Background
Breast cancer is the most commonly diagnosed cancer among women worldwide, with approximately 2.3 million new cases in 2022.1 In the United States, it has the highest incidence and the second-highest mortality rate among women across all racial and ethnic groups.2 In 2024, an estimated 310,720 new cases of breast cases and 42,250 breast cancer–related deaths are expected among women in the United States.2 A key decision for women diagnosed with breast cancer, made in consultation with their clinicians, is whether to undergo a course of systemic treatment.
For patients with early-stage breast cancer, adjuvant systemic treatment after surgery is aimed at reducing the risk of recurrence and mortality.3 Accurate estimates of survival and treatment benefit help ensure that potentially harmful therapies are directed toward those most likely to benefit. These estimates support oncologists in making optimal clinical decisions that minimize side effects and maintain patients’ quality of life.4 Prediction models such as PREDICT Breast,5 Adjuvant! Online,6 and CancerMath7 were developed to guide decisions on which adjuvant systemic therapy is most suitable based on patient and tumor characteristics, including tumor size, node status, and hormone receptor status.8 Adjuvant! Online is no longer available, and CancerMath has not been updated since its initial release. PREDICT Breast, introduced in 2011, has undergone regular updates, with the most recent version, PREDICT Breast version 3 (v3; https://breast.v3.predict.cam),9 released in May 2024. Version 3 incorporates both the benefits and harms of therapy and is trained on more recent data from England’s National Cancer Registration and Advisory Service (2000–2017). It is well-calibrated for breast cancer–specific mortality and performs strongly in validation datasets, with predicted deaths within 10% of observed outcomes. In contrast, version 2.210 consistently overpredicted deaths due to improvements in prognosis over time.9 Additionally, version 3.0 demonstrates slightly better model discrimination than version 2.2.
PREDICT Breast v1 and v2 have been validated in breast cancer cases from multiple countries, including the United Kingdom,11,12 Canada,13 Malaysia,14 the Netherlands,15,16 Japan,17 Indian,18 Spain,19 New Zealand,20 and the United States.21,22 However, PREDICT Breast v3 has only been validated in the United Kingdom—the population used to develop the model. Validating this version in other populations, including the United States, is essential. Given the diversity of the US population, it is also important to evaluate the model’s performance across different racial and ethnic groups. Thus, this study aims to address this gap by conducting an external validation of PREDICT Breast v3 using the latest release of the SEER data. This validation will assess the model’s accuracy in predicting outcomes for patients with breast cancer across diverse US populations. Additionally, we compared the performance of PREDICT Breast v3 with PREDICT Breast v2.2 and CancerMath.
Methods
Study Population
The study population was drawn from the SEER Research Plus data (November 2023 submission).23 SEER is a comprehensive cancer registry program in the United States that collects information on patients with cancer from multiple registries nationwide. The latest release recorded 1,291,324 breast cancer cases. The registry captures data on patient demographics, tumor site, time since initial diagnosis, tumor histology, and tumor behavior.
In this study, women aged 25 to 84 years diagnosed with primary breast cancer between 2000 and 2018 were included. The 2000–2018 inclusion period was chosen to ensure at least 5 years of follow-up, as patients diagnosed after this period would not provide sufficient information for assessing PREDICT model performance at 5 years. Exclusion criteria included patients with distant metastasis at diagnosis, tumor size >500 mm, or >50 positive lymph nodes, as well as those with missing information necessary for PREDICT Breast prognostic score calculations. We also excluded patients with missing data on survival time or cause of death. Ultimately, 615,865 female patients with breast cancer were included for further analysis.
SEER Prognostic Variables Used in the PREDICT Breast Model
The minimum set of input variables for PREDICT Breast v3 includes patient demographics (diagnosis year, age at diagnosis), tumor characteristics (tumor size, histologic grade, number of positive lymph nodes, estrogen receptor [ER] status), and treatment types (radiation therapy, adjuvant hormone therapy, adjuvant chemotherapy, trastuzumab, bisphosphonates). Adjuvant chemotherapy is classified as either standard anthracycline-based or high-dose anthracycline/taxane-based. Optional variables include mode of detection (clinical or screen-detected), tumor HER2 status, progesterone receptor (PR) status, and Ki-67 status. Ki-67 data are not available in SEER, so Ki-67 status was set to “unknown” for all cases. Mode of detection data were also unavailable; for cases aged 45–74 years, detection was assumed to be screening for approximately 49% of patients and clinically detected for all others, based on evidence from the literature and data from the US Census website.24–26
SEER also provides information on race and ethnicity, categorized as non-Hispanic White, non-Hispanic Black, non-Hispanic American Indian or Alaska Native, non-Hispanic Asian or Pacific Islander, and Hispanic. Each variable is either quantitative or categorical, as outlined in Appendix 1, available online in the supplementary materials.
Treatment data in SEER are limited to indicator variables for radiotherapy and adjuvant chemotherapy with no information on adjuvant hormone therapy, trastuzumab, or bisphosphonate therapy. We made the following assumptions: (1) all patients aged <65 years at diagnosis who received chemotherapy were given a high-dose anthracycline/taxane-based regimen, whereas those aged ≥65 years received a standard anthracycline-based regimen; (2) all patients with ER-positive breast cancer received hormone therapy; (3) patients diagnosed after 2000 with HER2-positive cancer were prescribed trastuzumab; and (4) no patients received bisphosphonate treatment.
Calculating PREDICT Breast v3 Predicted Survival Probabilities
Predicted all-cause mortality for PREDICT Breast v3 was calculated using a custom script based on the model described in Grootes et al.9 PREDICT is a competing risk Cox survival model with fractional polynomial baseline cumulative hazards. The competing risks in the model include breast cancer mortality and other causes of mortality.
HC(t) is the baseline hazard for breast cancer mortality.
HO(t) is the baseline hazard for other-cause mortality.
The baseline hazard functions, along with the coefficients and functions for all breast cancer and non–breast cancer risk factors, were taken from Grootes et al.9
Predicted all-cause mortality for PREDICT Breast v2.2 was calculated using the nhs.predict R package.16 The predicted all-cause mortality for CancerMath was calculated using a custom R script derived from the JavaScript extracted from the online tool. The output of the R script was verified by comparing it with the results generated by the online tool for a small set of cases.
Predictive Model Performance
Model performance was evaluated using calibration, goodness-of-fit, and discrimination. Calibration is given by the ratio of the observed number of events divided by the number of events predicted by the model. Goodness-of-fit was assessed graphically by plotting the observed number of deaths against the predicted number of deaths within quintiles of risk. Model discrimination was evaluated by calculating the area under the receiver operating characteristic curve (AUC), which measures the probability that the predicted mortality score for a randomly selected patient who died will be higher than that for a randomly selected patient who survived. An AUC value ranges from 0.5 to 1, with a higher AUC indicating a better model in identifying patients with worse survival. AUC statistics were calculated separately for different ER statuses and racial and ethnic groups.
This study adheres to the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) guidelines for the development and validation of prediction models.27 All analyses were conducted using R software28 implemented in R Studio version 4.3.3,29 along with the survival,30 pROC,31 and tidyverse packages.32
Results
Demographic and Clinical Characteristics
The study population included 615,865 women diagnosed with breast cancer between 2000 and 2018 in the SEER cancer registry, of whom 501,885 had ER-positive status and 113,980 had ER-negative status. Of these, 402,722 women had a minimum of 10 years of follow-up (320,878 ER-positive and 81,844 ER-negative), and 213,074 women had a minimum of 15 years of follow-up (166,150 ER-positive and 46,924 ER-negative). Table 1 summarizes patient demographics, tumor characteristics, and treatment types, stratified by ER status. Specifically, this dataset includes 61,360 (10%) Hispanic (all races), 50,390 (8.2%) non-Hispanic Asian or Pacific Islander, 56,299 (9.1%) non-Hispanic Black, and 447,816 (72.7%) non-Hispanic White individuals.
Patient Demographics, Tumor Characteristics, and Treatments, Stratified by ER Status
| Total n (%) |
ER-Positive n (%) |
ER-Negative n (%) | |
|---|---|---|---|
| Patients, N | 615,865 | 501,885 | 113,980 |
| Age at diagnosis, median (IQR), y | 60 (50–69) | 61 (51–70) | 56 (48–66) |
| Follow-up time, median (IQR), y | 10 (8.4–10) | 10 (8.2–10) | 10 (9.2–10) |
| Tumor size, median (IQR), mm | 16 (10–25) | 15 (10, 24) | 20 (13–31) |
| Person-years | 5,586,754 | 1,053,789 | 4,532,965 |
| Race and ethnicity | |||
| Hispanic (all races) | 61,360 (10.0) | 48,146 (9.6) | 13,214 (11.6) |
| NH Asian/Pacific Islander | 50,390 (8.2) | 41,379 (8.2) | 9,011 (7.9) |
| NH Black | 56,299 (9.1) | 38,235 (7.6) | 18,064 (15.8) |
| NH White | 447,816 (72.7) | 374,125 (74.5) | 73,691 (64.7) |
| Tumor grade | |||
| 1 | 142,823 (23.2) | 139,738 (27.8) | 3,085 (2.7) |
| 2 | 268,940 (43.7) | 246,924 (49.2) | 22,016 (19.3) |
| 3 | 204,102 (33.1) | 115,223 (23.0) | 88,879 (78.0) |
| HER2 status | |||
| Negative | 236,238 (38.4) | 205,091 (40.9) | 31,147 (27.3) |
| Positive | 37,450 (6.1) | 26,338 (5.2) | 11,112 (9.7) |
| Unknown | 342,177 (55.6) | 270,456 (53.9) | 71,721 (62.9) |
| PR status | |||
| Negative | 179,517 (29.1) | 72,459 (14.4) | 107,058 (93.9) |
| Positive | 436,348 (70.9) | 429,426 (85.6) | 6,922 (6.1) |
| Radiotherapy | |||
| No | 263,529 (42.8) | 210,148 (41.9) | 53,381 (46.8) |
| Yes | 352,336 (57.2) | 291,737 (58.1) | 60,599 (53.2) |
| Chemotherapy | |||
| No | 350,619 (56.9) | 317,956 (63.4) | 32,663 (28.7) |
| Yes | 265,246 (43.1) | 183,929 (36.6) | 81,317 (71.3) |
| Vital status | |||
| Alive | 449,500 (73.0) | 373,331 (74.4) | 76,169 (66.8) |
| Died of breast cancer | 70,977 (11.5) | 48,543 (9.7) | 22,434 (19.7) |
| Died of other causes | 95,388 (15.5) | 80,011 (15.9) | 15,377 (13.5) |
Abbreviations: ER, estrogen receptor; NH, non-Hispanic; PR, progesterone receptor.
Calibration
Overall, PREDICT Breast v3 was well-calibrated. The predicted number of deaths at 10 years was within 10% of the observed deaths for patients with ER-positive cancer (64,990 predicted/72,712 observed) and those with ER-negative cancer (25,537 predicted/25,021 observed). At 15 years, the predictions were within 5% (ER-positive: 58,792 predicted/61,490 observed; ER-negative: 19,686 predicted/19,033 observed). The observed and predicted numbers of deaths from all causes at 10 and 15 years, stratified by patient demographics, tumor characteristics, and treatment, are shown in Table 2 for ER-positive breast cancer (with the corresponding overall survival table and 95% confidence intervals provided in Appendix 2) and Table 3 for ER-negative breast cancer (with the corresponding overall survival table and 95% confidence intervals provided in Appendix 3). In most subgroups, the observed and predicted numbers of deaths were within 10%. However, calibration was poor for non-Hispanic Asians with ER-negative cancer, with PREDICT Breast v3 overpredicting the number of deaths at 10 and 15 years by >30%. Similarly, calibration in non-Hispanic Black women with ER-positive breast cancer was poor, with PREDICT Breast v3 underpredicting the number of deaths by ≥20% at 10 and 15 years. The observed and predicted numbers of deaths from breast cancer are shown in Appendix 4 for ER-positive breast cancer and Appendix 5 for ER-negative breast cancer, whereas deaths from other causes are shown in Appendix 6 for ER-positive breast cancer and Appendix 7 for ER-negative breast cancer. In general, breast cancer–specific mortality tended to be overestimated, whereas mortality from other causes tended to be underestimated.
Cumulative Observed and Predicted 10- and 15-Year All-Cause Mortality in ER-Positive Breast Cancer
| Characteristic | 10 Years | 15 Years | ||||||
|---|---|---|---|---|---|---|---|---|
| N | Pred | Obs | ΔPred−Obs n (%) |
N | Pred | Obs | ΔPred−Obs n (%) | |
| Patients, N | 320,878 | 64,990 | 72,712 | −7,722 (−10.6) | 166,150 | 58,792 | 61,490 | −2,698 (−4.4) |
| Age at diagnosis | ||||||||
| <36 y | 5,831 | 1,000 | 1,193 | −193 (−16.1) | 3,039 | 857 | 911 | −54 (−5.9) |
| 36–40 y | 11,685 | 1,535 | 1,726 | −191 (−11.1) | 6,324 | 1,420 | 1,424 | −4 (−0.3) |
| 41–50 y | 62,788 | 6,476 | 6,915 | −439 (−6.3) | 33,228 | 6,206 | 5,775 | 431 (7.5) |
| 51–60 y | 82,946 | 9,695 | 11,291 | −1,596 (−14.1) | 43,267 | 9,375 | 9,636 | −261 (−2.7) |
| >60 y | 157,628 | 46,283 | 51,587 | −5,304 (−10.3) | 80,292 | 40,933 | 43,744 | −2,811 (−6.4) |
| Race and ethnicity | ||||||||
| Hispanic (all races) | 27,620 | 5,155 | 5,691 | −536 (−9.4) | 12,752 | 4,170 | 4,255 | −85 (−2.0) |
| NH Asian/Pacific Islander | 24,282 | 4,175 | 3,767 | 408 (10.8) | 11,594 | 3,513 | 3,073 | 440 (14.3) |
| NH Black | 22,870 | 4,633 | 6,789 | −2,156 (−31.8) | 10,683 | 3,749 | 4,779 | −1,030 (−21.5) |
| NH White | 246,106 | 51,027 | 56,465 | −5,438 (−9.6) | 131,121 | 47,360 | 49,383 | −2,023 (−4.1) |
| Tumor size | ||||||||
| 0–10 mm | 91,139 | 13,795 | 14,724 | −929 (−6.3) | 46,500 | 13,320 | 13,935 | −615 (−4.4) |
| 11–20 mm | 128,288 | 23,177 | 25,743 | −2,566 (−10.0) | 68,325 | 22,207 | 23,417 | −1,210 (−5.2) |
| 21–50 mm | 86,769 | 22,542 | 26,101 | −3,559 (−13.6) | 44,242 | 19,197 | 20,145 | −948 (−4.7) |
| >50 mm | 14,682 | 5,476 | 6,144 | −668 (−10.9) | 7,083 | 4,068 | 3,993 | 75 (1.9) |
| Tumor nodes | ||||||||
| 0 | 217,994 | 37,458 | 41,490 | −4,032 (−9.7) | 110,634 | 34,907 | 36,942 | −2,035 (−5.5) |
| 1 | 44,080 | 8,770 | 10,118 | −1,348 (−13.3) | 23,373 | 8,056 | 8,443 | −387 (−4.6) |
| 2–4 | 35,223 | 8,858 | 9,990 | −1,132 (−11.3) | 19,219 | 7,968 | 8,202 | −234 (−2.9) |
| 5–9 | 14,578 | 5,292 | 6,087 | −795 (−13.1) | 8,008 | 4,411 | 4,503 | −92 (−2.0) |
| ≥10 | 9,003 | 4,613 | 5,027 | −414 (−8.2) | 4,916 | 3,451 | 3,400 | 51 (1.5) |
| Tumor grade | ||||||||
| 1 | 87,284 | 13,533 | 15,732 | −2,199 (−14.0) | 43,401 | 12,650 | 14,137 | −1,487 (−10.5) |
| 2 | 156,320 | 30,324 | 34,981 | −4,657 (−13.3) | 80,898 | 27,634 | 29,825 | −2,191 (−7.3) |
| 3 | 77,274 | 21,133 | 21,999 | −866 (−3.9) | 41,851 | 18,509 | 17,528 | 981 (5.6) |
| HER2 status | ||||||||
| Negative | 52,046 | 9,059 | 11,331 | −2,272 (−20.1) | — | — | — | — |
| Positive | 6,529 | 1,003 | 1,308 | −305 (−23.3) | — | — | — | — |
| Unknown | 262,303 | 54,928 | 60,073 | −5,145 (−8.6) | 166,150 | 58,792 | 61,490 | −2,698 (−4.4) |
| PR status | ||||||||
| Negative | 49,184 | 11,334 | 13,766 | −2,432 (−17.7) | 26,729 | 10,473 | 11,254 | −781 (−6.9) |
| Positive | 271,694 | 53,656 | 58,946 | −5,290 (−9.0) | 139,421 | 48,320 | 50,236 | −1,916 (−3.8) |
| Chemotherapy | 122,135 | 21,905 | 24,483 | −2,578 (−10.5) | 64,553 | 19,549 | 19,761 | −212 (−1.1) |
| Radiotherapy | 182,935 | 34,397 | 35,462 | −1,065 (−3.0) | 95,200 | 31,397 | 31,142 | 255 (0.8) |
Abbreviations: ER, estrogen receptor; NH, non-Hispanic; Obs, observed number of deaths; PR, progesterone receptor; Pred, predicted number of deaths.
Cumulative Observed and Predicted 10- and 15-Year All-Cause Mortality in ER-Negative Breast Cancer
| Characteristic | 10 Years | 15 Years | ||||||
|---|---|---|---|---|---|---|---|---|
| N | Pred | Obs | ΔPred−Obs n (%) |
N | Pred | Obs | ΔPred−Obs n (%) | |
| Patients, N | 81,844 | 25,537 | 25,021 | 516 (2.1) | 46,924 | 19,686 | 19,033 | 653 (3.4) |
| Age at diagnosis | ||||||||
| <36 y | 3,641 | 1,252 | 893 | 359 (40.2) | 2,219 | 926 | 606 | 320 (52.9) |
| 36–40 y | 5,333 | 1,497 | 1,245 | 252 (20.3) | 3,290 | 1,146 | 942 | 204 (21.7) |
| 41–50 y | 19,022 | 4,513 | 4,442 | 71 (1.6) | 11,419 | 3,525 | 3,321 | 204 (6.1) |
| 51–60 y | 23,574 | 5,628 | 5,785 | −157 (−2.7) | 13,470 | 4,396 | 4,327 | 69 (1.6) |
| >60 y | 30,274 | 12,646 | 12,656 | −10 (−0.1) | 16,526 | 9,693 | 9,837 | −144 (−1.5) |
| Race and ethnicity | ||||||||
| Hispanic (all races) | 8,885 | 2,735 | 2,561 | 174 (6.8) | 4,742 | 1,948 | 1,818 | 130 (7.2) |
| NH Asian/Pacific Islander | 6,006 | 1,763 | 1,314 | 449 (34.2) | 3,231 | 1,280 | 987 | 293 (29.7) |
| NH Black | 12,474 | 3,938 | 4,488 | −550 (−12.3) | 6,843 | 2,852 | 3,095 | −243 (−7.9) |
| NH White | 54,479 | 17,100 | 16,658 | 442 (2.7) | 32,108 | 13,606 | 13,133 | 473 (3.6) |
| Tumor size | ||||||||
| 0–10 mm | 14,126 | 2,499 | 2,581 | −82 (−3.2) | 7,880 | 2,211 | 2,391 | −180 (−7.5) |
| 11–20 mm | 27,392 | 6,855 | 6,899 | −44 (−0.6) | 16,182 | 5,760 | 5,704 | 56 (1.0) |
| 21–50 mm | 33,653 | 12,284 | 11,971 | 313 (2.6) | 19,234 | 9,185 | 8,706 | 479 (5.5) |
| >50 mm | 6,673 | 3,899 | 3,570 | 329 (9.2) | 3,628 | 2,530 | 2,232 | 298 (13.3) |
| Tumor nodes | ||||||||
| 0 | 52,444 | 12,260 | 12,239 | 21 (0.2) | 29,310 | 9,914 | 9,990 | −76 (−0.8) |
| 1 | 10,882 | 3,496 | 3,576 | −80 (−2.2) | 6,314 | 2,677 | 2,607 | 70 (2.7) |
| 2–4 | 10,027 | 4,281 | 4,087 | 194 (4.7) | 6,070 | 3,219 | 2,969 | 250 (8.4) |
| 5–9 | 4,828 | 2,803 | 2,638 | 165 (6.3) | 2,974 | 2,032 | 1,823 | 209 (11.5) |
| ≥10 | 3,663 | 2,697 | 2,481 | 216 (8.7) | 2,256 | 1,845 | 1,644 | 201 (12.2) |
| Tumor grade | ||||||||
| 1 | 2,434 | 523 | 522 | 1 (0.1) | 1,543 | 505 | 516 | −11 (−2.1) |
| 2 | 15,805 | 4,219 | 4,691 | −472 (−10.1) | 9,199 | 3,512 | 3,823 | −311 (−8.1) |
| 3 | 63,605 | 20,795 | 19,808 | 987 (5.0) | 36,182 | 15,669 | 14,694 | 975 (6.6) |
| HER2 status | ||||||||
| Negative | 8,826 | 2,354 | 2,797 | −443 (−15.9) | — | — | — | — |
| Positive | 2,925 | 669 | 699 | −30 (−4.3) | — | — | — | — |
| Unknown | 70,093 | 22,514 | 21,525 | 989 (4.6) | 46,924 | 19,686 | 19,033 | 653 (3.4) |
| PR status | ||||||||
| Negative | 76,941 | 24,131 | 23,729 | 402 (1.7) | 43,808 | 18,507 | 17,928 | 579 (3.2) |
| Positive | 4,903 | 1,406 | 1,292 | 114 (8.8) | 3,116 | 1,180 | 1,105 | 75 (6.8) |
| Chemotherapy | 56,335 | 16,523 | 15,878 | 645 (4.1) | 31,289 | 12,294 | 11,567 | 727 (6.3) |
| Radiotherapy | 42,795 | 12,531 | 11,940 | 591 (4.9) | 24,981 | 9,847 | 9,182 | 665 (7.2) |
Abbreviations: ER, estrogen receptor; NH, non-Hispanic; Obs, observed number of deaths; PR, progesterone receptor; Pred, predicted number of deaths.
Results of calibration at 5 years of follow-up for all-cause, breast cancer–specific, and other-cause mortalities are presented in Appendix 8 for ER-positive breast cancer and Appendix 9 for ER-negative breast cancer. PREDICT Breast v3 tended to underestimate both breast cancer–related and other deaths at 5 years, with a more substantial miscalibration observed for patients with ER-positive breast cancer.
Goodness-of-Fit
The comparison between predicted and observed all-cause mortality by quintiles of predicted risk is shown in Appendix 10. Overall, PREDICT Breast demonstrated good calibration across most quartiles, with the greatest miscalibration in patients at the highest risk. Goodness-of-fit plots for breast cancer–specific mortality and mortality from other causes at 10 and 15 years are shown in Appendix 11. Additionally, goodness-of-fit plots for all-cause mortality, breast cancer–specific mortality, and mortality from other causes are presented in Appendix 12.
Discrimination
Overall, model discrimination was very good in women with both ER-positive breast cancer (AUCs of 0.769 for 10-year follow-up and 0.794 for 15-year follow-up) (Appendix 13) and ER-negative breast cancer (AUCs of 0.738 for 10-year follow-up and 0.746 for 15-year follow-up) (Appendix 14). There was little difference in discrimination by race. Discrimination was slightly better for other-cause mortality than for breast cancer–specific mortality, as shown in Appendix 10. AUCs for 5-year mortality were similar (Appendix 15).
Performance of PREDICT Breast v3 Versus PREDICT Breast v2.2 and CancerMath
Calibration and discrimination for all 3 models for all-cause mortality at 10 and 15 years are shown in Appendix 16. Both PREDICT v2.2 and CancerMath substantially overpredicted the number of deaths, with calibration being particularly poor for CancerMath. Discrimination was good for all 3 models, with PREDICT v3 slightly outperforming the other 2 models.
Discussion
This study is the first to assess the performance of PREDICT Breast v3 in a non-UK population. Overall, the model performed well, demonstrating good calibration and discrimination at 10 and 15 years for patients with ER-positive and ER-negative breast cancers. The overall performance was similar to that observed in a large series of patients from the United Kingdom.9 Discrimination was generally very good across all populations for both ER-negative and ER-positive patients; however, calibration was poorer in specific populations. In particular, the model overestimated mortality in non-Hispanic Asian patients with ER-negative disease and underestimated mortality in non-Hispanic Black patients with ER-positive disease. The latter finding is consistent with results from a validation of PREDICT Breast v2 in the US population.22
The primary purpose of PREDICT Breast is to provide estimates of the absolute survival benefit associated with adjuvant therapies to aid shared decision-making between patients and their oncologists. Model performance indicates that PREDICT Breast v3 is sufficiently accurate in the US non-Hispanic White population to be incorporated into routine oncologic practice. However, the model is likely to overestimate the benefits of adjuvant therapy in non-Hispanic Asian patients with ER-negative disease and underestimate the benefits in non-Hispanic Black patients with ER-positive disease. These under/overestimates are approximately one-third at 10 years, and this should be considered when using the model for decision-making in these populations. PREDICT Breast consists of 2 main components: the baseline hazard and a set of coefficients (log hazard ratios) for each prognostic factor. Poor calibration is primarily due to misspecification of the baseline hazard, whereas discrimination depends on the set of coefficients. Given that discrimination was good across all population groups, completely refitting the model to generate different sets of population-specific coefficients is unlikely to improve the fit of the model substantially. However, improvements in calibration could be easily achieved by modifying the baseline hazard to be population-specific.
When comparing the performance of PREDICT Breast v3 in the United States to previous validations, several similarities and differences emerge. Our findings align with the UK v3 validation in demonstrating good calibration for breast cancer mortality, although we observed slightly poorer calibration in ER-negative patients, likely due to differences in population characteristics and treatment practices.33 The SEER dataset, which includes nearly 30% of patients from racial and ethnic minority groups, differs from the UK registry, which is predominantly White (100%). The greater homogeneity in the UK dataset likely contributed to stronger calibration and discrimination (AUC) across subgroups, whereas the more diverse SEER population, which includes racial and ethnic minorities, may have resulted in slightly lower performance. Compared with the validation of PREDICT Breast v2 in the US population,21 which reported 10-year estimated survival rates of 71.8% and 60% for patients with ER-positive and ER-negative breast cancer, respectively, our analysis yielded higher 10-year estimated survival rates of 79.7% and 68.8%, respectively (Appendices 2 and 3). This suggests that PREDICT Breast v3 not only provides accurate survival estimates but also demonstrates strong model performance in the US population. In this performance assessment of PREDICT Breast v3 using SEER data, calibration results for ER-positive patients closely matched those of the overall population, with predicted and observed survival rates well-aligned for both 10- and 15-year mortality, reflecting this subgroup’s homogeneity and responsiveness to hormonal therapies. For patients with ER-negative breast cancer, calibration remained reasonably good despite greater heterogeneity, demonstrating the model’s applicability to this subgroup. Among racial and ethnic groups, calibration was strong for non-Hispanic White and Hispanic patients but weaker for non-Hispanic Asian and Black patients, where observed survival rates were lower than predicted. These findings highlight the need for further refinement to address disparities in health care access, socioeconomic factors, and biologic variations in these populations.
Our findings have important clinical implications, given that breast cancer remains the most prevalent cancer among women globally and continues to be a major public health concern. Prognostic tools like PREDICT Breast v3 play a vital role in guiding personalized treatment decisions and addressing concerns about overtreatment and treatment risks. Despite advancements in cancer care having led to reduced mortality in the United States, these benefits are not evenly distributed. The performance gaps observed among non-Hispanic Black and non-Hispanic Asian patients in this study highlight the need for more inclusive models that better serve underrepresented populations. Widely used in clinical practice, the PREDICT Breast online tool—with >400,000 global sessions annually and multilingual support—provides a freely accessible decision-making aid for clinicians and patients. Expanding its applicability to diverse populations in future studies is essential to ensuring fairness and reducing health disparities, ultimately enabling more equitable and personalized cancer care.
However, several limitations exist regarding the overall study design, dataset used, and PREDICT Breast v3 model. First, the assumption that all patients aged <65 years receiving chemotherapy underwent ddAC-T chemotherapy overlooks the common use of TC chemotherapy in adjuvant settings.34 Additionally, assuming all patients with ER-positive breast cancer adhere to hormone therapy is less than ideal, because adherence rates in this group are typically <80%.35 These simplifications may compromise the model’s accuracy. To further evaluate the impact of varying assumptions underlying specific parameter settings, we conducted a sensitivity analysis focusing on 3 key parameters: the type of chemotherapy generation, the percentage of patients receiving hormone therapy, and the percentage of patients detected through screening. Appendix 17 presents the model’s discrimination performance across different parameter configurations, highlighting the robustness and consistency of the results. Second, limitations related to the choice of the SEER database include the lack of certain covariates and its reliance on self-reported race/ethnicity, which may introduce reporting biases and limit generalizability across diverse populations, compared with the National Cancer Database (NCDB), which was previously used to validate PREDICT Breast v2 in the US population.21 Third, a key limitation of the PREDICT Breast v3 model is its exclusion of certain prognostic markers, such as tumor gene expression profiles or genomic risk scores (GRS) such as EndoPredict (Myriad Genetics), MammaPrint (Agendia), and Oncotype DX (Exact Sciences). Although GRS has been validated in multiple studies, only one has evaluated its added value to clinical variables, showing minor improvements in discrimination and limited impact on reclassification.36 Thus, integrating GRS into PREDICT Breast v3 is unlikely to significantly enhance performance.37
Conclusions
We have demonstrated that PREDICT Breast v3 performs well for most patients with breast cancer in the United States. Future work will involve evaluating the benefit of incorporating GRS into the model and modifying the model to improve performance across all ancestries, ensuring it more accurately reflects the diversity of the US population.
References
- 1.↑
Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229–263.
- 3.↑
Early Breast Cancer Trialists’ Collaborative Group (EBCTCG). Effects of chemotherapy and hormonal therapy for early breast cancer on recurrence and 15-year survival: an overview of the randomised trials. Lancet 2005;365:1687–1717.
- 4.↑
Katz SJ, Morrow M. Addressing overtreatment in breast cancer: the doctors’ dilemma. Cancer 2013;119:3584–3588.
- 5.↑
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.
- 6.↑
Campbell HE, Taylor MA, Harris AL, Gray AM. 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;101:1074–1084.
- 7.↑
Chen LL, Nolan ME, Silverstein MJ, et al. The impact of primary tumor size, lymph node status, and other prognostic factors on the risk of cancer death. Cancer 2009;115:5071–5083.
- 8.↑
Liao GS, Chou YC, Hsu HM, et al. The prognostic value of lymph node status among breast cancer subtypes. Am J Surg 2015;209:717–724.
- 9.↑
Grootes I, Wishart GC, Pharoah PDP. An updated PREDICT breast cancer prognostic model including the benefits and harms of radiotherapy. NPJ Breast Cancer 2024;10:6.
- 10.↑
Candido Dos Reis F, Wishart GC, Dicks EM, et al. An updated PREDICT breast cancer prognostication and treatment benefit prediction model with independent validation. Breast Cancer Res 2017;19:58.
- 11.↑
Maishman T, Copson E, Stanton L, et al. An evaluation of the prognostic model PREDICT using the POSH cohort of women aged ≤40 years at breast cancer diagnosis. Br J Cancer 2015;112:983–991.
- 12.↑
Gray E, Marti J, Brewster DH, et al. Independent validation of the PREDICT breast cancer prognosis prediction tool in 45,789 patients using Scottish Cancer Registry data. Br J Cancer 2018;119:808–814.
- 13.↑
Wishart GC, Bajdik CD, Azzato EM, et al. A population-based validation of the prognostic model PREDICT for early breast cancer. Eur J Surg Oncol 2011;37:411–417.
- 14.↑
Wong HS, Subramaniam S, Alias Z, et al. The predictive accuracy of PREDICT: a personalized decision-making tool for Southeast Asian women with breast cancer. Medicine (Baltimore) 2015;94:e593.
- 15.↑
Van Maaren MC, Van Steenbeek CD, Pharoah PDP, et al. Validation of the online prediction tool PREDICT v. 2.0 in the Dutch breast cancer population. Eur J Cancer 2017;86:364–372.
- 16.↑
De Glas NA, Bastiaannet E, Engels CC, et al. Validity of the online PREDICT tool in older patients with breast cancer: a population-based study. Br J Cancer 2016;114:395–400.
- 17.↑
Zaguirre K, Kai M, Kubo M, et al. Validity of the prognostication tool PREDICT version 2.2 in Japanese breast cancer patients. Cancer Med 2021;10:1605–1613.
- 18.↑
Nair NS, Kothari B, Gupta S, et al. Validation of PREDICT version 2.2 in a retrospective cohort of Indian women with operable breast cancer. JCO Glob Oncol 2023;9:e2300114.
- 19.↑
Aguirre U, García‐Gutiérrez S, Romero A, et al. External validation of the PREDICT tool in Spanish women with breast cancer participating in population‐based screening programmes. J Eval Clin Pract 2019;25:873–880.
- 20.↑
Grootes I, Keeman R, Blows FM, et al. Incorporating progesterone receptor expression into the PREDICT breast prognostic model. Eur J Cancer 2022;173:178–193.
- 21.↑
Stabellini N, Cao L, Towe CW, et al. Validation of the PREDICT prognostication tool in US patients with breast cancer. J Natl Compr Canc Netw 2023;21:1011–1019.e6.
- 22.↑
Deng Z, Jones MR, Wolff AC, Visvanathan K. Evaluation of Predict, a prognostic risk tool, after diagnosis of a second breast cancer. JNCI Cancer Spectr 2023;7:pkad081.
- 23.↑
Murphy PK, Sellers ME, Bonds SH, Scott S. The SEER Program’s longstanding commitment to making cancer resources available. J Natl Cancer Inst Monogr 2024;2024:118–122.
- 24.↑
Puvanesarajah S, Gapstur SM, Patel AV, et al. Mode of detection and breast cancer mortality by follow-up time and tumor characteristics among screened women in Cancer Prevention Study-II. Breast Cancer Res Treat 2019;177:679–689.
- 25.↑
Miller JW, King JA, Trivers KF, et al. Vital signs: mammography use and association with social determinants of health and health-related social needs among women – United States, 2022. MMWR Morb Mortal Wkly Rep 2024;73:351–357.
- 26.↑
United States Census Bureau. Age and sex composition in the United States: 2023. Accessed January 14, 2025. Available at: https://www.census.gov/data/tables/2023/demo/age-and-sex/2023-age-sex-composition.html
- 27.↑
Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Br J Cancer 2015;112:251–259.
- 30.↑
Therneau T. A package for survival analysis in R. Accessed March 3, 2025. Available at: https://cran.r-project.org/web/packages/survival/vignettes/survival.pdf
- 31.↑
Robin X, Turck N, Hainard A, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011;12:77.
- 32.↑
Kabacoff RI. R in Action, Third Edition: Data Analysis and Graphics With R and Tidyverse. Manning Publications; 2022.
- 33.↑
Dunnwald LK, Rossing MA, Li CI. Hormone receptor status, tumor characteristics, and prognosis: a prospective cohort of breast cancer patients. Breast Cancer Res 2007;9:R6.
- 34.↑
Barcenas CH, Niu J, Zhang N, et al. Risk of hospitalization according to chemotherapy regimen in early-stage breast cancer. J Clin Oncol 2014;32:2010–2017.
- 35.↑
Cronin KA, Lake AJ, Scott S, et al. Annual report to the nation on the status of cancer, part I: national cancer statistics. Cancer 2018;124:2785–2800.
- 36.↑
Mihaescu R, van Zitteren M, van Hoek M, et al. Improvement of risk prediction by genomic profiling: reclassification measures versus the area under the receiver operating characteristic curve. Am J Epidemiol 2010;172:353–361.
- 37.↑
Pharoah PDP. Discussing validation of the PREDICT prognostication tool in patients with breast cancer. J Natl Compr Canc Netw 2023;21:1107–1108.
