A population-based validation of the prognostic model PREDICT for early breast

cancer
G.C. Wishart
a,g
, C.D. Bajdik
e
, E.M. Azzato
b,c
, E. Dicks
b
, D.C. Greenberg
d
, J. Rashbass
d
,
C. Caldas
a,b,f,g
, P.D.P. Pharoah
b,
*
a
Cambridge Breast Unit, Addenbrookes Hospital, Hills Road, Cambridge, CB2 2QQ, UK
b
Department of Oncology, University of Cambridge, Strangeways Research Laboratory, Worts Causeway, Cambridge CB1 8RN, UK
c
Genetic Epidemiology Branch, Division of Cancer Epidemiology and Genetics, National Cancer Institute, Bethesda, MD, USA
d
Eastern Cancer Registration and Information Centre (ECRIC), Unit C, Magog Court, Shelford Bottom, Hinton Way, Cambridge CB22 3AD, UK
e
Cancer Control Research Program, British Columbia Cancer Agency, West 10th Avenue, Vancouver, British Columbia, V5Z 1L3 Canada
f
Functional Breast Cancer Genomics Laboratory, Cancer Research UK Cambridge Research Institute, Li Ka-Shing Centre, Robinson Way, Cambridge CB2
0RE, UK
g
National Institute of Health Research (NIHR) Cambridge Biomedical Research Centre, Addenbrookes Hospital, Hills Road, Cambridge CB2 2QQ, UK
Accepted 7 February 2011
Abstract
Introduction: Predict (www.predict.nhs.uk) is a prognostication and treatment benet tool developed using UK cancer registry data. The
aim of this study was to compare the 10-year survival estimates from Predict with observed 10-year outcome from a British Columbia
dataset and to compare the estimates with those generated by Adjuvant! (www.adjuvantonline.com).
Method: The analysis was based on data from 3140 patients with early invasive breast cancer diagnosed in British Columbia, Canada, from
1989e1993. Demographic, pathologic, staging and treatment data were used to predict 10-year overall survival (OS) and breast cancer
specic survival (BCSS) using Adjuvant! and Predict models. Predicted outcomes from both models were then compared with observed
outcomes.
Results: Calibration of both models was excellent. The difference in total number of deaths estimated by Predict was 4.1 percent of
observed compared to 0.7 percent for Adjuvant!. The total number of breast cancer specic deaths estimated by Predict was 3.4 percent
of observed compared to 6.7 percent for Adjuvant! Both models also discriminate well with similar AUC for Predict and Adjuvant! respec-
tively for both OS (0.709 vs 0.712) and BCSS (0.723 vs 0.727). Neither model performed well in women aged 20e35.
Conclusion: In summary Predict provided accurate overall and breast cancer specic survival estimates in the British Columbia dataset that
are comparable with outcome estimates from Adjuvant! Both models appear well calibrated with similar model discrimination. This study
provides further validation of Predict as an effective predictive tool following surgery for invasive breast cancer.
2011 Elsevier Ltd. All rights reserved.
Keywords: Breast; Prognosis; Model validation
Introduction
Accurate survival estimates, and the likely benet of adju-
vant therapy, are key pieces of information when making
treatment decisions following surgery for invasive, early
breast cancer. Currently these decisions are based on known
pathological prognostic factors including tumour size,
tumour grade and lymph node status in addition to the relative
risk reductions of any adjuvant therapy. Several predictive
models are now available to help estimate the survival and
treatment benets for individual patients including the Not-
tingham Prognostic Index (NPI), Adjuvant! and Predict.
The NPI led the way in the UKin 1982 with a prognostic scor-
ing system based on tumour size, grade and lymph node sta-
tus
1
which has been prospectively validated
2,3
and updated
4
This model is based on treatment in a single institution and
may therefore not be relevant for more widespread utility.
* Corresponding author. Tel.: 44 1223 740166.
E-mail address: paul.pharoah@srl.cam.ac.uk (P.D.P. Pharoah).
0748-7983/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ejso.2011.02.001
Available online at www.sciencedirect.com
EJSO xx (2011) 1e7 www.ejso.com
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
Adjuvant!, a web-based (www.adjuvantonline.com) prog-
nostication and treatment benet tool has been widely used in
the UK to help clinicians and patients make decisions about
adjuvant therapy for breast cancer during the last ten years.
The mortality estimates used in Adjuvant! were based on
10-year observed overall survival (OS) of women aged
36e69 who were diagnosed between 1988e1992 and re-
corded in the Surveillance, Epidemiology and End Results
(SEER) registry.
5
Breast cancer specic survival (BCSS)
without adjuvant therapy was calculated based on estimates
of the number of patients likely to have received systemic
therapy and the risk reductions outlined in the Early Breast
Cancer Trialists Collaborative Group.
6,7
Although Adju-
vant! was successfully validated in a population-baseddataset
from British Columbia,
8
some UK oncologists have ex-
pressed concern about how applicable this model is to con-
temporary patients diagnosed in the UK. Furthermore,
a recent paper has shown that Adjuvant overestimated the
OS by 6 percent in a UK cohort of 1065 women with early
breast cancer treated in Oxford between 1986 and 1996
9
although some of these patients had what would now be
considered as sub-optimal treatment.
Predict is an online prognostication and treatment benet
tool recently developed in the UK, using cancer registration
and survival data recorded by the Eastern Cancer Registra-
tion and Information Centre (ECRIC) for 5694 women diag-
nosed in East Anglia from 1999e2003.
10
ECRIC provides
near complete breast cancer registration for the population
of the East of England region as well as information on
systemic treatment and mode of detection. The model was
validated in a second cohort of 5468 women from the West
Midlands Cancer Intelligence Unit and is now available on-
line (www.predict.nhs.uk) providing 5-and 10-year survival
estimates and treatment benet predictions.
The aim of this study was to compare the 10-year sur-
vival estimates from Predict with observed 10-year out-
come from the British Columbia dataset in which
Adjuvant! was rst validated. We also compared the esti-
mates from Predict with those from Adjuvant! in order to
provide a direct comparison of the two models.
Methods
Study population
The analysis was based on data from patients with Stage I
or II invasive breast cancer diagnosed in British Columbia,
Canada, from 1989e1993 who were identied from the
Breast Cancer Outcomes Unit (BCOU) of the British Co-
lumbia Cancer Agency (BCCA). BCCA covers a catchment
area population of approximately 4 million people and
treated approximately 75 percent of all newly diagnosed
breast cancers during the study period. The BCOU database
has previously been described
8
but, in brief, it prospectively
records demographic, pathologic, staging, initial treatment
and outcome information including rst loco-regional and
distant relapse as well as date and cause of death. Outcome
data were reported annually by the treating oncologist, fam-
ily physician or by monthly death certicate agging
through the British Columbia Cancer Registry and Depart-
ment of Vital Statistics for British Columbia. This study
uses anonymised patient data without individual patient
consent and was approved by the University of British
Columbia BCCA Institutional Review Board.
Information obtained from the BCOU database included
age at diagnosis, sex, menopausal status, year of diagnosis,
histology (ductal, lobular, other), histological grade (I, II,
III), tumour size (mm), number of lymph nodes sampled
and number of lymph nodes positive, lymphovascular inva-
sion (LVI) status (positive, negative, unknown), oestrogen
receptor (ER) status (positive or negative, unknown), infor-
mation on local therapy (wide local excision, mastectomy,
radiotherapy), and type of adjuvant systemic therapy (none,
chemotherapy, endocrine therapy, both). Chemotherapy
regimens were categorised as 4 cycles of doxorubicin
plus cyclophosphamide; 6 months of cyclophosphamide,
methotrexate and uorouracil or other chemotherapy dur-
ing this time period.
The original dataset used for the validation of Adjuvant!
(n 4083) excluded all males, women under 20 years or
older than 85 years, patients with incomplete local therapy
(wide local excision without radiotherapy), patients with
unknown nodal status and previous or synchronous contra-
lateral breast cancer.
8
For the purposes of this study, patients
with unknown ER status or inadequate node sampling (<4
nodes) if node negative were also excluded leaving a study
population of 3140 cases for OS and 3122 for BCSS as there
were 18 patients with unknown cause of death.
Study endpoints were overall survival (OS) and breast
cancer specic survival (BCSS). 10-year predicted OS
and BCSS were calculated for each patient using Adjuvant!
Standard Version 8, which has been updated since the val-
idation published in 2005, and Predict by investigators
blinded to the actual outcome data for each patient after en-
try of patient age, tumour size, number of positive nodes,
tumour grade, ER status and adjuvant systemic therapy.
The default comorbidity setting of minor symptoms
was used in Adjuvant! and the chemotherapy option chosen
was anthracycline-containing; 4 cycles. The second
generation chemotherapy button (anthracycline-containing
regimens) was used in Predict. The mode of detection input
was not used in Predict as this information was not avail-
able in the British Columbia dataset. 10-year predicted
OS and BCSS from both Predict and Adjuvant! were com-
pared with observed 10-year OS and BCSS.
Model calibration and discrimination
Model calibration is a comparison of the predicted mor-
tality estimates from each model with the total observed
mortality. Model discrimination was evaluated by calculat-
ing the area under the receiver-operator characteristic
2 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
(ROC) curve (AUC) calculated for 10-year breast cancer
specic and overall mortality. This is a measure of how
well the models identify those patients with worse survival.
The AUC is the probability that the predicted mortality from
a randomly selected patient who died will be higher than the
predicted mortality from a randomly selected survivor.
Model reclassication
This is a measure of what difference the model risk esti-
mates make to a particular classier. Patients treated in the
Cambridge Breast Unit are stratied for adjuvant chemother-
apy according to the following predicted 10-year absolute
survival benet (of adjuvant chemotherapy): <3 percent,
not recommended; 3e5 percent, chemotherapy discussed
as possible option with the patient; >5 percent, chemother-
apy recommended as best option for patient. We therefore
assigned patients from the BCOU dataset to one of the three
groups according to the predicted absolute benet from
chemotherapy using both Predict and Adjuvant! and results
compared.
Results
The demographic, pathologic and treatment characteris-
tics for 3140 women with stage I or II breast cancer are
shown in Table 1 together with observed 10-year overall
mortality from BCOU compared to predicted 10-year mor-
tality outputs from Predict and Adjuvant! Overall both
models performed well across all age groups except in
women aged 20e35 where they both under predicted the
actual number of deaths by 17.1 percent. The total number
of deaths predicted by Adjuvant! was within 0.7 percent of
observed compared to 4.1 percent for Predict. Predict per-
formed better than Adjuvant! with lower differences from
observed all cause mortality in Grade 3 tumours (2.7 per-
cent vs 4.7 percent), LVI positive tumours (3.6 percent
vs 9.0 percent) and node negative tumours (1.7 percent
vs 7.7 percent). Adjuvant! performed better than Predict
with lower differences from observed all cause mortality
in tumours >21 mm (4.2 percent vs 10.6 percent),
Grade 2 tumours (2.3 percent vs 10.8 percent) and ER
positive tumours (3.7 percent vs 9.2 percent).
The demographic, pathologic and treatment characteris-
tics for 3122 women with stage I or II breast cancer are
shown in Table 2 together with observed 10-year breast
cancer specic mortality from BCOU compared to pre-
dicted 10-year breast cancer specic mortality outputs
from Predict and Adjuvant! Overall both models performed
well across all age groups except in women aged 20e35
where they both under predicted the actual number of
deaths by 20.0 percent. Differences in observed and pre-
dicted BCSS were 1.31 percent for Adjuvant! and 0.67
percent for Predict. The total number of breast cancer spe-
cic deaths predicted by Adjuvant! was within 6.7 percent
of observed compared to 3.4 percent for Predict.
Predict performed better than Adjuvant! with lower dif-
ferences from observed breast cancer specic mortality in
node positive tumours (2.0 percent vs 10.3 percent), tu-
mours >21 mm (2.6 percent vs 4.0 percent) and ER
positive tumours (4.6 percent vs 14.4 percent). Adjuvant!
performed better than Predict with lower differences from
observed breast cancer specic mortality in Grade 3 tu-
mours (6.2 percent vs 14.5 percent), ER negative tumours
(7.3 percent vs 17.8 percent) and node negative tumours
(0.0 percent vs 13.4 percent).
Model discrimination
The AUC for 10-year OS (n 3140) and BCSS
(n 3122) are shown in Fig. 1. These demonstrate that
both models perform extremely well with similar AUC
for Predict and Adjuvant! respectively for both OS (0.709
vs 0.712) and BCSS (0.723 vs 0.727) (Fig. 2).
Model reclassication
The classication according to chemotherapy recom-
mendation with both models was compared and is shown
in Table 3. Thus only 3 patients who were recommended
chemotherapy using Adjuvant! would be regarded as get-
ting minimal benet according to Predict. Similarly only
6 patients who would be recommended chemotherapy un-
der Predict would not have chemotherapy recommended
using Adjuvant!.
Discussion
The results of this study conrm that both Predict and
Adjuvant! provide accurate 10-year survival estimates in
the British Columbia dataset with predicted and observed
outcomes that are within 1.5 percent for both OS (1.2 percent
vs 0.22 percent) and BCSS (0.67 percent vs 1.3 percent)
respectively. Model calibration conrmed that Predict
provided better all cause mortality estimates in patients
with Grade 3 tumours, LVI positive tumours and node nega-
tive tumours while Adjuvant! was better for large tumours
(>21 mm), Grade 2 tumours and ER positive tumours. A
comparison of breast cancer specic mortality estimates
showed that Predict was better for node positive tumours,
larger tumours (>21 mm) and ER positive tumours while
Adjuvant! was better for Grade 3 tumours, ER negative
tumours and node negative tumours.
Although model calibration demonstrates that the pre-
dicted and observed mortality is broadly similar with
both models, the real question is whether the patients
with predicted mortality are the patients who actually
died during the 10-year follow up period. The AUC is
the best way to examine this more closely and the results
presented in Fig. 1 demonstrate that both models perform
extremely well with similar AUC for Predict and
3 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
Adjuvant! respectively for both OS (0.709 vs 0.712) and
BCSS (0.723 vs 0.727).
Although the survival estimates from both Predict and
Adjuvant! are very similar there are marked differences in
the way that the models have been developed. In the Predict
model, based on women diagnosed in UK from
1999e2003, breast cancer specic mortality and competing
mortality were modelled separately. For breast cancer spe-
cic mortality, a Cox proportional hazards model was used
to estimate the hazard ratio associated with each prognostic
factor and as the effect of ER status varies over time
11
ER
negative and ER positive tumours were modelled
separately. In contrast the mortality estimates used in Adju-
vant! were based on 10-year observed overall survival (OS)
of women aged 36e69 who were diagnosed between
1988e1992 and recorded in the Surveillance, Epidemiol-
ogy and End Results (SEER) registry. Breast cancer spe-
cic survival (BCSS) without adjuvant therapy was
calculated based on estimates of the number of patients
likely to have received systemic therapy and the risk reduc-
tions outlined in the Early Breast Cancer Trialists Collab-
orative Group
6,7
whereas in Predict information on local
therapy (wide local excision, mastectomy, radiotherapy),
type of adjuvant systemic therapy (chemotherapy,
Table 1
Observed and predicted all cause mortality by patient characteristics.
Number cases Observed deaths Deaths predicted by PREDICT Death predicted by Adjuvant!
Number Difference (%) Number Difference (%)
Total 3140 936 898 38 4.1 929 7 0.7
Age group
20e35 108 35 29 6 17.1 29 6 17.1
36e50 873 198 208 10 5.1 195 3 1.5
51e65 1032 273 253 20 7.3 258 15 5.5
66e75 810 272 259 13 4.8 283 11 4.0
76 317 158 148 10 6.3 166 8 5.1
Menopausal status
Pre- 931 228 226 2 0.9 213 15 6.6
Post- 2117 692 500 192 27.7 493 199 28.8
Perie 92 16 28 12 75.0 30 14 87.5
Morphology
Ductal 2839 857 811 46 5.4 839 18 2.1
Lobular 247 67 71 4 6.0 76 9 13.4
Other 54 12 15 3 25.0 15 3 25.0
Grade
1 258 58 48 10 17.2 55 3 5.2
2 1336 342 305 37 10.8 350 8 2.3
3 1236 447 459 12 2.7 426 21 4.7
Unknown 310 88 85 3 3.4 98 10 11.4
LV invasion
Negative 1238 463 413 50 10.8 416 47 10.2
Positive 1712 411 426 15 3.6 448 37 9.0
Unknown 172 58 53 5 8.6 60 2 3.4
Node status
Neg 1910 418 425 7 1.7 450 32 7.7
Pos 1230 518 472 46 8.9 480 38 7.3
Tumour size
1e10 521 96 102 6 6.3 101 5 5.2
11e20 1403 349 351 2 0.6 360 11 3.2
21e50 1179 473 423 50 10.6 453 20 4.2
Unknown 37 18 22 4 0.3 15 3 4.9
ER status
ER- 900 294 314 20 6.8 311 17 5.8
ER 2240 643 584 59 9.2 619 24 3.7
Local Rx
BCS RT 1593 368 393 25 6.8 405 37 10.1
Mast RT 328 152 147 5 3.3 144 8 5.3
Mast 1219 416 357 59 14.2 381 35 8.4
Systemic Rx
None 1250 286 305 19 6.6 317 31 10.8
Hormone 1035 389 337 52 13.4 369 20 5.1
Chemo 535 153 159 6 3.9 154 1 0.7
Both 320 108 96 12 11.1 90 18 16.7
4 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
endocrine therapy, both) and mode of detection (screen-de-
tected or symptomatic) was available from the cancer reg-
istry database. Despite these differences in design and
development it is reassuring that both models perform
well in the BCOU dataset.
One key difference between the models is that Predict
includes mode of detection as one of the input parameters.
Until recently it was generally assumed that the survival ad-
vantage associated with breast screening was related to
a shift to an earlier stage at diagnosis and thus improved
survival. Two recent papers, have suggested that screen
detection not only confers an additional survival benet
beyond stage shift but also reduces the risk of systemic
recurrence when compared with symptomatic cancers of
a similar stage.
12,13
Recent studies have shown that al-
though the majority of the survival advantage associated
with breast screening can be explained by this shift to an
earlier stage at diagnosis, and more favourable prognostic
factors,
14
approximately 25 percent of the survival advan-
tage is still unexplained.
15
As screening data were not available for patients in the
British Columbia dataset it was not possible to use this fea-
ture when running the Predict model so the default setting
was symptomatic for all patients. There is some indirect
Table 2
Observed and predicted breast cancer specic mortality by patient characteristics.
Number cases Observed deaths Deaths predicted by PREDICT Death predicted by Adjuvant!
Number Difference (%) Number Difference (%)
Total 3122 614 635 21 3.4 573 41 6.7
Age group
20e35 108 35 28 7 20.0 28 7 20.0
36e50 872 188 192 4 2.1 176 12 6.4
51e65 1028 185 200 15 8.1 183 2 1.1
66e75 803 150 156 6 4.0 139 11 7.3
76 311 56 58 2 3.6 46 10 17.9
Menopausal status
Pre- 930 218 209 9 4.1 193 25 11.5
Post- 2103 386 408 22 5.7 362 24 6.2
Perie 89 10 18 8 80.0 17 7 70.0
Morphology
Ductal 2823 575 577 2 0.3 520 55 9.6
Lobular 245 34 44 10 29.4 41 7 20.6
Other 54 5 13 8 160.0 12 7 140.0
Grade
1 257 19 16 3 15.8 15 4 21.1
2 1328 208 178 30 14.4 184 24 11.5
3 1229 338 387 49 14.5 317 21 6.2
Unknown 308 49 54 5 10.2 57 8 16.3
LV invasion
Negative 1231 347 195 152 43.8 175 172 49.6
Positive 1701 230 452 222 96.5 404 174 75.7
Unknown 172 34 34 0 0.0 34 0 0.0
Node status
Neg 1900 216 245 29 13.4 216 0 0.0
Pos 1222 398 390 8 2.0 357 41 10.3
Tumour size
1e10 519 45 53 8 17.8 36 9 20.0
11e20 1393 204 222 18 8.8 188 16 7.8
21e50 1173 349 340 9 2.6 335 14 4.0
Unknown 37 16 20 4 25.0 14 2 12.5
ER status
ER- 896 219 258 39 17.8 235 16 7.3
ER 2226 395 377 18 4.6 338 57 14.4
Local Rx
BCS RT 1584 241 269 28 11.6 241 0 0.0
Mast RT 326 123 127 4 3.3 113 10 8.1
Mast 1212 250 239 11 4.4 220 30 12.0
Systemic Rx
None 1242 148 181 33 22.3 155 7 4.7
Hormone 1028 225 218 7 3.1 198 27 12.0
Chemo 534 148 149 1 0.7 142 6 4.1
Both 318 93 87 6 6.5 78 15 16.1
5 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
evidence to suggest that this may have disadvantaged the
Predict model in this comparison. In all age groups, except
age 20e35 where both models performed poorly, breast
cancer specic mortality estimates with Predict were closer
to observed mortality than Adjuvant! except for women
aged 51e65, the age range for breast screening in the UK
for the dataset used to develop the Predict model. In this
age group Predict overestimated the mortality by 8.1 per-
cent and this could be explained by the inability to apply
a screen-detected survival advantage to some of the study
patients aged 51e65 as mode of detection was unknown.
Mode of detection is thus a key piece of information re-
quired to optimise the output from the Predict model.
The major decisions facing most oncologists following
the local treatment of breast cancer is whether an individ-
ual patient will benet from chemotherapy and does that
benet outweigh potential side effects and toxicity. Both
Predict and Adjuvant! allow calculation of treatment ben-
et, including chemotherapy and hormone therapy, and the
information is presented in graphic and text format which
is both user- and patient-friendly. Estimation of absolute
survival benet with chemotherapy allows thresholds to
be set for use of chemotherapy. In this way the Cambridge
Breast Unit has dened chemotherapy treatment guide-
lines according to this benet with chemotherapy dis-
cussed for an absolute benet of 3e5 percent and
recommended for an absolute benet of >5 percent. Che-
motherapy benet estimates from Predict and Adjuvant!
were compared allowing assessment of reclassication
between chemotherapy groups (none, discussed, recom-
mended) (Table 3). This was very reassuring with com-
plete concordance between predictive models in 2574 of
3140 patients (87.7 percent) and only 9 patients (0.3 per-
cent) moved from a chemotherapy recommendation to
no chemotherapy. In the British Columbia dataset a total
of 1088 patients (34.6 percent) would have chemotherapy
discussed or recommended using Predict compared to
1180 patients (37.6 percent) with AOL based on the Cam-
bridge guidelines.
Conclusion
Both Predict and Adjuvant! provide accurate overall and
breast cancer specic survival estimates in the British Co-
lumbia dataset. Both models appear well calibrated and
model discrimination was also excellent and comparable.
The next challenge for both models is to nd the optimal
way to include HER2, and the benet of traztuzumab, in
future versions. The inclusion of 5-year survival estimates
in the Predict model is part of a plan to make the addition
of new prognostic factors to the model much easier in the
future.
Figure 1. Receiver-operator characteristic (ROC) curves for overall sur-
vival in 3140 patients based on Predict and Adjuvant! breast cancer prog-
nostic models.
Figure 2. Receiver-operator characteristic (ROC) curves for breast cancer
specic survival in 3122 patients based on Predict and Adjuvant! breast
cancer prognostic models.
Table 3
Classication of 3140 patients according to estimated absolute survival
benet from Predict and Adjuvant!
Adjuvant! Predict
<3% 3e5% >5% Total
<3% 1911 43 6 1960
3e5% 138 219 107 464
>5% 3 89 624 716
Total 2052 351 737 3140
6 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001
Authors contributions
GCW and PDPP conceived of the project and partici-
pated in the design, analysis and writing of the manuscript.
ED and CDB participated in the data analysis and writing
of the manuscript. EMA, ED, DCG, JR, and CC are mem-
bers of the Predict development team and participated in
the writing of the manuscript.
Funding
The development of the PREDICT model was funded by
an unrestricted educational grant from Pzer.
Conict of interest
The authors declare no conict of interest relating to this
manuscript.
Acknowledgements
GCW&CCreceive research funding fromthe Cambridge
NIHR Biomedical Research Centre. EMAwas supported by
the National Cancer Institute and the NIH-Cambridge Grad-
uate Partnership Program.
Abbreviations
AUC area under ROC curve
BCSS breast cancer specic survival
ECRIC eastern cancer registration and information centre
ER oestrogen receptor
NPI Nottingham prognostic index
OS overall survival
ROC receiver-operator characteristic
SEER surveillance, epidemiology and end results
WMCIUwest midland cancer intelligence unit
References
1. Haybittle JL, Blamey RW, Elston CW, et al. A prognostic index in pri-
mary breast cancer. Br J Cancer 1982;45:3616.
2. Todd JH, Dowle C, Williams MR, et al. Conrmation of a prognostic
index in primary breast cancer. Br J Cancer 1987;56:48992.
3. DEredita G, Giardina C, Martellotta M, et al. Prognostic factors in
breast cancer: the predictive value of the Nottingham prognostic index
in patients with a long-term follow-up that were treated in a single in-
stitution. Eur J Cancer 2001;37:5916.
4. Blamey RW, Ellis IO, Pinder SE, et al. Survival of invasive breast can-
cer according to the Nottingham prognostic index in cases diagnosed
in 1990e1999. Eur J Cancer 2007;43:154855.
5. Ravdin PM, Siminoff LA, Davis GJ, et al. Computer program to assist
in making decisions about adjuvant therapy for women with early
breast cancer. J Clin Oncol 2001;19:98091.
6. Early Breast Cancer Trialists Collaborative Group. Tamoxifen for early
breast cancer: an overview of the randomised trials. Lancet 1998;351:
145167.
7. Early Breast Cancer Trialists Collaborative Group. Polychemotherapy
for early breast cancer: an overview of the randomised trials. Lancet
1998;352:93042.
8. Olivotto IA, Bajdik CD, Ravdin PM, et al. Population-based validation
of the prognostic model ADJUVANT! for early breast cancer. J Clin
Oncol 2005;23:271625.
9. Campbell HE, Taylor MA, Harris AL, et al. An investigation into the
performance of the Adjuvant! online prognostic programme in early
breast cancer for a cohort of patients in the United Kingdom. Br J
Cancer 2009;101:107484.
10. Wishart GC, Azzato EM, Greenberg DC, et al. PREDICT: a new UK
prognostic model that predicts survival following surgery for invasive
breast cancer. Breast Cancer Res 2010;12. R1.
11. Azzato EM, Greenberg D, Shah M, et al. Prevalent cases in observa-
tional studies of cancer survival: do they bias hazard ratio estimates?
Br J Cancer 2009;100:180611.
12. Joensuu H, Lehtimaki T, Holli K, et al. Risk for distant recurrence of
breast cancer detected by mammography screening or other methods.
J Am Med Assoc 2004;292:106473.
13. Shen Y, Yang Y, Inoue LY, et al. Role of detection method in predicting
breast cancer survival: analysis of randomized screening trials. J Natl
Cancer Inst 2005;97:1195203.
14. Dawson SJ, Duffy SW, Blows FM, et al. Molecular characteristics of
screen-detected vs symptomatic breast cancers and their impact on
survival. Br J Cancer 2009;101:133844.
15. Wishart GC, Greenberg DC, Britton PD, et al. Screen-detected vs symp-
tomatic breast cancer: is improved survival due to stage migration
alone? Br J Cancer 2008;98:17414.
7 G.C. Wishart et al. / EJSO xx (2011) 1e7
Please cite this article in press as: Wishart GC et al., A population-based validation of the prognostic model PREDICT for early breast cancer, Eur J Surg
Oncol (2011), doi:10.1016/j.ejso.2011.02.001