1 blood bank, intensive care, ambulatory care, obstetrical and labor and delivery care, etc.), these monolithic systems have generally evolved over decades. In this model, the application programs are aware of the database structure when they query for particular items in the database; there- fore any changes in the database structure or the presentation layer usually mean that all the application programs need to be changed. It is also extremely difficult to switch platforms or vendors without signifi- cant effort and simultaneous disruption of enterprise operations (so called big-bang approach to change). In actuality, the architecture of the var- ious extant systems is not as pure as the above characterizations might suggest. If the monolithic system was originally archi- tected in a way that separated the presen- tation layer and the database layer from the actual application code (services act as intermediaries between the application code and the database or the presentation layer), then it is possible to make changes in the architecture without changing all of the programs. Even the most monolithic systems generally have some interfaced applications; they typically accept input from at least one modular component (generally a clinical laboratory system). As vendors have acquired specialty applica- tions (e.g. surgery scheduling, blood gas measurement systems. etc.), they have begun to interface them to the clinical information system. This purchase (rather than build) approach by vendors them- selves gives credence to the fact that it is often less expensive to buy and interface an existing system than to rebuild something that already exists from scratch. Building a Comprehensive Clinical Information System from Components The Approach at Intermountain Health Care P. D. Clayton 1, 2 , S. P. Narus 1, 2 , S. M. Huff 1, 2 , T. A. Pryor 1, 2 , P. J. Haug 1, 2 , T. Larkin 1 , S. Matney 1 , R. S. Evans 1,2 , B. H. Rocha 1 , W. A. Bowes, III 1 , F. T. Holston 1 , M. L. Gundersen 1 1 Intermountain Health Care, Salt Lake City UT, USA, 2 University of Utah, USA Introduction In this paper we discuss the benefits and drawbacks of a component-based, open architecture approach to building clinical information systems and describe the im- plementation of such a system at Inter- mountain Health Care (IHC) in Salt Lake City, Utah. By open architecture we mean the ability to interface multiple different individual software applications from a variety of sources (purchased, and inter- nally developed) to a central unified clini- cal repository. This foundation architecture enables an authorized user to see all patient data regardless of the application in which they were collected and removes the need to enter redundant data into multiple systems. In contrast to open architecture, we use the term monolithic to mean that the appli- cations and the database are tightly inte- grated because all software applications have been developed using the same common environment, development tools, database definitions, query processes and structures, presentation layer and user in- terface look and feel. This approach infers a single owner of the system development process and architecture. To many, the terms monolithic and integrated have simi- lar meanings (1). This monolithic approach is used by most commercial vendors, and, until recently, most internally developed systems (2-7). Because it takes so long to develop all of the programs necessary to support the provision of health care in a variety of settings (nurse documentation, surgery scheduling, patient visit scheduling, oncology treatment, nutrition support, pharmacy, clinical laboratory, pathology, Summary Objectives: To discuss the advantages and disadvan- tages of an interfaced approach to clinical information systems architecture. Methods: After many years of internally building almost all components of a hospital clinical informa- tion system (HELP) at Intermountain Health Care, we changed our architectural approach as we chose to encompass ambulatory as well as acute care. We now seek to interface applications from a variety of sources (including some that we build ourselves) to a clinical data repository that contains a longitudinal electronic patient record. Results: We have a total of 820 instances of inter- faces to 51 different applications. We process nearly 2 million transactions per day via our interface engine and feel that the reliability of the approach is accept- able. Interface costs constitute about four percent of our total information systems budget. The clinical database currently contains records for 1.45 m pat- ients and the response time for a query is 0.19sec. Discussion: Based upon our experience with both integrated (monolithic) and interfaced approaches, we conclude that for those with the expertise and resources to do so, the interfaced approach offers an attractive alternative to systems provided by a single vendor. We expect the advantages of this approach to increase as the costs of interfaces are reduced in the future as standards for vocabulary and messaging become increasingly mature and functional. Keywords HELP, IHC, CDR Methods Inf Med 2003; 42: 17 On the other hand, the interfaced systems are also somewhat monolithic. The developers of such systems generally try to avoid too much granularity, e.g. they gene- rally use components that encompass an entire discipline such as pharmacy, electro- cardiography, nurse charting or surgery scheduling. They dont use separate pro- grams to review problems, transcribed reports, medications and clinical laboratory test results; even though all of these data may come from separate sources. An institution starting from scratch to- day has three options. The first is to build all of the applications from scratch. This has been a successful model for most currently existing systems. If an individual institution were to pursue this approach today, it would effectively mean duplicating for a single site, the effort of a vendor that can spread development costs across many customers. It should be recognized that in addition to the expense of such an ap- proach, the time required until comprehen- sive functionality is obtained will be exces- sive. The second option is to purchase all ap- plications from a single vendor to insure ea- se of implementation and seamless integra- tion. This monolithic approach is probably the least expensive and least risky in the short run. It is not without problems how- ever. One cannot predict whether the ven- dor will stay in business over the 20-25 year system lifetime. Switching costs may perpe- tuate the use of a mediocre system long past the introduction of new technological breakthroughs. Perhaps the biggest im- mediate drawback is that most if not all of the user constituencies will feel somewhat dis- appointed with the selected system because of the need to compromise in the selec- tion of a single vendor. For example, the anesthesiologists may have just seen a won- derful new application at the most recent national meeting and the contemplated vendor offering in this area is somewhat pe- destrian. Even after the best overall vendor is selected, contracting, implementing, and training processes can extend for 5-10 years depending upon the size of the enterprise. Even if there were no financial limitations, human beings cannot accommodate chan- ges that are too rapid. The third option for people who are not satisfied with the quality and breadth of the offerings of a single vendor or the resulting dependency (all the eggs in one basket and no control over the pace of technological evolution) is to build the system from inter- faced, disparate components and have each component feed data into or retrieve data from a common longitudinal repository. In spite of years of experience and success in building an integrated monolithic system (HELP) (7), we have now undertaken the interface approach at Intermountain Health Care. In 1982, Don Simborg and his co- authors (8) at the University of San Fran- cisco demonstrated the use of multiple ope- rating systems on multiple processors that were bound together via a network to crea- te a working clinical information system. In 1988 Simborg (9) published a paper describing a standard known as HL7 for linking disparate systems. Since that time, systems that reflect this approach (10-21) have begun to emerge. We will now de- scribe our experience with this approach. Methods Intermountain Health Care (IHC) is a not for profit health care delivery system head- quartered in Salt Lake City Utah. It opera- tes 22 hospitals (2500 beds), 72 clinics, and an insurance plan that covers approxima- tely 500,000 people. It also coordinates insurance for another 500,000 people who are covered by affiliated insurers. Each year we admit over 100,000 patients to our hospitals which are located throughout the state of Utah (and one in Idaho) and pro- vide for more than 3,000,000 ambulatory clinic visits. Approximately half of the people living in Utah are served by IHC. Until the 90s IHC was primarily focused on delivering acute care. With the emer- gence of our own insurance plan and the creation of a physicians division, which staffed ambulatory clinics with employed physicians, we were directed to expand our in-formation systems beyond the walls of the hospitals. The options considered were to remodel the HELP system that had been developed at LDS hospital and installed at the other 21 IHC hospitals, purchase a new system from a commercial vendor, or build a system from components. For reasons that we explore in this paper, the last alter- native was chosen. Our system architecture is shown in Fig. 1. We call this system HELP2 to emphasize that the legacy and philosophy of the original HELP system are continued in this new environment. We will now discuss the various components depicted in this figure. Master Patient Index. The foundation of HELP2 is a central, patient-oriented, clini- cal data repository (CDR) that IHC co- developed with the 3M Corporation. All Clayton et al. 2 Methods Inf Med 1/2003 Fig. 1 An overview of the interfa- ced architecture used at In- termountain Health Care. Each arrow represents one or more types of interfaces between component appli- cations. In all there are 51 different components that are interfaced. Two-thirds of the interfaces are for cli- nical applications and the rest are for financial pur- poses. local identification numbers are maintai- ned in the master index so that local users continue to use their traditional numbering or identification processes to access the patient data but the data are stored in the repository using the unique identifier as the key. We interfaced the hospital and clinic patient registration systems to the 3M master patient index and used the associa- ted services to allow clerical personnel in those locations to query this index during the registration process. This connection allowed the clerks to ascertain whether the patient was uniquely new to the system or already existed in our enterprise master patient index. Clinical Data Repository Database. The second part of the clinical data repository is the patient data structure. The data are stored as defined codes according to data models defined for various types of clinical events. These models are structured using ASN.1 (22-24), a syntax very similar in con- cept to XML. Objects can be recursively nested within an event; this capability allows reuse of fundamental concepts in the process of creating broadly generic data models. In addition to the event object that is stored using SQL services to access a Unix-based Oracle database, we also store references to the individual concepts/ele- ments of each event in a separate table so that we can efficiently search for particular individual elements that may be part of a larger event. Pointers to waveforms, ima- ges, scanned documents etc, are stored as part of the patient record even though they may exist in a separate file (e.g. a Picture Archiving and Communication system). In order to preserve response time at the point of care, the CDR is tuned to opti- mize queries for a single patients data. When one wishes to search across a popula- tion of patients, we use the Enterprise Data Warehouse that contains the same clinical information as the CDR as well as addi- tional data from the billing and claims systems. Health term Data Dictionary. The Health term Data Dictionary (HDD) (25) was also developed in collaboration with the 3M Corporation and its content con- sists of the uninstantiated data models for each of the various types of clinical events, and a set of unique Numerically Coded Identifiers (NCIDs) for all medical con- cepts that can be stored as coded informa- tion in our database. The HDD also assigns a numerically coded identifier to each individual provider of services in our orga- nization. For each element/concept in the dictionary, synonyms and terms used by the interfaced application are stored. Interfaces. Each arrow in Fig. 1 repre- sents one or more interfaces between sepa- rate applications. These interfaces allow the native application to maintain its own data, and use its own local terminology. Thus when a message comes from one of the interfaced systems, the vocabulary in that message is translated from that of the origi- nating source to the canonical representa- tion of the central repository. Messages going in the reverse direction are translated back into the local vocabulary. The inter- face engine also changes the message struc- ture if necessary and then broadcasts the message to other subscribing applications as well as to the data store services for our clinical data repository. The eGate interface engine allows interfaces to be built that are largely soft-coded or table-driven. Soft- coded aspects of interfaces include message (record reformatting), routing, error hand- ling, and even I/O communication protocol translation. For clinical applications we use the HL7 message formats, and require that any products being considered for purchase provide a real time output using this stan- dard. There are error logs and queues which hold messages if a recipient applica- tion is unable to verify receipt and proper storage of the message. This approach does allow storage of redundant data in two or more places. For instance the clinical labo- ratory system has the patient registration information and a copy of the laboratory orders and keeps a copy of the results after it sends the result back to the CDR. Inference engine. We have purchased a commercial inference engine (Blaze Advisor) that runs in an XML and J2EE environment. This application is evoked by the services that store data in our repo- sitory or by sending a real-time message for synchronous activation. We have construc- ted data queries to the CDR that are used by this rules evaluation engine to retrieve data necessary to evaluate the rules and to store the results. We have also generated queries to the existing hospital based HELP system so that we can simultaneously query data from more than one system in order to evaluate a patients status. We will also use Blaze to send appropriate messages and in- sure that the messages have been read. Clinical Desktop. For some time we ha- ve been using a Windows-based clinical desktop shell to allow the user to seamless- ly navigate between various applications (order medications, see laboratory results, go to the literature, compose a clinic visit note). The context (user and patient) of these separate modules is maintained by this shell in an integrated way. We are now beginning to install a web-based clinical desktop that allows the user to Login via a call to our LDAP (Light weight Directory Access Protocol) compatible directory and to select a patient (single sign on) before choosing which application to run. After the user clicks to indicate the choice, this desktop can call various types of applica- tions (e.g. the original DOS based HELP application) as well as Windows and web- based applications. Until recently we have used hidden screens (web environment) or hard coded methods to evoke the various applications and to pass the context (user and patient). We are now able to use the Health Level Seven (HL7) Context Mana- gement Architecture version 1.3 (CCOW: Clinical Context Object Workgroup) pro- vided by a company known as Carefx in order to maintain seamless functionality as a user switches back and forth between applications on the desktop. Applications that do not use the interface engine. We have chosen to write several core applications for data entry and results review ourselves. These applications pro- vide high performance access to the data- base. The first of these programs is called Chart Notes. It is a form-based application that allows the user to enter data into fields on the form using text input or codes from pick lists. The data entry fields in the form are coded and directly mapped to the underlying data model. The Chart Note templates are constructed using a Win- dows-based editor and are stored in an XML format. The user can evoke this data Clinical Information System 3 Methods Inf Med 1/2003 entry tool in the Windows or the web envi- ronment without the need to rebuild the form. When the user hits save, the uniquely defined codes for each element are stored directly into the CDR in the appropriate instantiation of the data model. We also wrote in conjunction with 3M, a Windows-based, thick client application called the Clinical Workstation (CW). This application allows the physician to enter data (medication orders, visit notes, tem- plated data via Chart Notes, problems, allergies etc.) at the point of service. We actually run the application on Citrix servers and terminal emulation software to reduce the maintenance costs associated with installation of the thick client on every user desktop. Three features of the CW have enjoyed exceptional acceptance: Message Log, Hot Text, and Common Lists. The Message Log application lets mem- bers of a clinic staff communicate with each other asynchronously. A receptionist may write a note to the physician that says, Mrs. Jones called to ask for a refill of her xxx prescription. The physician can respond that the staff member should go ahead and notify the pharmacy to refill the prescrip- tion and call Mrs. Jones. Hot Text is used to construct narrative notes. A user can construct a variety of text macros to reflect common clinical situa- tions e.g. six week well-baby checkup or follow up visit for patient with diabetes. Rather than dictating of typing the entire note, the user activates the macro with a simple command and makes a small num- ber of changes (26), which reflect the actual patient status. Common Lists provide the ability to type a few letters and generate a dynami- cally pruned list of items that reflect what one is typing. For example, if one types the first few letters of a drug name, the list of fully qualified (dose, route, frequency, num- ber of tablets to dispense, and number of refills allowed are set to common default values) candidate prescriptions of that drug is displayed. Selection from the list finishes the prescription unless one wishes to modify any of these attributes. In addition to the CW applications that are primarily written to facilitate data entry, we have written a Results Review application to allow viewer access to the comprehensive clinical data that are avail- able in the CDR. This application allows us to restrict access based upon privacy con- siderations (users relationship to the pa- tient, location of the terminal, recent acti- vity of the patient, role of the user and relationship of the user to the patients primary care provider). In addition to our own displays of test results (including graphs over time) and narrative docu- ments, we link to commercially purchased applications as well. We show EKG wave- forms by linking to the GE/Marquette MUSE application and radiographs using the Amicus viewer. For problems, medica- tions or allergies we provide access to the literature by letting the user click on the infobutton (27). This feature generates the list of the answers to common que- stions for which our indexed/tagged literature resource can provide answers, e.g. is this drug safe to use during preg- nancy? The user selects the question and the answer appears. Results Our current clinical data repository con- tains the records for 1.48 million patients. The average record size is 24 kilobytes per Clayton et al. 4 Methods Inf Med 1/2003 Fig. 2 A comparison of the number of unique users who logon to one of our clinical applications each month. Most users still use the HELP hospital based system, but there has been steady growth of the use of HELP2 in the ambulatory setting. We will begin to move the hospital users to HELP2 in December of 2002. Comparison of Unique Users patient not counting waveforms, images and other scanned documents. The record size for the patient with the most data is presently 13 MB. Our system currently has 369 GB of storage capacity on a Storage Area Network (SAN). We have a redun- dant back up system and test and deve- lopment systems. We use UNIX for our database services (24 processors in 6 machines for Oracle and another 24 processors in 6 machines for the services). Two of these machines are for load ba- lancing. A query for patient data takes an average of 0.19 seconds during the busy part of the day. Our unscheduled downtime for the central database repository this year is 0.05% (3.58 hours/ 7200 hours). Our goal is to have less than one hour per year of unscheduled downtime. Our scheduled downtimes (if needed) occur once per month for approximately 3 hours (15 hours total through October of this year). We average 7,176,000 monthly queries against the database from 2416 individual clinical users who currently log on to the system at least once a month (Fig. 2 ). Most of these users are currently in the ambula- tory setting. In contrast, during the month of October 2002, 12,206 different individu- als used the hospital-based HELP system. The number of patients accessed via the new HELP2 system is currently (October 2002) 100,577 per month vs. 94,222 per month via the HELP hospital system. We will begin to switch users off of the original HELP system in December of 2002, although, by using the clinical desktop and CCOW, our users are free to seamlessly switch back and forth between HELP and HELP2 for the foreseeable future. Using our Clinical Workstation applica- tion, 981 individuals entered a prescription order during October 2002 ( Fig. 3 ). During this same time period 441 users entered a progress note and 302 individuals entered at least one coded problem into a patient record. We currently have interfaces to 51 diffe- rent unique applications. Sixteen of these are to external trading partners (financial transaction clearing houses, insurance com- panies, providers, etc.) Two-thirds of the interfaces are to clinical systems (clinical laboratory, PACS, blood bank, anatomic pathology, surgery scheduling, transcrip- tion, etc). The remaining third are to finan- cial systems. Each of these applications may require several interfaces. For example, we are working with the IDX CareCast product to implement computer-based phy- sician order entry and an evidence-based patient care management system. For these two applications, we actually needed to complete 18 interfaces (Staff in from our master provider tables, Admit discharge and transfer, orders out, orderable items in, problems in, problems out, nurse assess- ments, etc.). For this particular project, the costs of the interfaces amounted to 20% of the total application software costs. When we added the costs of education, know- ledge base creation, hardware etc., the in- terface costs comprised 4% of the entire projected cost of the project. This ratio applies on a larger scale as well. We spend 31% of our overall infor- mation systems budget on clinical applica- tions, 17% on administrative applications, 6% on administration, 42% on infrastruc- ture (machines, personal computers, ope- rating systems, and networks), and 4% on interfaces and vocabulary. For the entire group of interfaced appli- cations, there are 31 different types of inter- faces (for example, patient registration in- formation is one type) that are needed to ac- commodate the variety of data types (radio- logy orders, laboratory orders, medication administration, textual document, etc.). We have a total of 820 external network connec- tions to our interface engine and process nearly two million transactions per day. We currently have 19 full-time em- ployees who build and maintain interfaces (out of a total of 150 who work on clinical applications). Because of the order entry project and a new insurance claims product that we are installing, the number of people devoted to interfaces is approximately twice as large as our comparable staffing level of two years ago. At an average salary of $30 per hour and an estimated time of 650 hours to build an interface, we estimate that each interface costs about $20,000. For example, this means that the interface to the CareCast product costs us approxima- tely $350,000. These are our costs and the vendors typically charge for their side of the interface as well. We currently employ eight individuals who work full time to create and maintain the vocabulary codes that are stored in our dictionary. We have expended a total of approximately 36 person years to define terms, synonyms and hierarchy and to maintain the dictionary. We currently have 545,000 uniquely defined concepts in the dictionary. We import our drug information from First Data Bank and use other availa- ble data sets such as LOINC, CPT-4, ICD9, etc. As an example of the dictionary content, we have defined unique concepts for 4392 clinical observations, 2415 clinical laboratory tests, 3162 laboratory results, and 848 types of radiology examinations or supplies. We believe that we currently have approximately 70% of the content that will eventually be needed to support a compre- hensive clinical information system. This estimate will change as new observations Clinical Information System 5 Methods Inf Med 1/2003 Fig. 3 This graph shows the gro- wing use of the Clinical Workstation to enter data into the computer. Our users are not forced to use this application but have chosen to do so because of the benefits the application offers. This application has been primarily used in the ambulatory setting until now. Unique CW Users by Application and Role Each Month Physician Role from the field of genetics begin to be inclu- ded in the clinical record. This new content will include identifiers for genes and their phenotypic expressions as well as the names of tests to determine the presence to particular mutations. Although we have been running rules in the HELP system for three decades and will continue to do so for the foreseeable future, our HELP2 rules engine capability is not yet being used in a production envi- ronment. We are planning to begin clinical use of our decision support tool in March of 2003. Based upon our experience in the test environment, we anticipate that we will have sufficient performance to run the rules that are currently used in our hospital system as well as to expand the scope of our knowledge-based systems. Discussion For those institutions with the knowled- geable people and the resources to invest in the initially more expensive, yet ultimately more flexible approach, we believe that the interfaced approach is much more cost effective than the alternative approach over a thirty-year lifetime of such systems. We have adopted this belief because of the following observations and experience. In todays marketplace, no single vendor can satisfy all the desired functionality, and the monolithic approach to defining databases restricts the flexibility and scalability of such systems. As technology changes over a thirty-year period, it is the database content that remains valuable and should be defi- ned using a well-structured approach to a controlled medical vocabulary. The physi- cal implementation of the logical database model could change over time to take advantage of new database management applications. One would only have to chan- ge the services that store and retrieve the data and both versions of the database could simultaneously exist during a transi- tion much as we are doing with the current HELP system. When one thinks of the technological changes that have occurred in the last three decades (networks, perso- nal computers, handheld wireless com- puters, the concept of the world wide web and its associated information servers, the National Library of Medicines efforts to construct a common vocabulary, the emer- gence of HL7 standards, patient access to their medical records on-line etc.) and the improvement in commercially developed applications that has occurred in this time period, it is fairly evident that in the next two to three decades, there also will be dramatically enhanced capabilities when compared to the software that we are installing today. If we must accommodate this change by relying on a single vendor that may or may not stay in business or respond rapidly to these changes or offer the breadth that is needed, we place our initial investment at risk. How many vendors that are in business today were the leaders 30 years ago? The ability to gradually change or add one component at a time to the system allo- ws us to adapt to change incrementally. Experience at Columbia Presbyterian (28), and at IHC shows that getting the data into the database is the most difficult task. After the database exists, creating new applica- tions within a new technology (Web, hand held, patient home, etc.) is rather straight- forward. At both institutions, the creation of new web-based results review applicati- on took less than 4 person years. Another example of a gradual transition is our experience with the existing HELP application. Importing data from HELP into HELP2 via the interface shown in Fig. 1 illustrates the power of the interfaced approach. HELP simultaneously runs on the same clinical desktop as HELP2. This ability means that there is no need to train everybody at once and have a big bang switch over to a new and perhaps untested system. People are drawn to the new system because of the availability of longi- tudinal data, literature resources, Radio- graphs, EKG waveforms, ability to enter orders and data, etc. As the load increases we can monitor response time and titrate the user load gradually as we tune the sy- stem for optimal performance. One need not turn off an old system when implemen- ting the new. As people gradually switch to the new system and pieces of the old HELP system are replaced, it can atrophy. This approach makes the transition easy and allows us to tune the performance of the new system in a non-urgent setting as peo- ple gradually begin to use it. As bugs in the new system appear, the old system is still available as a back up. The biggest drawback to this approach is the initial cost of building the interfaces. It costs as much to build the interfaces in a small system as it does in the large system, so the proportionate costs of interfaces are higher. The cost of vocabulary definitions is less of an issue because vocabulary terms need to be defined in a monolithic system as well as in the interfaced system. In the monolithic system, it is a somewhat simpler task because there is no need to map syno- nyms as in the latter model. However, it should be kept in mind that the cost of choosing a monolithic system supplied by a vendor that does not stay in business far outweighs any interface costs. Usually the leading companies in one decade are not the leaders in subsequent decades due to a phenomenon known as installed user base anchor. If the interfaced model becomes widely adopted, the vendor of a safe mono- lithic choice today may not be competitive in the future. We believe that the efforts of the National Library of Medicine to define a common terminology and the continuing efforts of HL7 will drive down the costs of interfaces. Another hypothetical drawback that is often raised revolves around the issue of storing data in more than one place. Most people who raise this issue are referring to the case in which there is redundant data entry and there are subsequent discordan- ces because the separate databases contain different information. When examining this issue it should be kept in mind that, in our model, the data are not entered redundant- ly; that is the whole point of interfaces. Redundant data can be a blessing in some cases. One can ask the originating applica- tion to replay its log and resend data if there is a failure at the recipient end. If the CDR were to be unavailable, one could get laboratory results by going directly to the laboratory. This option was used at Colum- bia Presbyterian Medical Center (16). Some worry about the problems of keeping data perfectly synchronized. We have not Clayton et al. 6 Methods Inf Med 1/2003 found this to be a practical problem especi- ally when compared to the manual system in which there are many results that never get properly placed in the patient record. Another drawback of our approach is that less than accurate registration clerks use multiple registration programs that we cannot alter. This combination sometimes creates duplicate master index numbers for a single patient. This situation is dangerous because vital information might be stored in each fragment of the record, but the provider might only see one fragment. This problem with duplicates applies to any system with a longitudinal record, but is so- mewhat more challenging when the vendor registration systems are not expecting to go to a foreign enterprise master patient record. In summary, we have attempted to build upon an architecture where it is possible to change any component or application with- out the need to change any other compo- nent. We have used a dictionary and a data model that allow us to define the content of the patient record in a way that is indepen- dent of any one application. The flexibility and extensibility afforded by this approach lets us use additional applications regard- less of how the data might be collected in any other application. We are pleased with our decision to go in this direction. We feel that as the costs of interfaces are reduced, this will be a viable alternative to be consi- dered by all institutions or enterprises not just those large groups where interfaces are a smaller proportion of the overall costs of the system. References 1. Leguit F. Interfacing Integration in Hospital in- formation systems:Scope-Design-Architecture Bakker, AR, Bryant JR, Ehlers CT, Hammond WE, eds. North Holland, Amsterdam, 1992, pp 141-8. 2. Slack WV, Bleich HL.The CCC system in two teaching hospitals: a progress report. Int J Med Inf. 1999; 54 (3): 183-96. 3. Scherrer JR, Baud RH, Hochstrasser D, Ratib O. An integrated hospital information system in Geneva. MD Comput. 1990; 7 (2): 81-9. 4. Collen MF.A brief historical overview of hospi- tal information system (HIS) evolution in the United States. Int J Biomed Comput. 1991; 29 (3-4): 169-89. 5. Anderson CL, Meldrum KC.The VA compute- rized Patient Record a first look. Proc Annu Symp Comput Appl Med Care. 1994; 1048. 6. Overhage J, McDonald CJ, Suico JG. The REGENSTRIEF medical record system 2000:Expanding the breadth and depth of a community wide EMR Proc AMIA Symp. 2000, 1173. 7. Pryor TA, Gardner RM, Clayton PD, Warner HR. The HELP system. J Med Syst. 1983; 7 (2): 87-102. 8. Tolchin SG, Stewart RL, Kahn SA, Bergan ES, Gafke GP, Simborg DW, Chadwick MG, Whiting-OKeefe QE.A prototype generalized network technology for hospitals: initial imple- mentation. J Med Syst. 1982 Aug; 6 (4): 359-75. 9. Simborg DW. The case for the HL7 standard. Comput Healthc. 1988 Jan;9(1):39-40-2. 10. Scherrer JR, Spahni S. Healthcare information system architecture (HISA) and its middleware models. Proc AMIA Symp. 1999, 935-9. 11. Chu S, Cesnik B. A three-tier clinical informa- tion systems design model. Int J Med Inf. 2000; 57 (2-3): 91-107. 12. Van de Velde R. Framework for a clinical in- formation system. Int J Med Inf. 2000; 57 (1): 57-72. 13. Lowe HJ, Buchanan BG, Cooper GF, Vries JK. Building a medical multimedia database sy- stem to integrate clinical information: an appli- cation of high-performance computing and communications technology. Bull Med Libr As- soc. 1995; 83 (1): 57-64. 14. Greenes RA. A building block approach to application development for education and decision support in radiology: implications for integrated clinical information systems envi- ronments.J Digit Imaging. 1991; 4 (4): 213-25. 15. Glaser JP, Beckley RF 3rd, Roberts P, Marra JK, Hiltz FL, Hurley J. A very large PC LAN as the basis for a hospital information system. J Med Syst 1991; 15 (2): 133-7. 16. Clayton PD, Sideli RV, Sengupta S. Open archi- tecture and integrated information at Colum- bia-Presbyterian MedicalCenter. MD Comput. 1992; 9 (5): 297-303. 17. Clayton PD Current developments and future trends in Information technology in health care in Hospital Information Systems:Scope-De- sign-Architecture Bakker, AR, Bryant JR, Eh- lers CT, Hammond WE eds North Holland, Amsterdam 1992, pp 19-25. 18. Tarczy-Hornoch P, Kwan-Gett TS, Fouche L, Hoath J, Fuller S, Ibrahim KN, Ketchell DS, LoGerfo JP, Goldberg HI. Meeting clinician information needs by integrating access to the medical record and knowledge resources via the Web Proc AMIA Symposium 1997, 809-13. 19. Ferrara FM. The CEN healthcare information systems architecture standard and the DHE middleware. A practical support to the integra- tion and evolution of healthcare systems. Int J Med Inf. 1998; 48 (1-3): 173-82. 20. Van Mulligen EM, Timmers T, Brand J, Cornet R, van den Heuvel F, Kalshoven M, van Bem- mel JH. HERMES: a health care workstation integration architecture. Int J Biomed Comput. 1994; 34 (1-4): 267-75. 21. Patil RS, Silva JS, Swartout WR. An architectu- re for a health care providers workstation. Int J Biomed Comput. 1994; 34 (1-4): 285-99. 22. Huff S M, Rocha R A, Bray B E, Warner H R, Haug P J. An event model of medical informa- tion representation. J Am Medical Informatics Ass 1995; 2 (2): 116-28. 23. Huff S M, Rocha R A, Solbrig H R, Barnes M W, Schrank S P, Smith M. Linking a medical vocabulary to a clinical data model using Ab- stract Syntax Notation 1. Methods Inf Med 1998; 37 (4-5): 440-52. 24. Rocha RA, Huff SM. Development of a template model to represent the information content of chest radiology reports.Medinfo. 2001; 10 (Pt 1): 251-5. 25. Rocha R A, Huff S M, Haug P J, Warner H R. Designing a Controlled Medical Vocabulary Server: The VOSER Project. Computers and Biomedical Research.1994; 27 (6): 472-507. 26. Wilcox A, Narus SP, Bowes III WA, Using natu- ral language processing to analyze physician modifications to data entry templates. Proc 2002 AMIA. Kohane I, ed. Hanley and Belfus 2002, pp. 899-903. 27. Reichert JC, Glasgow M, Narus SP, Clayton PD. Using LOINC to link an EMR to the perti- nent paragraph in a structured reference know- ledge base. Proc 2002 AMIA. Kohane I,ed. Hanley and Belfus 2002, pp. 652-6. 28. Cimino JJ, Socratous SA, Clayton PD. Internet as clinical information system: application de- velopment using the World Wide Web. J Am Med Inform Assoc. 1995; 2 (5): 273-84. Correspondence to: Dr. Paul D. Clayton 4646 W Lake Park Blvd Salt Lake City UT 84120, USA E-mail: copclayt@ihc.com Clinical Information System 7 Methods Inf Med 1/2003