demand. Therefore,
mechanisms that can automate some of these manual tasks have
the potential to improve the efficiency and quality of patient care.
WSNs, comprising small, low-power, and wireless devices
equipped with sensors and actuators, are ideal candidates for
automating the repetitive procedures currently performed by
hospital staff. Some of the clinical tasks that WSNs can perform
include continuous monitoring of unattended patients during
routine and extreme event surges, as well as monitoring of intra-
hospital transports and pediatric patients. Although the need for
these applications is widely accepted, there has been little prac-
tical experience with deploying themin clinical environments.
In an effort to address this disconnection between technol-
ogy and practice, we assembled an interdisciplinary team of
academic researchers, industrial developers, and medical
practitioners to develop the MEDiSN wireless monitoring sys-
tem for tracking the vital signs of ambulatory patients. Rather
than an end product, MEDiSN is a research vehicle for under-
standing the challenges associated with deploying wireless
sensing applications in the hospital. At the same time, MED-
iSN is a fully functional system whose goal is to showcase the
benefits of wireless sensing to physicians.
We deployed MEDiSN as part of an IRB-approved study at the
Emergency Department, Johns Hopkins Hospital, to monitor
patients waiting to be seen at the emergency room (ER). Results
from this deployment combined with findings from a controlled
environment indicate that wireless sensing applications can suc-
cessfully meet many of the challenges related to automated patient
monitoring, including reliable data delivery in the face of adverse
radio-frequency (RF) environments and mobility support. At the
same time, the lessons we learned fromthis exercise, including the
feedback we collected frompatients and medical staff, have helped
us to identify outstanding issues that need to be resolved before
such systems can be commercialized and used on a daily basis.
WSNs in Clinical Environments
WSNs consist of low-power, embedded computing devices,
known as motes, that use sensors to collect measurements
from the physical world and its inhabitants. Wireless sensing
has been successfully employed in a variety of applications
including environmental monitoring [3], structural monitoring
[4], and surveillance [5]. These early successes have moti-
vated researchers to apply the same technology in clinical
environments in an effort to increase the quality of care on a
day-to-day basis as well as during extreme events. The subse-
quent sections outline a fewpotential applications and identify
the novel challenges these medical applications introduce.
Medical Sensing Applications
Monitoring Unattended Patients
Because of space and manpower limitations, many patients
visiting a hospitals ER spend multiple hours in waiting areas
before they are admitted to care units. Although wait times are
BY JEONGGIL KO, TIA GAO,
RICHARD ROTHMAN,
AND ANDREAS TERZIS
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Improving the Efficiency of the
Patient Monitoring Process
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
shorter for patients with urgent needs and waiting patients are
periodically retriaged, there have been reports of patients
whose condition deteriorated while waiting in the ER [6], [7].
Patient monitoring, whereby motes continuously record and
transmit the patients vital signs, can prevent tragic incidents
by alerting health-care providers of patient deterioration.
Surge Capacity Monitoring
Surge events during which a rapid influx of patients over-
whelms the hospitals capacity exacerbate the problem of
monitoring unattended patients. Systems that automate the
patient triage and monitoring processes can save peoples
lives during these critical times.
Monitoring of Intrahospital Patient Transports
Existing devices for monitoring patients vital signs are bulky
and cumbersome to use during patient transports within the
hospital. Therefore, a wireless vital signs monitoring system
with the added ability to track the location of patients inside
the hospital will improve patient safety during such transports.
Pediatric Patient Monitoring
Pediatric patients using patient-controlled analgesia (PCA)
devices are prone to overdose [8]. However, patients who
overdose can be automatically identified by continuously
monitoring their blood oxygen levels and the flowof analgesia
can be immediately discontinued while physicians are alerted
of the emergency.
Challenges of Medical Sensing Applications
Supporting Large-Scale
and Easy to Deploy Networks
In many scenarios, including natural and man-made disasters,
wireless sensing systems should be able to collect data from a
large number of patients dispersed around the hospital. There-
fore, the system should be able to scale to wide areas and be
easy to expand even to areas without preexisting wireless net-
work infrastructure.
High Data-Rate Networks
In addition to supporting large number of users, medical sens-
ing applications should be able to support high volumes of
traffic generated from high data-rate sensors such as electro-
cardiogram (ECG) and respiratory sensors.
Patient Mobility
Patient mobility is very common in many clinical scenarios
including patients in the ERand intrahospital patient transports.
For this reason, medical sensing applications should provide
high reliability without restricting the patients movements.
Patient Localization
Tracking patient locations is not only important during the
transport of patients admitted to the hospital but also simpli-
fies the process of locating patients in the ER waiting room.
Doing so would reduce staff workload and improve the effi-
ciency of patient-care workflow.
Soft Real-Time Data Delivery
with Any-to-Any Routing
Medical applications require low latency and high reliability
to ensure that medical staff can make clinical decisions based
on up-to-date patient information. Furthermore, the measure-
ments must be reliably delivered to all the caregivers who are
associated with a certain patient.
Security
According to U.S. law, medical devices must meet the privacy
requirements of the 1996 Health Insurance Portability and
Accountability Act (HIPAA). To meet these requirements, the
system must never broadcast identifiable patient data and
guarantee the authenticity of the data it delivers.
MEDiSN
We developed MEDiSN [9] as an experimental vehicle for
investigating the feasibility of wireless vital signs monitoring
in clinical environments. In collaboration with our partners at
Johns Hopkins Hospital, we selected the task of monitoring
unattended ER patients as the first application for MEDiSN.
From a technical standpoint, MEDiSN must securely and reli-
ably deliver the vital signs of ambulatory patients. Further-
more, MEDiSN must be able to easily extend to a larger area
(e.g., tents outside the hospital) during a surge event.
Figure 1 shows the three-core network components of
MEDiSN. Figure 1(a) is a picture of a patient monitor (PM),
known as a miTag. The miTags processing core is the Sentilla
Tmote Mini, which combines a Texas Instruments MSP430
microcontroller with a TI/Chipcon CC2420 IEEE 802.15.4
[10] radio in a mini-secure digital input/output (SDIO) form
factor [11]. The miTag includes an off-the-shelf sensor that
records pulse rate and blood oxygen levels [12] and integrates
actuators including a buzzer, a 1.8 in liquid-crystal display
(LCD) and five light-emitting diodes (LEDs). It collects a
patients vital signs through a finger clip shown in Figure 1.
Before transmitting these measurements using its on-board
radio, the miTag encrypts and signs them using the security
primitives that the CC2420 radio offers [13].
The PMs transmit the samples they collect to one of the net-
works relay points (RPs), shown in the center of Figure 1.
Unlike previous medical sensing designs (e.g., [14]), MED-
iSN includes a dedicated wireless backbone that forwards
the PMs data. Specifically, the RPs use the collection-tree
MEDiSNis a fully functional systemwhose
goal is to showcase the benefits of wireless
sensing to physicians.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
protocol (CTP) [15] to self-organize into a routing tree that
relays packets to the gateway [Figure 1(c)], where the data can
be delivered to any destination via the Internet.
While the performance of CTP is not within the scope of
this work, interested readers can find extensive evaluations of
this protocol in [9], [15], and [16]. Having this separate infra-
structure means that PMs do not need to relay other PMs data,
as in other ad hoc networks. As a result, PMs can aggressively
duty cycle (i.e., turn off) their radios to conserve energy. (The
current miTag can be used for approximately five days while
transmitting vital sign updates every 30 s without recharging
the battery. Our clinician partners require this device lifetime.)
Moreover, the dedicated backbone does not suffer from the
frequent path reconfigurations inherent in mobile ad hoc net-
works and thereby offers better service to mobile PMs.
Finally, having a multihop backbone simplifies the task of
expanding the systems coverage area. Doing so requires to
simply add more RPs to the existing backbone. These RPs
automatically join and seamlessly extend the existing MED-
iSN backbone.
Rather than broadcasting packets to all RPs in its communi-
cation range, each PM associates with one of the networks
RPs using a simple RP selection scheme (the interested reader
can find more details about the association mechanisms that
the PMs use in [9]). Doing so improves the networks effi-
ciency by suppressing the delivery of duplicate packets over
the RP backbone. The proposed RP selection scheme is robust
to PM movements, promptly allowing a PM to associate with
the next best RP after it disconnects from its previously
selected RP. To achieve high reliability, RPs and PMs in
MEDiSN perform hop-by-hop retransmissions. We evaluate
the performance of this two-tier network and the robust RP
selection scheme using an indoor test-bed experiment in the
Supporting Mobile PMs section.
The RPs not only forward the PMs data but also collect and
periodically report metadata regarding the quality of the back-
bones wireless links. This
information can be used to opti-
mize the networks perform-
ance by adjusting the number
and the locations of its RPs.
Deployment Results
Supporting Mobile PMs
Before deploying MEDiSN in
an actual clinical environ-
ment, we performed a test bed
experiment to validate its per-
formance with mobile PMs.
Since patient movement is
common in the clinical settings that MEDiSN targets, per-
forming well in this experiment was essential for the system to
be deployed in hospitals. This experiment was conducted on
two floors of a seven-story building. We deployed RPs on the
first and fourth floor hallways, each with an approximate
length of 170 ft. The RPs on the two floors were connected
using RPs on the buildings stairwell. In total, we used one
gateway and 25 RPs to cover the entire area.
We tested two types of configurations in this test bed. First,
we tested a two-tier MEDiSN network using RPs acting as the
dedicated wireless backbone network with PMs implementing
the RP selection mechanism described earlier. Second, a flat
network in which the PMs participate in the same ad hoc net-
work with the RPs was tested. We used CTP to route packets
over this single-tier (i.e., flat) network for both RPs and PMs.
The mobile PMs moved at a walking speed for approximately
8 min over a fixed route, spanning approximately 1,000 ft that
covered both floors. In each case, we formed a control group
by placing stationary motes at identical locations for 8 min.
All PMs sent one packet per second.
Figure 2 presents the distribution of the times that the PMs
were disconnected from the network. We determine that a PM
is disconnected from the network if packets from the PM are
not received at the gateway when packets are expected (the
gateway expects one packet each second from all PMs). In the
case of the MEDiSN network, there were ten PM disconnec-
tions due to mobility. Furthermore, all disconnections lasted
for less than 10 s. On the other hand, the mobile PM in the flat
network experienced 20 disconnections that lasted as long as
60 s. The overall packet reception ratio (PRR) was 96.06% for
the mobile MEDiSN PM and 62.45% for the mobile PM in the
flat network. We note that the PRR has been computed with
respect to the application-level performance (number of valid
packets received at the gateway divided by the total number of
packets that the PMsent) in all cases. Finally, while the mobile
MEDiSN PM made on average 1.9 transmission attempts per
MEDiSNis a research vehicle for
understanding the challenges associated with
deploying wireless sensing applications
in the hospital.
(a) (b) (c)
Fig. 1. (a) A medical information tag, or miTag for short, that collects patients vital signs, (b)
relay point that forwards vital sign measurements over the MEDiSN wireless mesh, and (c) the
gateway that collects all the measurements.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
successful packet delivery, the CTP-based mobile PM
required an average of 2.41 transmissions for each successful
packet delivery.
These results support our intuition that having a wireless
backbone improves data delivery as well as efficiency when
PMs are mobile. We also note that the reception ratios for the
stationary PMs were 100 and 99.80%for the flat and MEDiSN
networks, respectively. The discrepancy between the delivery
ratios for the static and mobile PMs in the case of the flat net-
work can be explained as follows. When the network is static
and CTP can collect all the link quality measurements for dif-
ferent neighbors, a node can find alternative high-quality end-
to-end paths when link failures occur [16]. On the other hand,
when movements occur in a mobile network, topology recon-
figurations happen faster than the time that CTP needs to col-
lect neighborhood information and update routes, leading to
severed end-to-end paths and lower data-delivery ratios.
Low-Power Wireless Performance at the ER
Since MEDiSN includes a dedicated wireless RP backbone, a
backbone setup phase was necessary at the beginning of our
hospital deployment. Although the deployment site (see Figure
3) is not large in terms of its total area, setting up a wireless
backbone in the hospital can pose several challenges. First, the
hospital is a clutter-rich environment with obstacles such as
large glass and aluminum dividers, thick steel doors, and lead
painted walls (see Figure 4), which severely distort RF signals.
Second, RP locations are constrained by the availability of wall
plugs, as RPs are powered by the hospitals electricity grid. The
alternative of using batteries to power the RPs would require
the hospital personnel to replace batteries frequently, as the RPs
cannot duty cycle their radios. Third, given the fact that the ER
is busy 24 h a day, movement by humans and equipment across
the deployment area leads to a highly variable RF environment.
To further evaluate the impact of these factors on the wire-
less channel environment, we compared the RF environment at
the ER to the environment of an indoor test bed located at an
office building. We performed tests for each of the two envi-
ronments on IEEE 802.15.4 channels 22 and 26 over the same
24-h period. (IEEE 802.15.4 channel 22 overlaps with WiFi
channel 11, which is active in both environments. On the other
hand, channel 26 is free from WiFi interference.) We selected
a transmitterreceiver (T-R) pair in the ER approximately 64 ft
apart from each other (see Figure 3). The transmitter broad-
casted one 111-B packet every 500 ms without performing any
CCA checks. [Carrier sense multiple access protocols perform
clear channel assessment (CCA) checks to make sure the wire-
less medium is idle before each transmission.] The receiver
logged the received signal-strength indicator (RSSI) values of
incoming packets. Furthermore, the receiver took signal-
strength measurements between packet receptions to measure
the ambient noise level of the environment. We performed the
same test in an indoor test bed using a link with the same dis-
tance and a line-of-sight path.
Figure 5 presents the 24-h plot of packet RSSI values and
the ambient noise levels collected at the ER [Figure 5(a) and
(b)] and the indoor testbed [Figure 5(c) and (d)] for the two
channels we tested. While differences in channel characteris-
tics from different environments are expected to a certain
degree, Figure 5 suggests that the ER environment exhibits
significantly higher RSSI variation compared with the indoor
test bed channel across both frequency channels. We conjec-
ture that changes in the physical environment (i.e., movement
of people and equipment) are the root cause of these variations
in signal strength.
Deploying a wireless network in an environment with such
variations in link conditions requires careful consideration. As
with any other wireless network deployment, a hospital
deployment starts with a site survey of the deployment area.
Through this process, we select the locations that minimize the
total number of deployed RPs, while ensuring good connectivity
throughout the ER waiting area. To do so, we set up a gateway
and gradually add more RPs to the site as we find coverage dead
spots. We use a combination of visual cues and network
measurements to identify these dead spots. The process of instal-
ling the RP backbone takes less than 30 min for an area of the
size shown in Figure 3 (approximately 70390 ft). Moreover,
Wireless sensing systems can overcome the
challenges introduced by clinical environments
and improve quality of care.
F
r
e
q
u
e
n
c
y
14
12
10
8
6
4
2
0
<10 <20 <30 <40 <50 <60
Disconnection Time (s)
Mobile Flat Network PM
Mobile MEDiSN PM
Stationary MEDiSN PM
Fig. 2. Comparing the performance of the two-tier MEDiSN
network to a flat ad hoc network that uses CTP. In both
cases, mobile PMs follow the same path over a multifloor
test bed, moving for approximately 8 min. While PMs that
use MEDiSN experience a small number of short disconnec-
tions, PMs in the flat network experience multiple and pro-
longed disconnected periods. The static PM in the flat
network experienced no disconnections.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
one can use the same method
to extend the networks cover-
age or to accommodate varia-
tions in wireless link quality by
adding more RPs to the back-
bone. The fact that we can
easily improve network per-
formance and extend the wire-
less networks coverage is a
big advantage of having a
wireless backbone.
After deploying the RP
backbone, we admitted a total
of 46 patients to a pilot study
(see the Patient and Staff
Acceptance at the ER sec-
tion). During the deployment,
a maximum of three patients
wore miTags at any one time.
An average of 5,341 packets
were transmitted from each
patients miTag. Of these, an
average of 244 packets were
lost, leading to an average
application-level PRR of
95.43% with a standard devia-
tion of 5.41. We conjecture
that the high variation in PRR
is due to the significantly dif-
ferent behaviors among the
patients that wore the miTags.
While most patients used the miTag properly, some held it
tight in their hands or covered it with thick jackets. Changing
the antenna design to overcome these adverse conditions is part
of our future work.
Patient and Staff Acceptance at the ER
Our IRB-approved pilot study took place at the ERof Johns Hop-
kins Hospital, Baltimore, Maryland, spanning a period of ten
days during November/December 2008. The objective of this
study was to measure the level of acceptance of a wireless sens-
ing system such as MEDiSN by ER staff and patients. At the
same time, this study provided us with the opportunity to evalu-
ate the applications performance in a realistic environment.
All adult patients with triage-level assignments of 35 (on a
15 scale, where 1 is the highest acuity), who were sent to the
ERs waiting roomafter their initial triage were eligible to par-
ticipate in the study. A total of 46 patients volunteered to par-
ticipate. All of them had triage level 3, the mean age was 48.2,
and 33% of the volunteers were male. Moreover, 33.3% of
them were African American, 0.6% were Asian, and the rest
were Caucasians or Latin Americans. Chest pain was the most
common chief complaint (52.2%) among the volunteers. Fig-
ure 6 illustrates how the patients wore the device around their
neck using a lanyard. Patients wore the device throughout
their stay in the waiting area. We collected pulse rate and
blood oxygen level measurements every second using single-
use, disposable pulse oximeter clips. At the end of their stay,
volunteers were asked to rate their satisfaction level on a scale
of 14, 4 being most satisfied with MEDiSN. The average
patient satisfaction level was 3.47, and 91% of them indicated
that they would be willing to use the device in the future.
We also surveyed a total of 21 staff members, 38% of whom
were resident nurses and the average numbers of years that
they worked at the hospital was 8.5 years. Computers on the tri-
age desks and nursing stations were capable of displaying
MEDiSNs graphical user interface (GUI), and staff members
were able to use the GUI to retrieve patients information. We
gathered staff-satisfaction surveys when staff members entered
the rest area during the day. Overall, the average staff satisfac-
tion score for our systemwas 3.11 on a 14 scale.
Fig. 4. View of the waiting area at Johns Hopkins Hospital
adult ER as seen from the triage desks. Aluminum dividers and
glass walls act as obstacles, disrupting RF signal propagation.
Johns Hopkins Hospital
Emergency Room
T
R
Gateway
Relay Point
Patient Monitor (miTag)
Fig. 3. Floor plan of Johns Hopkins Hospital ER. RPs are placed in the waiting area, while the
gateway is located at the nurse-triage desks. MEDiSN was deployed to cover an area of
70 390 ft.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Discussion
Next, we present some of the comments provided by the staff
members that we surveyed during the pilot deployment.
Work-Flow Integration
Many staff members commented that the current ERworkload
is already overwhelming. Thereby, any device introduced at
the ER should not impose additional burden to overworked
staff members. One way to achieve this goal is by properly
integrating new devices to existing workflows.
Staff Education
Another common concern from the staff was that the GUI at the
monitoring station was not intuitive. Some of these comments
were because staff members were not trained to use the inter-
face. Nonetheless, developing easy-to-use interfaces for envi-
ronments with high-cognitive overload is an
ongoing challenge.
Localization
Many staff members indicated that a method
for tracking the patients location would be
broadly useful. For example, nurses spend a
nontrivial amount of time trying to identify
and locate patients in the hospitals waiting
room. While indoor localization is generally
considered a challenging task, ranging-based
approaches using ultrawide band (UWB)
transceivers offer promising results. Consid-
ering that a recent amendment to the IEEE
802.15.4 standard describes a radio that
combines communication and UWBranging
and that early implementations of the new
standard are starting to emerge [17], practi-
cal indoor localization may be within reach.
Avenues for Commercialization
We envision that hospitals will be the primary
users of wireless vital sign monitoring systems,
such as MEDiSN, because of their potential for decreasing staff
workload, increasing patient safety, and decreasing the duration of
hospital stays. Furthermore, we expect that these systems will be
used to monitor disaster victims at the scene and at field hospitals
in war zones.
Nevertheless, a number of issues need to be resolved before
WSNs for medical sensing can be productized. First, there is
currently a lack of sensors (e.g., blood pressure and ECG) that
can be used with low-power wireless devices. Existing sensors
are too bulky and power-hungry to be used with portable devi-
ces such as the miTag. More importantly, longer clinical stud-
ies are necessary to quantify the benefits that WSNs bring to
applications such as those described in the WSNs in Clinical
Environments. These studies should also produce amend-
ments to existing clinical workflows that incorporate wireless
medical sensing systems with the minimumlevel of disruption.
Summary
In this article, we examine the potential of
WSN technologies to improve the effi-
ciency of the patient-monitoring process in
clinical environments. We describe several
applications that can benefit from low-
power WSNs and introduce the technical
challenges that such systems should over-
come before they are widely deployed.
With these challenges in mind, we design
MEDiSN, a WSN-based research vehicle
that monitors the vital signs of unattended
patients and present performance results
from both test bed experiments and a pilot
study performed at Johns Hopkins Hospital
Emergency Department. Despite the chal-
lenging wireless channel conditions preva-
lent in clinical environments, we show that
MEDiSN can successfully accomplish its
task of continuously monitoring unattended
patients vital signs. These encouraging
results along with the positive feedback
Fig. 6. A triage nurse in the ER wear-
ing a miTag. A lanyard connects
the miTag to the patients neck and
a single-use, disposable pulse oxim-
eter clip is attached to the finger.
Channel 22 - Emergency Room
Time (min)
d
B
m
110
100
90
80
70
60
50
0 200 400 600 800 1,000
Channel 22 - Indoor Test Bed
Time (min)
d
B
m
110
100
90
80
70
60
50
0 200 400 600 800 1,000
Channel 26 - Emergency Room
Time (min)
(a) (b)
d
B
m
110
100
90
80
70
60
50
0 200 400 600 800 1,000
Channel 26 - Indoor Test Bed
Time (min)
(c) (d)
d
B
m
110
100
90
80
70
60
50
0 200 400 600 800 1,000
Noise Level RSSI
Noise Level RSSI
Noise Level RSSI
Noise Level RSSI Noise Level RSSI Noise Level RSSI
Noise Level RSSI Noise Level RSSI
Fig. 5. Packet RSSI and ambient noise levels collected at (a), (b) the ER and (c), (d) the indoor test bed.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
from the user-satisfaction surveys suggest that WSN-based
systems can contribute to the greater goal of improving the
efficiency of clinical workflows.
JeongGil Ko received the B.Eng. degree in
computer science and engineering from
Korea University, Seoul, South Korea, in
2007 and a M.S.E. degree in computer sci-
ence at Johns Hopkins University, Baltimore,
Maryland, in 2009. He is a Ph.D. student in
the Department of Computer Science at Johns
Hopkins University. He is also a member of
the Hopkins Internetworking Research Group (HiNRG) lead by
Dr. Andreas Terzis. His research interests include wireless sensing
systems for health care, embedded network system design, and
the deployment of such systems to real environments.
Tia Gao received a M.S. degree in electri-
cal engineering and expects a M.B.A.
degree, both from Stanford University. She
is the founder and president of Aid Net-
works. Previously, she was an engineering
manager at Johns Hopkins University Ap-
plied Physics Laboratory and managed
research projects in the field of patient
tracking solutions for mass casualty response. Before that, she
was at Medtronics Cardiac Rhythm Management division
and developed implantable cardiac-monitoring devices. She
served as a program manager at Microsoft Corporation and
developed products for Windows Mobile and Office. She has
authored more than 20 peer-reviewed technical publications
on the topic of wireless mesh sensor networks.
Richard Rothman received his M.D. and
Ph.D. degrees from Cornell University and
the University of California, respectively.
He is an associate professor in the Depart-
ment of Emergency Medicine, Department
of Medicine and Division of Infectious Dis-
eases at Johns Hopkins University School of
Medicine. He is the research fellowship di-
rector and chair of the research committee in emergency medi-
cine. He has served as a consultant for the biotechnology
industry in molecular assay development for infectious diseases.
Currently, he is a principal investigator for the diagnostics pro-
gram for the Mid-Atlantic Regional Center for Excellence for
Biodefense and Emerging Infectious Disease Research. He has
received numerous awards including the Society for Academic
Emergency Medicine Young Investigator Award. His research
interests include basic molecular innovations for rapid detection
of biologic agents in acute care settings.
Andreas Terzis received his Ph.D. degree
in computer science from the University of
California at Los Angeles. He is an assistant
professor in the Department of Computer
Science at Johns Hopkins University, where
he leads the HiNRG Group. He is a recipi-
ent of the National Science Foundation
(NSF) CAREER award. His research inter-
ests include the broad area of WSNs, including protocol design,
system support, and data management.
Address for Correspondence: JeongGil Ko, Department of
Computer Science, Johns Hopkins University, 3400 N. Charles
St., 224 New Engineering Building, Baltimore, MD 21218,
USA. E-mail: jgko@cs.jhu.edu.
References
[1] American Association of Colleges of Nursing. (2004). Nursing shortage fact sheet
[Online]. Available: http://www.aacn.nche.edu/Media/Backgrounders/shortagefacts.htm
[2] The Center for Disease Control. (2005, May). Press release: Visits to U.S.
emergency departments at all- time high; number of departments shrinking
[Online]. Available: http://www.cdc.gov/od/oc/media/pressrel/r050526.htm
[3] R. Musaloiu-E, A. Terzis, K. Szlavecz, A. Szalay, J. Cogan, and J. Gray,
Life Under your Feet: A Wireless Sensor Network for Soil Ecology, in Proc.
of the Workshop on Embedded Networked Sensors (EmNets), May 2006.
[4] S. Kim, S. Pakzad, D. Culler, J. Demmel, G. Fenves, S. Glaser, and M. Turon,
Health monitoring of civil infrastructures using wireless sensor networks, in
Proc. Int. Conf. Information Processing in Sensor Networks (IPSN), Apr. 2007.
[5] T. He, S. Krishnamurthy, J. A. Stankovic, T. F. Abdelzaher, L. Luo,
R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh, An energy-efficient surveil-
lance system using wireless sensor network, in Proc. Int. Conf. Mobile Systems,
Applications, and Services (MobiSys), June 2004.
[6] CNN, Death after two-hour ER wait ruled homicide, Sept. 15, 2006.
[7] P. Stark. (2006, July). Press release: Stark opening remarks at emergency care
hearing [Online]. Available: http://www.house.gov/stark/news/109th/pressreleases/
20060727_emergencycare.htm
[8] M. Weinger, Dangers of postoperative opioids, APSF Newslett., vol. 21,
no. 4, pp. 6168, 2007.
[9] J. Ko, J. H. Lim, R. Musaloiu-E, A. Terzis, G. Masson, T. Gao, W. Destler,
T. Gao, L. Selavo, and R. Dutton, MEDiSN: medical emergency detection in
sensor networks, ACM Trans. Embedded Computing Systems (Special Issue on
Wireless Health Systems), to be published.
[10] (2003, May) IEEE Standard for Information TechnologyTelecommunica-
tions and Information Exchange Between SystemsLocal and Metropolitan Area
Networks. Specific Requirements, Part 15.4: Wireless Medium Access Control
(MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal
Area Networks (LR-WPANs), IEEE Standard 802.15.4 [Online]. Available: http://
www.ieee802.org/15/pub/TG4.html
[11] MoteIV Corporation. (2007, May). Tmote Mini [Online]. Available: http://
blog.moteiv.com/archives/2007/05/introducingtmo.php
[12] Nellcor Puritan Bennet Inc. (2006). OxiMax NELL-1: OEM Pulse Oximetry
Module [Online]. Available: http://www.nellcor.com/Serv/manuals.aspx?ID=291
[13] Hopkins Internetworking Research Group (HiNRG). (2008). CC2420 security sup-
port for TinyOS 2 [Online]. Available: http://hinrg.cs.jhu.edu/git/?p=jgko/tinyos-2.x.git
[14] D. Malan, T. Fulford-Jones, M. Welsh, and S. Moulton, CodeBlue: An ad
hoc sensor network infrastructure for emergency medical care, in Proc. Int.
Conf. Mobile Systems, Applications, and Services (MobiSys) Workshop on Appli-
cations of Mobile Embedded Systems (WAMES), June 2004.
[15] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, Collection tree proto-
col, in Proc. ACM Conf. Embedded Networked Sensor Systems (SenSys), Nov. 2009.
[16] J. G. Ko, T. Gao, and A. Terzis, Empirical study of a medical sensor appli-
cation in an urban emergency department, in Proc. Int. Conf. Body Area Net-
works (BodyNets), Apr. 2009.
[17] Nanotron Technologies. (2007). Nanotron nanoLOC TRX Datasheet
[Online]. Available: http://www.nanotron.com/EN/PRnlTRX.php
Existing devices for monitoring patients vital
signs are bulky and cumbersome to use during
patient transports within the hospital.
___________
______________
___
___
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Real-Time Analysis
for Intensive Care
T
he lives of many thousands of children born premature
or ill at term around the world have been saved by those
who work within neonatal intensive care units (NICUs).
Modern-day neonatologists, together with nursing staff
and other specialists within this domain, enjoy modern technol-
ogies for activities such as financial transactions, online pur-
chasing, music, and video on demand. Yet, when they move
into their workspace, in many cases, they are supported by
nearly the same technology they used 20 years ago. Medical
devices provide visual displays of vital signs through physio-
logical streams such as electrocardiogram (ECG), heart rate,
blood oxygen saturation (SpO
2
), and respiratory rate. Electronic
health record initiatives around the world provide an environ-
ment for the electronic management of medical records, but
they fail to support the high-frequency interpretation of stream-
ing physiological data. Recent medical research has reported
that potentially life-threatening conditions such as nosocomial
infection [1], pneumothorax [2], intraventricular hemorrhage
[3], [4], and periventricular leukomalacia [5] exhibit early indi-
cators in physiological data (see Conditions Affecting Patients
in an NICU). These indicators precede the detection of the
medical conditions using existing clinical practices.
We have taken a collaborative research approach to address
this need to provide a flexible platform for the real-time online
analysis of patients data streams to detect medically significant
conditions that precede the onset of medical complications. The
platform supports automated or clinician-driven knowledge dis-
covery to discover new relationships between physiological data
stream events and latent medical conditions as well as to refine
existing analytics. Patients benefit fromthe systembecause earlier
detection of signs of the medical conditions may lead to earlier
intervention that may potentially lead to improved patient out-
comes and reduced length of stays. The clinician benefits from a
decision support tool that provides insight into multiple streams
of data that are too voluminous to assess with traditional methods.
The remainder of this article summarizes the strengths of
our research collaboration and the resulting environment
known as Artemis, named after the Greek goddess associated
with protecting child-bearing women and young children,
which is currently being piloted within the NICU of The Hos-
pital for Sick Children (SickKids) in Toronto, Ontario, Canada.
Although the discussion in this article focuses on a NICU, the
technologies can be applied to any intensive care environment.
Research Collaboration Teams
Artemis was designed, built, and deployed by a multiinstitu-
tional, multidisciplinary research team. The research team from
IBM T.J. Watson Research Center has an average of more than
15 years of industrial research experience in distributed comput-
ing, ubiquitous computing, pervasive health care, and machine
learning. They also bring five years experience in building
health-care solutions, including two years experience building
health-care solutions using a state-of-the-art stream computing
platform, which was the result of a five-year research project.
The University of Ontario Institute of Technology (UOIT)
team brings to the collaboration expertise in health infor-
matics with more than ten years of research collaboration with
clinicians on information technology use in NICUs, event
stream processing, and acquisition of data from medical sen-
sors, and 20 years of expertise in data warehousing and data
mining. Their research on temporal abstraction is particularly
relevant for the real-time processing and temporal data-
mining components of the project [6].
The team from the SickKids and the Department of Pedia-
trics, University of Toronto, has more than 20 years of experi-
ence in neonatology and clinical research. This team provides
the clinical and medical expertise, guides in interpreting the
analytic results, and leads in designing the clinical deployment
of Artemis at the SickKids.
Collaborations like this involve more than just a shared
technical vision and complementary research teams. The teams
had to put in place intellectual property agreements and institu-
tional Research Ethics Board applications that were mutually
agreeable and that met the guidelines at each institution.
Evolution of the Artemis Design
The Artemis system design evolved over time as the research
teams joined together to define the research project. The infor-
matics research team from the UOIT has many years of experi-
ence in collaborating with clinicians on the application of
information technologies in NICUs. They had started a path of
research on collecting patient data froma large group of patients,
BY MARION BLOUNT, MARIA R. EBLING,
J. MIKAEL EKLUND, ANDREW G. JAMES,
CAROLYN MCGREGOR, NATHAN PERCIVAL,
KATHLEEN P. SMITH, AND DABY SOW
EYEWIRE
Development and Deployment
of the Artemis Analytic System
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
and given the known outcomes of the patients, to perform analy-
sis to find correlations between distinctive patterns in the physio-
logical data streams and the onset of medical conditions. One
goal of the Artemis project is to assist them in taking the steps in
their research to perform temporal data mining on a broad set of
physiological data frominfants in a NICU. To do this, we need to
capture all of the raw physiological data streams from a large
number of infants over time. This means that the Artemis system
must be capable of 1) interfacing with a broad set of medical
devices and 2) storing the raw physiological data from multiple
infants at the rate the data are generated.
A second major goal of the Artemis project is to run clinical
rules, some of which are derived from the research described
earlier, online and in real time. In this way, knowledge of early
indicators of medical conditions can be made available to clini-
cians as soon as they are detected. This means the data must not
Conditions Affecting Patients in an NICU
Nosocomial Infection
Infection is a very common cause of morbidity and an
important cause of mortality for the newborn infant.
Although many infants acquire their infection around
the time of delivery, others acquire an infection while
receiving intensive care in the NICU. Nosocomial infec-
tions, also called hospital-acquired infections, are infec-
tions that are secondary to the original cause for
admission into the NICU. The early diagnosis of a noso-
comial infection is difficult, because the clinical signs of
infection are usually subtle, vague, andnonspecific until
the infection is well established. Such infections occur
48 h or more after birth and are caused by pathogens
not associated with the mother. Data from the neona-
tal network indicate that almost 30% of infants born at
2528 weeks gestation and more than 45% of infants
born prior to 25 weeks gestation will experience a seri-
ous nosocomial infection while in the NICU [S1]. Earlier
detection and intervention would be expected to
reduce morbidity andmay reduce mortality.
Pneumothorax
One to 2% of all newborns have air or gas in the pleural
cavity that separates the visceral from the parietal
pleura. The lungs are surrounded by a membrane that
folds back on itself, with one layer attached to the
chest wall and one layer attached to the lungs. The
membrane produces a fluid that acts as lubrication so
that these layers move smoothly when we inhale and
exhale. When air or gas accumulates between these
two layers, it is called pneumothorax. Goldberg [S2]
showed that the recognition of subtle clinical signs,
including increased in systolic arterial blood pressure as
well as an increased heart rate and pulse pressure, can
lead to earlier recognition of pneumothorax, and
therefore, earlier intervention.
Intraventricular Hemorrhage
Intraventricular hemorrhage (IVH) is another common
cause of morbidity and mortality for the newborn infant.
Approximately 20% of preterm infants less than 1,500 g
birthweight develop an IVH. The incidence and severity
are inversely proportional to gestational age. The
hemorrhages occur during the first few days of life. More
than 90% of the IVHs have occurred by the third day of
life. Nearly 10% of IVHs occur before delivery. Important
risk factors for IVH include extreme immaturity, birth
asphyxia, asynchronous breathing of ventilated preterm
infants, pneumothorax, and sudden increase in arterial
blood pressure [S3]. Fluctuations in cerebral blood flow,
and especially blood flow through the delicate, fragile
blood vessels of the germinal matrix layer, a centrally
located region of the brain, are considered to be the
dominant cause of IVH. There are many events around
the time of birth and during the first week of life, which
are associated with fluctuations in cerebral blood flow.
The likely cause of most IVHs is a rapid increase in cere-
bral bloodflowoccurringafter a periodof reducedflow.
Periventricular Leukomalacia
Periventricular leukomalacia (PVL) refers to the death
of white matter near the cerebral ventricles: 34% of
premature, very low birthweight (1,500 g or 3 lb 5 oz)
infants and 410%of those born prior to 33 weeks gesta-
tion will develop PVL [S4]. The white matter is the inner
part of the brain, and periventricular refers to the part
of the white matter that surrounds the ventricles. Leuko-
malacia refers to the softening of the white matter,
which quickly leads to death of the brain tissue. PVL
typically occurs when a fetus or newborn experiences
oxygen deprivation during labor and delivery or at any
time after birth. Variations in the oxygen and carbon
dioxide content of the blood are involved in the causa-
tion of PVL. Stream analysis of physiological data has
the potential to detect these variations.
References
[S1] R. Polin and L. Saiman, Nosocomial infections in the neonatal
intensive care unit, NeoReviews, vol. 4, no. 3, pp. e81e89, 2003.
[S2] R.N. Goldberg, Sustained arterial blood pressure elevation
associated with pneumothoraces: Early detection via continuous
monitoring, Pediatrics, vol. 68, no. 6, pp. 775777, 1981.
[S3] J.J. Volpe, Intracranial hemorrhage: Germinal matrix-intraven-
tricular hemorrhage of the premature infant, in Neurology of the
Newborn, 4th ed. Philadelphia, PA: Saunders, 2001, pp. 428493.
[S4] J.J. Volpe, Hypoxic-ischemic encephalopathy: Clinical
aspects, in Neurology of the Newborn, 4th ed. Philadelphia, PA:
Saunders, 2001, pp. 331394.
Infection is a very common cause of morbidity
and an important cause of mortality for the
newborn infant.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
only be stored in real time, but the data must be processed in
real time. IBM Research entered the collaboration with the goal
of exploiting a novel streaming middleware system developed
in a five-year, multidisciplinary, multiparty research project. It
was decided to include the streaming middleware component in
the Artemis system to provide an online, real-time processing
run-time environment. If we want to execute clinical rules, the
neonatologist pointed out that we must incorporate patient data
from a number of data sources; streaming physiological data
frommedical devices was necessary but not sufficient for clinical
rules. Therefore, another Artemis system requirement was the
need to stream the most up-to-date data from the clinical infor-
mation management system(CIMS) and the laboratory system.
Artemis would have a set of components to aid in discover-
ing clinical rules and a set of components for executing clini-
cal rules. The Artemis team decided on a goal of creating a
closed-loop system, whereby the new clinical rules, parameter
values, and clinical rule refinements can be immediately
deployed in the run-time component of Artemis. The key to
the closed-loop system is to create an ontological relationship
between the output of the knowledge extraction component of
Artemis and the clinical rule execution component of Artemis.
A number of Artemis system requirements are generated from
this goal. We need a way to efficiently integrate newly
captured patient data into the data mining repository. We need
an ontology relating data mining outputs to clinical rules.
These high-level system goals generated the following
Artemis design requirements:
support real-time processing of multiple high-rate physio-
logical data streams using a novel stream processing system
interface to and stream data from medical devices, the
clinical information management system CIMS, and the
laboratory system
store all of the raw physiological data from devices con-
nected to an infant and selected patient data from other
sources
support temporal data mining and other data mining
techniques to find relationships, particularly time-based
relationships, between patterns and correlations in multi-
ple patient data streams and medical conditions
scale with respect to the number of data streams and the
number of patients connected to the system.
The next section will describe the system that was imple-
mented to meet these requirements.
The Artemis Framework
The Artemis platform supports the acquisition and storage of
patients physiological data streams and clinical information
The Artemis platform supports the acquisition
and storage of patients physiological data
streams and clinical information system data for
the purposes of online real-time analytics,
retrospective analysis, and data mining.
Data Integration Mgr Knowledge Extraction
Data Miner
HIR
Data Mover
Deployment Server
Ontology-Driven
Rule Modifier
Deployment Server
Alert Sink
Op
QRS
BP
RR PT
FA
WT
AR
Sepsis
BPA
EP
WTA
HR Source Op
SpO2 Source Op
BP Source Op
CIS Source Op
Patient Stream
Medical
Data
Hub
CIS Adapter
Configuration
Server
CapsuleTech
Server
MP50
Data Captor
Terminal Unit
Clinical
Information
System
ECG
SpO2
BP HR
U
s
e
r

I
n
t
e
r
f
a
c
e
InfoSphere Streams Run Time
DB2
Fig. 1. Artemis system architecture.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
systemdata for the purposes of online real-time analytics, retro-
spective analysis, and data mining. The Artemis system archi-
tecture is illustrated in Figure 1.
To meet the goal of being able to interface Artemis to the
myriad of medical devices used in intensive care environ-
ments, Artemis employs a set of hardware and software ele-
ments from Capsule Tech Inc. [7]. Capsule Tech has cables
and device drivers for interfacing with more than 450 different
types of devices from all of the major vendors. A DataCaptor
terminal unit shown in Figure 2, which is located near the
medical devices, can connect to eight devices and convert the
devices RS232 output to an Internet Protocol (IP) stream.
Alternate RS232 to IP converters have also been tested and
found to be compatible with this configuration. By converting
the data stream to an IP stream, only the terminal unit has to be
located in the busy and crowded intensive care unit. The other
Artemis computers can be placed in a secure area of the hospi-
tal. All data are forwarded to a Capsule DataCaptor Interface
Server that can support up to 500 simultaneously connected
devices [8]. Using the Capsule application programming inter-
face (API) and software development tool kit, we imple-
mented a server-based function to filter the data received at
the server to extract only data streams necessary for the study,
format it for use by Artemis, and send it to the Medical Data
Hub. We implemented a configurable Medical Data Hub sys-
tem consisting of a set of data hubs that receive the aggregated
data item from the server and create concurrent data streams
for the streaming system. Although a single Capsule Server is
sufficient for our deployment, multiple servers can be used to
achieve system scalability.
The clinical information system (CIS) adapter interfaces with
the clinical information management systemCIMS to access the
SickKids CIMS patient data and stream the data to the Artemis
clinical rules. Information-use protocols at
the SickKids do not allow nonclinical ac-
cess to the CIMS or laboratory database, of
which both are Oracle based. If clinical
researchers at SickKids require such data
for patients satisfying their research study,
a database for enrolled/qualifying patients
is created in which CIMS data and labora-
tory database data are replicated every
30 min. Researchers then use these shadow
databases for their approved research. As
per this protocol, a replica subset database
for Artemis was created that is populated
for enrolled patients. The CIS adapter ex-
tracts the selected data for infants enrolled
in the Artemis project and maintains the
data in a set of database tables accessible
by Artemis. We have implemented system operators to access
these interface database tables and stream the updated data to
the clinical rule applications.
The core of Artemis is a stream computing middleware
component, IBM InfoSphere Stream Computing System, which
provides scalable processing of multiple streams of high-volume,
high-rate data [9], [10]. The conventional approach to processing
data streams is to store the data and then immediately analyze the
data in near real time. If the processing cannot keep up with the
rate, the data are stored and will eventually be analyzed. Stream
processing systems are compute-first, store-second systems. In
Artemis, processing of data streams and storing of the data are
done concurrently. The stream computing system can run on a
range of systems fromnotebooks to supercomputers; thus, it pro-
vides Artemis with a very scalable real-time execution environ-
ment. An application in streams consists of a set of operator
nodes interconnected in a graph. Each operator node inputs one
or more streams and produces one or more output streams.
The programming language for the stream computing sys-
tem is stream processing application declarative engine
(SPADE). SPADE is a high-level declarative language for
programming the streaming system [11]. It allows a program-
mer to specify the data streams, operators, and connections
between the operators and streams. SPADE language con-
structs have many similarities with higher-level programming
languages like stream structured query language (StreamSQL),
yet provide mechanisms to interact with lower-level system
programming APIs if needed. SPADE has operators that are
specialized for ingesting data from varied data sources, for
interacting with external entities through protocols like IP and
simple mail transport protocol (SMTP), and for coordinating
data streams during processing. Indeed, while SPADE pro-
vides a set of built-in stream-relational operators able to per-
form relational query on data streams,
SPADE can also be extended. It allows the
developer to specify user-defined opera-
tors typically written in C++ or Java, when
the need for complex operators arises.
These operators can be integrated seam-
lessly with built-in operators to compose
applications. An application specified in
the SPADE language is compiled into
directed graph processing elements, which
can then be instantiated on the system run
time. Artemis clinical rules are imple-
mented as SPADE programs.
The data integration manager (DIM)
consists of a set of SPADE operators that
have the specialized function to interact
with an open database connectivity
Fig. 2. Capsule DataCaptor terminal
unit.
The data integration manager allows Artemis
to interact with an open database
connectivity system.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
(ODBC) system. Some operators are specialized to insert a
steam element into the database and some are specialized to
stream data from a database table to the application. The
requirement to store all raw data into the database is achieved
by the selective placement of DIM operators in the application
graph. Periodically, the data are moved fromthe DIMdatabase
systemto a data repository in the knowledge extraction compo-
nent by the data mover. The data mover gives us the ability to
control the frequency of the data movement and to perform
some data transformations in preparation for data mining.
The knowledge extraction component uses new, multidimen-
sional, temporal, data stream data mining frameworks and tech-
niques [12][14]. These new data mining approaches can be
automated but also enable clinicians to perform scientific
method-based hypothesis research as an active participant in the
multidimensional temporal data mining process. The output of
the knowledge extraction component can be a new correlation
between a pattern detected in the streaming physiological data
and a medical condition or a newset of parameters for an existing
SPADE application. We are developing an ontology that relates
the output of the knowledge extraction component and the
expression of the clinical rules. The ontology-driven rule modi-
fier (ODRM) component, using the ontology, supports the trans-
lation of the new rule into a SPADE program as well as the
modification of the parameters in an existing SPADE applica-
tion. This closed-loop feature permits us to dynamically update a
running clinical rule or replace the clinical rule with the one that
has been shown through analysis and testing to be better.
The deployment server supports the deployment of the new
SPADE application into the stream computing run time and
maintains information about which SPADE applications are
active at any given time. The deployment history can be queried
to determine which version of what clinical rule was active for
a given patient at a given time. The deployment server has both
a graphical user interface and a programmatic interface. The
ORDM uses the programmatic interface to drive the redeploy-
ment of clinical rules based on the output of the knowledge
extraction component. When a patient is enrolled in Artemis,
the deployment server deploys the clinical rule applications. All
clinical rule applications allowper-patient parameterization.
Programming Clinical Rules in Artemis
A clinical rule is a specification as detailed within an existing
clinical guideline, defined anecdotally by a clinician as part of
clinical research, or proposed through data mining of a set of
conditions in the physiological data streams, laboratory results,
and observations of a patient, which if found to be present,
holds a strong correlation as a predictor for an impending clini-
cal event and as such should be reported by some means to a
nurse or physician. Some clinical rules can be processed by the
people. An example of such a rule is notify me when an
hourly systolic blood pressure (BP) reading exceeds 140. The
clinician who takes the patients BP can easily monitor this
clinical rule. However, when the clinical rule involves specifi-
cation of second-to-second changes across multiple data sour-
ces of physiological streams, it becomes impractical and
sometimes impossible for a human to detect these subtle sig-
nals of condition onset [15]. In some cases, the conditions are
episodic and do not occur often, and it is unlikely a clinician
will be observing when the condition occurs.
Although Artemis can easily monitor the simple clinical
rules, it is ideally suited for clinical rules that specify condi-
tions involving multiple data sources and/or high-rate data
sources. Data sources such as ECGs and electroencephalo-
grams are time sampled at 5001,000 Hz and require nontri-
vial signal processing techniques for analysis. As mentioned
in the previous section, Artemis clinical rules are implemented
as streaming system applications encoded in the SPADE
language. The following paragraphs provide an example of a
clinical rule and explain portions of the program for imple-
menting the clinical rule.
Although our testing encompasses simple and complex
rules, we provide further details via the explanation of a simple
clinical rule that we used to test Artemis. The clinical rule if
mean arterial BP (MBP) is less than the neonatal patients cur-
rent gestational age (e.g., 24 mmHg for 24 weeks gestation) for
20 s or more, and if SpO
2
is less than 85% for the same period
of 20 s or more, then a reportable condition is present. Prior to
commencing our pilot study with SickKids, we utilized data
from the companion Targeted Delivery Intervention Study
(TARDIS) [14], a randomized control trial of volume-targeted
ventilation for resuscitation of preterm babies intubated in the
delivery suite. The data were collected as discontinuous seg-
ments of at least 2 h of data at a time commencing en route to
the delivery suite and then at 12, 24, 36, and 48 h after delivery.
For that study, each data set contains multiple concurrent
streams of physiological data; however, for the purpose of this
case study demonstration, we focused on six, 5-min, patient
segments extracted at a reduced data set containing four of
these temporal data streams, specifically: 1) ECG; 2) SpO
2
; 3)
MBP; and 4) BP. The team performing the test of Artemis was
blinded to the characteristics of these chosen segments.
We will describe this clinical rule using the four stages of
clinical rules that have been observed. Figure 3 illustrates the
four stages.
The Adaptation Stage
In this stage, SPADE source operators listen for streaming
events sent by the medical data hubs, which in this study are
simulators which replay the TARDIS data in faster-than real
time. This scenario uses two source operators: one for SpO
2
data and the other for the MBP data.
The intensive care unit is a very dynamic
environment where the care of the patient
is paramount.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
The Intrastream
Analysis Stage
The SpO
2
module and the
MBP module operate in paral-
lel. The SpO
2
module contains
logic to count the number of
times SpO
2
events are below
85% in a 20-s sliding window
and issues SpO
2
alerts if this
count is greater than a prede-
fined threshold. Meanwhile,
the MBP module contains log-
ic to count the number of
times MBP events fall below
the patients gestational age,
in a 20-s sliding window, and
issues MBP alerts if this count
is greater than a threshold.
The gestational age is a pa-
rameter obtained from the
CIMS. To access it within this
SPADE application, we use
the DIMenrich operator.
The Fusion and
Scoring Stage
In the fusion and storage step,
the SpO
2
and MBP features are merged using a SPADE join
operator. Its output is consumed by another operator that deter-
mines whether this is a reportable event for this clinical rule.
The Delivery Stage
The delivery stage externalizes reportable analytic results. In
our test setup, we wrote the result to a database table and sent
a short message service/e-mail message.
By replaying these traces into the SPADE application
described earlier, we accurately detected which patients exhib-
ited the clinical rule. The clinical rule was present for two
patients within these six traces: patients 2 and 5. Both were cor-
rectly identified by the SPADE application. For the remaining
normal patients, no alerts were generated by the SPADE appli-
cation, as expected.
Figure 4(a)(c) shows the plots of the results obtained for
patient 5. We can clearly see in these graphs that the long dips
in MBP and SpO
2
were accurately identified and resulted in
instability alerts. For patient 5, we may note that the SPADE
code also detected the number of physiological streams that
are out of range, as shown in the Figure 4(a).
While clinical rules can be implemented in the SPADE
language, there are approaches to integrating other rule repre-
sentation languages with SPADE. The first approach lever-
ages standard rule representation languages, thus allowing our
infrastructure to be compliant with the existing rule languages.
One example is the predictive modeling markup language
(PMML) [16], a well-established extensible markup language
(XML) dialect used to represent prediction rules. The streams
teamhas implemented an operator that accepts PMML models
and scores them in real time.
The second approach to represent clinical rules can be used
by developers who have proprietary rule representation
schemes. In this case, one can leverage the extensibility of the
SPADE language to develop user-defined operators capable
of interpreting such proprietary rules. SPADE user-defined
operators can be implemented in either C/C++ or Java, thus
facilitating the integration of legacy clinical rule systems.
Security and Privacy
Because the data collected is personal health-care data, we are
bound by the health-care privacy laws of Canada, the United
States, and the province of Ontario. The Research Ethics
Boards at the three institutions all mandated a plan to ensure
Alerts for Patient 5
2
1
0
Instability Alert State
Time (s)
Time (s)
Time (s)
(a)
(b)
(c)
I
n
s
t
a
b
i
l
i
t
y

A
l
e
r
t
s
S
p
O
2

M
B
P
0 50 100 150 200 250 300 350
0 50 100 150 200 250 300 350
0 50 100 150 200 250 300 350
100
90
85
80
75
32
30
20
MBP Alert
State
SpO
2
Alert
State
Fig. 4. Alerts generated by a sample clinical rule.
Source
MBP
Source
SpO
2
Aggregate
Aggregate
f (x)
Functor
f (x)
Functor
f (x)
Functor
f (x)
Functor
f (x)
Functor
Join
Database
At
SMTP
SpO
2
Module
Fusion and Scoring
Module
Delivery
Module
Delivery
Stage
Fusion and Scoring
Stage
Intratype Analysis
Stage
Adaptation
Stage
Adaptation
Module
MBP
Module
Fig. 3. Examples of stages of sample clinical rule.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
compliance with the laws. We do not collect or store any data
that could directly identify the infants. We deidentify the data
prior to it entering Artemis; Artemis processes and stores only
deidentified patient data. An Artemis identifier is generated for
each infant when the infant is enrolled. An unique identifier of
the patient monitor associated with each data element is trans-
mitted fromthat patient monitor. The hospital tracks an associa-
tion between a bed and the patient monitor associated with that
bed. During enrollment processing, an association between the
patient monitor identifier and the Artemis identifier is placed in
a mapping database table. As the data streams into the Artemis
system, one of the initial operators has the task to pick up the
patient identifier from the data elements and insert the associ-
ated Artemis identifier, thereby, deidentifying the data.
The Artemis system operates on the hospital network that
has been secured for the transmission of all types of medical
data. The processing components and Artemis database sys-
tem are located in a physically secure location accessible only
by Artemis team members. Using the built-in authentication
system of the operating system and the authorization mecha-
nism and access control system of the database system, Arte-
mis controls the operations a user can perform on the
physiological data streams stored in the database. Because the
system is separate from all other hospital systems, we can
administer access to the Artemis database to ensure there is no
unauthorized access, even from within the hospitals network.
The interactions between the deployment server and the
streaming system are done over the network, so steps have
been taken to close the security exposure. First, access to the
deployment server is password protected. Furthermore, the
deployment server authenticates with the streaming system
for each remote operation it attempts to performon the stream-
ing system.
The UOIT creates a mirror copy of the Artemis database by
doing incremental downloads on a periodic basis. The connec-
tion between the computers at the SickKids and UOIT is made
using a secure tunnel. The tunnel is implemented using the tun-
neling capabilities of the secure shell 2 (SSH2) protocol. The
tunnel is encrypted with a 4,096-b Rivest, Shamir, Adleman
public encryption algorithm (RSA) key. At the UOIT, the mir-
ror copy is maintained on a dedicated computer in a secure
locked roomaccessible to members of the UOIT teamonly.
The IBM team maintains a mirror by performing incremen-
tal periodic downloads from the UOIT mirror. The same
secure tunnel mechanisms are used to securely transmit the
data. The IBM mirror is stored on a password-protected
machine stored in a secure laboratory environment. The data
are available only to members of the IBM team.
Deployment
Artemis has been deployed in the NICU of the SickKids in
Toronto, Ontario, Canada, since early August 2009. In this
phase of deployment, we are capturing physiological data
streams and electronic health record information forwarded
from the CIMS data for the infants within the study. The
physiological data being collected contains ECG, heart rate,
respiratory rate, and blood SpO
2
; BP may be streamed or
obtained as CIMS observations. We also monitor these data
using Artemis and an initial version of a clinical rule for the
early detection of nosocomial infection, which represents a
complex rule that will be reported in future research publica-
tions. In this phase, we have monitored as many as four infants
at the same time. Because the stream computing middleware
is scalable, we believe our approach could easily be used to
monitor all patients within the NICU, but we have not tested
this configuration. As of early December 2009, 19 infants
have been enrolled in the research study.
A Capsule DataCaptor terminal unit is located in each of
two NICU bed bays. The placement of a unit in a bed bay is
illustrated in Figure 5. Figure 6 shows a typical bed in a NICU
with the associated equipment. The DataCaptor terminal unit
transmits data streams from Philips IntelliVue MP70 patient
monitors to the Capsule server. Each DataCaptor terminal unit
can handle up to eight patient monitors. The dimensions of the
unit are 9.3 in by 10.6 in by 2.1 in. To evaluate the processor
and memory requirements of the Artemis server components,
we partitioned the Artemis server among a Windows PC and
two Linux notebooks.
Before Artemis could be deployed at the SickKids, we had
to perform some tests to provide assurances to the hospital
staff. All interfaces to the existing NICU devices and hospital
systems had to be designed so that no data or process used by
the NICU staff would be disrupted. Hardware deployed in the
NICU could not interfere with clinical processes and had to be
effectively invisible to the medical staff. In the NICU, there is
a patient monitor (Philips IntelliVue MP70) per patient, and
Fig. 6. Example of devices and equipment attached to an
infant in an NICU (photo used with permission from Terry
Tremethick).
Local
Area
Network
Local
Area
Network
Bed D Bed E Bed F
Bed A Bed B Bed C
Capsule
Data
Captor
Philips
MP90
Philips
MP90
Philips
MP90
Philips
MP90
Philips
MP90
Philips
MP90
Fig. 5. NICU bay floor layout.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
the devices connected to the patient are connected to the
patient monitor. There are two ports on the patient monitor for
externalizing the data streams. One port is currently being used
for clinical purposes, and we were able to demonstrate that we
can get the streamfromthe alternate port without interference.
We also performed extensive testing to ensure that the
deployment would not interfere with the existing network traf-
fic or with the existing slower feeds of data to the CIMS data-
base. The initial development testing was all performed away
from the NICU setting utilizing a Philips IntelliVue MP50
running in demonstration mode located within the Health
Informatics Research laboratory located at UOIT. Once alpha
component testing and integrated system beta testing were
completed, we commenced testing utilizing an Philips Intelli-
Vue MP70 located within the NICU in demonstration mode
and then later with a patient simulator. Our initial tests show
that the network traffic generated is below 0.2% (based on a
100 Mb/s network) for each Philips IntelliVue MP70 con-
nected to the Artemis system.
The deidentified data are collected by the deployed Artemis
system at the SickKids into the local Artemis relational data-
base. Periodically, the UOIT team, using secure connections,
performs an incremental download from the Artemis database
to a database at UOIT.
The nosocomial infection clinical rules will be refined over
time with both new feature sets and fusion rules based on data
mining results and review of the correlation of computed fea-
tures with infants who actually develop nosocomial infection.
Lessons Learned
Even in this early phase of deployment, we have learned some
lessons that should be of interest to other researchers and to us
as we move forward.
Expect Unforeseen Situations in NICUs
We realized early that it was imperative to develop mecha-
nisms that allow us to dynamically change the configuration
of the system. In particular, mechanisms have been built to
allow the dynamic deployment of data sources and the rede-
ployment of parts of applications on a running instance of the
Artemis system. Even though it was not a primary design
requirement, the decision to implement these mechanisms has
been quite important for the success of the deployment. On
several occasions, we had to either change the configuration
of the system or modify a running application and redeploy it.
Minimize Data Source Dependency for Clinical Rules
The intensive care unit is a very dynamic environment where
the care of the patient is paramount. Patients are disconnected
and reconnected to devices for various reasons. The staff is
more concerned with the generation of data they need to care
for the patient than they are in making sure the research system
is receiving the data it needs. A clinical rule should be
designed to generate analytic results as long as the requisite
data streams are flowing. The temporary loss of some data
streams may reduce the set of features available, but, the pres-
ence of the requisite streams ensures the generation of a set of
critical features required to generate results. This causes some
complication in the implementation of clinical rules in gen-
eral. The initial nosocomial clinical rule design assumed the
presence of all seven data streams; however, we determined
that four of the streams were always present and the other
three were less likely to be present. We reimplemented the
clinical rule to work under these conditions.
Conclusions
We have described the design, implementation, and initial
deployment of Artemis. Artemis has the potential to revolu-
tionize intensive care medicine, because it gives clinicians a
way to discover early indicators of medical conditions and to
encode these in clinical rules. Artemis also provides a way for
clinicians to have online, real-time execution of the clinical
rules in an intensive care environment. We have also shown
the need for a diverse collaborative team to address the broad
set of issues involved in such a system.
While many research and industry-based initiatives exist to
propose clinical rules for specific condition onset or propose
methods for detecting features in certain streams, these
approaches do not apply a systems-based approach for the
support of multiple streams, from multiple patients relating to
multiple diagnoses [17].
The UOIT and SickKids-based teams have membership
with the Canadian Neonatal Network (a network of 30 level 3
and above NICUs in Canada) and envision an ultimate
national rollout of the technology introduced here.
Marion Blount received his B.S. degree in
electrical engineering from Duke University
and his M.S. and Ph.D. degrees from Stan-
ford University. Over the course of his career
at IBM T.J. Watson Research Center, he
has conducted research in the areas of distrib-
uted computing, memory coupled systems,
pervasive computing, and utility computing.
In recent years, his interest has been in the application of stream
computing technologies to the health-care domain.
Maria R. Ebling received her B.S. degree
from Harvey Mudd College and her M.S.
and Ph.D. degrees in computer science
from Carnegie Mellon University. She is a
research staff member and senior manager
at the IBM T.J. Watson Research Center.
She manages a team, building systems
capable of detecting events in the real world
and reacting to those events in a timely and appropriate man-
ner. Her interests are in distributed systems supporting mobile
and pervasive computing, privacy, humancomputer interac-
tion, and applications of these technologies to health care.
J. Mikael Eklund received his B.Sc. and
M.Sc. degrees in engineering in 1989 and
1997, respectively, and Ph.D. degree from
the Queens University, Kingston, in 2003.
Between his bachelors and masters stud-
ies, he was a simulator flight control sys-
tems engineer with CAE Electronics in
Montreal, Quebec. He is an assistant profes-
sor and program director for electrical and software engineering
at the UOIT. Until August 2006, he was a visiting postdoctoral
scholar in the Department of Electrical Engineering and
Computer Sciences at the University of California, Berkeley.
His research is in the areas of time-series analysis, nonlinear
system identification, intelligent control, and data knowledge
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
discovery with applications in smart medical systems, robotics,
and other autonomous systems.
Andrew G. James received his graduate
degree in biomedical informatics from the
Oregon Health and Science University, Port-
land, Oregon, and a medical degree from the
University of Auckland, Auckland, New
Zealand. He is a staff neonatologist at the
SickKids and an associate professor of pedia-
trics at the University of Toronto, Toronto,
Canada. He is interested in the application of information tech-
nologies for the care of the newborn infant. His primary focus is
clinical decision support and biomedical knowledge representa-
tion. He is the site investigator at SickKids for the Canadian Neo-
natal Networks evidence-based quality improvement research
program. He is chair of the Informatics Research Committee and
Canadian Neonatal Network and cochair of the Concept Model
Special Interest Group, and International Health Terminology
Standard Development Organization, Copenhagen, Denmark. He
is the foundation chief medical editor of AboutKidsHealth.ca, an
evidence-based Web site for children, parents, and families.
Carolyn McGregor received her bachelor
of applied science in computer science hon-
ors degree and a Ph.D. degree in computing
science from the University of Technology
in Sydney, Australia, where she developed
new ways to use intelligent decision support
systems (IDSSs) to assist organizations in
the monitoring of business performance.
This research included an innovative IDSS framework that she
extended, resulting in the solution manager service (SMS). The
SMS contained pioneering research in Web services, event
stream processing, multiagent data mining, and nearreal-time
data warehouse loading. Her recent research focus has been the
improvement and extension of that research within the context
of health and medicine for advanced support for clinical man-
agement and research, together with new remote patient care
frameworks and patient journey modeling and reengineering
approaches. She is the Canada research chair in health infor-
matics based at the UOIT, Canada.
Nathan Percival received his bachelor of
mathematics in computer science, infor-
mation systems option from the Faculty of
Mathematics at the University of Waterloo
and a masters degree in information
technology security from the Faculty of
Business and Information Technology at
the UOIT. He is currently a staff member
at UOIT supporting the Faculty of Engineering and Applied
Science and the Health Information Research Group.
Kathleen P. Smith is a masters of health
science (health informatics) candidate
and a research associate at the UOIT
under the supervision of Dr. Carolyn
McGregor and Dr. J. Mikael Eklund. Her
research interests include real-time event
processing, temporal data mining, clini-
cal rule ontologies, and the application of
services computing to critical care and women in computing
and engineering.
Daby Sow received his bachelor of science
in electrical engineering from the Univer-
site Laval, Quebec, Canada, and his master
of science and Ph.D. degrees in electrical
engineering from Columbia University.
He was a graduate research assistant at
Columbia University in the Department of
Electrical Engineering from 1996 to 2000,
working on lossy source coding theoretical problems with
applications to image and video processing. He is a research
staff member at IBM T.J. Watson Research Center. His
research interests are at the intersection of several fields of
computer science and electrical engineering, including stream
computing, complex event processing, applied machine learn-
ing, signal processing, and information theory.
Address for Correspondence: Marion Blount, IBM T.J.
Watson Research Center, 19 Skyline Drive, Hawthorne, NY
10532 USA. E-mail: mlblount@us.ibm.com.
References
[1] P. Griffin and R. Moorman, Toward the early diagnosis of neonatal sepsis and
sepsis-like illness using novel heart rate analysis, Pediatrics, vol. 107, no. 1,
pp. 97104, 2001.
[2] N. McIntosh, J.-C. Becher, S. Cunningham, B. Stenson, A. Laing, A. J. Lyon,
and P. Badger, Clinical diagnosis of pneumothorax is late: Use of trend data and
decision support might allow preclinical detection, Pediatric Res., vol. 48, no. 3,
pp. 408415, 2000.
[3] J. Fabres, W. A. Carlo, V. Phillips, G. Howard, and N. Ambalavanan, Both
extremes of arterial carbon dioxide pressure and the magnitude of fluctuations in
arterial carbon dioxide pressure are associated with severe intraventricular hemor-
rhage in preterm infants, Pediatrics, vol. 119, no. 2, pp. 299305, 2007.
[4] V. Tuzcu, S. Nas, U. Ulusar, A. Ugur, and J. R. Kaiser, Altered heart rhythm
dynamics in very low birth weight infants with impending intraventricular
haemorrhage, Pediatrics, vol. 123, no. 3, pp. 810815, 2009.
[5] S. Shankaran, J. C. Langer, N. Kazzi, A. R. Laptook, and M. Walsh, Cumulative
index of exposure to hypocarbia and hyperoxia as risk factors for periventricular leuko-
malacia in low birth weight infants, Pediatrics, vol. 118, no. 4, pp. 16541659, 2006.
[6] M. Stacey and C. McGregor, Temporal abstraction in intelligent clinical data
analysis: A survey, Artif. Intell. Med., vol. 39, no. 1, pp. 124, 2007.
[7] Home page of Capsule Tech Inc. (2009). Medical device connectivity with
capsule [Online]. Available: http://www.capsuletech.com/index.htm
[8] Web page describing solutions of Capsule Tech Inc. (2009). Point of care con-
nectivity [Online]. Available: http://www.capsuletech.com/our-solution-
products.htm
[9] L. Amini, H. Andrade, R. Bhagwan, F. Eskesen, R. King, P. Selo, Y. Park,
and C. Venkatramani, SPC: A distributed, scalable platform for data mining,
Proc. SIGKDD 2006, Workshop Data Mining Standards, Services, and Platforms,
Aug.20, 2006pp. 2737.
[10] N. King. (2008). Instant information: Stream computing [Online]. Available:
http://www.ibmdatabasemag.com/story/showArticle.jhtml?articleID=211300227
[11] B. Gedik, H. Andrade, K. Wu, P. S. Yu, and M. Doo, SPADE: The system
S declarative stream processing engine, in Proc. Int. Conf. Management of Data,
ACM SIGMOD, 2008, pp. 11231134.
[12] M. Stacey, C. McGregor, and M. Tracy, An architecture for multi-dimen-
sional abstraction and its application to support neonatal intensive care, in Proc.
29th Annu. Int. Conf. IEEE Engineering in Medicine and Biology Society
(EMBC07), Lyon, France, 2007, pp. 37523756.
[13] C. McGregor and M. Stacey, High frequency distributed data stream event
correlation to improve neonatal clinical management, in Proc. Inaugural Int.
Conf. Distributed Event-Based Systems (DEBS07), Toronto, ON, Canada, CD-
ROM, 2007, p. 6.
[14] C. Catley, K. Smith, C. McGregor, and M. Tracy, Extending CRISP-DM to
incorporate temporal data mining of multi-dimensional medical data streams: A
neonatal intensive care unit case study, in Proc. 22nd IEEE Int. Symp.
Computer-Based Medical Systems (CBMS2009), pp. 15.
[15] G. A. Miller, The magical number seven, plus or minus two: Some limits
on our capacity for processing information, Psychol. Rev., vol. 63, no. 2,
pp. 8197, 1956.
[16] Data Mining Group. (2009). Data mining groupWelcome to DMG
[Online]. Available:http://www.dmg.org
[17] M. Stacey and C. McGregor, Survey paper: Temporal abstraction in intelli-
gent clinical data analysis, Artif. Intell. Med., vol. 39, no. 1, pp. 124, 2007.
_______________
______
____________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
____________
EYEWIRE
Clustering Techniques
T
he objective of this article is to use clustering technique
to characterize the providers of maintenance services in
a health-care institution according to their performance.
A characterization of the inventory of equipment from
seven pilot areas was carried out first (including 264 medical
devices). The characterization study concluded that the inven-
tory on a whole is old [exploitation time (ET)/useful life (UL)
average is 0.78] and has high maintenance service costs rela-
tive to the original cost of acquisition (service cost /acquisition
cost average 8.61%). A monitoring of the performance of
maintenance service providers was then conducted. The varia-
bles monitored were response time (RT), service time (ST),
availability, and turnaround time (TAT). Finally, the study
grouped maintenance service providers into clusters according
to performance. The study grouped maintenance service pro-
viders into the following clusters. Cluster 0: Identified with
the best performance, the lowest values of TAT, RT, and ST,
with an average TAT value of 1.46 days; Clusters 1 and 2:
Identified with the poorest performance, highest values of
TAT, RT, and ST, and an average TAT value of 9.79 days;
and Cluster 3: Identified by medium-quality performance,
intermediate values of TAT, RT, and ST, and an average TAT
value of 2.56 days.
Outsourcing can be defined . . . as a business strategy
through which one organization formally delegates critical and
noncritical operations to a third party . . . [1], [2]. Mainte-
nance outsourcing has become a prevalent practice in many
industries, such as medical equipment manufacturing and avia-
tion. For example, in a recent survey, it was reported that 84%
of the companies surveyed employ maintenance contractors
[3]. In another study, it was reported that annual medical equip-
ment maintenance outsourcing is worth US$26 billion [4].
There have been four traditional methods for providing
equipment maintenance service: an in-house service by the
clinical or biomedical department, the original equipment
manufacturer (OEM), a third party (TP), and insurance mainte-
nance (IM) [5]. Obviously, outsourced maintenance services
are provided by OEM, TP, and IM, which sometimes share
service tasks with the service department in the hospital. The
IM combines the use of any of the service groups with financial
protection against unusually high service costs. Maintenance
service by the OEM may be on demand or contract, whether
during or after warranty that requires overseeing the quality
of the maintenance service performed. A TP is used when the
OEM delegates its maintenance service to a TP. A TP has the
same requirements as those of the OEM, with respect to moni-
toring and quality measurement of maintenance service per-
formance. There is not a full consensus on which type of service
contract is best. It often depends on many factors (e.g., budgets,
personnel training level, etc.). For example, for developed
countries, the most common practice is to use a combination
(see Figure 1).
However, for many health-care institutions in undeveloped
countries, contracting maintenance services is inevitable,
because they often neither have properly trained staff nor the
material resources available to handle these functions on their
own. In terms of monetary volume, it has been reported that
the budget spent on maintenance service contracts is approxi-
mately 64% of the total technology management budget [6],
[7]. The Emergency Care Research Institute (ECRI) stated
that the application of mechanisms to control the level of qual-
ity of maintenance service providers could reduce the total
maintenance costs by at least 20% [8], [9]. Thus, one of the
main functions of the clinical engineer is to oversee the
maintenance services executed by external companies through
outsourcing [10], [11]. Even so, health-care institutions con-
tinue to lack adequate methods for evaluating the performance
of maintenance service providers. In many cases, the informa-
tion required to complete such studies is simply unavailable.
Overseeing the quality of outsourced equipment mainte-
nance service is perhaps the most difficult element of manag-
ing contract services [5]. The clinical engineering and
enterprise community should welcome any approach with the
goal of evaluating maintenance service quality with possible
practical applications. However, this is still an unsolved prob-
lem. Although several academic research papers address
maintenance outsourcing [12][16], none specifically address
classifying providers according to quality. Examining a
related issue, this article is the first stage of a comprehensive,
novel approach that allows clinical engineers to predict TAT
as a function of a service providers feature. The predicting
tool will help to assess the quality and effectiveness of a
BY ANTONIO MIGUEL CRUZ,
SANDRA PATRICIA USAQU
EN PERILLA,
AND NIDIA NELLY VANEGAS PAB
ON
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Measuring the Performance of
Contract Service Providers
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
maintenance service provider tasks in the health-care environ-
ment and identify areas for improvement in those processes.
As a contributory measure, this article focuses on the appli-
cation of clustering techniques to measure and classify the
performance of maintenance service providers in a university
hospital with a total of 251 licensed beds. Of these, 209 beds
are in the general care units, 11 are located in the adult inten-
sive care unit (ICU), and 31 beds and incubators are in the neo-
natal care unit. The total inventory is 1,050 medical devices.
The purpose of the study is to provide health-care institutions
with a method for evaluation and control of contracted mainte-
nance services, contributing to an area of clinical engineering
with a lack of published works. Specifically, the two objec-
tives of this work are as follows:
1) to offer to the clinical engineering community a general
methodology, allowing the characterization of mainte-
nance service providers and medical equipment invento-
ries in their charge
2) to apply clustering techniques for the classification of mainte-
nance service providers according to their performance.
Clustering is one of the several data mining tools used to
extract patterns from data. Data mining is classified into two
main categories: supervised learning and unsupervised learning
techniques. Supervised learning is a machine learning method
for deducing a function from training data. The unsupervised
learning technique is the opposite. The supervised learning
techniques are numeric prediction (both linear simple and
multiple regression) and classifications (decision tables, deci-
sion trees, classification rules, etc.). Linear numeric prediction
models fit a linear equation that combines attribute values to
predict a numeric target attribute. Classification arranges the
data into predefined groups. It is a procedure in which individ-
ual items are placed into groups based on quantitative informa-
tion on one or more characteristics inherent in the items.
With unsupervised learning, one can find two main techni-
ques: clustering and link analysis. Link analysis is a descriptive
procedure for exploring data to identify relationships among
values. It can be divided into two main portions: association
discovery and sequence discovery. Association discovery finds
rules about items that appear together in an event such as pur-
chase transactions [17], [18], while sequence discovery is very
similar, in that it yields associations related over time [18].
Clustering divides data into different groups. These groups
are in someway very different from each other, but the
composite members are very similar. Clustering allows the
researcher to find distinguishing characteristics in the data not
readily noticeable under standard statistical analyses [18].
With this method, it is not known what the clusters will be at
the start or by which attributes the data will be clustered. Clus-
tering is like classification, with the principal difference being
that clustering of the groups is not predefined.
Because the research in this article is in its first stage, a pre-
dictive study is not possible at this time, nor can association
rules be obtained. Clustering is often followed by a stage
where a decision tree or rule set is inferred, which allocates
each instance to the cluster it belongs to. As such, the cluster-
ing method is just one step on the way to a structural descrip-
tion of the knowledge representation. Therefore, clustering
was selected to tackle the problem of classifying maintenance
service providers according to their performance.
For the application of the clustering method, three techni-
ques were taken into consideration: 1) the k-means algorithm;
2) the incremental algorithm; and 3) the statistical clustering
method [18]. The k-means choice for clustering is suitable
when all variables analyzed are numeric. The k-means follows
a simple and easy way to classify a given data set through a
certain number of clusters (assume k clusters) fixed a priori.
The main idea is to define one k centroid for each cluster. (The
centroid of a plane figure X is the intersection of all straight
lines that divide X into two parts of equal moments about the
line. That is, the average of all points of X.) However, when
39.00
68.00
80.00
43.00
46.00
6.00 2.00
8.00
87.00
4.00
15.00
55.00
16.00
2.00
3.00 8.00
21.00
49.00
10.00
120.00
100.00
80.00
60.00
P
e
r
c
e
n
t
a
g
e
40.00
20.00
0.00
A
n
e
s
t
h
e
s
i
a

U
n
i
t
s
V
e
n
t
i
l
a
t
o
r
s
P
a
t
i
e
n
t

M
o
n
i
t
o
r
s
2
-
D

U
l
t
r
a
s
o
u
n
d

S
c
a
n
n
e
r
s
D
i
a
l
y
s
i
s

U
n
i
t
s
L
a
s
e
r
s
F
e
t
a
l

M
o
n
i
t
o
r
s
80.00
37.00
Equipment Categories
Types of Maintenance
Service for Common Equipment Categories
Third Party
OEM
In-House Service
Fig. 1. Types of maintenance service for common equip-
ment categories [5]. Note 1: Does not total 100%; for clarifi-
cation, please refer to [5].
Maintenance outsourcing has become a
prevalent practice in many industries, such as
medical equipment manufacturing
and aviation.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
attributes or variables analyzed are not numeric, this algorithm
does not work well (obviously, the centroid calculations are
difficult to obtain because the distance between elements is
not evident). To solve the problem mentioned above, the
incremental and statistical clustering methods were created.
Both methods form clusters as a probability distribution. The
best-known incremental clustering algorithms are COBWEB
(conceptual cluster algorithm). However, one problem with
the incremental clustering method is overfitting. The expecta-
tion-maximization algorithm (EM) is better than the incre-
mental methods. The EM algorithm is used for finding the
maximum likelihood estimates of parameters in probabilistic
models. The EM is better because it is the only clustering
method that generates an explicit knowledge structure that
describes the clustering in a way that can be readily visualized
[18]. The use of EM algorithm is preferable when analyzed
variables waiting to be clustered are both numeric and nonnu-
meric, and/or when there are missing values in the sample
[18]. A tutorial on the EM algorithm can be found at [19]. This
algorithm works iteratively in two steps:
1) The E-step (expectation): In which the missing data are
estimated, and the probability of the possible groups are
calculated following the Bayer theorem:
P(A=x) =
P(x=A) + P(A)
P(x)
=
f (x; l
A
, r
A
) + P
A
P(x)
, (1)
where A is a cluster, x is a member of cluster A, P
A
is the
probability of the cluster A, f (x; l
A
,r
A
) is the function of
the normal distribution of the cluster A, P(x/A) is the proba-
bility of member x belonging to cluster A, l
A
is the average,
and r
A
is the standard deviation.
2) The M-step (maximization): In which l
A
, r
A
parameters
distribution of every cluster are calculated by the maxi-
mization of the likelihood function (the likelihood func-
tion is a measure of how good is the cluster obtained).
Distribution parameters are calculated as follows:
l
A
=
P
N
i=1
w
i
+ x
i
P
N
i=1
w
i
, r
A
=
P
N
i=1
w
i
+ (x
i
l)
P
N
i=1
w
i
, (2)
where N is the total number of elements in the cluster, w
i
is
the probability of member x belonging to cluster A, l
A
is
the average and r
A
: is the standard deviation.
The EM algorithm stops when the maximization condition
is reached, that is: the incremental values of every iteration is
negligible (3) and (4)
MF =
Y
M
i=1
X
M
j=1
p
j
!
+ P(x
i
=j), (3)
lim
i
(MF
i
MF
i1
) ~ 0, (4)
where i is the number of iterations, j is a specific cluster, M is
the total number of clusters, MF is the maximization function
of iteration i, p
j
is the probability of cluster existence, and
P(x/j) is the probability of member x belonging to cluster j.
For managers of clinical engineering departments, the
application of this research will have a significant impact in
decision-making processes. It will allow them to cluster
maintenance service providers, taking into account both their
institutional features and performance variables that charac-
terize maintenance service quality (values of the RT, ST, and
TAT indicators). Classifying maintenance service providers
as such, managers can make important decisions on their rela-
tionships with maintenance service providers, such as decid-
ing to maintain, stop, or strictly monitor existing services.
Methods
First, a study characterizing the hospitals inventory was
conducted. The study was carried out by characterizing the
inventory according to the pilot area and equipment type. Seven
pilot areas were selected: diagnostic images (BB), surgical unit
(CC), ICU (DD), neonatal care (FF), emergency (AA), interme-
diate care unit (EE), immunology laboratory (HH), clinical
laboratory (GG), and microbiology (II). The authors selected
these areas owing to the technical complexity of their equip-
ment, which makes them the most likely to have maintenance
services contracted to an external provider. The equipment
types included were diagnostics (A), treatment (B), monitoring
(C), support (D), and laboratory (E) [23].
Twelve variables for inventory characterization were studied:
total number of equipment; the ratio equipment ET to UL; total
inventory value; the value of equipment inventory as a percent-
age of the total analyzed inventory (also known as acquisition
cost penetration); cost of contracted maintenance service pro-
viders; cost of in-house maintenance; the ratio comparing cost
of in-house maintenance to the cost of contracted maintenance
services; total maintenance cost, which adds the total of con-
tracted and in-house maintenance service costs; percentage of
maintenance cost by area to total maintenance cost; and the ratio
of service cost to the acquisition cost (SC/AC).
Next, a survey was designed and conducted to characterize
maintenance service providers. A total of 26 variables were
studied (see Table 1). These variables are the most typical
for completing an adequate characterization of maintenance
service providers according to ECRI [5]. Additionally, as a vali-
dation test, the survey was submitted to review by experts.
Thus, the survey satisfies the construct validity criteria require-
ments. [24]. The most important variables are company name;
contract service start date; contract service end date; contract
duration (in years); total quantity of equipment included in the
contract (NEquip); whether the contract is a guarantee contract
(Guaranty); whether the contract includes replacements parts
(REParts); number of annual visits stipulated in the contract
(NvisitYear); equipment types involved in the maintenance
contract service (EquipTypes); total number of the companys
engineers and technicians (TotalHumResCompany); number of
engineers and technicians working on the contract (TotalHum-
ResContract); experience of the companys human resources,
in years (ExpPersonalYear); distance of the service provider
from the hospital, in kilometers (HospDist); quantity of con-
tracts managed by the company (TotNumContracts); experi-
ence of the company in the industry, in years (ExpCompany);
whether online services are offered (ServLine); whether
replacement stocks exist in the country (BackUp); and whether
training is provided to users and operators (UserTraining and
TechTraining).
A monitoring of each service provider, including in-house
maintenance service, was then completed for the following
variables: RT, ST, TAT, and availability of equipment [6],
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
[7]. The purpose of this monitoring was to observe service
quality and find the possible categories for grouping these var-
iables for each maintenance service provider.
Software Packages
As the sources of data used in this research are both numerical and
categorical, the k-means algorithm for clustering was discarded.
Because of the known limitations of the COBWEBalgorithm, this
was not taken into consideration. Thus, the EM algorithm was the
only technique used in this research for classifying maintenance
service providers according to their performance. There are many
commercial products that include a EM clustering algorithm tool
(CART/MARS/TreeNet/RF, SPSS Clementine, SPSS, KXEN,
SAS, E-Miner, WEKA, R, MATLAB, etc.). In this research,
WEKA software V. 3.6 was used [20]. WEKA is an open-source,
free software with excellent performance in its calculations (reli-
ability and speed) [21] (Table 2).
In this research, three types of data sources were used: two
secondary and one primary [22]. The secondary sources were
the medical device inventory and the records from the mainte-
nance work order databases in the clinical engineering depart-
ment. The primary source of information was a survey
conducted on external maintenance service providers.
Results
Characterization of Equipment Inventory
A characterization of the equipment inventory was completed
with a population of 264 pieces of equipment. The 264 pieces
of equipment were selected to assure that 100% of the sample
of service providers would be surveyed. This phase of the
study found the following results:
1) The equipment can be cataloged as old: the ET/UL ratio had
an average of 0.78 units. The areas that showed notably high
Table 1. Clusters obtained.
Clusters
Variable Type 0 1 2 3
Cluster attributes
Providers Code NA V, M, K, Z, T, Y O, l, C, B J I, S, H, F, G
Total number of service providers by cluster and % Numeric 6 (38%) 4 (25%) 1 (6%) 5 (31%)
Contract length (years) Numeric 1.11 0.92 5.00 1.68
Nequip Numeric 1.32 41.49 266.00 4.75
NvisitY Numeric 4.86 3.25 2.00 3.59
TotalHumResCompany Numeric 10.96 5.75 3.00 20.09
TotalHumResContract Numeric 3.68 3.75 3.00 3.00
TotalHumResEqCont Numeric 3.19 0.26 0.01 1.15
ExpPersonalYears Numeric 12.72 8.00 8.00 7.82
HospDist Numeric 11.02 3.32 0.00 10.70
TotNumContracts Numeric 96.17 10.00 0.00 94.03
ExpCompany Numeric 34.82 7.87 7.50 20.29
ServLine Nominal (Dummy)
Yes 6.00 4.00 0.00 5.09
No 1.00 1.00 0.00
Backup Nominal (Dummy)
Yes 0.00 4.00 0.00 5.00
No 5.00 0.00 1.00 0.00
RepP Nominal (Dummy)
Yes 6.00 5.00 0.00 4.09
No 0.00 01.00 1.00 1.00
UserTraining Nominal (Dummy)
No 0.00 4.00 0.00 1.00
Yes 6.00 1.00 2.00 4.00
TechTraining Nominal (Dummy)
Yes 2.00 2.00 2.00 4.00
No 4.00 3.00 0.00 1.00
Service performance Numeric
RT (hours) 32.98 49.72 43.33 54.46
ST (hours) 2.08 185.25 2,396.15 6.97
TAT (hours) 35.06 234.99 2,439.48 61.44
Availability (%) 87.00 98.00 98.00 93.00
Explanatory notes:
1. Nominal or ordinal variables become dummy variables. They can take on one of two values, 0 or 1, TRUE or FALSE, or,
YES or NO, to indicate the presence or absence of some categorical effect that may be expected to shift the outcome.
2. NA: not applicable.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
ET/UL ratios were neonatal care,
microbiology, and the surgical unit,
with 1.01, 0.92, and 1.19 units,
respectively.
2) Contracted services represented
97.57% of the total cost of mainte-
nance services. The ratio comparing
the cost of in-house maintenance to
the cost of contracted maintenance
services represents only 2.43% of
total maintenance costs.
3) The areas of diagnostic images
(BB), surgical unit (CC), and ICU
(DD) represent 90.30% of the total
cost of the equipment inventory in
the seven analyzed areas. Coinci-
dentally, they exhibit 87.72% of
the total cost of maintenance serv-
ices, while representing only 50%
of the total pieces of equipment
(see Table 3).
4) The average value of the SC/AC
ratio is 8.61% per year. It is impor-
tant to stress that notably high
results were found in ICU (DD),
microbiology (II), and the clinical
laboratory, with values of 27.8%,
23.81%, and 21.38% respectively.
In the authors opinion, the values
of SC/AC ratio obtained are unacceptable. For example,
Cohen in [6] reported and referenced values of SC/AC
ratio that run from 3.2% to 4.9%. In a more comprehen-
sive studies conducted as a survey of eight hospitals,
Cohen found that the average value of the SC/AC ratio is
3.90% [6]. In the same publication, it was reported that
the average value of the SC/AC ratio for clinical labora-
tory and ICU pilot areas are 5.20% and 4.5%, respectively
[7]. Finally, doing a benchmark between the hospital
coded as H in the Cohen survey and our hospital, we can
find that the total average value of the SC/AC ratio is
3.90%; and for clinical laboratory and ICU pilot areas are
2.80% and 6.1%, respectively. (Hospitals were codified
from letters AH.) Again, these are low values in compari-
son with our hospital.
Characterization of Maintenance Service Providers
A total of 26 maintenance service providers (coded with letters
from A to Z, 100% of the sample) were characterized. In-house
maintenance services were included as an additional provider.
This phase produced the following results:
1) Contracted equipment exhibits 46.69% of the total equip-
ment population. Of this 46.69%, 62.23% are held by
maintenance service providers.
2) The most common distance of a company to the hospital
is 10 km (mode of variable HospDist).
3) The most common length of a maintenance service con-
tract is one year (mode of variable contract length).
4) The companies can be cataloged as experienced and small,
with an average value of 25 years experience in the industry,
and a human resources total of three personnel, including
technicians and engineers, 15.38% of companies had no
equipment replacement stock held in the country, and 11.50%
did not offer online assistance. Companies designated three
specialists between technicians and engineers, as the mode
value, to service equipment under their maintenance contracts.
Clustering Study of Maintenance Service Providers
Figure 2 shows the ordering of maintenance service providers
according to the TAT variable (in days). Notice how the TAT
values are divided into three large groups: High values (TAT >
6 days), including impermissible values in the order of 101 days
(more than three months) (providers O, l, C, B, and J); medium
values (1 < TAT < 6) (less than one week) (providers I, S, H,
F, and G); and low values (TAT < 1) (less than one day), which
is an extremely desirable TAT indicator (providers V, M, K, Z,
T, and Y).
Table 2. Pseudocode of the EM algorithm.
1. Stop_cond e //stop condition is an entry parameter, (e.g., 10
4
)
2. M Num-Cluster // number of clusters (M) is an entry parameter
3. For I 0 to (M-1) //loop in which each belongings probability, r, and l
of each cluster are initialized to zero. In C-code arrays, index starts in 0
index
r[I] 0
l[I] 0
Prob[I] 0
I = I 1
End for
4. J 1 //control variable for each iteration
5. Calculation of MF[J] //first calculation with belongings probability = r
= l = 0; see (3)
6. Do
For I 1 To (M-1) //real calculation of r, l, Prob using x
1
. . .x
n
of each cluster;
see (1) and (2)
Calculation of r [I]
Calculation of l [I]
Calculation of Prob[I]
EndFor
J J 1
Calculation of MF[J] // Calculation of next MF
While (MF[J]-MF[J-1]) > Stop_cond // loop is repeated until stop condi-
tion is reached (e.g., 10
4
); see (4)
120.00
1
0
1
.
6
5
2
8
.
4
6
8
.
0
3
7
.
6
2
6
.
0
3
4
.
3
8
1
.
6
3
1
.
4
6
1
.
2
1
1
.
1
1
0
.
2
9
0
.
2
6
0
.
1
4
0
.
0
9
0
.
0
9
0
.
0
4
100.00
80.00
60.00
40.00
20.00
0.00
J O V L H I B C G S
Service Provider Codes
F Y K M Z T
T
A
T

(
d
a
y
s
)
Fig. 2. Analysis of the TAT variable by service provider. Note 1:
Letters in the figure are maintenance service providers with
preventive or corrective works orders in the analyzed period.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
A comprehensive analysis of maintenance service
provider characteristics can be completed through a grouping
study. Such a study allows for the grouping of providers by
identification of their fundamental characteristics. Four clus-
ters or groups were identified, classifying service providers
as follows:
1) Group 0: Best performance; the service providers are V, M,
K, Z, T, Y; they represent 38% of the total; with an average
TAT value of 1.46 days.
2) Groups 1 and 2: Poorest performance, with the highest
TAT values. In this group, we find the service provider,
coded as J, with an extreme TAT value of 101 days; and
service providers coded as O, l, C, B with TAT an aver-
age value of 9.79 days.
3) Group 3: Medium quality performance, intermediate val-
ues of TAT, RT, and ST, and an average TAT value of
2.56 days; the service providers are I, S, H, F, G; they rep-
resent 31% of maintenance service providers (Figure 3).
Discussion
From a managerial point of view, the significance of the clus-
ters is clear. As mentioned before, by classifying maintenance
service providers, managers can make decisions on their rela-
tionships with these groups. For example, with Group 0, it is
Table 3. Inventory characterization.
Area Code
Total Number
of Equipment ET/UL
Acquisition Cost
Penetration (%)
In-House
Maintenance
Cost Versus
Maintenance
Contract Cost
(%)
Total
Maintenance
Cost
Maintenance
Cost
Penetration (%)
SC/AC
(%)
(a) Areas
BB 22 0.68 63.60 0.11 496,585,745 49.47 6.7
CC 75 0.92 20.10 1.13 168,744,218 16.81 7.2
DD 35 0.57 6.60 0.29 215,158,674 21.44 27.8
FF 56 1.01 5.50 0.31 78,256,440 7.80 12.3
AA 18 0.72 1.50 0.17 8,747,420 0.87 4.9
EE 18 0.75 1.10 0.39 16,445,052 1.64 12.6
HH 28 0.50 1.20 0.01 9,146,115 0.91 6.7
GG 10 0.69 0.30 0.02 6,833,914 0.68 21.4
II 2 1.19 0.10 0.00 3,804,548 0.38 23.8
Totals 264 Ave: 0.78 100.00 2.43 1,003,722,126 100.00 Ave: 8.6
Equipment
Type Code
Total Number
of Equipment ET/UL
Acquisition Cost
Penetration (%)
In-House
Maintenance
Cost Versus
Maintenance
Contract Cost
(%)
Total
Maintenance
Cost
Maintenance
Cost
Penetration (%)
SC/AC
(%)
(b) Equipment Types
A 31 0.83 66.20 0.23 528,717,788 52.68 6.8
B 76 0.94 20.60 0.59 168,446,728 16.78 7.0
C 65 0.58 8.80 1.44 54,217,090 5.40 5.3
D 52 1.00 2.90 0.13 232,555,943 23.17 69.4
E 40 0.58 1.60 0.04 19,784,577 1.97 10.7
Total 264 Ave: 0.79 100.00 2.43 1,003,722,126 100.00 Ave: 8.6
Explanatory notes for Table 3(a) and (b).
1. The ET-to-UL ratio ( ET/UL) is an indicator of medical device obsolescence. An ET/UL > 1 would suggest removing the
equipment from inventory.
2. Acquisition cost penetration represents the percentage of acquisition cost of a category in comparison with the total
acquisition cost for all categories.
3. In-house maintenance cost (including labor, parts, and components) versus maintenance contract cost is the ratio
comparing cost of in-house maintenance to the cost of contracted maintenance services.
4. Total maintenance cost, which adds the total of contracted and in-house maintenance service costs (including labour,
parts, and components).
5. Maintenance cost penetration represents the percentage of maintenance cost of a category in comparison with the
total maintenance cost for all categories.
6. The service cost (SC) to acquisition cost ratio (AC) (SC/AC) is used many times as a standardized indicator that
measures how effectively outside services are managed, as well as measuring the cost effectiveness of in-house mainte-
nance service.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
recommendable to maintain the contractual
relationship, while with Groups 1 and 2, it is
more advisable to discontinue the arrange-
ment. Finally, a manager may decide to strictly
monitor existing services with Group 3.
The availability of equipment was studied as
a variable by monitoring the quality of mainte-
nance service providers (availability: understood
as the probability that a piece of equipment is
operational when service is needed. It can be cal-
culated as the ratio of Hope/(Hope-Hnope) [25],
where Hope is the total hours of operation or
functioning period of the equipment, Hnope is
the total hours of nonfunctioning or nonopera-
tion period of the equipment). Excellent values
for this variable were obtained (more than 90%
for three of the four service provider clusters).
However, this value owes to the high number of
hours the pieces of equipment are in operation to
complete their functional purpose within health-
care services (reaching 10,240.05 h over the ana-
lyzed period). Thus, it can be said that the defi-
cient TAT values are masked by the desirable
findings for equipment availability.
Unfortunately, service provider J is the worst maintenance
service provider (in-house maintenance service). The reasons
influencing these results are related to the high level of obsoles-
cence of equipment, including a high number of equipment fail-
ures and time-consuming maintenance tasks. New strategies
must be developed to improve the quality of in-house mainte-
nance service, such as increasing the number of full-time tech-
nicians hired by the clinical engineering department, improving
the level of training for users and technicians, and applying
adequate techniques for the prioritization of medical devices.
Where Are We Going Next?
1) First, monitoring should be continued to incorporate into
the study those maintenance service providers that have
not reported service orders, either because equipment
under their charge has not failed or because maintenance
programs did not exist during the analyzed period.
2) Second, increase the hospitals data sample to obtain a
greater number of service providers. It will allow research-
ers to conduct benchmarking and sophisticated inferential
studies (ANOVA, one and multiple way) to identify differ-
ences in the quality of maintenance service providers by
health-care institution.
3) Third, monitoring of the TAT variable should continue,
through observation of work orders, to apply a predictive
technique for obtaining TAT values using the characteris-
tics of maintenance service providers and the RT and ST
variables. That is, to predict or calculate TAT as follows:
TAT = function(ST, RT, NEquip, Guaranty, REParts,
NvisitYear, EquipTypes, TotalHumResCompany, Total-
HumResContract, ExpPersonalYear, HospDist; ExpCom-
pany, ServLine, BacUp, UserTraining, TechTraining).
4) Fourth, after obtaining the TAT predictor, exploiting its
potential as an online service tool for the calculation of
TAT, through the implementation of the algorithm in an
application service provider platform (ASP). (The business
model first appeared in the late 1990s and it has generally
been said that there is a good amount of consensus that
ASP means remotely hosted applications management and
onside hosting [26].) This proposed centralized system
architecture has shown promise in reducing network infra-
structure labor and equipment costs, benchmarking of
equipment performance indicators, and the development of
avenues for proper and timely problem reporting. Using
this new service online, users can connect and calculate the
predicted TAT as a function of the data the user enters.
Finally, this research, in it first views could see to be of mildly
application. However, authors of this article believe that this
TAT in hours
2439.48
1220.24
Cluster 3
Cluster 2
Cluster 1
Cluster 0
1
Fig. 3. Groups or clusters identified by service provider, with the TAT variable
in hours. Note 1: The values by cluster are given as an average. Note 2: Clus-
tering was obtained using WEKA software V. 3.6 [20].
Data
Information
Managers
Actions
Maintain Services
Strictly Monitor Services
Stop Services
Service Providers Cluster
BP
MP
PP
Data Mining
(Clustering
Technique)
Output
Making Decision
1) Service Providers [Best Performance (BP)]
2) Service Providers [Medium Performance (MP)]
3) Service Providers [Poor Performance (PP)]
1) Equipment Hospital Inventory
2) Maintenance Service Providers
Input
Fig. 4. The added value of this research.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
research has an added value, that is: to become the data recov-
ered from maintenance service providers into information to
make managerial decision. This latest idea is the most impor-
tant contribution that this article makes to the clinical engi-
neering community (see Figure 4).
Conclusions
1) The service providers with the poorest performance are
those which have the most equipment under their charge
and which the equipment is of highest technological
complexity (diagnostic equipment).
2) Data mining constitutes an effective means for evaluating
the performance of maintenance service providers.
Acknowledgments
We thank COLCIENCIAS for the resources provided in Grant
Announcement 459/08, which financed the study. Thanks
also to University Hospital La Samaritana for its support and
collaboration in completing the study. Finally, a special
thanks to Gregory L. Haugan for his assistance and collabora-
tion in the translation and review of this article.
Antonio Miguel Cruz holds a nuclear
engineering degree and an M.Sc.degree in
bioengineering, both from the University of
Nuclear Sciences and Jose Antonio Eche-
verr a (ISPJAE) Technical University,
Havana, Cuba, in 1995 and 1997, respec-
tively. He also holds a Ph.D. degree in bio-
engineering from (ISPJAE) Technical
University, Havana, Cuba, in 2003. He is a senior lecturer and
course tutor in clinical engineering, object-oriented program-
ming, medical informatics, and advanced maintenance sub-
jects. He is currently a professor of biophysics at Rosario
University Medicine School. His primary research interests are
in health-care knowledge management, clinical engineering,
data mining, fuzzy logic, intelligent computational methods for
planning, and medical informatics. He has published more than
25 technical papers. He is a Member of the IEEE and Ad-
vanced Association for Medical Instrumentation (AAMI).
Sandra Patricia Usaquen Perilla grad-
uated with a masters degree in biomedical
engineering at Rio de Janeiro Federal
University, Brazil, and obtained the title of
biomedical engineering. She was a benefi-
ciary of Brazilian government scholarship,
program student-commitment postgradua-
tion (PEC-PG). The research was devel-
oped about clinical engineering, the main approaching topics
were the performance evaluation of mechanicals ventilators
and technovigilance. She has participated as a speaker at a
lecture on the biomedical engineering Brazilian Conference
in 2006 and 2008; likewise, in 2007, she participated in the
Intensive Therapy Conference at Rio de Janeiro State. She
presented in Colombia Bioengineering Conference on 2008
and Argentinean Bioengineering Conference on 2009. She is
a current researcher at Rosario University Medicine School.
Nidia Nelly Vanegas Pabon became a biomedical engineer
from Manuela Beltran University in December 2002, a special-
ist in Institutions Providing Health Services Management from
Rosario University in March 2005, with studies in (SOGCS)
Obligatory System for Quality Assurance in Health and (SUA)
Unique Accreditation System. He has experience in biomedical
technology assessment, project design for quality improvement
in health, strategical planning and project development for
technology management and clinical engineering, performance
in positions as coordinating quality assurance, and as a plan-
ning manager of Une Cundinamarca. He is currently coordina-
tor of medical equipment in the University Hospital of the
Samaritana (HUS), a member of the research group of clinical
engineering for HUS.
Address for Correspondence: Antonio Miguel Cruz, Facul-
tad de Medicina, Dpto. de Ciencias Basicas Universidad del
Rosario, Calle 63D # 2431, 7 de Agosto, Bogota DC, Colom-
bia. E-mail: antonio.cruz43@urosario.edu.co.
References
[1] (2009, June). Ministerio de la Proteccion Social. Modelo de Evaluacion y Gestion
del Equipamiento Biomedico [Online] Available http://www.minproteccionsocial.gov.
co/VBeContent/Library/documents/DocNewsNo14710DocumentNo5095.pdf
[2] G. Duarte, P. y Sackett, and S. Evans, Migration and outsourcing, IEEE.
Trans Manufact. Engineer., vol. 3, no. 4, pp. 4448, 2004.
[3] D. Jensen, Contract worker services: An inside view, Aviation Manage.,
vol. 192, no. 1, pp. 138150, July 1, 2006.
[4] J. Tieman, In need of repairs, Mod. Healthcare, vol. 32, no. 2, pp. 3032, 2002.
[5] ECRI, Types of services: Their advantages and disadvantages, Health.
Technol., vol. 3, no. 4, pp. 920, 1889.
[6] T. Cohen, Benchmark indicators for medical equipment repair and mainte-
nance, Biomed. Instrum. Technol., vol. 29, no. 4, pp. 308320, 1995.
[7] T. Cohen, Validating medical equipment repairs and maintenance metrics:
Parts II, Biomed. Instrum. Technol., vol. 32, no. 2, pp. 136144, 1998.
[8] K. Segupta and M. Zviran, Measuring user satisfaction in an outsourcing
environment, IEEE. Trans. Eng. Manage., vol. 44, no. 4, pp. 414421, 1997.
[9] ECRI, In house servicing of X-ray and CT equipment, Health. Technol.
Manage., vol. 21, no. 6, pp. 231246. Winter, 1993.
[10] L. S. Grimes, The future of clinical engineering: The challenge of change,
IEEE Eng. Med. Biol. Mag., vol. 22, no. 2, pp. 9199, 2003.
[11] E. Antunez, M. Do Vale, P. Mordelet, and V. Y Grabois, Gestao da Tecno-
logia Biomedica, Tecnovigilancia e Engenharia Cl nica, Ediciones cient ficas
ACODESS. France, p. 210, 2002.
[12] D. N. P. Murthy and E. Asgharizadeh, Optimal decision making in a
maintenance service operation, Eur. J. Operation. Res., vol. 116, no. 2, pp. 259
273, 1999.
[13] E. Asgharizadeh and D. N. P. Murthy, Service contracts: A stochastic
model, Math. Comput. Model., vol. 31, no. 3, pp. 1120, 2000.
[14] E. L. Plambeck and S. A. Zenios, Performance-based incentives in a
dynamic principal-agent model, Manufact. Service Oper. Manage., vol. 2, no. 3,
pp. 240263, 2000.
[15] H. Tarakci, K. Tang, H. Moskowitz, and R. Plante, Incentive maintenance
outsourcing contracts for channel coordination and improvement, IIE Trans.,
vol. 38, no. 2, pp. 671684, 2006.
[16] H. Tarakci, K. Tang, H. Moskowitz, and R. Plante, Maintenance outsourc-
ing of a multi-process manufacturing system with multiple contractors, IIE
Trans., vol. 38, no. 1, pp. 6778, 2006.
[17] Two Crows Corporation, Introduction to Data Mining and Knowledge Dis-
covery, 3rd ed. MD, 1999.
[18] H. W. Ian and F. Eibe, Data Mining: Practical Machine Learning Tools and
Techniques with Java Implementations. San Mateo, CA: Morgan Kaufmann Pub-
lishers, 2000.
[19] S. Borman. (2004, July 18). The expectation maximization algorithm: A
short tutorial [Online]. Available http://www.isi.edu/natural-language/teaching/
cs562/2009/readings/B06.pdf
[20] (2009, June). WEKA v 3.6. [Online]. Available: http://www.cs.waikato.ac.
nz/ml/weka/index_downloading.html
[21] Available: http://www.kdnuggets.com/polls/2009/data-mining-tools-used.htm
[22] R. H. Sampieri, C. Fernandez-Collado, and P. Baptista, Metodologa de la
investigaci on, 4th ed. Mexico, D.F.: McGraw Hill, 2006, p. 102.
[23] INVIMA. (2009, June). Decreto 4725 de 2005 [Online]. Available http://
www.minproteccionsocial.gov.co/VBeContent/library/documents/DocNewsNo15472
DocumentNo2802.PDF
[24] G. J. Bostwick and N. S. Kyte, Measurement, in Social Work: Research
and Evaluation. Quantitative and Qualitative Approaches, R. Grinnell, Jr., Ed.,
7th ed. NY: Oxford Univ. Press, 2005, pp. 97101.
[25] C. G. Mosquera, Availability and Reliabiluty of Industrial Systems, 3rd ed.
Valencia, Venezuela: Ediciones Universitarias UGMA, 1985, pp. 5260.
[26] S. Schuster, Defining the PACS ASP model, Med. Imag., vol. 37, no. 3,
pp. 110117, Nov. 2002.
___________________________________
_____________________
___________
______________
__________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
___
EYEWIRE
Clinical Appointment
Process
M
ost clinics rely on the appointment scheduling sys-
tems to manage patients access to their health-care
providers. Conventional appointment scheduling
processes have intrinsic inefficiency because of the
tendency to generate fragmented time slots. In this article, a
solution, which considers patientprovider mutual preference,
is provided to guide the appointment scheduling process by
means of schedule defragmentation. Computer simulation
shows that patientprovider cooperation can effectively reduce
schedule fragmentation, yielding higher appointment accep-
tance rate and clinic time utilization rate at
given appointment demand matched by serv-
ice supply. When service time distribution can
be accurately estimated, decreasing unit time
slot size may further improve the appointment
scheduling efficiency. In addition, a clinical
survey was conducted to identify the opportu-
nities and challenges of applying the proposed
defragmentation method in clinical appoint-
ment scheduling practice.
Health-Care Reform
The ongoing national debate on health-care
reform has become the focus of daily coverage
in the media. Despite the highest spending in
health care, the United States ranks lower than
other countries on most dimensions of
performance [1], [2]. To reign in costs while
improving health outcomes or to move toward
a value-based health-care system [3], higher
efficiency and timely access to health services
are necessary. Clinical appointment scheduling
lies at the intersection of these two demands.
Since the seminal work by Bailey in the early
1950s [4], a number of studies have evaluated the
appointment scheduling process using engineer-
ing tools such as queuing models [5], [6],
computer simulations [7], [8], and optimization
theory [9]. These studies mainly focus on two
objectives: 1) to maximize the physicians time
utilization and 2) to minimize the patients waitingtime. The former
is physician-centered and impacts the clinics financials, whereas
the latter is patient-centered and affects the experience of care.
Quite often, difficult managerial decisions have to be made by con-
sidering various tradeoffs between these goals [10]. Yet one topic
that has received little attention is how to increase the acceptance
rate for appointment requests, which has the potential to improve
both physicians time utilization and patients access to care.
Typically, a physicians schedule is divided into a finite num-
ber of nonoverlapping time slots, where each time slot has fixed
duration, which is the shortest time interval that
can be allocated. On receiving an appointment
request, the scheduler first estimates the num-
ber of consecutive time slots (K) needed for
this appointment and then searches the physi-
cians schedule for all options that contain at
least K consecutive open time slots. If there are
no available time slots, then the appointment
cannot be made. Otherwise, the scheduler pro-
vides the patient a list of available time slots. If
the patient has no time preference, then usually
the first matching or arbitrarily chosen time
slots are allocated to the patient. If the patients
preferred time slots are available, then they are
allocated to the patient. If the patient preferred
time is not on the list, then the patient has to
accept the nonpreferred time slots or the
appointment cannot be made.
The appointment scheduling practice de-
scribed earlier has an intrinsic limitation in
efficiency because of the tendency of schedule
fragmentation, which can be caused by the
allocation of patient-preferred time slots and/
or the cancelation of previously allocated time
slots. As a result, the fragmented schedule
may become difficult to accommodate new
appointment requests and compromise pro-
viders time utilization. For example, an ap-
pointment requesting two consecutive time
slots may be rejected, despite the presence of
two separated open time slots, which will be
wasted if left unallocated (Figure 1).
BY JIE LIAN, KORI DISTEFANO, SOMER D. SHIELDS,
CINDY HEINICHEN, MELISSA GIAMPIETRI,
AND LIAN WANG
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Improvement Through Schedule
Defragmentation
Busy
Busy
Open
Open
Open
Busy
Busy
Busy
Open
Busy
Busy
Busy
Busy
Busy
Busy
Open
Open
Busy
Busy
Open
Busy
Open
Busy
Open
Busy
Busy
Open
Busy
Busy
Busy
Busy
Busy
Busy
Busy
Busy
Busy
Open
Open
Open
Open
(a) (b)
Fig. 1. Schematic illustration of
(a) a more-fragmented sched-
ule, which has seven sepa-
rated open time slots, and (b)
a less-fragmented schedule,
which also has seven open
time slots that are organized
into three open blocks.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Two issues have been overlooked in the conventional
appointment scheduling process. First, there is no formal guid-
ance on how to allocate the time slots to minimize the frag-
mentation. Although, in practice, a scheduler may intuitively
encourage patients to pick certain time slots (e.g., by disclos-
ing some available time slots to but withholding others from a
patient), there is no quantitative assessment of each possible
time slot allocation option and its potential impact on the
schedule fragmentation status. Second, customers preference
is usually not binary. Instead, it is a continuous variable that
spans the spectrum from highly like to highly dislike, with
many intermediate preference levels. Moreover, a patients
decision to select time slots may be affected by external fac-
tors other than his/her preference level. For example, because
of the established patientphysician relationship, a patient
who has a slight preference of a morning appointment may
instead gladly accept (and may even prefer) an afternoon
appointment, after being informed that choice may help clinic
improve efficiency and reduce waiting time.
In this article, a method that considers patientprovider
mutual preference is proposed to guide the appointment schedul-
ing process by means of schedule defragmentation. A computer
model is developed to assess the impact of schedule defragmen-
tation on the appointment acceptance rate and physicians time
utilization rate. The effects of changing time slot size and vary-
ing statistical profiles of the appointment requests on the
appointment scheduling efficiency are also investigated. In addi-
tion, a clinic survey was conducted to identify potential opportu-
nities and challenges of applying the schedule defragmentation
method to real-world appointment scheduling practice.
Methods
PatientProvider Mutual Preference
For patient-centered care, the appointment scheduling process
must strive to satisfy patients preference. On the other hand, a
less-fragmented schedule with large blocks of contiguous
open slots is clearly preferred by the provider. Therefore, to
improve both patient satisfaction and clinical efficiency, the
appointment scheduling needs to adopt a new process model
based on patientprovider mutual preference.
According to this new process model, for each patient
appointment request, the list of all available time slots (if any)
is ranked from the most provider preferred to the least provider
preferred. Options ranked higher on the list would cause less
schedule fragmentation than those ranked lower on the list.
Patient is encouraged but not obliged to choose the provider-
preferred slots. Hence, the ranked list provides the guidance on
allocating time slots to minimize the schedule fragmentation,
while both customer and providers preferences can be taken
into consideration in appointment scheduling. The provider
customer mutual preference could be established through nego-
tiation, incentives, or other means. For example, the options of
available time slots can be presented to the patient in a sequen-
tial order so that an option ranked higher on the list is presented
first, and an option ranked lower on the list is presented only if
the higher ranked options are rejected by the patient.
Schedule Defragmentation
Figure 2 shows a flowchart illustrating the improved appoint-
ment scheduling process. On receiving an appointment
request from the patient, an estimate is made on the number of
time slots needed. Then the providers schedule is searched to
find all options of time slots that can serve the patients
request. In the absence of available time slots, the appointment
request is denied. Otherwise, for each option of available time
slots, a mock allocation is performed, and the respective frag-
mentation index is calculated as the number of resulting open
blocks, where the open block is defined as one isolated open
time slot or multiple adjacent open time slots [e.g., the frag-
mentation indexes are seven and three, respectively, for the
schedules shown in Figure 1(a) and (b)]. Then all options are
sorted based on the calculated fragmentation indexes, where
the top option has the smallest fragmentation index, and the
bottom option has the largest fragmentation index. If there are
options with equal fragmentation index, then option with early
time is ranked higher (i.e., prefer keeping open slots at the
later time) so that the later appointment requests could be bet-
ter served. Finally, the ranked list of available time slots is pro-
vided to the patient who is encouraged to choose an option as
high as possible on the list.
Computer Modeling and Simulation Protocols
A computer model is developed to assess the efficiency of the
proposed appointment scheduling process and investigate the
effects of time slot size, number of appointment requests, and
the distribution of service time.
Two metrics are used for measuring efficiency: 1) accep-
tance rate, which is the ratio between the number of accepted
appointments and the total number of appointment requests;
2) utilization rate, which is the providers actual service time
divided by the total work time. In this study, both metrics are
measured on a daily basis. The total number of appointment
Receive
Service Request
Determine the Number
of Options of Available
Time Slots
Select One Option of
Available Time Slots
Mock Allocation of
Selected Option of
Available Time Slots
Calculate Fragmentation
Index by Counting
Open Blocks
All
Options Examined?
Sort All Options
According to
Fragmentation Indexes
Provide Patient the
Ranked List of Available
Time Slots
Determine Needed
Number of Time Slots
Search Providers
Schedule
Find
Slots That Can
Serve the
Request?
Service Request
Is Denied
Y
Y
N N
Fig. 2. Flowchart diagram of the new appointment schedul-
ing process based on schedule defragmentation.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
requests per day varied from 11 to 20 (step 1). The providers
daily schedule consists of four morning work hours and four
afternoon work hours, separated by 1 h break (i.e., total work
time 8 h/day, see Figure 3).
The skew normal distribution is used to model the needed
service time of the appointment requests, and its probability
density function (PDF) is defined by [11]:
f (x) =
1
rp
e
(xl)
2
2r
2
Z
a
xl
r
t
2
2
dt, (1)
where l is the location parameter that controls the shift of the
distribution, r is the scale parameter that determines the spread
out of the distribution, and a is the shape parameter that
measures the asymmetry of the distribution: it becomes normal
distribution when a = 0, and the distribution is right skewed if
a > 0 and is left skewed if a < 0 [11]. In this study, l is varied
between 10 and 50 min (step 5 min), r is varied from 1 to
10 min (step 1 min), and a is varied from3 to 3 (step 1).
Three different time slot sizes, 5, 10, and 15 min, are simu-
lated for appointment scheduling. Appointment requests with
service time characterized by (1) need to be converted to multi-
ples of time slot. For example, when time slot size is 15 min, all
services that need _15 min will require one time slot (15 min),
all services that need 1630 min will require two time slots (30
min), and so on. Therefore, for each time slot size and a chosen
service time distribution, the probability of appointment
requests that need n time slots (n = 1, 2, 3, . . .) can be calculated
by integrating the PDF and shown in a histogram. In the follow-
ing simulations, the shortest appointment time is one time slot,
and the longest appointment time is limited to 60 min (i.e., serv-
ice time > 60 min in PDF is integrated to 60-min appointment
0.1
0.05
P
D
F
P
r
o
b
a
b
i
l
i
t
y
0
0.8
0.6
0.4
0.2
0
40
35
S
e
r
v
i
c
e

T
i
m
e

(
m
i
n
)
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9
Work Hour
Co-Op Ratio = 1.0
40
35
S
e
r
v
i
c
e

T
i
m
e

(
m
i
n
)
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9
Work Hour
Co-Op Ratio = 100
40
35
S
e
r
v
i
c
e

T
i
m
e

(
m
i
n
)
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9
Work Hour
Co-Op Ratio = 2.0
40
35
S
e
r
v
i
c
e

T
i
m
e

(
m
i
n
)
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9
Work Hour
Co-Op Ratio = 1.5
20 0 20 40 60
Service Time (min)
15 30 45 60
Appointment Time (min)
(b)
(f) (e) (d)
(a)
80
40
35
S
e
r
v
i
c
e

T
i
m
e

(
m
i
n
)
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9
Work Hour
Co-Op Ratio = 1.25
(c)
Fig. 3. Exemplary illustration of schedule defragmentation by promoting patientprovider mutual preference. The providers
daily schedule consists of eight working hours (04 and 59 h) with 1 h break (hour 45). Time slot size is set to 15 min. Subplot
(a) shows the probability density function (PDF) of the needed service time and the histogram of appointment probability.
Subplots (b)(f), respectively, show the resulting schedule maps after ten appointment requests (six need one time slot each,
and four need two time slots each) corresponding to five different co-op ratios.
Acomputer model is developed to assess the
impact of schedule defragmentation on the
appointment acceptance rate and physicians
time utilization rate.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
time, which is divided into 12, six, and four time slots, respec-
tively, for the three time slot sizes).
A cooperation (co-op) ratio is introduced to simulate the
degree of patientprovider mutual preference. Assume OA
and OB are two adjacent options of available time slots in the
ranked list and OA ranks higher than OB (i.e., OA is more pre-
ferred by the provider), the co-op ratio is defined as the proba-
bility of patient choosing OA over the probability of patient
choosing OB. Hence, the higher the co-op ratio, the more
likely patient will choose the providers preferred option. On
the other hand, a co-op ratio 1.0 implies that the patient
ignores providers preference and will choose either option
with equal likelihood. For simplicity purpose, it is assumed
that the same co-op ratio applies to all pairs of adjacent
options in a ranked list. In this study, seven co-op ratios are
simulated: 1.0, 1.25, 1.5, 2.0, 2.5, 5.0, and 100.
By varying the number of total appointment requests, serv-
ice time distribution (l, r, and a), time slot size, and co-op
ratio, a total of 10 39 310 37 33 37 = 132,300 configura-
tions are modeled. Each configuration is simulated by 100 ran-
dom runs, and the average utilization rate and average
acceptance rate are reported.
Clinical Survey
To identify the opportunities and challenges of applying the
proposed schedule defragmentation method in real-world clini-
cal appointment practice, a survey was conducted in four clinics
of various types located in Portland, Oregon: one obstetrics
gynecology clinic, one pediatric clinic, one primary care clinic,
and one mental health clinic. The survey includes an onsite
interviewwith each clinics operation manager on the following
focus topics: 1) What is the current appointment scheduling
process? 2) What are the major problems of the current schedul-
ing practice? 3) Will schedule defragmentation be helpful to
improve the operational efficiency? 4) What are the perceived
challenges to implement schedule defragmentation in the
clinic? The survey results are consolidated and communicated
to the interviewees for further feedback, and the final consensus
opinions are presented.
Results
Simulation Results
Figure 3 shows a representative example of schedule defrag-
mentation based on patientprovider mutual preference. In
this example, time slot size is set to 15 min. The PDF of
the needed service time is characterized by l = 15 min,
r =5 min, and a =1 and is shown in the top panel of Figure
3(a). By integrating the PDF, the histogram of appointment
probability is obtained and shown in the bottom panel of Fig-
ure 3(a), where about three fourth of the appointments need
15 min (one time slot), and about one quarter of the appoint-
ment requests need 30 min (two time slots).
Corresponding to five different co-op ratios, the resulting
schedule maps after ten appointment requests are, respectively,
shown in Figure 3(b)(f), and the fragmentation index (i.e., the
number of open blocks) is 10, 6, 5, 3, and 2, respectively. Clearly,
the higher the co-op ratio, the less fragmentation of the schedule
map. Note that slight increase of the co-op ratio from 1.0 to 2.0
causes sharp decrease of the fragmentation index from ten to
three, while further increase of the co-op ratio up to 100 yields
only small additional gain in defragmentation. Therefore, com-
pared with conventional appointment scheduling practice (co-op
ratio 1.0), slight consideration of the patientprovider mutual
preference can significantly reduce schedule fragmentation.
In Figures 411, (a) shows the utilization rate when there
are different number of appointment requests (N) and at vari-
ous co-op ratios and (b) shows the acceptance rate correspond-
ing to different N and co-op ratios.
Figures 49 show the effects of varying service time distri-
bution (through permutation of three PDF parameters l, r,
and a) on appointment scheduling efficiency, with time slot
size set to 10 min. Two general trends are evident across all
six conditions: 1) more appointment requests results in higher
utilization rate and lower acceptance rate at given co-op ratio
(as expected) and 2) increasing co-op ratio leads to both higher
1
0.95
Utilization Rate Acceptance Rate
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
1
0.9
0.8
0.7
0.6
0.5 N = 20
N = 18
N = 16
N = 14
N = 12
Fig. 4. (a) The utilization rate and (b) the acceptance rate
for different numbers of appointment requests and at vari-
ous co-op ratios. Time slot size is 10 min, and PDF parameters
are l = 30 min, r = 10 min, and a = 0.
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
1
0.9
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
utilization rate and higher acceptance rate at given number of
appointment requests.
With fixed location parameter l and shape parameter a, de-
creasing scale parameter r (i.e., less spread distribution) can slight-
ly increase acceptance rate when there are fewer appointment
requests, but tends to decrease acceptance rate with lower utili-
zation rate when Nincreases (Figure 4 versus Figure 5).
On the other hand, with fixed scale parameter r and shape
parameter a, increasing location parameter l (i.e., longer
appointments) leads to significant drop of acceptance rate,
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
Aless-fragmented schedule with large blocks
of contiguous open slots is clearly preferred
by the provider.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
irrespective of increase (for smaller N) or decrease (for larger
N) of the utilization rate (see Figures 4, 6, and 7).
Furthermore, with fixed location parameter l and scale
parameter r, left-skewed distribution (a < 0, i.e., more shorter
appointments) increases, whereas right-skewed distribution
(a > 0, i.e., more longer appointments) decreases the accep-
tance rate. The increase of acceptance rate is associated with
either decrease of the utilization rate for smaller N or increase
of the utilization rate for larger N (see Figures 4, 8, and 9).
Finally, the effects of time slot size on appointment sched-
uling efficiency can be appreciated by comparing Figure 4
(10 min time slot) with Figure 10 (5 min time slot) and Fig-
ure 11 (15 min time slot). Given the same service time distri-
bution (PDF), decreasing (versus increasing) time slot size
leads to finer (versus coarser) histogram of appointment
probability. Consequently, smaller (versus larger) time slot
size yields higher (versus lower) acceptance rate, despite the
decrease (versus increase) of the utilization rate.
Survey Results
Survey results show that the appointment scheduling practice
varies markedly from clinic to clinic. For example, while most
appointments are made several months in advance at obstet-
ricsgynecology and mental health clinics, same-day appoint-
ments are quite common for pediatric and primary-care
clinics. Different clinics also differ in appointment times,
which are typically set to 10/20/30/40 min in the obstetrics
gynecology clinic, 5/10/15/20/30 min in the pediatric clinic,
15/20 min for the primary care clinic, and 30/90 min for the
mental health clinic, respectively.
One common problem faced by all clinics is no shows or
late arrivals by the patient, which is beyond the scope of this
study but addressed elsewhere [12]. The problem of schedule
fragmentation was also acknowledged by all interviewees.
However, it is less obvious in the obstetricsgynecology clinic
where physicians schedules are almost always full because of
high demand or in the mental health clinic where most
appointments need only one time slot (30 min). On the other
hand, the schedule fragmentation issue is more pronounced in
the pediatric clinic where the same-day appointments of vari-
ous durations are common. Interestingly, the pediatric clinic
has recently experimented to reduce the time slot size from
10 to 5 min. Despite the absence of quantitative analysis, anec-
dotal evidence has indicated improved operational efficiency.
All interviewees agree the schedule defragmentation will be
an attractive add-on feature for the existing appointment sched-
uling software. The method will be mostly helpful for clinics
that accept same-day appointments and/or online booking and
have a distribution of different appointment times. The per-
ceived challenges of implementing this new method include,
on technical side, to integrate the schedule defragmentation
module into various commercial scheduling software from dif-
ferent vendors, whereas on clinical side, to demonstrate its
operational efficacy (the gains in utilization rate and/or accep-
tance rate) in a controlled longitudinal comparison study,
which requires commitment from the clinics management
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
1 1.25 1.5
Co-Op Ratio
2 2.5 5 1 1.25 1.5
Co-Op Ratio
(a) (b)
2 2.5 5
0.9
1
0.8
0.7
0.6
0.5
N = 20
N = 18
N = 16
N = 14
N = 12
N = 20 N
N = 18 N
N = 16 N
N = 14 N
N = 12 N
Conventional appointment scheduling
processes have intrinsic inefficiency because
of the tendency to generate
fragmented time slots.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
team to budget and allocate resources for study planning, mon-
itoring, data collection, and reviewprocess.
Discussion
Defragmentation has been an important tool in various engineer-
ing disciplines to effectively manage limited resource (e.g.,
computer disk, device memory, and communication channels)
so as to optimize system performance. However, this engineer-
ing concept has not been applied to clinical appointment sched-
uling, a task which in essence is to optimally manage the limited
clinical resource (physicians schedule) to better serve patients
appointment requests.
In this article, a new appointment scheduling process is pro-
posed that takes into account both patients and providers
preferences. The patientprovider mutual preference could be
established through negotiation, incentive, or other means,
and provides the guidance on allocating time slots to reduce
schedule fragmentation. Consequently, the reserved time slots
are aggregated, leaving more open time slots bundled
together, thus allowing easier accommodation of new appoint-
ment requests (see Figure 3).
As illustrated in Figures 411, simulation results showed
that, in principle, with fixed co-op ratio, greater service demand
(more appointment requests and/or longer service time)
increases utilization rate and decreases acceptance rate. On the
other hand, for any given service demand, fostering patient
provider mutual preference (increasing co-op ratio) can boost
acceptance rate by improving the efficiency of service time
utilization. The latter is evidenced by either increasing utiliza-
tion rate when demand is high (i.e., less-fragmented schedule
can accommodate more appointment requests) or decreasing
utilization rate when demand is low (i.e., serve patients with
less time and have more free time slots available).
Computer simulations also found that the statistical distribution
of the service time, thus the needed appointment time, has signifi-
cant impact on the appointment scheduling efficiency. No schedul-
ing system can work efficiently if there is a mismatch between
demand and supply [13]. Generally, when demand is greater than
supply, both acceptance rate and utilization rate will approach satu-
ration at high co-op ratios and have little roomfor further improve-
ment (e.g., N = 20 in Figures 411). Therefore, given a fixed
service time distribution, a fully defragmented schedule can theo-
retically provides the maximum capacity to accommodate the
appointment requests. On the other hand, whenthere is less demand
than supply, all appointments will be accepted while schedule is
under utilized, regardless of the co-op ratio (e.g., N=12 in Figures
6 and 8). Evidently, the co-op ratio also has no impact on appoint-
ment scheduling efficiency if most or all appointments need only
one time slot since they could be easily accommodated, despite
schedule fragmentation (data not shown). Moreover, when service
time distribution can be accurately estimated, decreasing unit time
slot size can further improve the scheduling efficiency by serving
more patients at shorter duration (see Figures 4, 10, and 11).
The survey results further confirmed the importance of
maintaining the demand-supply balance. The schedule defrag-
mentation will more likely benefit clinics that accept same-day
appointments and have a distribution of different appointment
times, although the challenges remain on integration of the fea-
ture into scheduling software systems and proof of its clinical
cost-effectiveness.
Appointment scheduling is an integral activity in patient care.
Admittedly, the deployment of patient scheduling software has
been slower compared with other service industries due to
under-funded capital budget, limited demand, lack of educa-
tional support, and interoperability gaps [14], [15]. However,
advanced appointment scheduling will likely surface as a top
priority for the health-care industry, fueled by the quest of maxi-
mizing efficiency and enabled by the rapid growth of informa-
tion technology (IT). Indeed, recent analysis from Frost and
Sullivan projected that the market size of U.S. patient scheduling
software systems would increase from US$75.6 million in 2003
to US$182.9 million in 2007 [15], and the market revenues of
European patient e-booking systems would grow from US$58.4
million in 2005 to reach US$107.3 million by 2012 [16].
The schedule defragmentation method presented in this article
is simple to implement and holds promise for better matching
patient requirements with health-care delivery. We envision
close collaboration with both health-care IT providers (who
develop integrated scheduling software systems) and clinical
partners (who commit to experimental studies on process
improvement) to speed adoption and diffusion of the technology.
The present study has a number of limitations. First, it is
limited by the nature of computer simulation, and practical vali-
dation of the model is warranted to confirm its concrete behav-
ior. Second, the appointment scheduling is modeled as a single-
batch process, which assumes all service requests are known
before allocating the time slots [9]. Third, it is assumed that, if
there are available time slots, patient will choose one option
(i.e., no declination). In addition, the simulation assumes that the
service time can be estimated with reasonable accuracy and does
not consider other postscheduling factors such as patient arrival
time and no shows. Despite these simplified conditions, the
present study clearly demonstrated that clinical appointment
scheduling process can be improved with the guidance of
schedule defragmentation by promoting patientprovider
mutual preference. This new appointment scheduling process
has the potential to achieve win-win by improving patient access
to health care and improving clinics operational efficiency.
Acknowledgments
The authors are indebted to Dr. R. Sakaguchi, Dr. J. Hunt-
zicker, Dr. N. Steckler, Dr. D. Pollack, Dr. D. Handel, Dr. J.S.
Amethod that considers patient-provider
mutual preference is proposed to guide the
appointment scheduling process by means of
schedule defragmentation.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Lou, Mr. M. Neal, Ms. N. Rich, and Ms. V. Miller of Oregon
Health & Science University, Ms. K. Johnson of Option Care
at Legacy, Mr. G. S. Pawson of Womens Health Associates,
and Ms. Ann OConnell of OHSU Richmond Clinic for help-
ful discussions on the clinical appointment scheduling process
and/or facilitation of the clinic survey.
Jie Lian received his Ph.D. degree in bio-
engineering from the University of Illinois
at Chicago in 2002. He then joined Micro
Systems Engineering Inc., part of the
worldwide Biotronik Group, where he is
currently a staff clinical research biomedi-
cal engineer. He is a Senior Member of the
IEEE. His research interests include feature
design and algorithm development for implantable pace-
makers and defibrillators, cardiac and neural electrophysiol-
ogy, physiological system modeling and computer simulation,
and biomedical signal analysis and interpretation. He is also
interested in broad areas of health-care management, opera-
tional decision making, and medical design innovation. He has
published three book chapters, more than 30 peer-reviewed
journal articles, and more than 40 conference papers or
abstracts. In addition, he has 19 issued/pending U.S. patents.
He has been an expert reviewer for a number of scientific jour-
nals, and has served as an editorial board member of The Open
Pacing, Electrophysiology & Therapy Journal.
Kori DiStefano received her B.A. degree
in political science from the University of
Portland in 2003. She was a member of Pi
Sigma Alpha, National Political Science
Honor Society, and a charter member of
Alpha Beta Alpha Chapter. She also
proudly served six years in the U.S. Navy,
earning numerous achievements for per-
formance excellence. She started her career in the pediatrics
field in 2001 at Westside Pediatric Clinic P.C., working part
time while completing her degree. Since graduation, she has
worked full time through virtually every department, being
promoted from EMR manager to Director of Operation, and
has recently been promoted to Practice Administrator over-
seeing both of Westside Pediatric Clinic locations.
Somer D. Shields received her B.S. degree
in business management with minor in
prelaw from Oregon State University in
1999. She has ten years experience in the
health-care field and had been a practice
manager at Perterkort North & NW Gyne-
cology Center, Portland, Oregon. She was
responsible for oversight and management
of clinic operations and personnel to support 13 clinicians in
OB/GYN specialties for Womens Healthcare Associates, LLC.
She currently works as a medical practice management consul-
tant for Shepperd Consulting performing consultative services
in practice management for a variety of clinics in the area. She
has served on the board of Portland Metro Community Manag-
ers (PMCM) and was the 2006 president of the board. She has
been a Medical Group Management Association (MGMA)
member since 2004 and became certified through American
College of Medical Practice Executives (ACMPE) in 2006.
Cindy Heinichen received her B.A.
degree in arts and letters from Portland
State University in 2002. She is currently a
patient accounts specialist at Oregon
Health & Science University Richmond
Clinic, Portland, Oregon.
Melissa Giampietri received her degree
in psychology from Eastern Washington
University in 2005. She is currently a
medical case management supervisor at
LifeWorks Northwest, Portland, Oregon.
She is responsible for supervising support
staff (Qualified Mental Health Associates)
for the psychiatrists and psychiatric mental
health nurse practitioners in the agency. She also works with
the medical director in the development and implementation
of policies and procedures and for building administrative
supports and system efficiencies for medical staff.
Lian Wang received her M.S. degree in
bioengineering at the University of Illinois,
Chicago, in 2002. She is currently a biomed-
ical engineer at Oregon Medical Laser Cen-
ter of Providence St. Vincent Medical
Center. Her research interests include regen-
erative medicine, system and cellular physi-
ology, and biomedical signal processing.
Address for Correspondence: Jie Lian, Micro Systems Engi-
neering Inc., 6024 SW Jean Road, Lake Oswego 97035, OR,
USA. E-mail: jie.lian@biotronik.com.
References
[1] K. Davis, Closing the quality chasm: Opportunities and strategies for moving
toward a high performance health system, Invited Testimony, Senate Committee
on Health, Education, Labor, and Pensions, Hearing on Crossing the Quality
Chasm in Health Care Reform, Jan. 29, 2009.
[2] G. F. Anderson and B. K. Frogner, Health spending in OECD countries: Obtain-
ing value per dollar, Health Affairs, vol. 27, no. 6, pp. 17181727, Nov./Dec. 2008.
[3] M. E. Porter, A strategy for health care reformToward a value-based sys-
tem, N. Engl. J. Med., vol. 361, no. 2, pp. 109112, July 2009.
[4] N. T. J. Bailey, A study of queues and appointment systems in hospital out-
patient departments, J. R. Stat. Soc. (Ser. B), vol. 14, no. 2, pp. 185198, 1952.
[5] J. P. Birchall, T. F. Cox, and H. Wong, Computer modeling of ENT outpa-
tients, Clin. Otolaryngol., vol. 8, no. 6, pp. 411415, Dec. 1983.
[6] M. Brahimi and D. J. Worthington, Queuing models for outpatient appointment
systems: A case study, J. Operat. Res. Soc., vol. 42, no. 9, pp. 733746, Sept. 1991.
[7] S. Taylor and J. Kuljis, Simulation in health care management: modelling an
outpatient clinic, OR Insight, vol. 11, no. 3, pp. 711, July 1998.
[8] P. R. Harper and H. M. Gamlin, Reduced outpatient waiting times with
improved appointment scheduling: A simulation modeling approach, OR
Spectrum, vol. 25, no. 2, pp. 207222, May 2003.
[9] D. Gupta and B. Denton, Appointment scheduling in health care: challenges
and opportunities, IIE Trans., vol. 40, no. 9, pp. 800819, Sept. 2008.
[10] T. R. Rohleder and K. J. Klassen, Rolling horizon appointment scheduling: A
simulation study, Health Care Manage. Sci., vol. 5, no. 3, pp. 201209, Aug. 2002.
[11] A. Azzalini, A class of distributions which includes the normal ones,
Scand. J. Stat., vol. 12, no. 2, pp. 171178, 1985.
[12] L. R. Laganga and S. R. Lawrence, Clinic overbooking to improve patient access
and increase provider productivity, Decis. Sci., vol. 38, no. 2, pp. 251276, May 2007.
[13] M. Murray and D. M. Berwick, Advanced access: Reducing waiting and
delays in primary care, JAMA, vol. 289, no. 8, pp. 10351040, Feb. 2003.
[14] K. Godfrey, Delays in implementing e-booking threaten patient choice
agenda, Brit. Med. J., vol. 330, no. 7484, p. 166, Jan. 2005.
[15] (2004, Dec. 7). Automation in healthcare gains ground, facilities eye patient
scheduling software systems. Business Wire [Online]. Available: http://findarticles.
com/p/articles/mi_m0EIN/is_2004_Dec_7/ai_n7639276/
[16] K. John. (2006, Aug. 27). European patient e-booking systems: Scheduling
administrative success. Frost & Sullivan Res. Service [Online]. Available: http://
www.frost.com/prod/servlet/report-brochure.pag?id=M01F-01-00-00-00
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_______________
___
EYEWIRE
Dysphonic Speech
Reconstruction
V
erbal communication is one of the most influential
and effective way of social communication. Voiced
sounds are produced when the vocal cords vibrate;
thus, the flow of air from the lungs to the vocal tract
interrupts, and quasi-periodic pulses of air are produced dur-
ing excitation. Dysphony is a functional disorder of larynx as
a result of pathologic vibration in vocal cords. Chronic dys-
phonia occurs in the presence of organic lesions (such as
polyp, nodule, and Reinkes edema) in the vocal cords, lethal
larynx diseases, throat cancer, neurological disorders, and
chronic irritation due to smoking. The breathed air, which is
send to the trachea to produce voice, could make the vocal
cords to vibrate barely or not at all because of the pathological
formation in the vocal cords of the patients. As a result, voice
comes out as a low whisper and more cracked than usual.
After total laryngectomy, there are three well-established
methods to fix the voice. The first one is alaryngeal speaking.
Through this method, the patient can speak by using an elec-
trolarynx. The second method is training the patient and help-
ing him speak in esophageal speech. The third method is
speaking by using tracheoesophageal voice prostheses. Even
though the speech is not as qualified as the previous, patients
can speak through artificial larynx, voice prosthesis, or
esophageal speech. However, these methods cannot be applied
to the patients with apoplectic chordae vocalis, organic lesions
of vocal cords, or who suffer from dysphony due to a partial
laryngectomy in which some parts of the larynx and vocal
cords are removed. Solutions such as voice therapies and/or
operations to help patients to speak again may not work at all.
Several systems that analyze and enhance the characteris-
tics of the esophageal speech and speaking using electrolarynx
have been designed so far [1][5]. However, there is no
reported research in the literature that produces synthetic
voice digitally based on the patients voice in cases where the
patients were treated with partial laryngectomy or had com-
pletely lost speech as a result of organic lesions on the vocal
cord or of vocal-cord paralysis.
In this article, we present a novel system that delivers syn-
thetic speech with a quality close to natural by reconstruct-
ing dysphonic speech. We believe that it will be an
important improvement in the social patients for effective
and efficient communication.
Acoustic Characteristics of Dysphonic Speech
Chronical dysphonia mainly occurs because of the malfunc-
tioning of the vocal cords. Voice formed this way demon-
strates whisperlike characteristics. Dysphonic speech differs
from normally phonated speech in terms of voicing, pitch, and
formant structure. Spectrograms of normal and dysphonic
speech for the Turkish word C al+ma (IPA codes of charac-
ter c =t
R
and + =
R
) are given in Figure 1.
Figure 1 clearly shows that, contrary to the voiced pho-
nemes of normal speech, there is no perceivable pitch period
or voicing observed in the voiced phonemes of dysphonic
speech. In addition to this, voiced phonemes of dysphonic
speech differ from the voiced phonemes of normal speech in
terms of formant distortion. Bandwidths of dysphonic pho-
nemes are larger, and their formant frequencies are greater.
However, in unvoiced phonemes of dysphonic speech, there
is no significant formant distortion observed [5]. Differences
between dysphonic speech and normal speech are summar-
ized in Table 1 in terms of pitch, voicing, and formant distor-
tion characteristics.
According to Table 1, it was determined that no modifica-
tion should be done for the unvoiced phonemes of a dys-
phonic speech.
Data Collection
The voices of dysphonic patients come out as whispers
because their vocal cords cannot function properly. On the
other hand, evaluating both the dysphonic voice and its orig-
inal form before the disorder is essential to choose the appro-
priate method for normal speech reconstruction. Since
accessing a dysphonic patients original voice recordings is
rather difficult, normal voices and whispers of healthy
speakers were used to choose the proper method. For this
purpose, a database consisting of recordings of normal voi-
ces and whispers of 30 men and 20 women speakers aged
2550 was established.
There is no public database of dysphonic speech in litera-
ture; so, a dysphonic speech database containing 22 patients
speech recordings was created to appraise the success of the
BY H. IREM TURKMEN
AND M. ELIF KARSLIGIL
N
A
V
I
G
A
T
I
N
G
T
E
C
H
N
O
L
O
G
Y
T
R
A
N
S
F
E
R
Construction of a Novel System for
Effective and Efficient Communication
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
system. The recordings in the created database are of eight
men and five women with apoplectic chordae vocalis, who
were aged 3555, and of six men and three women having par-
tial laryngectomy, who were aged 4565. The voice record-
ings were done in a silent room with a microphone that was
placed 10-cm away from each patients mouth. The sampling
frequency was set at 8 kHz. The patients were made to read
sentences that contained a total of 80 different Turkish words
with normally distributed phonemes.
Dysphonic Speech-Enhancement System
Taking into consideration the acoustic differences of dys-
phonic and normal speech, a mixed excitation linear predictive
coding (MELP) [6] based system was designed to enhance
the dysphonic speech. Figure 2 illustrates the block diagram
of the system.
First, speech of a dysphonic patient is recorded, and the
recorded speech is analyzed by using MELP. Second, each
phoneme is classified as voiced or unvoiced. To this end, mel-
frequency cepstral coefficients are used to extract the features
of voiced and unvoiced phonemes, and principle component
analysis [7] is applied to reduce the dimensions of the features.
Then, voiced and unvoiced phoneme groups are classified by
using support vector machine [8].
Next, pitch, formant, and voicing parameters of voiced pho-
nemes are modified. Synthetic pitch production is done by
using the relation between pitch and gain values [9], [10].
Line spectrum frequency-based formant structure modifica-
tion is applied to obtain narrow bandwidths and altered fre-
quencies [11]. Line spectrum pair trajectories are smoothed by
median filter during the vowels without destroying the rapidly
varying spectral content of the phonemes [9].
Pulse excitation is produced for the voiced phonemes at a
frequency band of 05 kHz. For this reason, the lower four
frequency bands (03 kHz) are fixed as voiced and upper band
(34 kHz) are fixed as unvoiced [9].
Results and Observations
Log spectral distances were used to test the spectral differen-
ces between the normal and synthetic speech. The average
spectral enhancement obtained was calculated as 25%. Figure
3(c) and (d) shows two different synthetic speech spectro-
grams for the Turkish word Ta+ spoken by a patient with
partial laryngectomy. In Figure 3(d), all the phonemes, voiced
and unvoiced, were modified, whereas in Figure 3(c), only the
voiced ones.
As can be seen in Figure 3(c), with the help of the modifica-
tions, synthetic speech was given the attribute of periodicity
20
0
20
40
60
80
100
A
m
p
l
1
0
Time (s)
(a)
(b)
1
0.0 0.2 0.4 0.6 0.8 1.0 1.2
MA
S ALI
20
10
0
10
20
30
40
50
60
A
m
p
l
1
0
1
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
MA
S ALI
Fig. 1. Spectrogram of (a) normal speech and (b) dysphonic
speech for C alsma.
Table 1. Acoustic differences of dysphonic
and normal speech.
Pitch Voicing
Formant
Distortion
Dysphonic speech
Unvoiced phonemes 3 3 3
Voiced phonemes 3 3
_
Normal speech
Unvoiced phonemes 3 3 3
Voiced phonemes
_ _
3
_
: Presence of the related feature; 3: absence of the
related feature.
Dysphony is a functional disorder of larynx
as a result of pathologic vibration
in vocal cords.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
similar to the ones of normal
speech [Figure 3(a)]. The re-
sults show that preservation
of the acoustic features of
unvoiced phonemes produce a
synthetic speech that is closer
to the normally phonated one.
To determine the preferen-
tiality of the synthesized speech
versus dysphonic speech, sub-
jective listening tests were con-
ducted. Fifteen listeners were
requested to listen to the sam-
ples and then rate them within
a scale of 15 (1 being the
worst and 5 the best) accord-
ing to timber, intelligibility,
and naturalness. The calcu-
lated average points are given
in Table 2.
As can be seen in Table 2, synthetic speech produced by the
use of either approaches is recognized better than dysphonic
speech, especially in terms of timber and naturalness.
Future Work
There are currently no commercially available devices to use
normal speech reconstruction from dysphonic speech. Our
normal speech reconstruction software can be implemented in
MELP
Analysis
Unvoiced
Phoneme
Detection
Pitch
Modification
Format
Modification
MELP
Synthesis
Dysphonic
Speech
Enhanced
Speech
Synthetic
Pitch
Modified
Spectrum
Voicing
Parameters of
Dysphonic
Speech
Voiced Phonemes'
Acoustic Features
Unvoiced Phonemes'
Acoustic Features
Voiced Phonemes'
Modified Acoustic
Features
Fig. 2. Block diagram of the proposed system.
0
20
40
60
80
100
A
m
p
l
1
0
Time (s)
(a)
(c)
1
0.0 0.1 0.2 0.3 0.4 0.5
S A T
20
10
0
10
20
30
40
50
60
70
A
m
p
l
1
0
1
0.0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45
S A T
10
0
20
10
30
50
40
60
70
80
A
m
p
l
1
0
Time (s)
(b)
(d)
1
0.0 0.05 0.10 0.15 0.20 0.25 0.30
S A T
20
10
0
10
20
30
40
50
60
70
A
m
p
l
1
0
1
0.0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45
S
A T
A
m
p
l
1
0
Tim
1
.0 0.1 0.2 0.3 0.4 0.5
S SS A T
Fig. 3. Spectrograms of Turkish word TA*: (a) normal speech, (b) dysphonic speech, (c) synthetic speech produced by modifi-
cation of voiced phonemes, and (d) synthetic speech produced by modification of each phoneme.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
an embedded system of the size of a mobile MP3/4 player or a
cellular phone, which patients could carry with them all the
time. One can also think of implementing it directly on a com-
mercially available cellular phone as a software add on. Con-
sidering the capabilities of cellular phones like recording and
playback of streaming audio and video, this processing power
can be used for normal speech reconstruction. It is obvious
that the presented algorithm should be tailored and/or opti-
mized for the aforementioned resource constrained-embedded
devices, which, we believe, is not of great concern. In addi-
tion, cellular phones are already equipped with the right voice
recording and playback hardware. Therefore, the normal
speech reconstruction software installed on a cellular phone
can intercept all outgoing speech data during an active call,
reconstruct the dysphonic speech, and then have it sent to the
calling party. Depending on the processing power and/or
memory capacity of the cellular phone, it may be necessary
for the speaker to give brief pauses every one or two sentences
for the software to do the reconstruction.
As stated earlier, the normal speech reconstruction software
could also be implemented on a dedicated embedded hardware
as a standalone mobile device. This standalone device could
have a built-in speaker for close-range conversations and/or both
analog and digital sound outputs to be tied to a sound system for
using it when the speaker needs to address bigger crowds. In this
scenario, the mobile device could be equipped with necessary
processing power and memory capacity to enable a streaming
speech reconstruction as the speaker speaks freely.
Especially, both the mobile communications and entertain-
ment industries are the driving force for the hardware manu-
facturers to produce more powerful CPUs and higher density
memories with lesser power consumption available at lower
prices. Therefore, we believe that it is justified to think that
the normal speech reconstruction system could be made avail-
able to the patients at very reasonable prices both as a soft-
ware-only solution available in the form of a software add on,
e.g., cellular phones, as a dedicated mobile device.
Conclusions
In this article, we present a MELP-based real-time system that
produces synthetic speech for patients with chronic dysphonia
while preserving their continual structure of speech.
Dysphonic patients, even though strained, could express
themselves within face-to-face communications. However,
when speaking in a group or using communication devices
such as mobile phones, they face serious difficulties in vocal
communication since it becomes harder to understand their
speech. Our system will make the speech of such a patient
much more understandable by reconstructing the patients
dysphonic speech in real time. Our system can easily be fit
into an embedded system of the size and shape of a mobile
phone, so that it could be carried all the time by the patients.
Acknowledgment
We express our appreciation to Istanbul University Cerrah-
pasa Medical FacultyEar Nose Throat and Head & Neck
Surgery Department for their support in this work.
H. Irem Turkmen received her B.Sc. and
M.Sc. degrees in computer engineering
from Yildiz Technical University, Istanbul,
Turkey, in 2005 and 2008, respectively.
She is a Ph.D. student at the Computer
Engineering Department of Yildiz Techni-
cal University. Her research interests in-
clude pattern recognition, speech processing,
and machine learning.
M. Elif Karsligil received her B.Sc.,
M.Sc., and Ph.D. degrees in computer
engineering from Yildiz Technical Univer-
sity, Istanbul, Turkey, in 1988, 1990, and
1998, respectively. She is currently an
assistant professor at the Computer Engi-
neering Department of Yildiz Technical
University. From October 2001 to Novem-
ber 2002, she worked as senior researcher at NTT Communi-
cation Science Laboratories, Kyoto, Japan. Her research
interests include machine learning, pattern recognition,
speech processing, and digital video processing.
Address for Correspondence: H. Irem Turkmen, Yildiz
Technical University, Computer Engineering Department,
34349 Yildiz, Istanbul, Turkey. E-mail: irem@ce.yildiz.edu.tr.
References
[1] G. Aguilar and M. Nakano-Miyatake, Alaryngeal speech enhancement using
pattern recognition techniques, IEICE Trans. Inform. Syst., vol. 88, no. 7,
pp. 16181622, 2005.
[2] N. Bi and Y. Qi, Speech conversion and its application to alaryngeal speech
enhancement, in Proc. 3rd Int. Conf. Signal Processing (ICSP96), 1997,
pp. 15861589.
[3] A. Pozo and S. Young, Continuous tracheoesophageal speech repair, in Proc.
European Signal Processing Conf. (ICSP96), Florence, Italy, Sept. 48, 2006.
[4] Y. Qi, B. Weinberg, and N. Bi, Enhancement of female esophageal and
tracheoesophageal speech, J. Acoust. Soc. Amer., vol. 98, no. 5, pp. 24612465,
1995.
[5] H. Sawada, N. Takeuchi, and A. Hisada, A real-time clarification filter of a
dysphonic speech and its evaluation by listening experiments, in Proc. Int. Conf.
Disability, Virtual Reality and Associated Technologies (ICDVRAT04), 2004,
pp. 239246.
[6] A. V. McCree and T. P. Barnwell, III, A mixed excitation LPC vocoder
model for low bit rate speech coding, IEEE Trans. Speech Audio Processing,
vol. 3, no. 4, pp. 242250, July 1995.
[7] V. Tran, G. Bailly, H. Lvenbruck, and T. Toda, Predicting F0 and voicing
from NAM-captured whispered speech, in Proc. 4th Conf. Speech Prosody,
Campinas, Brazil, May 2008, pp. 107110.
[8] V. N. Vapnik, An overview of statistical learning theory, IEEE. Trans.
Neural Networks, vol. 10, no. 5, pp. 998999, Sept. 1999.
[9] R. W. Morris and M. A. Clements, Reconstruction of speech from whis-
pers, Med. Eng. Phys., vol. 24, no. 7, pp. 515520, 2002.
[10] M. Higashikawa, K. Nakai, A. Sakakura, and H. Takahashi, Perceived pitch
of whispered vowels- relationship with formant frequencies: A preliminary
study, J. Voice, vol. 10, no. 2, pp. 155158, June 1996.
[11] I. V. McLoughlin and R. J. Chance, LSP-based speech modification for
intelligibility enhancement, in Proc. 13th Int. Conf. DSP, vol. 2, pp. 591594.
Table 2. Comparison of dysphonic speech
with the synthetic speech.
Timber Intelligibility Naturalness
Dysphonic speech 0.3 2.3 1.1
Synthetic speech
produced by modifi-
cation of each
phoneme 2.8 2.4 3.0
Synthetic speech
produced by modifi-
cation of voiced
phonemes 2.9 2.8 3.1
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
______________
Couple Meet,
Fall in Love
Through IEEE EMBS
Conferences
W
hen Farah Corona Chavez and Daniel Strauss
met at the IEEE EMBS Conference in Cancun,
Mexico, in 2003, they did not know it was to be a
fateful moment in their lives and in fact did not
even speak. Strauss, a native of Germany, and Corona, a
native of Mexico, went their separate ways until 2005, when
Strauss, who by then had completed his second Ph.D. degree,
came to Mexico again as a visiting professor in the Biomedical
Engineering Department of Mexico Citys Universidad Auton-
oma Metropolitana, where Corona was working on her mas-
ters degree. She was also working in the research laboratory
of Innovamedica, which Strauss toured during his stay, and
they talked about the project.
Later that year, at the IEEE EMBS Conference in Shang-
hai, they met again, and this time they got to know one
another better. Corona spoke very little German and
Strauss even less Spanish, and so they spoke primarily in
English as they went to dinner, took walks, and began to
fall in love. What followed was a very long-distance rela-
tionship that lasted well into the following year, when at
last Corona packed her bags and boarded a plane for Ger-
many, where she and Strauss were married in a civil cere-
mony a few months later. The next year, they went back
around the globe to Mexico, where they were married in a
religious ceremony.
Today, the couple works closely together in a laboratory
affiliated with the University Hospital in Saarland, Ger-
many. Strauss is a full-time professor of medical engineer-
ing and director of the Computational Diagnostics and
Biocybernetics (CDB) Unit at the Saarland University Hos-
pital and Saarland University of Applied Sciences. He is
BY JESSE JAYNE RUTHERFORD
DIGITAL STOCK
Fig. 1. Corona and Strauss at their religious ceremony in
Mexico, 2007. (Courtesy of Farah Corona-Strauss and
Daniel Strauss.)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
also the technical head of the Center for Research in Medi-
cal Communication Disorders at the Saarland University
Hospital and a research associate of the Leibniz Institute for
New Materials in Saarbruecken, Germany. Farah Corona-
Strauss has completed her Ph.D. degree and is a lecturer in
the masters program in biomedical engineering at the
Saarland University of Ap-
plied Sciences. She also
works in the same program
and is a researcher in the
CDB Unit. In 2010, Corona-
Strauss will changeover to
working part time at Key
Numerics, a small software
development company the
couple owns. Finally, the cou-
ple are both visiting profes-
sors in Malaysia.
In their free time, although
there is not much of it, the
couple enjoys jogging to-
gether and touring vine-
yards. Corona-Strauss adds
that she has tried to teach her
husband to dance, but so far,
the results have been less
than optimal. They also en-
joy seeing friends at the
EMBS Conferences and oth-
er work-related events; one
reason they love the field of
biomedical engineering is
because the variety of disci-
plines and backgrounds that
come together lend a great
deal of vibrancy to their
work. Professional collabo-
ration, a rigorous schedule,
and little time for a social
life could be a recipe for
disaster, but the couple shares
a joyous life together. Please
make sure that it is written
somewhere howmuch I love Farah, Strauss adds.
Jesse Jayne Rutherford is an award-winning ghostwriter
and a freelance writer. She lives near San Diego, California.
For more information, please visit her Web site at
www.JesseRutherford.com.
Fig. 2. The couple at the IEEE EMBS Neural Engineering Conference in Antalya, Turkey, in 2009.
(Courtesy of Farah Corona-Strauss and Daniel Strauss.)
Professional collaboration, a rigorous schedule,
and little time for a social life could be
a recipe for disaster, but the couple shares
a joyous life together.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Senior Design
developing teamwork skills in capstone design courses
T
he majority of our biomedical
engineering graduates will even-
tually work in an industry where
they will be part of multidiscipli-
nary teams that use the collective skills,
expertise, experience, and training of
each team member. Diversity within
these teams provides different perspec-
tives, opinions, and ways of viewing
problems, leading to a larger set of
potential solutions. Successful careers
require engineers to be able to function
on multidisciplinary teams.
Most surveys I have read list team-
work skills and the ability to work on
teams as one of the most important
characteristics that employers look for
when hiring. Ideally, engineering stu-
dents will have had experience with
team projects prior to the senior year as
part of a class, an extracurricular project
activity (such as a national design com-
petition), a summer job, or a coop job
experience. If not, the senior capstone
design course can provide students with
the opportunity to develop the team-
work skills needed for successful ca-
reers in engineering.
According to studies of effective new
product development teams, team
members should possess the following
components:
technical competence and prob-
lem-solving skills
a commitment to common goals
that are clear and measurable
shared values and priorities
a feeling of mutual accountability
a willingness to contribute and
cross boundaries when needed.
Additional components of effective
capstone design teams include the
following:
clearly defined roles and account-
abilities based on the strengths of
each team member
a collaborative climate where each
member contributes and benefits
from the team project experience
an effective communication system
between team members, with advi-
sers, sponsors, course instructors,
and class
a method for monitoring individual
performance and providing feed-
back to team members.
The personal characteristics and
skills needed by capstone design team
members are as follows:
the ability to get along with others
conflict management skills
interpersonal skills
a willingness to attend team meet-
ings and contribute ideas
respect for teammates and their
contributions
a sense of obligation to complete
their fair share of the work.
The capstone design course can help
the student learn about teamwork and
conflict management. This can be done
through course lectures, team-building
activities, and coaching on how to deal
with conflicts that typically occur
within project teams.
Lectures that deal with the impor-
tance of teams, requirements for ef-
fective teams, common team-related
conflicts and how to deal with them,
expectations of student teams, and how
to run effective team meetings can be
very helpful in preparing students for
the team-based project experience
required by the capstone design course.
Ideally, these lectures would be pre-
sented prior to or just after the forma-
tion of project teams and include a
discussion of the stages of team devel-
opment (forming, storming, norming,
performing, and adjourning) and con-
flict management within teams.
Conflict management is an extremely
valuable skill not only for the capstone
design project but also for successful
careers. It is important for the students
to understand that conflict is a normal
part of teamwork and, if managed
properly, can result in more effective
collaborations among team members
and better design solutions. I find it
helpful to discuss common problems
that often occur within capstone design
teams along with the strategies for deal-
ing with the resulting conflict. These
problems include a team member 1) not
completing his/her fair share of work, 2)
insisting that his/her ideas are the best
and not listening to others ideas, 3) not
getting along with other team members
(bossy, disrespectful, and argumenta-
tive), and 4) failing to contribute to the
team by not participating in team meet-
ings or offering ideas or opinions.
There are a variety of team-building
activities that have been used in indus-
try and other sectors to create a sense
of teamwork among team members.
Faculty in schools of business adminis-
tration can be a helpful resource in find-
ing a team-building exercise that might
be suitable for a particular capstone
design course. Asking (or requiring) all
teams to develop a team logo for their
project can help create team identity
and team spirit and could be part of a
logo design competition. At Marquette
University, each member of the team
with the best logo design wins a golf
shirt containing their embroidered team
logo. These teams typically wear their
shirts as a team uniform during team
oral presentations, conveying a sense of
team unity and spirit. After the logo
design competition, the logos begin
appearing on all official team docu-
ments and course deliverables, estab-
lishing a team brand identity. Poster
competitions at the end of the capstone
design course give the students an expe-
rience with another technical communi-
cation tool and provide the team with a
forum for showcasing their project
work. Designing the poster and discus-
sing their work with attendees can
Jay R. Goldberg
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
develop a sense of pride among the stu-
dents as well as strengthen their sense
of team unity.
Team size can have a significant
effect on how the students benefit from
the teamexperience and what they learn
from it. The word team implies more
than one person working on the project.
A project run by one student does not
provide the student with a valuable
team experience. A project with a team
consisting of six or more students may
not provide enough work for each
student and can create a confusing, inef-
fective team environment. A team con-
sisting of two students can work well if
the students are well matched. How-
ever, if one student does not complete
their fair share of the project work, then
the other student will be unfairly over-
burdened. During my ten years of
teaching the biomedical engineering
senior capstone design course, I have
found that project teams consisting of
three to five students work well within
the structure of my course (two-semes-
ter sequence combining biomedical,
electrical, mechanical, and computer
engineering students).
Team composition can also play a
major role in the success of the team.
Team members should have an interest
in the project, a willingness to work with
the other team members, and the techni-
cal skills and knowledge base required
for the successful completion of the
project. I have found that allowing stu-
dents to form teams or rank their top
choices of projects increases the proba-
bility of a successful project outcome.
When finalizing the team roster, I con-
sider their engineering disciplines (bio-
mechanical, bioelectrical, biocomputer,
electrical, computer, mechanical engi-
neering), technical skills (AutoCAD,
ProE, SolidWorks, machine shop expe-
rience, etc.), and professional experien-
ces (coop job, research, internships,
summer jobs). I try to create teams
whose members collectively possess as
many of the technical skills and knowl-
edge areas that will be needed for the
successful completion of the project.
Personalities of the team members
play an important role in the success of
the project. Some design course instruc-
tors use the Myers-Briggs type indica-
tor (MBTI) assessment (or modified
version) to determine personality types
and attempt to create teams with an
optimal balance of different, comple-
mentary personalities. My experience
with student-formed teams is that stu-
dents tend to choose teammates that
they know and get, increasing the
probability of compatibility between
personalities on the team. However, this
can also result in the formation of teams
of students who get along well but lack
the technical expertise and other skills
needed for the successful completion of
the project.
I find that making students aware of
what is expected of them as a team
early in the course can be very helpful
in avoiding team-performance prob-
lems. Students are provided with a
student handbook that discusses how
teams are expected to function and pro-
vides suggestions for how to manage
themselves. Students are expected to be
self-managed (without prodding from
faculty advisers to meet course dead-
lines), respect their project advisers
and sponsors time, and communicate
often with project advisers and spon-
sors. The following are recommended
for the teams:
assess team members strengths
and weaknesses and assign work in
areas of competency and strength
help each other and ask for help
when needed
support and respect each other
share opinions and ideas
arrive at a true consensus
conduct themselves as professionals.
Another helpful lecture topic involves
how to run effective meetings. This in-
cludes determining whether a meeting is
necessary, distributing and using meeting
agendas, confirming responsibility for
action items, and sending a summary of
what was discussed and decided at the
meeting.
According to the Accreditation Board
for Engineering and Technology (ABET),
undergraduate engineering programs
must demonstrate that their students attain
11 specific outcomes. Outcome D is an
ability to function on multidisciplinary
teams. There are a few ways that a pro-
gram can show that this outcome is being
met in their capstone design course. First,
students can complete a peer evaluation
of their fellowteammembers. This can be
used to determine how well each team
member functions on the team and has
developed their teamwork skills and
serves as a measure of individual student
performance. Second, grading of group
deliverables assesses the quality of work
that the teamis capable of completing as a
team and is a measure of team perform-
ance. Third, a team evaluation from each
member helps assess the ability of the
teamto work as a team.
In summary, the capstone design
course can be used to provide students
with the opportunity to develop the
teamwork skills needed for successful
careers. This can be done through course
lectures, team-building activities, and
coaching on how to deal with conflicts
that typically occur within project teams.
Team size, composition, and personal-
ities play an important role in improving
the effectiveness of capstone design
project teams. The earlier in the curricu-
lum that students are exposed to project
team experiences, the better prepared
they will be for the senior capstone
design course and their careers.
Most surveys I have
read list teamwork
skills and the ability
to work on teams as
one of the most
important
characteristics that
employers look for
when hiring.
Senior Design (continued)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
______________________________
_________________________________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Point of View
medical devices in cinema
I
s it time to take a more sophisticated
approach to the way we portray medi-
cal devices in American cinema?
Back when I was still an engineer
in training in 1982, the Extraterrestrial
(E.T.) was resuscitated using Physio-
Controls portable Lifepak 5 defibrilla-
tor (Figure 1). Admittedly, at a tender
age of 19 years, I bought it. E.T.s skin
had ashened, and he could not stand up
in Eliotts closet. Then E.T. was quaran-
tined in a specially setup emergency
room by the Feds. When E.T. lost con-
sciousness, they defibrillated him. Since
we still do not fully understand the
mechanism behind external defibrilla-
tion, how did we know that E.T. would
survive? Did he have a four-chambered
heart? Did he even have a heart?
Outbreak was released in 1995, which
was unfortunately not one of Dustin Hoff-
mans best films. An ebolalike virus breaks
out in Africa and is mistakenly transported
to the United States via a monkey in a
freighter (you cannot make this stuff up).
The monkey transmits the virus to humans,
so the Centers for Disease Control (CDC)
has to prevent the epidemic. Luckily, the
U.S. Army virologist Dustin is the ex-hus-
band of the CDC scientist investigating
this. He isolates the virus and is able to
manufacture a vaccine in about 24 h. We
are stuck with the old-egg technology for
manufacturing influenza vaccine, but Dus-
tin has a secret he has not yet shared.
Undoubtedly, Dustin probably had the
time to design and validate the correspond-
ing in vitro diagnostic test the day after.
But my all-time favorite plot is from
Greys Anatomy in 2005. OK, it is not a
movie, but the story line is heart-valve
heavy. Orthodox JewDevo has a congen-
ital von Willebrand disease, which affects
her ability to clot. She requires a heart-
valve replacement, but because of her dis-
order, she cannot have a preferred (at
least at fictional Seattle Grace hospital)
mechanical valve implanted. Intern Alex
and cardiac surgeon Burke tell Devo that
they will implant a pig valve instead.
Devo goes ballistic because she cannot
ingest pork. Alex investigates and sug-
gests that Burke will implant a pericardial
valve instead. That is a tissue valve
shaped from cow, not pig, pericardium.
Burke is hesitant because pericardial
valves are more difficult to implant (at
least at Seattle Grace) than pig valves.
Specific models are not mentioned, but I
would swear that the physicians were dis-
cussing the Star-Edwards ball (not
retired until 2007) and Carpentier-
Edwards S.A.V porcine and Perimount
Magna Ease valves. Edwards Lifescien-
ces can not buy this kind of publicity.
In 2008, Flash of Genius was released
in theaters. The main character was real-
life Wayne State University mechanical
engineering Prof. Robert Kearns, who
won multimillion-dollar judgments
against Ford and Chrysler for stealing
his intermittent windshield wiper patent.
Are not medical device events just as
compelling? It is the late 1950s. Cardiol-
ogist Paul Zoll, who demonstrated that
patients in asystole could be paced exter-
nally across the chest in 1952, is nowin a
race against engineer Wilson Greatbatch
and cardiac surgeon William Chardack.
Who will successfully (i.e., for at least
two months) implant the first pacemaker
in a human patient? But then, surgeon
Robert Rubio in Uruguay (yes, Uru-
guay!) implants engineer Rune Elmqv-
ists pacemaker into his patient first in
February 1960 [1], a few months before
Greatbatch and Chardack reach their
goal. Earl Bakken then licenses Great-
batchs pacemaker design for Bakkens
fledgling company Medtronic [2].
Or how about the physician Willem
Kolff building the first artificial kidney for
acute renal patients during World War II
in the Nazi-occupied Netherlands? Kolff
has to scrounge-up sausage casing, an alu-
minum drum and an enamel tank to build
his dialyzer [3]. When nephrologist Beld-
ing Scribner connects a patients artery
and vein through a U-shaped shunt to keep
the circulatory access open in 1960, his
patients can now receive repeated hemo-
dialysis for the first time from a dialyzer-
related in design to Kolffs dialyzer. Scrib-
ner establishes the worlds first outpatient
dialysis center but is overwhelmed by the
number of end-stage renal patients requir-
ing dialysis. So, he creates a committee of
seven anonymous Seattle citizens who
pick the patients. This could be called the
first death panel in the United States,
except that it is really a life panel [4].
For me, these are just two of the count-
less medical device dramas that would
make fascinating viewing. If Leslie
Geddes was still with us, he would have
even more exciting material for scripts.
Hollywood, we are ready for our close up.
Gail D. Baura is a professor at Keck
Graduate Institute of Applied Life Sci-
ences, a member of the Claremont Col-
leges. She became interested in medical
device history while reading Leslie
Geddes inspiring historical articles and
Retroscope columns in IEEE Engineer-
ing in Medicine and Biology Magazine
from1990 to 2009. She can be contacted
at gbbook@mindspring.com.
References
[1] O. Fiandra, The first pacemaker implant in
America, Pacing Clin. Electrophysiol., vol. 11,
no. 8, pp. 12341238, Aug. 1988.
[2] W. Greatbatch and C. F. Holmes, History of
implantable devices, IEEE Eng. Med. Biol. Mag.,
vol. 10, no. 3, pp. 3841, 1991.
[3] W. J. Kolff, Lasker Clinical Medical Research
Award. The artificial kidney and its effect on the
development of other artificial organs, Nat. Med.,
vol. 8, no. 10, pp. 10631065, Oct. 2002.
[4] B. H. Scribner, Lasker Clinical Medicine
Research Award. Medical dilemmas: The old is new,
Nat. Med., vol. 8, no. 10, pp. 10661067, Oct. 2002. Digital Object Identifier 10.1109/MEMB.2009.935724
Gail D. Baura
Fig. 1. The Physio-Control Lifepak 5
saved E.T.s life (Medtronic Physio-Con-
trol, Redmond, WA).
________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Book Reviews
Biological Database Modeling
Jake Chen and Amandeep S. Sidhhu,
Artech House, 2008. ISBN: 978-1-
59693-258-6, 224 pages, Hardcover,
US$110.34.
Current biomedical research in areas
such as drug discovery and/or microar-
ray produces a large volume of data, and
the appropriate utilization of scientific
data modeling tools and strategies can
help researchers maximize the benefits,
standardization, and interoperability of
modern biomedical informatics data
management systems in practical ways
never imaginable before. This would
call for collaborative efforts in future-
generation bioinformatics data model-
ing techniques, which needs to be based
on innovative concepts, platform-inde-
pendent standard, methods, and applica-
tions, with case studies in genome,
microarray, proteomics, and drug dis-
covery projects to develop more power-
ful and intensive data management
systems in any domain. Focused on the
intersection of postgenomic life scien-
ces and data management, the book
introduces the readers to core biological
data modeling techniques, biological
database resources, ontology concepts,
and data management challenges for
high-throughput data. From a practical
perspective, the omics sciences and feder-
ated database of genomics, functional
genomics, proteomics, and metabolomics
will continue to thrive and mature in the
next millennium and will enable the
pharmaceutical scientists or researchers
design target-specific drugs toward cer-
tain human proteins with high therapeutic
values and lowtoxicological effects.
This book is written by a large number
of biomedical professionals with hands-
on bioinformatics database development
experience, including both developers
and researchers in industry and aca-
demic settings. The leading author and
editor of the book, Jake Chen, is a Senior
Member of the IEEE and has had combi-
national bioinformatics data-modeling
experience in industry and academics.
The book comprises 11 chapters. The
chapters have been prepared by selected
subject expert teams active in the research
of respective topics with varying degrees
of balance between computational data
modeling theories and real-world case
presentations.
The focus of Chapters 15 is to present
basic biological database concepts and
general data representation practices
essential to postgenome biology. Chapter
1 introduces biomedical data modeling
concepts, including generic modern
markup language XML, modeling with
complex data structures or XML, ontol-
ogy, semantic modeling, subject index,
query language, and views. Chapter 2
summarizes the major public database
efforts in omics and systems biology
studies, including the application of
genomics, proteomics, metabolomics,
pharmacogenomics, and systomics data-
base. Chapter 3 introduces biomedical
data modeling techniques in the en-
hanced-ER model with three constructs:
ordered, process, and molecular spatial
relationships. Chapter 4 describes the
gene ontology as an established basic set
of controlled vocabulary in genome
database annotations, including applica-
tions of gene ontology in biological and
medical science. Finally, Chapter 5
presents the recent progress on protein
ontology and the use of related semantic
Web technologies to enable the readers
to make the connection between accu-
mulating biological data collection and
integration trends.
The focus of Chapters 69 is to exam-
ine in detail how to develop the biologi-
cal data management techniques to
process and analyze high-throughput
biomedical data via case studies. Chap-
ter 6 describes the quality control techni-
ques to reduce variations during the
execution of transcriptomics and proteo-
mics data collection steps. Chapter 7
discusses the biological sequence man-
agement experience for a fungi genomics
project and emphasizes the development
keys: iterative development approach,
intensive discussion with laboratory sub-
ject-matter experts, and movement
beyond data management to knowledge
management. Chapter 8 investigates data
management and integration methodolo-
gies for microarray-based functional
genomics studies with important chal-
lenges in integration and interoperability
among different data repositories or data-
bases. Finally, Chapter 9 presents the
data management challenges and oppor-
tunities for mass spectrometry-based ex-
pression proteomics.
The focus of Chapters 10 and 11 is to
engage in practical aspects to demonstrate
howto apply bioinformatics data manage-
ment for drug discoveries. First, Chapter
10 introduces the fundamental drug dis-
covery concepts based on macromolecular
structural modeling; then, Chapter 11 dis-
cusses the data management system with
high-throughput drug compound screen-
ing to conclude the book.
Paul King
The book introduces
the readers to core
biological data
modeling
techniques,
biological database
resources, ontology
concepts, and data
management
challenges for high-
throughput data.
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Overall, the book is quite extensive,
covering aspects from basic bioinfor-
matics concepts to biological data model-
ing techniques and to real-world
integration/system interoperability chal-
lenges. Most of the topics in the text
emphasize the theory, current research
and literature findings, and future direc-
tions associated with biological data
modeling and data management imple-
mentations that may require skills and
knowledge in bioinformatics, molecular
biology, database design, data modeling,
and XML at a graduate level. The re-
viewer, therefore, believes that this book
is an excellent reference for bioinfor-
matics graduate students, researchers,
clinical scientists, clinical data architects,
and drug discovery professionals inter-
ested in managing postgenome biological
data and systems.
Zhengwu Lu
San Jose, CA
Information Retrieval in
Biomedicine: Natural
Language Processing for
Knowledge Integration
Violaine Prince and Mathieu Roche, IGI
Global, 2009. ISBN: 978-1-60566-274-
9, 432 pages, Hardcover, US$225.00.
Natural language processing (NLP)
is a subfield of computational sciences
addressing the operation and manage-
ment of texts as inputs or outputs of
computational devices. As such, this
domain includes a large amount of dis-
tinct topics, depending on which a par-
ticular service is considered. In todays
Internet-based knowledge-sharing world,
NLP is highly solicited by various scien-
tific communities as a tremendous help
for the following endeavors: 1) informa-
tion retrieval and knowledge extraction;
2) knowledge integration to existing
devices; 3) using and applying existing
knowledge structures for services in in-
formation retrieval.
This book is written by a large num-
ber of subject experts, comprises six
sections, and provides relevant theoreti-
cal frameworks and the latest empirical
research findings in the area of BioNLP
according to a linguistic granularity. As
a critical mass of advanced knowledge,
this book presents original applications,
going beyond existing publications
while opening up the road for a broader
use of NLP in biomedicine. The chapter
Text Mining in Biomedicine is an
introduction written by Sophia Anania-
dou, one of the most renowned leaders
in BioNLP, and it presents the ground-
ing principles of NLP and emphasizes
the needs, the requirements, the nature
of the issues and their stakes, and the
achievements of the state of the art. The
core sections of the book are four: three
that follow the NLP granularity scope,
ranging fromthe lexical level in Section
I to the sentence level in Section II and
to the discourse level in Section III, and
the one devoted to selected existing
software tools (Section IV). Chapters
belonging to these sections are numbered
fromII to XIX.
Section I, named Works at a Lexical
Level: Crossroads Between NLP and
Ontological Knowledge Management,
includes nine chapters and is by far the
largest, since the research on words, con-
cepts, and word-to-word relations is well
established and has reached a recognized
maturity. The order in which the chap-
ters have been organized is set up by the
following pattern: 1) using existing
resources to perform document-process-
ing tasks: indexation (Chapter II), cate-
gorization (Chapter III), and information
retrieval (Chapter IV). Indexation and
categorization could be seen as a prereq-
uisite to an intelligent information
retrieval, as they prestructure textual
data according to topics, domain, key-
words, or interest areas; 2) dealing with
the cross-linguistic terminological prob-
lem: from a specialists language to gen-
eral language within one tongue
(Chapter V) or across different tongues
(Chapter VI); 3) enriching terminology:
the beginning of a strong lexical NLP
involvement (Chapter VII); 4) increasing
lexical NLP involvement in biomedical
application (Chapter VIII).
Section II, titled Going Beyond
Words: NLP Approaches Involving the
Sentence Level, includes three chap-
ters and concentrates on research that
looks at words in context and investi-
gates broader linguistic granularity to
overcome ambiguity of terms. Chapters
VII and VIII have shown the effects of
ambiguity: intrinsic ambiguity (local-
ized in ontologies themselves) and
induced ambiguity detected in texts.
The sentence context might erase lexical
ambiguities effect in some cases. Chap-
ter IX, named Information Extraction
of Protein Phosphorylation from Bio-
medical Literature, presents a rule-
based system to capture the lexical, syn-
tactical, and semantic constraints found
in sentences expressing phosphorylation
information from MEDLINE abstracts,
since isolated words or phrases could
potentially be thematic clues but can by
no means account for the different
stages of a process. Chapter X, Cor-
Tag: A Language for a Contextual
Tagging of the Words Within Their
Sentence, complements the preced-
ing one, presents a design to extract
contextual knowledge from senten-
ces, deals with knowledge discovery
and how to induce relations between Digital Object Identifier 10.1109/MEMB.2009.935707
Natural language
processing (NLP)
is a subfield of
computational
sciences
addressing the
operation and
management of
texts as inputs or
outputs of
computational
devices.
Book Reviews (continued)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
concepts recognized in texts. Chapter
XI, titled Analyzing the Text of Clini-
cal Literature for Question Answer-
ing, introduces a method, design, and
system to not only deal with complex
information at the sentence level but
also present argumentative articulation
between fragments of sentences and
intersection between sentence-level and
discourse-level investigations. This
chapter is also highly recommended to
readers who need an accurate survey of
current question-answering systems.
Section III, titled Pragmatics, Dis-
course Structures and Segment Level as
the Last Stage in the NLP Offer to Biome-
dicine, groups three chapters and pres-
ents works on mining techniques, which
focus more on structures than on words,
on the use of neural network architecture
to improve retrieval, and the need in clini-
cal domains, taking the reader from the
theoretical to the practical applications of
mining techniques. Chapter XII, Dis-
course Processing for Text Mining,
presents discourse processing, the nature
and effects of text structures, as well as a
broad panel of intersentence relations to
readers and writers to organize informa-
tion and knowledge in texts. Chapter XIII,
A Neural Network Approach Imple-
menting Nonlinear Relevance Feedback
to Improve the Performance of Medical
Information Retrieval Systems, attempts
to mathematically ground information-
reliability hypothesis and pragmatics.
Chapter XIV, Extracting Patient Case
Profiles with Domain-Specific Semantic
Categories, is focused on increasing
pragmatics involvement via tackling a
particular issue: the roles of fragments of
patient medical records as diagnosis or
symptomclues.
Section IV, named NLP Software
for IR in Biomedicine, is dedicated to
software tools dealing with information
or knowledge extraction or verification
(Chapters XVXIX). Reviews on dif-
ferent NLP software developed and
used for IR in biomedicine should ena-
ble readers awareness of the software
state and tailor/adapt/compare it with
more theoretical and methodological
studies to assess the gap between semi-
automatic and automatic productions.
Chapters IIXIX have presented,
through four main sections, the ongoing
NLP research, theoretical advances, and
existing software tools, produced by
BioNLP, which is a multidisciplinary
area where artificial intelligence (AI),
NLP, statistics, applications domains,
data management, coding, biology, and
clinical medicine interact and evolve.
Finally, the last section, Conclusion
and Perspectives, with its sole chapter,
the twentieth titled, Analyzing Clinical
Notes for Translation Research: Back to
the Future, presents a window open on
the future. The chapter chooses, as a data
field to plough, a type of text that is highly
operational: no academic style but some-
thing very heterogeneous, mixing images,
sound, text, and in text, introducing abbre-
viations, acronyms, and a real-world
scenario of organizing knowledge. In this
conclusion, shifting frombiology to medi-
cine is a way to shift fromacademic genre
to a practitioner type of writing in which
general language is disturbed and noisy.
Overall, the book describes the differ-
ent contributions in the light of our
approach to NLP interaction with the
application domains in biomedicine.
Authors of these volume chapters are very
representative of the current research in
the field and spread across the globe.
Though prepared for advanced readers
with various challenging discussions, this
book presents a thoughtful arrangement of
works that provide great insight into the
field: achievements, current challenges,
and the future of dependable automated
information retrieval. The reviewer, there-
fore, believes that this is an excellent book
for researchers or graduate students inter-
ested in NLP, AI, and linguistics, as well
as those interested in the most recent
advances in information management and
retrieval in a health-care setting.
Zhengwu Lu
San Jose, CA
Enhanced Visualization:
Making Space for 3-D Images
Barry G. Blundell, Wiley, Hoboken,
NJ, 2007. 425 pages, US$112.50.
This book, Enhanced Visualization:
Making Space for 3-D Images, has a
peculiar title and subject: three-dimen-
sional (3-D) images. It is peculiar
because we all know how 3-D images
are generated; so, howcan Blundell write
a 425-page book about 3-D images?
Well, it turns out that we are wrong.
Blundell received a Ph.D. degree in
physics from the University of Man-
chester, England. He has obviously
devoted many years to the development
of 3-D systems.
The book has nothing to say about the
3-D stereoscopic images that are viewed
through redgreen eye filters looking at
appropriate superimposed outputs of a
pair of redgreen cameras. This binocu-
lar vision is important for 3-D material
that is a few meters or less distant from
the eyes; beyond that, the red and green
images are almost identical (except for
the exaggerated views generated by the
photographer). There is, of course, no
true color reproduction. More serious is
the fact that, if the observer moves his/
her head, the images do not change, in
contrast with true 3-Dexperiences.
Holographic techniques are also
absent in this book.
Blundells book is about true 3-D
volumes that show voxels in contrast
with two-dimensional (2-D) screens
that show pixels.
There is an important biomedical
engineering application for voxels: in
tomography, we get thin slices of body
tissues; this information would be
greatly enhanced if the slices could be
used to build up a 3-D view. One could
then see, at a glance, how the organs are
related to each other.
Another application is for aircraft
flight controllers, who would see the
airplanes as flying in a volume that has
height, width, and distance coordinates.
Alas, one cannot realistically pur-
chase a 3-D equipment. You can try the
two Web sites that Blundell mentions:
www.actuality-systems.com and www.
lightspacetech.com. However, engi-
neers should read this book because it
gives the technical details about the
ingenious solutions to the 3-D image
problem. You could be the one to Digital Object Identifier 10.1109/MEMB.2009.935705
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
____
realistically harness 3-D images for the
benefit of mankind!
Listed as part of references, the
volume ends with a compilation of 77
patents: 57 U.S., 12 U.K., six European,
and two Japanese patents. In each case,
the number, name of the inventor(s),
year granted, and invention title are
given. These patents are the foundation
upon which the book is constructed.
We have an image of the inventor
building a prototype (we hope) and then
dashing off to the patent office with an
application. But after the patent was
granted, almost all of these people did
not attempt to manufacture the 3-D
invention because, realistically, one
could not recover its cost. None of the
invention led to a mass-produced box
for home entertainment.
Blundell centers a discussion around
each representative invention by repro-
ducing the appropriate figures from the
patent application, often embellished
with a figure or two of his own. He has
thus performed a valuable service for
any future would-be 3-D inventors, by
exposing them in advance to many of
the possible pitfalls.
The text frequently refers to two other
books: Volumetric Three-Dimensional
Display Systems by Barry G. Blundell
and Adam J. Schwarz, Wiley (2000),
and Creative 3-D Display and Interac-
tion Interfaces: A Trans-Disciplinary
Approach by Blundell and Schwarz,
Wiley (2006). There is a mystery here
why did Wiley publish three books in
rapid succession (2000, 2006, 2007) on
the same narrowsubject?
The volume contains 12 chapters.
Each begins with an abstract (called an
Introduction), followed by text, followed
by conclusions (called Discussion), and
ending (except for Chapter 12) with
homework problems (called Investiga-
tions). Here are some of the brief com-
ments about each chapter:
Chapter 1: Setting the Scene. There
are two main 3-D paradigmsvolumet-
ric and varifocal. Ideally, in volumetric
displays, we have an image that appears
in a volume that has width, height, and
depth, but it can sometimes have the
round shape of a disc. Generally, the vox-
els that comprise an image emit light and
are translucent or transparent. In a varifo-
cal display, depth is generated by varying
the focal length of a lens. For example,
the lens can consist of a stretched Mylar
mirrored surface sheet, whose curvature
is modified by a rapidly moving loud-
speaker drive. A minimum of 30 Hz is
needed to avoid the perception of a
flicker. Looking at the front surface, the
virtual image moves back and forth; it
appears to focus behind the sheet.
Chapter 2: Aspects of the Visual
System, is an excellent review of the
characteristics of the human visual sys-
temthat are important in 3-Dviewing.
Chapter 3: Creative 3-D Display
Techniques. The chapter is restricted
to the volumetric approach. Various
arrangements and their connections to
stereoscopic vision are considered.
Chapter 4: The Swept-Volume
Approach, is restricted to volumetric
designs in which the display employs a
mechanical system that causes a surface
(or 3-D structure) to rapidly and repeat-
edly sweep through a physical volume.
Three systems are featured (among
others): 1) the rotational motion of a pla-
nar screen (here, the screen rotates around
an axis in the plane of the screen); 2) the
translational motion of a planar screen
(yes, the screen flops back and forth at 30
Hz!); and 3) the rotational motion of a 3-
Dstructure (think of a rotating translucent
or transparent disc with voxels that light
up in accordance with external control).
Chapter 5: The Static-Volume
Approach. Here, we do not rely on
mechanical motion for imagespace
creation. There is a complete discussion
of fluorescence: two beams intersecting
to give the stepwise excitation of
fluorescence.
Chapter 6: Swept-Volume Systems:
Limited Viewing Freedom. According
to an Appendix Table A.1, 17 different
systems are discussed.
Chapter 7: Low Parallelism Swept-
Volume Systems. According to Table
A.2, 19 different systems are reviewed.
Chapter 8: Highly Parallel Swept-
Volume Systems. According to Table
A.3, 11 different systems are considered.
Chapter 9: Static-Volume Systems:
Example Implementations, highlights
the aspects of the display techniques of
Chapter 5.
Chapter 10: Varifocal Mirror
Techniques.
Chapter 11: The Graphics Pipeline
and Interaction Issues. The graphics
pipeline is the source of data to be dis-
played on a 3-D image. The pipeline
output should be optimized to be
compatible with the 3-D system. Vari-
ous arrangements are considered.
Chapter 12: General Discussion:
Suggestions du Jour. Blundell writes
that . . . we focus primarily on a number
of display paradigms that do not appear
to be the subject of current research but
which seem to the author to warrant
further investigation. As he says else-
where, Many of these patents have now
lapsed, and approaches have returned to
the public domain. (p. 188.)
Blundell is addicted to footnotes. It is
scarcely possible to read a paragraph of
text without being interrupted by a foot-
note. Of course, Blundell would say that
the reader is not being forced to shift
down to the bottom of the page, but he
would be wrong. Most of the footnotes
convey useful information; almost all
of them should be moved up, modified,
and seamlessly incorporated into the
text. This reviewer has written many
books without finding it necessary to
use footnotes.
There are practically no typographical
errors. The list of referenced publica-
tions is extensive. The three-page index,
however, is inadequate; one complaint is
that many of the acronyms that appear in
the book are not listed in the index.
A perpetual problem, especially with
the patent drawings, is that the observer
is omitted. It is not clear where the
observers eyes are in relation to the
equipment. I suspect that, if the equip-
ment is actually constructed, an observer
would occasionally be guillotined!
Sid Deutsch
Book Reviews (continued)
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
_______________________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
Conference Calendar
1418 APRIL 2010
Seventh IEEE International
Symposium on Biomedical
Imaging, ISBI 2010
Rotterdam, The Netherlands
Contact: EMBS Executive Office
Phone: 31 732 981 3451
Fax: 31 732 465 6435
E-mail: emb-conferences@ieee.org
14 SEPTEMBER 2010
32nd Annual International
Conference of the IEEE Engineering
in Medicine and Biology Society
Buenos Aires, Argentina
Phone: 54 732 981 3451
Web: http://embc2010.embs.org
2023 APRIL 2011
Eighth IEEE International
Symposium on Biomedical
Imaging, ISBI 2011
Chicago, Illinois
Phone: 1 732 981 3451
Fax: 1 732 465 6435
27 APRIL1 MAY 2011
5th International IEEE EMBS
Conference on Neural
Engineering
Cancun, Mexico
Phone: 1 732 981 3451
14 SEPTEMBER 2011
33rd Annual International
Conference of the IEEE
Engineering in Medicine
and Biology Society
Boston, Massachusetts
Phone: 1 732 981 3451
29 AUGUST2 SEPTEMBER 2012
34th Annual International
Conference of the IEEE
Engineering in Medicine
and Biology Society
San Diego, California
Phone: 1 732 981 3451
_________________ _________________ _________________
_________________
_________________
_________________
_____________________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
____________
ADVERTISERS INDEX
The Advertisers Index contained in this issue is compiled as a service to our readers and advertisers: the publisher is not liable for errors or omissions
although every effort is made to ensure its accuracy. Be sure to let our advertisers know you found them through IEEE Engineering in Medicine and
Biology Magazine.
Advertiser Page
Biopac Systems, Inc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CVR 4
URL Phone
. . . . . . . . . . . . . www.biopac.com . . . . . . . . . . . . . . . . . . . . . 1 805 685 0066
Coming in May
Wearable Technology
Call your local advertising sales representative today!
Space reservations: April 9
Material deadline: April 14
Susan E. Schneiderman +1 732 562 3946 Tel.
Business Development Manager +1 732 981 1855 Fax.
ADVERTISING SALES OFFICES
IEEE MAGAZINES
445 Hoes Lane, Piscataway, NJ 08854
www.ieee.org/ieeemedia
James A. Vick
Staff Director,
Advertising Business
+1 212 419 7767;
Fax: +1 212 419 7589
jv.ieeemedia@ieee.org
Susan E. Schneiderman
Business Development Manager
+1 732 562 3946;
Fax: +1 732 981 1855
Marion Delaney
Advertising Sales Director
+1 415 863 4717;
Fax + 1 415 863 4717
md.ieeemedia@ieee.org
Product Advertising
Midatlantic
Lisa Rinaldo
+1 732 772 0160;
Fax: +1 732 772 0164
lr.ieeemedia@ieee.org
NY, NJ, PA, DE, MD,
DC, KY, WV
New England/
Eastern Canada
Jody Estabrook
+1 774 283 4528;
Fax: +1 774 283 4527
je.ieeemedia@ieee.org
CT, ME, VT, NH, MA, RI
Canada: Nova Scotia, Prince
Edward Island, Newfoundland,
New Brunswick, Quebec
Southeast
Thomas Flynn
+1 770 645 2944
Fax: +1 770 993 4423
tf.ieeemedia@ieee.org
VA, NC, SC, GA, FL, AL, MS, TN
Midwest/Central Canada
Dave Jones
+1 708 442 5633
Fax: +1 708 442 7620
dj.ieeemedia@ieee.org
IL, IA, KS, MN, MO, NE,
ND, OH, SD, WI,
Canada: Manitoba, Saskatchewan,
Alberta
Midwest/Ontario, Canada
Will Hamilton
+1 269 381 2156
Fax: +1 269 381 2556
wh.ieeemedia@ieee.org
IN, MI Canada: Ontario
Southwest
Shaun Mehr
+1 949 923 1660
Fax: +1 775 908 2104
sm.ieeemedia@ieee.org
AR, LA, OK, TX
West Coast/Northwest/
Western Canada
Marshall Rubin
+1 818 888 2407
Fax: +1 818 888 4907
mr.ieeemedia@ieee.org
AZ, CO, HI, NM, NV, UT,
AK, ID, MT, WY, OR, WA, CA.
Canada: British Columbia
Europe/Africa/Middle East
Heleen Vodegel
+44 1875 825 700;
Fax: +44 1875 825 701
hv.ieeemedia@ieee.org
Europe, Africa, Middle East
Asia/Far East/Pacific Rim
Susan Schneiderman
+1 732 562 3946;
Fax: +1 732 981 1855
Asia, Far East, Pacific Rim,
Australia, New Zealand
Recruitment Advertising
Midatlantic
Lisa Rinaldo
+1 732 772 0160
Fax: +1 732 772 0164
lr.ieeemedia@ieee.org
NY, NJ, CT, PA, DE, MD, DC,
KY, WV
New England/Eastern Canada
John Restchack
+1 212 419 7578;
Fax: +1 212 419 7589
j.restchack@ieee.org
ME, VT, NH, MA, RI
Canada: Quebec, Nova Scotia,
Prince Edward Island,
Newfoundland, New Brunswick
Southeast
Cathy Flynn
+1 770 645 2944
Fax: +1 770 993 4423
cf.ieeemedia@ieee.org
VA, NC, SC, GA, FL, AL, MS, TN
Midwest/Texas/Central Canada
Darcy Giovingo
+1 847 498 4520
Fax: +1 847 498 5911
dg.ieeemedia@ieee.org
AR, IL, IN, IA, KS, LA, MI, MN,
MO, NE, ND, SD, OH, OK, TX,
WI,
Canada: Ontario, Manitoba,
Saskatchewan, Alberta
West Coast/Mountain States/
Southwest/Asia/Far East/
Pacific Rim
Tim Matteson
+1 310 836 4064;
Fax: +1 310 836 4067
tm.ieeemedia@ieee.org
AK, AZ, CA, CO, HI, ID, MT, NM,
NV, OR, UT, WA, WY
Canada: British Columbia.
Asia, Far East, Pacific Rim
Europe/Africa/Middle East
Heleen Vodegel
+44 1875 825 700;
Fax: +44 1875 825 701
hv.ieeemedia@ieee.org
Europe, Africa, Middle East
______________
____________
___________
___________
____________
____________
____________
____________
____________
____________
____________
____________
____________
____________
____________
____________
____________
___________
___________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
[os
/
te
e
blast
/
]
Etymology: Gk, osteon blastos, germ
A bone-forming cell that is derived from the embry-
onic mesenchyme and, during the early development of
the skeleton, differentiates from a fibroblast to function
in the formation of bone tissue. Osteoblasts synthesize
the collagen and glycoproteins to form the osteoid
matrix and, with growth, develop into osteocytes.
From Mosbys Medical Dictionary, 8th ed. ' 2009,
Elsevier.
Editors Note
In recent weeks, Prof. Tayfun Akg
ul and I have
gotten to know each other through e-mails, and
we have exchanged ideas for some of his car-
toons that will be appearing in this magazine. A
few of these, such as the one on this page, may
distort the meaning of important terms and con-
cepts in biomedical engineering and beyond
just a little bit. So I thought that we might also
include a more acceptable definition as well just
to set the record straight.
MRN
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
2010 EMBS ADMINISTRATIVE COMMITTEE
(Term expires December 31 of year indicated)

Executive Committee

Bin He
President 09-10
University of
Minnesota
binhe@umn.edu
+1.612.626.1115

Zhi Pei Liang
President-Elect 10
University of Illinois,
Champaign
z-liang@illinois.edu
+1 217.244.4023

Craig Hartley
V.P. Finances 10
Baylor College of
Medicine
cjhartley@ieee.org
+1.713.798.4195

Andrew Laine
V.P. Publications 10
Columbia University
Ak418@columbia.edu
+1.212.854.6539

Nigel Lovell
V.P. Conferences 11
University of Sydney
n.lovell@unsw.edu.au
+1.217.244.4023

Gudrun Zahlmann
V.P. Member &
Student 11
Roche Switzerland
Gudrun.zahlmann@
roce.com
+54.3783.423126

Laura J. Wolf
Executive Director
IEEE EMBS
emb-exec@ieee.org
+1.732.981.3433
Administrative Committee

North America
Robert Butera 10
Georgia Tech
rbutera@gatech.edu

Paolo Bonato 10
Harvard Medical School
pbonato@partners.org

Steve Johnson 11
Amgen Inc.
sfjohnso@ieee.org

Michael Khoo 11
University of Southern California
khoo@usc.edu

Carolyn McGregor 11
University of Ontario Institute
of Technology (UOIT)
c.mcgregor@ieee.org

Nick Chbat 12
Philips NA
chbat@philips.com

Mike McShane 12
University of Texas, A&M
mcshane@bme.tamu.edu

Latin America
Martha Zequera Daz 11
Pontifica Universidad J averiana
mzequera@javeriana.edu.co

Europe
Christian Roux 10
ENST Brest
c.roux@ieee.org

Christopher James 12
University of Southampton
c.j.james@ieee.org

Ilkka Korhonen 12
ilkka.korhonen@vtt.fi

Middle East & Africa
Ahmed Morsy 10
Cairo University
amorsy@ieee.org

Asia Pacific
Kenji Sunagawa 10
Kyushu Univ School of Medicine
sunagawa@cardiol.med.kyushuu
.ac.jp

Makoto Yoshizawa 11
Tohoku University
yoshizawa@ieee.org

Shaun Cloherty 12
University of Sydney NW
s.cloherty@ieee.org

Student Rep
Cristian A. Linte 10
University of Western Ontario
clinte@imaging.robarts.ca

GOLD Rep
Matthias Reumann 11
IBM
mkreuman@us.ibm.com

Editors
EMB Magazine
Michael Neuman
Michigan Technical University
mneuman@mtu.edu

Transactions on Neural
Systems & Rehabilitation
Engineering
Nitish Thakor
tnsreditor@gmail.com

Transactions on Biomedical
Engineering
Bruce Wheeler
Office@tbme.embs.org

Transactions on
Information Technology in
Biomedicine
Yuan-Ting Zhang
ytzhang@ee.cuhk.edu.hk

Reviews in Biomedical
Engineering
J ose Principe
principe@cnel.ufl.edu

EMBS Book Series
Metin Akay
metin.akay@asu.edu

Committee Chairs

Awards: J . Principe
Conference: N. Lovell
Constitution & Bylaws: Z.P. Liang
Chapter Development: M. Zequera Diaz
Distinguished Lecturer: Z.P. Liang
EAB/CEAA: P. Benkeser
Education: R. Butera
Emerging Technologies: A. Dhawan
Ethics: Z. Taqvi
Fellows: R. Barr
Finance: C. Hartley
History: D. Rhees
Industry Relations: D. Panescu
Infostructure: H. van Oostrom
Member: G. Zahlmann
Nominating: Y. Kim
Publications: A. Laine
Standards: C. Carey
Strategic Planning: Z.P. Liang
Student Activities: M. McShane
WIE Rep: S. Demir
EMBS Webmaster: H. van Oostrom

Technical Activities: Z.P. Liang
Biomedical Informatics: N. Saranummi
Bionanotechnology: C. Ruggiero
BioRobotics: J . Patton
BioSignal Processing: Y. Kahya
Cardio-Pulmonary Systems: P. Iaizzio
Clinical Engineering: P. Bonato
Computational Biology & Physiome:
P. Hunter and N. Lovell
Medical Imaging: S. Wright
Microtechnology & BioMEMS:
A. Khademhosseinni
Neuroengineering: M. Akay
Wearable Sensors: T. Tamura
Therapeutic Systems: D. Haemmerich
COMAR: R. Tell
Grant Development: C. McGregor
Member Recognition: A. Dhawan
Student Paper Competition: L. Ding
Public Outreach & Relations: S. Demir

IEEE-USA Policy Committees
Transportation & Aerospace: R. Gantenbein
Communication & Info: E. Topsakal
Critical Infrastructure: G. Zouradakis
Defense & Engineering: N. Diakides
Medical Technology: D. Sigg

AIMBE: B. He, Z.P. Liang, A. Laine, N. Thakor
EAMBES: C. Roux, I. Korhonen, C.J . J ames
IFMBE: B. He, N. Saranummi, Y.T. Zhang
Nanotechnology Council: L. Nagahara, J .P. Wang
Sensors Council: W. Besio
Biometrics Council: J . Principe, N. Chbat
IEEE RFID Technical Committee: C.J . J ames

Conferences
2010: R. Armentano, D.Hudson, J . Monzon,
J . Patton Buenos Aires, Argentina
2011: P. Bonato, C. Brenan, A. Laine, M. Akay
Boston, MA USA
2012: San Diego, CA USA
2013: K. Sunagawa, Osaka, J apan

___________
_________ __________ ________ __________
__________
__________
__________
___________ ______________
_________
______________
__________
_________
________
_______
________________
____________
_________
_____________
____________
___
__________
__________
___________
__________
____________
___________
__________
_____________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
___________
______________
_______________
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F
IEEE ENGINEERINGIN
MEDICINEANDBIOLOGY
B
A
M S a
G E
F

EMB 20100301 Mar 2010

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EMB 20100301 Mar 2010

Uploaded by

Copyright:

Available Formats

IEEE ENGINEERINGIN

C, FRANCIS J. DOYLE, III,

You might also like