PLOS ONE RESEARCH ARTICLE Developing and validating an individualized breast cancer risk prediction model for women attending breast cancer screening Javier Louro ID1,2,3,4, Marta Román ID1,2,3*, Margarita Posso1,2,3, Ivonne Vázquez5, Francina Saladié ID6, Ana Rodriguez-Arana7, M. Jesús Quintana8,9, Laia Domingo1,2,3, Marisa Baré2,10, Rafael Marcos-Gragera ID9,11, Marı́a Vernet-Tomas12, Maria Sala1,2,3, Xavier Castells1,2,3, on behalf of the BELE and IRIS Study Groups¶ 1 IMIM (Hospital del Mar Medical Research Institute), Barcelona, Spain, 2 Research Network on Health Services in Chronic Diseases (REDISSEC), Barcelona, Spain, 3 Servei d’Epidemiologia i Avaluació, Hospital a1111111111 del Mar, Barcelona, Spain, 4 European Higher Education Area (EHEA) Doctoral Programme in Methodology a1111111111 of Biomedical Research and Public Health in Department of Pediatrics, Obstetrics and Gynecology, a1111111111 Preventive Medicine and Public Health, Universitat Autónoma de Barcelona (UAB), Bellaterra, Barcelona, a1111111111 Spain, 5 Servei de Patologia, Hospital del Mar, Barcelona, Spain, 6 Cancer Epidemiology and Prevention a1111111111 Service, Hospital Universitari Sant Joan de Reus, Institut d’Investigació Sanitària Pere Virgili, Universitat Rovira i Virgili, Reus, Spain, 7 Servei de Diagnòstic per la imatge, Hospital del Mar, Barcelona, Spain, 8 Department of Clinical Epidemiology and Public Health, University Hospital de la Santa Creu i Sant Pau, IIB Sant Pau, Barcelona, Barcelona, Spain, 9 CIBER of Epidemiology and Public Health (CIBERESP), Barcelona, Spain, 10 Clinical Epidemiology and Cancer Screening, Parc Taulı́ University Hospital, Sabadell, Spain, 11 Department of Health, Epidemiology Unit and Girona Cancer Registry, Oncology Coordination OPEN ACCESS Plan, Autonomous Government of Catalonia, Catalan Institute of Oncology, Girona, Spain, 12 Servei d’Obstetricia i Ginecologia, Hospital del Mar, Barcelona, Spain Citation: Louro J, Román M, Posso M, Vázquez I, Saladié F, Rodriguez-Arana A, et al. (2021) ¶ Complete membership of the author group can be found in the Acknowledgments Developing and validating an individualized breast * mroman@parcdesalutmar.cat cancer risk prediction model for women attending breast cancer screening. PLoS ONE 16(3): e0248930. https://doi.org/10.1371/journal. pone.0248930 Abstract Editor: Erin J. A. Bowles, Kaiser Permanente Washington, UNITED STATES Background Received: September 23, 2020 Several studies have proposed personalized strategies based on women’s individual breast can- Accepted: March 8, 2021 cer risk to improve the effectiveness of breast cancer screening. We designed and internally vali- Published: March 23, 2021 dated an individualized risk prediction model for women eligible for mammography screening. Peer Review History: PLOS recognizes the benefits of transparency in the peer review Methods process; therefore, we enable the publication of Retrospective cohort study of 121,969 women aged 50 to 69 years, screened at the long- all of the content of peer review and author standing population-based screening program in Spain between 1995 and 2015 and fol- responses alongside final, published articles. The editorial history of this article is available here: lowed up until 2017. We used partly conditional Cox proportional hazards regression to esti- https://doi.org/10.1371/journal.pone.0248930 mate the adjusted hazard ratios (aHR) and individual risks for age, family history of breast Copyright: © 2021 Louro et al. This is an open cancer, previous benign breast disease, and previous mammographic features. We inter- access article distributed under the terms of the nally validated our model with the expected-to-observed ratio and the area under the Creative Commons Attribution License, which receiver operating characteristic curve. permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Results Data Availability Statement: Wwe have uploaded During a mean follow-up of 7.5 years, 2,058 women were diagnosed with breast cancer. All the database to the Harvard Dataverse online three risk factors were strongly associated with breast cancer risk, with the highest risk PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 1 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening repository. The data is accessible with DOI: https:// being found among women with family history of breast cancer (aHR: 1.67), a proliferative doi.org/10.7910/DVN/3T7HCH. benign breast disease (aHR: 3.02) and previous calcifications (aHR: 2.52). The model was Funding: This study was supported by grants from well calibrated overall (expected-to-observed ratio ranging from 0.99 at 2 years to 1.02 at 20 Instituto de Salud Carlos III FEDER [PI15/00098 years) but slightly overestimated the risk in women with proliferative benign breast disease. and PI17/00047]; the Research Network on Health Services in Chronic Diseases [RD12/0001/0015]; The area under the receiver operating characteristic curve ranged from 58.7% to 64.7%, and the Spanish Society of Epidemiology (SEE) [XV depending of the time horizon selected. Alicia Llacer grant for the best research by a young researcher]. Competing interests: The authors have declared Conclusions that no competing interests exist. We developed a risk prediction model to estimate the short- and long-term risk of breast Abbreviations: aHR, Adjusted hazard ratio; AUC, cancer in women eligible for mammography screening using information routinely reported Area under the receiving operating characteristic at screening participation. The model could help to guiding individualized screening strate- curve; BI-RADS, Breast Imaging Reporting and Data System; BBD, Benign breast disease; DCIS, gies aimed at improving the risk-benefit balance of mammography screening programs. Ductal carcinoma in situ; E/O, Expected to observed; SNPs, Single Nucleotide Polymorphisms; 95%CI, 95% Confidence intervals. Introduction There is ongoing debate on the benefits and harms of breast cancer screening [1–3]. To improve this balance, current evidence supports personalized screening [4,5]. Modeling stud- ies have shown that modifying the screening interval, screening modality, or age range of the target population based on women’s individual risk yielded greater benefit than conventional standard strategies [5–7]. Several risk models have been designed to estimate women’s individ- ual breast cancer risk based on their personal characteristics [8–15]. However, most of these models have not been specifically developed to estimate the risk of women targeted for breast cancer screening in order to offer them personalized strategies. A recent consensus statement of the European Conference on Personalized Early Detection and Prevention of Breast Cancer (ENVISION) [16] stated the need to develop breast cancer risk prediction models based on data from large screening cohorts and including risk factors easily obtainable at screening participation, such previous mammographic features and prior benign breast disease. To date, only one model has specifically aimed to predict women’s individual risk looking to personalize breast cancer screening strategies [17]. Although highly valuable, the model was based on short-term risk estimates and did not account for relevant characteristics of prospec- tive studies such as internal time-dependent covariates. This model only estimates the two- year risk, which could lead to bias as one of the aims proposed in breast cancer screening personalization is to see which women are at a lower risk in order to extend their screening period to three or four years. Therefore, if new breast cancer risk models are developed with the aim of analyzing the possibilities offered by personalized screening strategies, it would be interesting to estimate the biennial risk of each woman, in other words, to obtain estimators not only at 2 years, but also every two years (2, 4, 6, 8. . . up to 20 years, which is the total time a woman is screened). This will help to better understand the different possibilities of screen- ing strategies and will allow to observe the differences in the validation of the model estimators for the different time horizons. There is therefore a need for breast cancer risk prediction mod- els, with risk estimates in the short- and long-term, and based on data from large screening cohorts. These new risk models should include a limited and feasible number of variables for the proposed objective, for example, detailed information on the type of previous benign PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 2 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening breast disease or previous mammographic characteristics, which existing risk models tend not to use. We aimed to design and validate an individualized risk prediction model to estimate the biennial risk of breast cancer in women eligible for mammography screening by using data from the long-standing population-based screening program in Spain. Materials and methods Setting and study population Breast cancer screening in Spain started in 1990 in a single setting and expanded until it became nationwide in 2006. This program follows the recommendations of the European Guidelines for Quality Assurance in Breast Cancer Screening and Diagnosis [18]. Women aged 50 to 69 years are invited to biennial screening mammography by written letter. Screen- ing mammograms are interpreted according to the Breast Imaging Reporting and Data System (BI-RADS) scale by trained breast radiologists [19]. Women with an abnormal mammo- graphic feature are recalled for further assessments to confirm or rule out malignancy. Women without a breast cancer diagnosis are invited again for routine screening at 2 years. Overall, breast cancer screening in Spain has a recall rate of 43.0, a detection rate of 4.0, and an interval cancer rate of 1.1 per 1,000 mammographic examinations [20]. The positive predictive value is 9.8% for recalls and 38.9% for recalls involving invasive procedures. Overall, 16.8% of all screen-detected cancers are ductal carcinoma in situ (DCIS). More details of breast cancer screening in Spain are described elsewhere [21]. We analyzed data from two centers forming part of the Spanish breast cancer screening program in the Metropolitan Area of Barcelona. These centers routinely gather information on family history of breast cancer, previous benign breast disease (BBD), and previous mam- mographic features. The centers collect information on screening mammography examina- tions, recalls, further assessments, and diagnostic results in their defined catchment areas. The cohort included all 123,251 women screened at least once between 1995 and 2015 and fol- lowed-up until December 2017. We excluded 758 women diagnosed with breast cancer at the first screen, 210 women with missing information on family history, 213 women with missing information on previous BBD, and 101 women with missing information for both family his- tory and previous BBD. The study population for the analysis consisted of 121,969 women who underwent 437,540 screening mammograms during the study period. Definition of study variables Information on family history and history of prior breast biopsies was self-reported and col- lected from face-to-face interviews conducted by trained professionals at the time of mam- mography screening. This information was consistently collected over the 20 years study period. A family history of breast cancer was defined as having at least one first-degree relative with a history of breast cancer. Breast biopsy results were classified by a community pathologist at each center using SNOMED codes [22]. Pathological diagnoses were grouped following the benign breast disease classification proposed by Dupont and Page [23–25] into non-proliferative and proliferative disease. Proliferative lesions with and without atypia were combined into a single category due to the small number of subsequent breast cancer cases among those with a proliferative lesion with atypia. If women reported having had a biopsy before the start of the screening but no pathology results were available, the biopsy was classified as having a prior biopsy, unknown diagnosis. PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 3 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening A community radiologist routinely reported on mammographic features found at mam- mography screening interpretation. We classified as mammographic features any mass, calcifi- cation, asymmetry or architectural distortion reported by radiologists at mammographic interpretation. Findings were assigned to the category of multiple mammographic features if more than one of the previous mammographic features had been reported simultaneously at screening interpretation. We included both invasive breast cancers and DCIS for the analysis. Model design We built the risk prediction model using a random sample of 60% of the study population (estimation subcohort). The remaining 40% was used for an internal validation (validation subcohort). We estimated the age-adjusted hazard ratios (aHR) and the 95% confidence intervals (95%CI) for the breast cancer incidence for each category of family history, previous BBD, and previous mammographic features with the estimation subcohort. Age was included in the model as a con- tinuous variable. We used partly conditional Cox proportional hazards regression, an extension of the standard Cox model, to incorporate changes in these risk factors over time. Robust standard errors were used to estimate 95% confidence intervals using the Huber sandwich estimator [26]. If a woman has had a diagnosis of cancer, she will contribute women-years at risk from the date of her first mammogram to the diagnosis of cancer. Since we can identify all interval cancers, a woman who has not had a diagnosis of cancer at the end of her follow-up will contribute women- years at risk from the first mammogram to the last mammogram plus 2 years of follow-up. We tested whether family history, previous BBD, and previous mammographic features interacted among themselves or with age. The interaction terms were not significant and were therefore not included in the model. The proportional hazards assumption was assessed by plotting the log-minus-log of the survivor function against log time for each predictor variable. The proportional hazards assumption appeared to be reasonable for all predictors. Model validation We calculated the absolute breast cancer risk estimates for each 2-year interval over the 20-year lifespan covered by screening (ages 50 to 69 years) for each individual in the validation subcohort. As proposed by Zheng and Heagerty, we used a general hazard function to predict the absolute risk of breast cancer diagnosis based on length of follow-up, prediction time, and women’s risk profile [27]. We conducted an internal validation of the model to evaluate its predictive performance by assessing its calibration and discrimination. To assess calibration, we calculated the ratio between the expected breast cancer rate in the validation subcohort versus the observed rate in the estimation subcohort. To account for censoring, the observed rate was estimated using the Kaplan-Meier estimator. The expected breast cancer rate was calculated as the average of the risk estimates in the validation subcohort. The expected breast cancer rate in a specific risk group was calculated as the average of the risk estimates for each woman in that risk group of the validation subcohort. The expected-to-observed (E/O) ratio assessed whether the number of women predicted to develop breast cancer from the model matched the actual number of breast cancers diagnosed in the validation subcohort. An E/O ratio of 1.0 indicates perfect cali- bration. We calculated the E/O ratio 95% confidence intervals (95% CI) using the formula of the standardized mortality ratio proposed by Breslow and Day [28]. The discriminatory accu- racy of our model was assessed by estimating the area under the receiving operating character- istic curve (AUC) for each 2-year interval based on the predicted risks for each woman and PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 4 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening whether she developed breast cancer during the time interval or not [29]. The predicted risks were calculated using the model coefficient estimates at the baseline mammogram for those women in the validation cohort who have been followed for a time greater than or equal to the time horizon being estimated. The AUC measured the ability of the model to discriminate between women who will develop breast cancer from those who will not. We calculated the 95% CI using the approach proposed by Hanley and McNeil [30]. Statistical tests were two-sided and all p-values <0.05 were considered statistically signifi- cant. All analyses were performed using the statistical software R version 3.4.3 (Development Core Team, 2014). The study was approved by the Clinical Research Ethics Committee of Hospital del Mar Medical Research Institute (2015/6189/I). The review boards of the institutions providing data granted approval for data analyses. This is an entirely registry-based study that used anon- ymized retrospective data and hence there was no requirement for written informed consent. The authors declare that they have no conflicts of interest. Results During a mean follow-up of 7.52 years, breast cancer was diagnosed in 2,058 out of the 121,969 women in the study population. The mean follow-up was shorter in women with a breast cancer diagnosis than in those without (5.8 years vs 7.6 years, p-value < 0.05). Women with breast cancer were more likely to have a family history of breast cancer (18.32% vs 13.86%), biopsies with unknown diagnosis (23.76% vs 21.72%), non-proliferative and prolifer- ative BBD (5.59% vs 3.20%, and 1.60% vs 0.45%, respectively), masses (20.51% vs 18.12%), and calcifications (6.85% vs 2.71%) (Table 1). Breast cancer was strongly associated with previous benign breast disease, with the highest risk being found among women with a proliferative BBD (aHR, 3.02; 95% CI: 1.75, 5.21) com- pared with those without a BBD (Table 2). Family history was also associated with breast can- cer (aHR, 1.67; 95% CI: 1.41, 1.98). Among women with previous mammographic features, the highest risks were found in calcifications (aHR, 2.52; 95% CI: 1.93, 3.29) and architectural distortions (aHR, 2.07; 95% CI: 1.27, 3.38). Overall calibration of the model was accurate across all 2-year time horizons. The E/O ratio ranged from 0.99 at 2 years to 1.02 at 20 years and was never significantly different than 1 (Table 3). The AUC was lowest at the 4-year risk estimate (AUC, 58.7%; 95%CI: 55.9%-61.5%) and highest at the 18-year risk estimate (AUC, 64.7%; 95%CI: 62.5%-66.9%) and were signifi- cantly higher than 50% for all the time horizons. Estimates for the 10-year time horizon showed that the model slightly overestimated breast cancer rates in women with masses (E/O ratio, 1.18; 95%CI: 1.02–1.37) and in women aged 55–59 years (E/O ratio, 1.15; 95%CI: (1.03–1.29) (Table 4). The model also underestimated breast cancer rates in women aged 50–54 years (E/O ratio, 0.83; 95%CI: 0.75–0.94). Because of the small number of breast cancer cases, calibration was overestimated among women with proliferative BBD (E/O ratio, 1.85; 95%CI: 1.00–3.40). Distribution of the absolute cumulative risk estimates at 2-, 10- and 20-year time horizons are shown in Fig 1. The 10-year risk was between 1.5% and 2% in 60% of the women and was higher than 2% in 35%. The 20-year risk was lower than 3% in only 4% of the women, between 5% and 7% in 17% of the women, and was higher than 7% in approximately 9% of the women. Discussion We used individual-level data from a large cohort of women regularly screened in Spain to design and validate a risk prediction model to estimate the biennial risk of breast cancer in PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 5 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening Table 1. Baseline characteristics of the study population. No breast cancer (n = 119,911) Breast cancer (n = 2,058) p-value Mean follow- up 7.6 years 5.8 years <0.001 Age (years) 50–54 63,507 (52.96%) 1,149 (55.83%) 0.010 55–59 25,738 (21.46%) 542 (26.34%) <0.001 60–64 22,796 (19.01%) 325 (15.79%) <0.001 65–69 7,870 (6.56%) 42 (2.04%) <0.001 Family history of breast cancer No 103,296 (86.14%) 1,681 (81.68%) <0.001 Yes 16,615 (13.86%) 377 (18.32%) <0.001 Benign breast disease None 89,500 (74.64%) 1,421 (69.05%) <0.001 Prior biopsy, unknown diagnosis 26,042 (21.72%) 489 (23.76%) 0.028 Non-proliferative 3,832 (3.20%) 115 (5.59%) <0.001 Proliferative 537 (0.45%) 33 (1.60%) <0.001 Mammographic features None 86,326 (71.99%) 1,283 (62.34%) <0.001 Mass 21,728 (18.12%) 422 (20.51%) <0.001 Calcifications 3,246 (2.71%) 141 (6.85%) <0.001 Asymmetry 3,371 (2.81%) 56 (2.72%) 0.858 Architectural distortion 1,249 (1.04%) 29 (1.41%) 0.129 Multiple features 3,991 (3.33%) 127 (6.17%) <0.001 Differences in mean of follow-up were tested by Mann–Whitney U test. Differences in qualitative variables were tested by two-sided test of equality for column proportions (z-test). Tests adjusted for all pairwise comparisons within each tumor characteristic using the Bonferroni correction. https://doi.org/10.1371/journal.pone.0248930.t001 Table 2. Partly conditional Cox proportional hazards model results showing the hazard ratios of the risk factors on breast cancer. Women-years Breast cancer cases aHR� (95%CI) Family history of breast cancer No 471,552 976 Ref. Yes 79,471 227 1.67 (1.41–1.98) Benign breast disease No 408,883 832 Ref. Prior biopsy, unknown diagnosis 118,010 286 1.36 (1.16–1.59) Non-proliferative 21,123 67 1.41 (1.02–1.94) Proliferative 3,007 18 3.02 (1.75–5.21) Mammographic features No 380,314 752 Ref. Mass 110,597 239 1.32 (1.11–1.57) Calcifications 17,160 81 2.52 (1.93–3.29) Asymmetry 17,526 38 1.66 (1.16–2.39) Architectural distortion 6,287 20 2.07 (1.27–3.38) Multiple features 19,140 73 1.86 (1.43–2.43) aHR: Adjusted Hazard Ratio. 95%CI: 95% Confidence Interval. � Model adjusted by age, family history, previous benign breast disease and previous mammographic features. https://doi.org/10.1371/journal.pone.0248930.t002 PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 6 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening Table 3. E/O ratio and area under the ROC curve of the model for each time horizon. Observed events E/O ratio (CI95%) AUC 2-year risk 188 0.99 (0.86–1.14) 63.0 (59.1–66.9) 4-year risk 455 1.01 (0.92–1.11) 58.7 (55.9–61.5) 6-year risk 685 1.00 (0.92–1.07) 59.5 (57.2–61.8) 8-year risk 853 1.02 (0.95–1.09) 61.0 (58.9–63.0) 10-year risk 1,000 1.01 (0.95–1.08) 60.9 (59.0–62.8) 12-year risk 1,092 1.03 (0.97–1.09) 60.5 (58.6–62.4) 14-year risk 1,165 1.01 (0.96–1.07) 62.4 (60.5–64.3) 16-year risk 1,195 1.00 (0.95–1.06) 64.3 (62.4–66.3) 18-year risk 1,201 1.01 (0.96–1.07) 64.7 (62.5–66.9) 20-year risk 1,203 1.02 (0.97–1.08) 63.8 (61.3–66.3) E/O: Expected observed. 95%CI: 95% Confidence Interval. https://doi.org/10.1371/journal.pone.0248930.t003 women aged 50 to 69 years eligible for mammography screening. We tested a model that uses only variables easily obtainable at screening participation. The model showed very good cali- bration but only modest discrimination. Our model calculates the risk of breast cancer for each 2-year time horizon during a wom- an’s screening lifespan. Until now, the 5-year risk estimate has been the standard since the BCRAT model used a 5-year risk time horizon for decision making about chemoprevention. The BCRAT model was the basis for enrolment into the two major US prevention trials Table 4. Calibration of the 10-year estimates from the model in risk factor subgroups. Observed events E/O ratio (95%CI) Overall 1,000 1.01 (0.95–1.08) Family history No 824 1.00 (0.93–1.07) Yes 176 1.10 (0.95–1.28) Benign breast disease No 709 1.02 (0.95–1.10) Prior biopsy, unknown diagnosis 238 1.02 (0.90–1.16) Non-proliferative 43 1.17 (0.87–1.57) Proliferative 10 1.85 (1.00–3.40) Mammographic features No 661 1.03 (0.96–1.11) Mass 175 1.18 (1.02–1.37) Calcifications 60 1.07 (0.83–1.38) Asymmetry 29 1.14 (0.79–1.63) Architectural distortion 15 1.05 (0.63–1.73) Multiple features 60 0.90 (0.70–1.16) Age (years) 50–54 296 0.83 (0.75–0.94) 55–59 296 1.15 (1.03–1.29) 60–64 278 0.96 (0.85–1.08) 65–69 130 1.02 (0.86–1.21) E/O: Expected observed. 95%CI: 95% Confidence Interval. https://doi.org/10.1371/journal.pone.0248930.t004 PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 7 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening Fig 1. Distribution of the absolute cumulative risk estimates. https://doi.org/10.1371/journal.pone.0248930.g001 PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 8 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening [31,32]. However, as stated in the statements of the last European Conference on Risk-Strati- fied Prevention and Early Detection of Breast Cancer, there is a need for risk models specifi- cally designed for women eligible for breast cancer screening, based on data from large screening cohorts [16]. A previous model was designed to estimate the risk of breast cancer in women eligible for mammography screening [17]. The model used the Karma cohort from Sweden and included information on mammographic features. That study focused solely on estimating the short- term risk of breast cancer over the next mammographic examination. In addition, it used a case-control design to establish risk factors, which may bias the estimates of the short-term association with breast cancer risk. Our model adds to the breast cancer risk prediction models currently available and can be used to help guide personalized screening strategies by employ- ing information easily obtained at screening participation. Additional useful information from our model is estimation of a woman’s risk for breast cancer at 2-yearly intervals. Our model was further developed by adding the effect of mammographic features, such as masses, calcifications, asymmetries, and architectural distortions. Previous studies have shown that mammographic features increase the subsequent risk of breast cancer [33]. In our model, the strongest influence on risk was conferred by calcifications. The biology behind calcifica- tions is not well established. It has been suggested that mammary cells may acquire some mes- enchymal characteristics, being able to contribute to the production of breast calcifications as a sign of carcinogenic transformation [34]. The role of BBD as a risk factor for breast cancer is well established [9,33,35]. However, its inclusion in breast cancer risk prediction models is rare, mainly because available information on BBD in large cohorts of women is uncommon. Only one previous risk model included dif- ferent estimates for the different categories of the Dupont and Page BBD pathological classifi- cation [23–25]. The Breast Cancer Surveillance Consortium model was updated to include BBD, which led to only minimal improvement in discrimination [9]. This lack of significant improvement could be due to the absence of pathology results for most women who reported breast biopsies prior to their first screening round, as was also the case in our study. However, the addition of BBD to the model markedly increased the proportion of women identified as being at high risk for invasive breast cancer. We assessed the internal validity of the model by means of its calibration and discrimina- tory accuracy. To perform internal validation we split our cohort in two sets, the estimation subcohort, to perform the analysis and development of the model and the validation subco- hort, to perform the internal validation of the model. This technique known as split validation is common for this type of models [9] but cross validation or bootstrapping could also have been performed [36,37]. The model showed accurate calibration, neither overestimating nor underestimating the overall risk through the different years. In Table 4 we saw the calibration of the 10-year estimates from the model in risk factor subgroups. We also performed the E/O ratio estimates in risk factor subgroups for each one of the time horizons proposed. We only showed the 10-year estimates since showing all of them could be confusing. We showed the 10-year estimates since they have a good balance between the number of events observed (in the first time horizons some subcategories have a low number of observed events was lower) and the number of people observed (in the last time horizons we have some lost to follow–up, as the mean time of follow-up is 7.5 years). Nonetheless, the E/O ratio was overestimated for women with a proliferative BBD, due to the small number of cases among this subgroup. The model showed modest discrimination with a maximum AUC of 64.7%. Discriminatory accuracy in breast cancer risk prediction models is usually low because a substantial propor- tion of cases are diagnosed in women with no known risk factors and the AUC of the different models vary between 60 and 70% [14]. This is clearly in contrast with prediction models for PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 9 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening other diseases, such as cardiovascular disease, which achieve good discrimination [38,39]. However, the model presented in this paper performed as well as other models that include many other risk factors that were not available in this study. As one of the reasons why the existing risk models have not been implemented for personalized screening is that it is difficult to collect all of the necessary risk factors in practice, a simpler model like the one we present could be useful. We tested other approaches to validate our model, such as the AUC estimation proposed by Li et al [40]. This estimation uses weights to calculate the contribution in the esti- mates of those women without a breast cancer diagnosis who were censored before reaching the time horizon. However, this approach produced no substantial differences in our validation. A major strength of our model is that we used individual-level data from more than 120,000 women participating in a large, well-established, population-based screening program in Spain from 1995 to 2015, with a mean follow-up of more than 7.5 years and a maximum of 20 years. The program has a participation rate of 67% and a re-attendance rate of 91.2% [19]. This study also has some limitations. First, a major weakness is the lack of information on breast density, which was not systemically collected as part of screening data in the participat- ing centers. Previous models estimating individual breast cancer risk have shown that the addi- tion of breast density improved the discriminatory power of the models [9,17,41,42]. Dense breasts confer women a higher risk of breast cancer and are also associated with a higher risk of false-positive results, masking, and interval cancers [43]. In addition, we had no information on common genetic variants, which has been added to other breast cancer risk prediction models [44,45]. However, the discriminatory accuracy of the models was scarcely improved by the inclusion of information on single nucleotide polymorphisms (SNPs). This lack of both variables may be useful for some institutions where these risk factors are not available. Second, the number of breast cancer cases among women with a proliferative BBD was small, which reduced our ability to accurately predict the expected number of cases across risk factor subgroups. Nevertheless, the overall calibration of the model across the time horizons assessed was highly accurate. Also, as a consequence of the small number of subsequent breast cancer cases among those women with a proliferative BBD with atypia, we merged prolifer- ative BBD with and without atypia into a single category which might make the model less usable in practice. Third, our model was based on a large set of representative data from the Breast Cancer Screening Program in Spain, which provides good generalizability. However, external valida- tion of the results is needed to verify the predictive performance of our risk model. Another limitation might be the reason for censoring. Over 52% of women in the cohort had their last mammogram in the last two years of the study follow-up and 17% of women had their last mammogram at ages 68 or 69 years. Most of the remaining 31% are women who did not participate in the 2014–2015 round or who have changed health areas and thus are not in our study population. The screening program does not have an exhaustive record of which women die and, therefore, we cannot differentiate them from non-participating women. Finally, we were unable to analyze the association between the laterality of the BBD with the subsequent risk of breast cancer. In a previous analysis, we found that 40% of incident breast can- cer cases in women with BBD were contralateral to the prior BBD, suggesting that a large propor- tion of benign lesions may be risk markers rather than precursors of subsequent cancer [46]. Conclusions We designed and internally validated a risk prediction model to estimate the short- and long- term risk of breast cancer in women eligible for mammography screening based on their age, PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 10 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening family history, previous benign breast disease, and previous mammographic features. The model showed good calibration and modest discriminatory power, and could be improved by adding further variables such as breast density and polygenic risk scores. The model can be used biennially to predict a woman’s breast cancer risk during her screening lifespan (age 50 to 69 years) using information easily obtained at screening participation. Risk prediction mod- els specifically designed for women eligible for breast cancer screening are key to guide indi- vidualized screening strategies aiming to improve the risk-benefit balance of mammography screening programs. Acknowledgments The authors acknowledge the dedication and support of the Benign Lesion (BELE) Study Group leaded by Xavier Castells (xcastells@parcdesalutmar.cat) and listed here in alphabetical order and grouped by institution: (a) IMIM (Hospital Del Mar Medical Research Institute), Barcelona, Spain: Andrea Burón, Xavier Castells, Merce Comas, Jose Maria Corominas, Javier Louro, Ana Rodrı́guez-Arana, Marta Román, Maria Sala, Sonia Servitja, Mar Vernet-Tomas; (b) Corporació Sanitària Parc Taulı́, Sabadell, Spain: Marisa Baré, Nuria Tora; (c) Catalan Institute of Oncology, Barcelona, Spain: Llucia Benito, Carmen Vidal (d) Hospital Santa Cate- rina, Girona, Spain: Joana Ferrer; (e) Catalan Institute of Oncology, Girona, Spain: Rafael Mar- cos-Gragera; (f) Hospital de la Santa Creu i Sant Pau, Barcelona, Spain: Judit Solà-Roca, Marı́a Jesús Quintana; (g) General Directorate of Public Health, Government of Cantabria, Spain: Mar Sánchez; (h) Principality of Astúrias Health Service, Spain: Miguel Prieto; (i) Fundació Lliga per a La Investigació i Prevenció Del Cáncer, Tarragona, Spain: Francina Saladié, Jaume Galceran; (j) Hospital Clinic, Barcelona, Spain; Xavier Bargalló, Isabel Torá-Rocamora; (k) Vallés Oriental Breast Cancer Early Detection Program, Spain; Lupe Peñalva; (l) Catalonian Cancer Strategy, Barcelona, Spain: Josep Alfons Espinàs. The authors also acknowledge the dedication and support of the Individualized Risk (IRIS) Study Group leaded by Marta Román (mroman@parcdesalutmar.cat) and listed here in alpha- betical order and grouped by institution: (a) IMIM (Hospital Del Mar Medical Research Insti- tute), Barcelona, Spain: Rodrigo Alcantara, Xavier Castells, Laia Domingo, Javier Louro, Margarita Posso, Maria Sala, Ignasi Tusquets, Ivonne Vazquez, Mar Vernet-Tomas; (b) Cor- poració Sanitària Parc Taulı́, Sabadell, Spain: Marisa Baré, Javier del Riego; (c) Catalan Insti- tute of Oncology, Barcelona, Spain: Llucia Benito, Carmen Vidal (d) Hospital Santa Caterina, Girona, Spain: Joana Ferrer; (e) Catalan Institute of Oncology, Girona, Spain: Rafael Marcos- Gragera; (f) Hospital de la Santa Creu i Sant Pau, Barcelona, Spain: Judit Solà-Roca, Marı́a Jesús Quintana; (g) General Directorate of Public Health, Government of Cantabria, Spain: Mar Sánchez; (h) Principality of Astúrias Health Service, Spain: Miguel Prieto; (i) Fundació Lliga per a La Investigació i Prevenció Del Cáncer, Tarragona, Spain: Francina Saladié, Jaume Galceran; (j) Hospital Clinic, Barcelona, Spain; Xavier Bargalló, Isabel Torá-Rocamora; (k) Vallés Oriental Breast Cancer Early Detection Program, Spain; Lupe Peñalva; (l) Catalonian Cancer Strategy, Barcelona, Spain: Josep Alfons Espinàs. The authors also acknowledge the help of the (l) Biomedical Informatics Research Unit (GRIB) of the UPF; Alfons Gonzalez-Pauner, Ferran Sanz and (m) the Cardiovascular epide- miology and genetics group of the IMIM; Jaume Marrugat, Isaac Subirana, Joan Vila. Javier Louro is a Ph.D. candidate at the Methodology of Biomedical Research and Public Health program, Universitat Autònoma de Barcelona (UAB), Barcelona, Spain Author Contributions Conceptualization: Javier Louro, Marta Román, Margarita Posso, Xavier Castells. PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 11 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening Data curation: Javier Louro, Marta Román, Ivonne Vázquez, Ana Rodriguez-Arana. Formal analysis: Javier Louro. Funding acquisition: Marta Román, Xavier Castells. Investigation: Javier Louro, Marta Román, Margarita Posso, Laia Domingo, Rafael Marcos- Gragera, Marı́a Vernet-Tomas, Xavier Castells. Methodology: Javier Louro, Marta Román, Margarita Posso, Laia Domingo, Xavier Castells. Project administration: Javier Louro, Xavier Castells. Resources: Javier Louro, Francina Saladié, M. Jesús Quintana, Marisa Baré, Rafael Marcos- Gragera. Supervision: Marta Román, Marı́a Vernet-Tomas, Maria Sala. Validation: Maria Sala. Writing – original draft: Javier Louro. Writing – review & editing: Javier Louro, Marta Román, Xavier Castells. References 1. Bleyer A, Welch HG. Effect of three decades of screening mammography on breast-cancer incidence. N Engl J Med. 2012; 367(21):1998–2005. https://doi.org/10.1056/NEJMoa1206809 PMID: 23171096 2. Paci E, Broeders M, Hofvind S, Puliti D, Duffy SW, Group EW. European breast cancer service screen- ing outcomes: a first balance sheet of the benefits and harms. Cancer Epidemiol Biomarkers Prev. 2014; 23(7):1159–63. https://doi.org/10.1158/1055-9965.EPI-13-0320 PMID: 24991022 3. Welch HG, Passow HJ. Quantifying the benefits and harms of screening mammography. JAMA Intern Med. 2014; 174(3):448–54. https://doi.org/10.1001/jamainternmed.2013.13635 PMID: 24380095 4. Onega T, Beaber EF, Sprague BL, Barlow WE, Haas JS, Tosteson AN, et al. Breast cancer screening in an era of personalized regimens: a conceptual model and National Cancer Institute initiative for risk- based and preference-based approaches at a population level. Cancer. 2014; 120(19):2955–64. https://doi.org/10.1002/cncr.28771 PMID: 24830599 5. Vilaprinyo E, Forne C, Carles M, Sala M, Pla R, Castells X, et al. Cost-effectiveness and harm-benefit analyses of risk-based screening strategies for breast cancer. PLoS One. 2014; 9(2):e86858. https:// doi.org/10.1371/journal.pone.0086858 PMID: 24498285 6. Trentham-Dietz A, Kerlikowske K, Stout NK, Miglioretti DL, Schechter CB, Ergun MA, et al. Tailoring Breast Cancer Screening Intervals by Breast Density and Risk for Women Aged 50 Years or Older: Col- laborative Modeling of Screening Outcomes. Ann Intern Med. 2016; 165(10):700–12. https://doi.org/10. 7326/M16-0476 PMID: 27548583 7. Schousboe JT, Kerlikowske K, Loh A, Cummings SR. Personalizing mammography by breast density and other risk factors for breast cancer: analysis of health benefits and cost-effectiveness. Ann Intern Med. 2011; 155(1):10–20. https://doi.org/10.7326/0003-4819-155-1-201107050-00003 PMID: 21727289 8. Gail MH, Brinton LA, Byar DP, Corle DK, Green SB, Schairer C, et al. Projecting individualized probabili- ties of developing breast cancer for white females who are being examined annually. J Natl Cancer Inst. 1989; 81(24):1879–86. https://doi.org/10.1093/jnci/81.24.1879 PMID: 2593165 9. Tice JA, Miglioretti DL, Li CS, Vachon CM, Gard CC, Kerlikowske K. Breast Density and Benign Breast Disease: Risk Assessment to Identify Women at High Risk of Breast Cancer. J Clin Oncol. 2015; 33 (28):3137–43. https://doi.org/10.1200/JCO.2015.60.8869 PMID: 26282663 10. Tyrer J, Duffy SW, Cuzick J. A breast cancer prediction model incorporating familial and personal risk factors. Stat Med. 2004; 23(7):1111–30. https://doi.org/10.1002/sim.1668 PMID: 15057881 11. Barlow WE, White E, Ballard-Barbash R, Vacek PM, Titus-Ernstoff L, Carney PA, et al. Prospective breast cancer risk prediction model for women undergoing screening mammography. J Natl Cancer Inst. 2006; 98(17):1204–14. https://doi.org/10.1093/jnci/djj331 PMID: 16954473 12. Gail MH, Costantino JP, Pee D, Bondy M, Newman L, Selvan M, et al. Projecting individualized absolute invasive breast cancer risk in African American women. J Natl Cancer Inst. 2007; 99(23):1782–92. https://doi.org/10.1093/jnci/djm223 PMID: 18042936 PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 12 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening 13. Rosner B, Colditz GA, Iglehart JD, Hankinson SE. Risk prediction models with incomplete data with application to prediction of estrogen receptor-positive breast cancer: prospective data from the Nurses’ Health Study. Breast Cancer Res. 2008; 10(4):R55. https://doi.org/10.1186/bcr2110 PMID: 18598349 14. Louro J, Posso M, Hilton Boon M, Román M, Domingo L, Castells X, et al. A systematic review and qual- ity assessment of individualised breast cancer risk prediction models. British Journal of Cancer. 2019; 121(1):76–85. https://doi.org/10.1038/s41416-019-0476-8 PMID: 31114019 15. Antoniou AC, Pharoah PP, Smith P, Easton DF. The BOADICEA model of genetic susceptibility to breast and ovarian cancer. Br J Cancer. 2004; 91(8):1580–90. https://doi.org/10.1038/sj.bjc.6602175 PMID: 15381934 16. Pashayan N., Antoniou A.C., Ivanus U. et al. Personalized early detection and prevention of breast can- cer: ENVISION consensus statement. Nat Rev Clin Oncol (2020). 17. Eriksson M, Czene K, Pawitan Y, Leifland K, Darabi H, Hall P. A clinical model for identifying the short- term risk of breast cancer. Breast Cancer Res. 2017; 19(1):29. https://doi.org/10.1186/s13058-017- 0820-y PMID: 28288659 18. Perry N, Broeders M, de Wolf C, Tornberg S, Holland R, von Karsa L. European guidelines for quality assurance in breast cancer screening and diagnosis. Fourth edition—summary document. Ann Oncol. 2008; 19(4):614–22. https://doi.org/10.1093/annonc/mdm481 PMID: 18024988 19. Ascunce N, Salas D, Zubizarreta R, Almazan R, Ibanez J, Ederra M, et al. Cancer screening in Spain. Ann Oncol. 2010; 21 Suppl 3:iii43–51. https://doi.org/10.1093/annonc/mdq085 PMID: 20427360 20. Domingo L, Hofvind S, Hubbard RA, Roman M, Benkeser D, Sala M, et al. Cross-national comparison of screening mammography accuracy measures in U.S., Norway, and Spain. Eur Radiol. 2016; 26 (8):2520–8. https://doi.org/10.1007/s00330-015-4074-8 PMID: 26560729 21. Castells X, Sala M, Ascunce N, Salas D, Zubizarreta R, Casamijana M. Descripción del cribado del cán- cer en España. Proyecto DESCRIC. Madrid: Plan de Calidad para el Sistema Nacional de Salud. Minis- terio de Sanidad y Consumo: Agència d’Avaluació de Tecnologia i Recerca Medı̀ques de Cataluña 2006 [Available from: http://www.cribadocancer.es/images/archivos/mama/documentacion/informes_ tecnicos/archivos/InformeDESCRIC.pdf. 22. Ruch P, Gobeill J, Lovis C, Geissbuhler A. Automatic medical encoding with SNOMED categories. BMC Med Inform Decis Mak. 2008; 8 Suppl 1:S6. https://doi.org/10.1186/1472-6947-8-S1-S6 PMID: 19007443 23. Dupont WD, Page DL. Risk factors for breast cancer in women with proliferative breast disease. N Engl J Med. 1985; 312(3):146–51. https://doi.org/10.1056/NEJM198501173120303 PMID: 3965932 24. Page DL, Dupont WD, Rogers LW, Rados MS. Atypical hyperplastic lesions of the female breast. A long-term follow-up study. Cancer. 1985; 55(11):2698–708. https://doi.org/10.1002/1097-0142 (19850601)55:11<2698::aid-cncr2820551127>3.0.co;2-a PMID: 2986821 25. Page DL, Schuyler PA, Dupont WD, Jensen RA, Plummer WD Jr., Simpson JF. Atypical lobular hyper- plasia as a unilateral predictor of breast cancer risk: a retrospective cohort study. Lancet. 2003; 361 (9352):125–9. https://doi.org/10.1016/S0140-6736(03)12230-1 PMID: 12531579 26. Freedman DA. On the So-Called ‘Huber Sandwich Estimator’ and ‘Robust Standard Errors.’ The Ameri- can Statistician, vol. 60, no. 4, 2006, pp. 299–302. JSTOR, www.jstor.org/stable/27643806. Accessed 23 Feb. 2021. 27. Zheng Y, Heagerty PJ. Partly conditional survival models for longitudinal data. Biometrics. 2005; 61 (2):379–91. https://doi.org/10.1111/j.1541-0420.2005.00323.x PMID: 16011684 28. Breslow NE, Day NE. Statistical methods in cancer research. Volume II—The design and analysis of cohort studies. IARC Sci Publ. 1987(82):1–406. 29. Harrell FE Jr., Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluat- ing assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996; 15(4):361–87. https://doi.org/10.1002/(SICI)1097-0258(19960229)15:4<361::AID-SIM168>3.0.CO;2-4 PMID: 8668867 30. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982; 143(1):29–36. https://doi.org/10.1148/radiology.143.1.7063747 PMID: 7063747 31. Fisher B, Costantino JP, Wickerham DL, Redmond CK, Kavanah M, Cronin WM, et al. Tamoxifen for prevention of breast cancer: report of the National Surgical Adjuvant Breast and Bowel Project P-1 Study. J Natl Cancer Inst. 1998; 90(18):1371–88. https://doi.org/10.1093/jnci/90.18.1371 PMID: 9747868 32. Vogel VG, Costantino JP, Wickerham DL, Cronin WM, Cecchini RS, Atkins JN, et al. Effects of tamoxi- fen vs raloxifene on the risk of developing invasive breast cancer and other disease outcomes: the PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 13 / 14 PLOS ONE Developing a breast cancer risk prediction model for women attending breast cancer screening NSABP Study of Tamoxifen and Raloxifene (STAR) P-2 trial. JAMA. 2006; 295(23):2727–41. https:// doi.org/10.1001/jama.295.23.joc60074 PMID: 16754727 33. Castells X, Tora-Rocamora I, Posso M, Roman M, Vernet-Tomas M, Rodriguez-Arana A, et al. Risk of Breast Cancer in Women with False-Positive Results according to Mammographic Features. Radiol- ogy. 2016; 280(2):379–86. https://doi.org/10.1148/radiol.2016151174 PMID: 26878225 34. Scimeca M, Giannini E, Antonacci C, Pistolese CA, Spagnoli LG, Bonanno E. Microcalcifications in breast cancer: an active phenomenon mediated by epithelial cells with mesenchymal characteristics. BMC Cancer. 2014; 14:286. https://doi.org/10.1186/1471-2407-14-286 PMID: 24758513 35. Hartmann LC, Sellers TA, Frost MH, Lingle WL, Degnim AC, Ghosh K, et al. Benign breast disease and the risk of breast cancer. N Engl J Med. 2005; 353(3):229–37. https://doi.org/10.1056/NEJMoa044383 PMID: 16034008 36. Steyerberg EW, Harrell FE Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y, Habbema JD. Internal valida- tion of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epide- miol. 2001; 54(8):774–781. https://doi.org/10.1016/s0895-4356(01)00341-9 PMID: 11470385 37. Efron B, Tibshirani R. Improvements on cross-validation: The .632+ bootstrap method. J Amer Statist Assoc. 1997; 92(438):548–560. 38. Wilson PW, D’Agostino RB, Levy D, Belanger AM, Silbershatz H, Kannel WB. Prediction of coronary heart disease using risk factor categories. Circulation. 1998; 97(18):1837–47. https://doi.org/10.1161/ 01.cir.97.18.1837 PMID: 9603539 39. Marrugat J, D’Agostino R, Sullivan L, Elosua R, Wilson P, Ordovas J, et al. An adaptation of the Fra- mingham coronary heart disease risk function to European Mediterranean areas. J Epidemiol Commu- nity Health. 2003; 57(8):634–8. https://doi.org/10.1136/jech.57.8.634 PMID: 12883073 40. Li L, Greene T, Hu B. A simple method to estimate the time-dependent receiver operating characteristic curve and the area under the curve with right censored data. Stat Methods Med Res. 2018; 27(8):2264– 78. https://doi.org/10.1177/0962280216680239 PMID: 27895266 41. Kerlikowske K, Gard CC, Sprague BL, Tice JA, Miglioretti DL. One versus Two Breast Density Mea- sures to Predict 5- and 10-Year Breast Cancer Risk. Cancer Epidemiol Biomarkers Prev. 2015; 24 (6):889–97. https://doi.org/10.1158/1055-9965.EPI-15-0035 PMID: 25824444 42. Armero C, Forne C, Rue M, Forte A, Perpinan H, Gomez G, et al. Bayesian joint ordinal and survival modeling for breast cancer risk assessment. Stat Med. 2016; 35(28):5267–82. https://doi.org/10.1002/ sim.7065 PMID: 27523800 43. Posso M, Louro J, Sanchez M, Roman M, Vidal C, Sala M, et al. Mammographic breast density: How it affects performance indicators in screening programmes? Eur J Radiol. 2019; 110:81–7. https://doi.org/ 10.1016/j.ejrad.2018.11.012 PMID: 30599878 44. Zhang X, Rice M, Tworoger SS, Rosner BA, Eliassen AH, Tamimi RM, et al. Addition of a polygenic risk score, mammographic density, and endogenous hormones to existing breast cancer risk prediction models: A nested case-control study. PLoS Med. 2018; 15(9):e1002644. https://doi.org/10.1371/ journal.pmed.1002644 PMID: 30180161 45. Shieh Y, Hu D, Ma L, Huntsman S, Gard CC, Leung JW, et al. Breast cancer risk prediction using a clini- cal risk model and polygenic risk score. Breast Cancer Res Treat. 2016; 159(3):513–25. https://doi.org/ 10.1007/s10549-016-3953-2 PMID: 27565998 46. Castells X, Domingo L, Corominas JM, Tora-Rocamora I, Quintana MJ, Bare M, et al. Breast cancer risk after diagnosis by screening mammography of nonproliferative or proliferative benign breast dis- ease: a study from a population-based screening program. Breast Cancer Res Treat. 2015; 149 (1):237–44. https://doi.org/10.1007/s10549-014-3208-z PMID: 25503778 PLOS ONE | https://doi.org/10.1371/journal.pone.0248930 March 23, 2021 14 / 14