Evaluating mammographic density′s contribution to improve a breast cancer risk model with questionnaire-based and polygenic factors

Abstract

Incorporation of mammographic density into breast cancer risk models may improve risk stratification for tailored screening and prevention. We evaluated the added value of Breast Imaging Reporting and Data System (BI-RADS) breast density to a validated model combining questionnaire-based risk factors and a 313-variant polygenic risk score (PRS), using the Individualized Coherent Absolute Risk Estimator (iCARE) tool for risk model building and validation. Calibration and discrimination were assessed in three prospective cohorts of European-ancestry women (1468 cases, 19,104 controls): US-based Nurses’ Health Study (NHS I and II) and Mayo Mammography Health Study (MMHS); and Sweden-based Karolinska Mammography Project for Risk Prediction of Breast Cancer (KARMA) study. Analyses were stratified by age (<50, ≥50 years). Adding density modestly improved discrimination: among younger women, AUC increased from 65.6% (95% CI: 61.9–69.3%) to 67.0% (95% CI: 63.5–70.6%); among older women, from 65.5% (95% CI: 63.8–67.2%) to 66.1% (95% CI 64.4–67.8%). Among US women aged 50–70 years, 18.4% were identified at ≥3% 5-year risk with density included, capturing 42.4% of future cases; 7.9% were reclassified, identifying 2.8% more future cases. In Sweden, 10.3% were identified at elevated risk, capturing 29.4% of cases, with 5.3% reclassified and 4.4% more cases identified. Integrating density with established risk factors and PRS may enhance breast cancer risk stratification among European-ancestry women, supporting its potential for clinical utility.

Introduction

Clinical application of risk-stratified breast cancer prevention strategies in the population requires the development and robust prospective validation of flexible and comprehensive models for absolute risk prediction to provide accurate individualized risk estimates, in particular for women at high risk for whom such applications have the greatest potential impact^1,2. Several risk models have been developed incorporating different sets of risk factors and targeting different clinical scenarios^3,4; however, further work is needed to demonstrate whether improvements in risk stratification of current models may be achieved by incorporating additional risk factors, ultimately enhancing our ability to identify women at the extremes of risk distribution.

In our previous work, we built and validated a literature-based 5-year breast cancer prediction model incorporating reproductive, lifestyle and behavioral factors, family history and the recently developed PRS composed of 313 common variants⁵ with the Individualized Coherent Absolute Risk Estimator (iCARE) software tool⁶. This tool provides a flexible framework for absolute risk model development, aggregating information on risk-factor associations, population-based age-specific disease incidence rates and competing mortality rates and the risk-factor distributions from multiple data sources, and further implements standardized model validation methods. The model with questionnaire-based risk factors and PRS showed good calibration in multiple populations of European-ancestry women⁷. Moreover, we also predicted that adding mammographic breast density to this model could further improve risk stratification⁸.

Since the discovery by Wolfe in 1976⁹, density has been consistently shown to be a strong risk factor for breast cancer. The radio-opaque structures on a mammogram indicate stromal and epithelial tissue, while the radiolucent area indicates adipose tissue¹⁰. Currently, the most widely used clinical system to classify density is the Breast Imaging-Reporting and Data System (BI-RADS), where density is visually assessed by a radiologist and categorized into four levels: almost entirely fatty, scattered areas of fibro-glandular density, heterogeneously dense, or extremely dense¹¹. Population-wide studies have demonstrated that approximately 50% of the US female population aged 40–74 years have heterogeneously or extremely dense breasts, with extremely dense breasts conferring a two- to four-fold higher relative risk compared to almost entirely fatty breasts^12–15.

Breast density has been incorporated into established risk models like the Gail model, the Breast Cancer Surveillance Consortium (BCSC) model, the Tyrer-Cuzick (IBIS) model, Rosner-Colditz model and the Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm (BOADICEA)^16–23. The BCSC, Tyrer-Cuzick and BOADICEA models also incorporate density into their clinical risk calculator tools^1–3. A robust evaluation of the added value of mammographic breast density to a model with comprehensive information on questionnaire-based risk factors and the most recent polygenic risk scores are needed to determine its effectiveness in improving risk stratification of models.

In our current work, we validate the fully integrated model with questionnaire-based risk factors, the 313-SNP PRS and density for calibration and discrimination in three prospective cohorts of European-ancestry women (two from the US and one from Sweden), totaling 1468 cases and 19,104 controls. Our risk projection and reclassification calculations in the populations of European-ancestry women aged 50–70 years from the US and Sweden show the improvements in risk stratification attainable by incorporating density to the current literature-based model with questionnaire-based risk factors and PRS.

Results

Model validation results

For women younger than 50 years, the integrated model with questionnaire-based risk factors, PRS and density was generally well calibrated in terms of relative and absolute risk across the risk categories in NHS II and KARMA (Fig. 1 and Fig. S1A, B). In the meta-analysis across studies, the area under the curves (AUC) with and without density were 67.0% (63.5–70.6%) vs. 65.6% (61.9–69.3%), respectively (Fig. 3a and Table S6A).

Fig. 1. Calibration and discrimination of 5-year risk predictions of breast cancer for women aged younger than 50 years.

Calibration and discrimination of 5-year risk predictions of breast cancer for women aged younger than 50 years in the nested case-control sample of NHS II and KARMA with risk categories based on deciles of predicted 5-year absolute risk. Validation results are shown for the extended iCARE model that incorporates questionnaire-based risk factors with a PRS based on 313 common germline variants and density. Estimates and 95% CI of the calibration slope and intercept are reported based on a linear regression of the decile-specific observed proportion of cases within 5 years and the average of the predicted 5-year absolute risk. AUC area under the curve, c²χ² goodness-of-fit test, CI confidence interval, E/O expected to observed number of cases, KARMA Karolinska Mammography Project for Risk Prediction of Breast Cancer, NHS II Nurses’ Health Study II, PRS polygenic risk score, QRF questionnaire-based risk factors.

Fig. 3. Risk discrimination is measured by the model area under the curve (AUC) of the iCARE-Lit models.

Risk discrimination measured by the model area under the curve (AUC) of the iCARE-Lit models for a women younger than 50 and b 50 years or older based on a meta-analysis across studies for the risk-factor combinations: incorporates (i) questionnaire-based risk factors only, (ii) questionnaire-based risk factors with density, (iii) questionnaire-based risk factors with a PRS based on 313 common germline variants, (iv) the 313-variant PRS only, (v) the 313-variant PRS and density and (vi) the fully integrated model incorporating questionnaire-based risk factors, the 313-variant PRS and density. Colored dots were used to denote estimates, and colored horizontal lines denote the 95% confidence intervals. iCARE-Lit model based on literature review, MBD mammographic breast density, PRS polygenic risk score, QRF questionnaire-based risk factors.

For women 50 years or older, the integrated model showed some underestimation of overall risk in NHS I and KARMA (NHS I: E/O = 0.87 (0.78-0.98) and KARMA: E/O = 0.87 (0.80-0.94), Fig. 2 and Table S5). This was mainly noticeable in the lowest risk decile (KARMA: E/O = 0.51 (0.28-0.94), NHS I: E/O = 0.41 (0.19–0.88), MMHS: E/O = 0.28 (0.12–0.67), Fig. 2). For this age group, in the meta-analysis across studies, the AUCs with and without density were 66.1% (95% CI 64.4–67.8%) and 65.5% (95% CI: 63.8–67.2%) (Fig. 3b and Table S6B).

Fig. 2. Calibration and discrimination of 5-year risk predictions of breast cancer for women aged 50 years and older.

Calibration and discrimination of 5-year risk predictions of breast cancer for women aged 50 years and older in the nested case-control sample of KARMA, NHS I, and MMHS with risk categories based on deciles of predicted 5-year absolute risk. Validation results are shown for the extended iCARE model that incorporates questionnaire-based risk factors with a PRS based on 313 common germline variants and density. Estimates and 95% CI of the calibration slope and intercept are reported based on a linear regression of the decile-specific observed proportion of cases within 5 years and the average of the predicted 5-year absolute risk. AUC area under the curve, c²χ² goodness-of-fit test, CI confidence interval, E/O expected to observed number of cases, KARMA Karolinska Mammography Project for Risk Prediction of Breast Cancer, MMHS Mayo Mammography Health Study, NHS I Nurses’ Health Study I, PRS polygenic risk score, QRF questionnaire-based risk factors.

Risk stratification and reclassification

The extended model with questionnaire-based risk factors, the 313-variant PRS and density identified 18.4% of the US non-Hispanic White population women 50–70 years old ≥3% predicted 5-year risk, the cut-off used for recommending risk-reducing medication in the United States (Fig. 4 and Table S7). This group includes 42.4% of predicted future cases (Fig. 4 and Table S7). The addition of density led to the reclassification of 7.9% of US non-Hispanic White women aged 50–70 years, with 4.1% moving from below the ≥3% predicted 5-year risk threshold to above and 3.8% moving in the opposite direction. This resulted in the identification of 2.8% of additional future cases (Fig. 4 and Table S8). At and above the 6% risk threshold, the high-risk cutoff in the WISDOM trial²⁴, the fully integrated model identified 3.0% of the US non-Hispanic White population above 50–70 years. This group is expected to include 12.0% of future cases (Fig. 4 and Table S7). Among these women, 1.7% were reclassified, with 1.1% moving from below the ≥6% predicted 5-year risk threshold to above and 0.6% moving in the opposite direction. This led to the identification of 2.2% additional future cases (Table S8 and Fig. 4).

Fig. 4. Reclassification of US women at two high-risk thresholds.

Women of European-ancestry aged 50–70 years in the general population of the USA expected to be identified at high risk of breast cancer according to two risk thresholds, and the incident cases of breast cancer expected to occur in these groups within a 5-year interval, comparing the (i) questionnaire-based risk factor and PRS model with (ii) the fully integrated model with questionnaire-based risk factors, PRS and density. The expected number of women is calculated using 2020 (N = 43,718,160) population estimates from the US Census Bureau for the USA. The expected numbers of cases are estimated using the average predicted 5-year risk in the US population, calculated using the US breast cancer incidence rates and risk-factor distributions (Supplementary data). The 3% threshold is used by the US Preventive Services Task Force for recommending risk-reducing medications, and 6% is used by the WISDOM trial as a cutoff for very high risk^27,60. AR absolute risk, BMI body mass index, MBD mammographic breast density, PRS polygenic risk score, QRF questionnaire-based risk factors, US United States.

In the Swedish population, the integrated model identified 10.3% of women aged 50–70 years ≥3% predicted 5-year risk with 29.4% of future cases expected to occur in this group (Table S7 and Fig. 5). With the addition of density, 5.3% of women of European-ancestry were reclassified, with 3.3% from below to above and 2.0% in the opposite direction. This identified an additional 4.4% of future cases (Table S8 and Fig. 5). At and above the 6% risk threshold, the fully integrated model identified 1.4% of women of European-ancestry aged 50–70 in the Swedish population. 6.7% of future cases are expected to occur in this group (Table S7 and Fig. 5). With the addition of density, 0.9% of all women were reclassified above the ≥6% risk threshold, with 0.7% of women moving from below the threshold to above, and 0.2% moving in the opposite direction. This resulted in the identification of 2.5% additional future cases (Table S8 and Fig. 5).

Fig. 5. Reclassification of Swedish women at two high-risk thresholds.

Women of European-ancestry aged 50–70 years in the general population of the USA expected to be identified at high risk of breast cancer according to two risk thresholds, and the incident cases of breast cancer expected to occur in these groups within a 5-year interval, comparing the (i) questionnaire-based risk factor and PRS model with (ii) the fully integrated model with questionnaire-based risk factors, PRS and density. The expected number of women is calculated using 2016 population estimates (N = 1,249,695) from Statistics Sweden for Sweden. The expected numbers of cases are estimated using the average predicted 5-year risk in the Swedish population, calculated using the SE breast cancer-incidence rates and risk-factor distributions (Supplementary Data). The 3% threshold is used by the US Preventive Services Task Force for recommending risk-reducing medications, and 6% is used by the WISDOM trial as a cutoff for very high risk^27,60. AR absolute risk, BMI body mass index, MBD mammographic breast density, PRS polygenic risk score, QRF questionnaire-based risk factors, SE Sweden.

Discussion

We investigated the added value of incorporating breast density to the iCARE-Lit questionnaire-based risk factor and PRS model to predict 5-year absolute risk of breast cancer among women of European-ancestry. The addition of density to the model resulted in a modest improvement in risk stratification and reclassification. For instance, incorporating density identified an additional 2.8 and 4.4% of future cases at and above a 3% predicted 5-year risk threshold and an additional 2.2 and 2.5% of future cases at and above a 6% predicted 5-year risk threshold in populations of European-ancestry in the USA and Sweden, respectively. The model’s ability to stratify more women above and below clinically relevant risk thresholds would lead to more women being rightfully allocated to high and low-risk categories and, therefore, able to qualify for risk-reducing strategies.

Our model, with an AUC of 67.1% (Table S6A) in women younger than 50 years and 66.1% (Table S6B) in women 50 years or older, demonstrates modest risk discrimination. A key challenge in risk stratification based on models with modest risk discrimination is that a substantial percentage of breast cancer cases occur among women not classified to be at high risk. While the high-risk women could be targeted for certain interventions with inherent risks and benefits, the high proportion of cases outside the high-risk groups underscores the need for implementing broad, globally beneficial public health measures to reduce the population burden of breast cancer. These include lifestyle modifications such as maintaining a healthy BMI, increasing physical activity, and limiting alcohol intake, as advocated by the American Cancer Society and other public health organizations.

The implementation of such risk-stratified screening and prevention strategies of breast cancer in the population is a key goal of risk prediction efforts. Currently, two non-inferiority trials, WISDOM^24,25 and MyPeBS²⁶, are underway investigating the potential of risk-based personalized screening as a safe alternative to mammographic screening programs²⁷. As present screening programs are based solely on age as an entry criterion; comprehensive risk models can pave the way for individualized risk-based screening strategies. For countries where density is collected routinely in the clinical setting, this would be a relatively simple addition to risk models. However, prior to implementing a model incorporating comprehensive questionnaire-based risk factors, PRS and mammographic breast density in clinical and screening practices, a formal clinical utility assessment should be carried out to assess effectiveness, cost-effectiveness, psychological, ethical and legal implications of an intervention.

Our model incorporating density showed some signs of miscalibration at the extremes of the distribution. Since the model is literature-based with the relative risks derived using a literature review, it is subject to structural assumptions about the underlying model. For women younger than 50 years, we obtained an age-adjusted relative risk for density from ref. ¹⁹, which did not account for the correlation of density with BMI²⁸. Moreover, in both NHS and KARMA, mammographic density was measured using semi-automated software, CUMULUS and STRATUS, respectively, and converted to a four-level variable to approximate density. There might be some misclassification due to the conversion from percent density to the four-category visually assessed density. Conversely, in MMHS, the density categorization might have contributed to misspecification as it was visually assessed by radiologists, and although broadly accepted, the BI-RADS reporting system is subject to substantial intra- and inter-observer variability between radiologists, with Kappa values²⁹ ranging between 0.4 and 0.7^4–7. Several studies show a stronger relationship between breast cancer and percent density compared to BI-RADS density^30–33. Incorporating automated quantitative mammographic features of mammographic images into risk models could address these limitations. Interestingly, ref. ³⁴ found similar discriminatory accuracy between their model with automated or clinical density and ref. ³⁵ found that adding both a clinical and automated volumetric density measure improved risk stratification.

Our validation was conducted using three relatively old prospective cohorts of women of European-ancestry in the USA and Sweden, which may not be fully representative of the general population. In particular, MMHS and KARMA cohorts consist of women participating in annual or biannual screening programs, respectively, and the NHS cohort is comprised of nurses, both of which are groups that tend to have higher educational attainment and healthier lifestyles compared with the general population. While this could influence the calibration of absolute risk, the calibration of relative risk estimates is less likely to be substantially affected. Importantly, since the primary application of our model is within breast cancer screening contexts, these cohorts remain highly relevant for our aims. Nevertheless, further validation is needed in more representative and contemporary populations, particularly those more closely aligned with the target settings of integrated healthcare systems, such as the CONNECT for Cancer Prevention Study³⁶.

This paper has a notable strength in the breadth of the validation, as it assesses the model’s performance across three different studies with participants coming from two populations of European-ancestry women in the US and Sweden. Other validation studies either included a smaller number of participants or used combinations of a limited set of risk factors and polygenic risk scores (PRS) with fewer genetic variants (Table S9)^{16,20,37–39}. Recently, the extended version of the BOADICEA model incorporating questionnaire-based risk factors, the 313-SNP PRS, density and rare moderate- and high-risk variants has also been externally validated in KARMA, similarly showing the importance of reclassification and risk stratification³⁷. Other established models have also shown improvement of model discrimination after the incorporation of density^{35,38,40–43}, however most have yet to be externally validated in fully independent cohorts. The independent external validation of risk models is critical before clinical applications; however, a potential barrier is the availability of large prospective cohorts with comprehensive genetic, mammography and risk-factor information. The Tyrer-Cuzick (v 7.02) model and the BCSC model (v2) with mammographic density have been validated in large independent cohorts; however, both these versions do not include genetic information^16,22. In contrast to the three aforementioned validation studies, other established models like the Gail model have only been internally validated after the addition of breast density information^21,40. Moreover, there is an increasing interest in using fully automated density measures using deep learning algorithms^44,45. Using artificial intelligence on mammograms to aid risk prediction has the benefit of moving beyond density by also characterizing (micro)calcifications, focal masses, left-right asymmetry and a landscape of radiomic features^46–48. However, many of these algorithms have yet to be independently validated in epidemiological and clinical studies.

In addition to incorporating additional mammographic features, future advancements in breast cancer risk prediction should aim to distinguish between less and more aggressive subtypes, in situ vs. invasive lesions, and ultimately, between cancers that are likely to be lethal and those that are not. Since subtypes of breast cancer differ in their risk-factor associations and preventive strategies are often tailored to specific subtypes (e.g., based on estrogen receptor status), subtype-specific risk models could improve the identification of women most likely to benefit from targeted interventions. In contrast, our current model—like many widely used tools—estimates the absolute risk of developing any breast cancer, irrespective of subtype or prognosis. While it does not predict breast cancer–specific mortality, it remains a valuable resource for guiding primary prevention and early detection. Although not all incident breast cancers are life-threatening, reducing overall incidence, including less aggressive forms, may help mitigate the burden of overdiagnosis, overtreatment, and associated morbidity. Ultimately, while models predicting the risk of lethal disease represent a critical future direction⁴⁹, incidence-based models are currently more validated and actionable, and thus continue to play an essential role in risk-stratified screening and prevention.

To summarize, the incorporation of density to questionnaire-based risk factors and PRS results in modest improvements in the identification of European-ancestry women at elevated breast cancer risk. Additional prospective validation in diverse populations, in particular of non-European-ancestry women, are needed to ensure equitable clinical application.

Methods

Study populations

Model validation analyses were performed in three case-control samples nested in prospective cohort studies of European-ancestry women: the US-based Nurses’ Health Study (NHS I and II) and Mayo Mammography Health Study (MMHS), and the Sweden-based Karolinska Mammography Project for Risk Prediction of Breast Cancer (KARMA) study. In total, analyses were carried out in 1468 cases and 19,104 controls and were done separately for women younger (NHS II, KARMA; 280 cases, 5037 controls) and older than 50 years (NHS I, MMHS, KARMA; 1188 cases, 14,067 controls). Women with a prior history of breast and other cancers, except for nonmelanoma skin cancer, were excluded from the study. Women consented to the use of their genetic material, mammogram with density and completed a risk-factor questionnaire. For women with multiple mammograms or multiple questionnaires, the data closest to the DNA collection were used. In MMHS, density was obtained from routine clinical examination by attending radiologists. All four mammogram views (craniocaudal and mediolateral oblique for ipsilateral and contralateral sides) contribute to the assessment of density⁵⁰. In both NHS and KARMA, mammographic density was measured using semi-automated software, Cumulus⁵¹ and STRATUS⁵², respectively, and converted to a four-level variable to approximate density categories following BI-RADS⁵³. With Cumulus software, an area-based measure of mammographic density is estimated with user-defined thresholds to define dense tissue, and percent density (i.e., dense tissue area/total breast area) was categorized using the thresholds <10%, 10–24% 25–49%, and >50%^54,55. With STRATUS, an area-based measure of mammographic density is estimated using a machine learning method. Thereafter, the percentage of MD is calculated as the ratio of dense tissue to the total area. This measure was then categorized to approximate BI-RADS categories using the thresholds <2%, 2–8%, 8–49%, and >49%⁵⁶. Characteristics of the prospective cohort studies used for model validation and the distribution of risk factors is provided in the supplementary material (Tables S1, S2, respectively).

Breast cancer outcome was ascertained through linkage to SEER registries, state tumor registries, and pathology databases. A woman was considered a case when she developed incident primary breast cancer, either in situ or invasive, during the follow-up period. The follow-up for all subjects began one year after study entry. This was designed to reduce the impact of prevalent cancers, those already present at the time of screening, on the validation analyses. Follow-up ended at the last record of cancer registry linkage or 5 years, whichever came first.

Risk model development

We used iCARE⁶ to build a model for the 5-year absolute risk of developing breast cancer, integrating questionnaire-based risk factors, PRS and density separately for women younger and older than 50 years. iCARE implements a flexible approach of estimating absolute risk by combining information on relative risks of the risk factors, age-specific incidence and competing mortality rates from multiple data sources, and an individual-level reference dataset of risk factors representing the underlying population. Age was used as a timescale for disease incidence modeling. The conditional relationship between age-specific incidence rates and the risk factors were assumed to follow a Cox proportional hazards model.

The questionnaire-based risk factors included were ages at menarche, first birth and menopause, parity, height, BMI, alcohol intake, family history (i.e., presence/absence of breast cancer in at least one first-degree relative), history of benign breast disease, oral contraceptive use, menopausal hormone therapy (MHT) use, and current MHT type. Our previous works^7,8 describe the integration of questionnaire-based risk factors and the 313-variant PRS⁵, and here we further extend the model incorporating density. The relative risks for the questionnaire-based factors and the 313-variant PRS were previously obtained from a literature review as described by ref. ⁸ (supplementary Table 1). The relative risk estimate of the four-level density variable was obtained through a literature search and was integrated assuming a multiplicative joint association with the other factors⁵⁷. For women younger than 50, the estimate was obtained from ref. ¹⁹, which was adjusted for age. For women over 50 years, the estimate was obtained from ref. ⁵⁸, and was adjusted for age, BMI, menopausal status, surgical menopause, family history, age at childbirth, HRT use, prior breast procedures, race, Hispanic ethnicity, and previous mammographic outcome.

An individual-level reference dataset of risk factors, representative of the underlying country’s population, was used to estimate the joint distribution of risk factors and absolute risk projections in the respective population. For the US population, the majority of the risk factors were derived from the National Health and Nutrition Examination Survey (NHANES) from 2008, 2010, and 2012 as previously described⁸. The PRS was simulated assuming an independent relationship with the questionnaire-based risk factors, conditional on family history. Density was incorporated into the existing reference dataset by estimating the joint distribution of classical risk factors and density using proportional odds regression methods for ordinal data. Analyses were performed in the Breast Cancer Surveillance Consortium (BCSC) dataset, and data were collected from six participating registries: KP Washington, New Hampshire, North Carolina, San Francisco, Vermont and Chicago. Our dataset included 815,968 US women of European-ancestry aged 30–70 with screening mammograms and diagnostic mammograms, collected from 2005 through 2017. For each woman, the first record and corresponding covariates collected at that visit were used, and only records with complete covariate information were included in the analysis. The majority of the risk factors in the BCSC dataset were obtained through a questionnaire filled out by patients (available from www.bcsc-research.org). The variable on benign breast disease was derived from pathology reports and included the diagnosis with the highest severity. Finally, 144,037 women were included with full risk-factor information, from which we simulated the density variable for the reference dataset.

For the Swedish population, the reference dataset was generated using data on subjects with no diagnosis of breast cancer and with no missing information for the risk-factor variables from the prospective cohort KARMA. The sources for building the reference datasets for each country may be found in Table S4.

Lastly, the marginal age-specific disease incidence rates are provided by the United States’ Surveillance Epidemiology and End Results (SEER) cancer registry, 2008–2012. Competing mortality rates due to causes other than breast cancer was obtained from the CDC WONDER database 2008–2012.

Model validation

We evaluated models predicting 5-year absolute breast cancer risk prospectively for calibration and discriminatory accuracy. We considered models with only questionnaire-based risk factors, questionnaire-based risk factors and density, questionnaire-based risk factors and PRS, a model with density and PRS, and lastly, a fully integrated model with questionnaire-based risk factors, PRS and density.

Calibration refers to the model’s ability to accurately predict the absolute risk of breast cancer. We evaluate calibration by estimating the ratio of expected-to-observed (E/O) number of cases, the calibration slope and intercept. To assess calibration, the individual subjects were stratified in low to high-risk categories based on deciles of both predicted 5-year absolute risk and their relative risk score. For each category, the average predicted absolute risk and relative risk scores were computed. The E/O ratio was calculated overall (and within each category), first using the predicted absolute risk. The expected-to-observed ratio is defined as the ratio of the average expected absolute risk and the proportion of observed cases, overall and by category. We assessed model calibration using the Hosmer-Lemeshow (HL) goodness-of-fit test, where the HL statistic approximately follows a χ² distribution. We also estimated a calibration slope and intercept from a linear regression of the observed and expected proportion of cases within risk deciles.

The discriminatory accuracy of the model is assessed using the area under the receiver operating characteristic curve (AUC) based on the 5-year absolute risk and relative risk scores. AUC is defined as the probability that the risk for a randomly selected case is higher than the risk for a randomly selected control. An AUC of 50% corresponds to a model with no discriminatory power, while 100% corresponds to perfect discrimination. The AUC was estimated empirically as the proportion of case-control pairs in which the risk of the case is higher than that of the control. The 95% Wald-based confidence intervals for AUC using the asymptotic variance formula were computed⁵⁹.

Adjustment for nested sampling

The fully integrated model incorporated a 313-SNP PRS and BI-RADS breast density. In all validation cohorts, BI-RADS density and the genotype data on SNPs used to calculate the PRS was available for a nested case-control study. To reflect the underlying full cohort, all calculations are adjusted using sampling weights using the inverse-probability weighted (IPW) approach. Weights are defined by the probability of selection of subjects to the nested case-control sample from the full cohort. The probability of selection is estimated by fitting a logistic regression model with age of study entry, follow-up time, the matching factors, and interactions of case/control status with age of study entry and follow-up time to the entire underlying cohort, where the outcome is defined as whether the individual is included or not. All standard errors and 95% Wald-based confidence intervals were calculated using influence function-based variance estimators that account for sampling weights.

Risk reclassification

Improvements in risk stratification resulting from the incorporation of density to a model with questionnaire-based risk factors and PRS were assessed among women of European-ancestry aged 50–70 years in the populations of the US and Sweden by calculating the number of women and future cases identified to be at high-risk based on pre-specified 5-year absolute risk thresholds. We used two high-risk thresholds: 3%, which corresponds to the United States Preventive Services Task Force (USPSTF) recommendation for risk-reducing interventions⁶⁰, as well as 6%, which corresponds to the breast cancer risk of a BRCA mutation carrier and is used as a cutoff for very high risk by the WISDOM Trial^24,25. We also calculated the number of women and future cases reclassified at the high-risk thresholds to further quantify the improvements in risk stratification after incorporating density into a model with questionnaire-based risk factors and PRS.

All analyses were performed using R version 3.6.2 (www.r-project.org).

Ethics declaration

All contributing studies received approval from their respective local institutional review boards. Study participants provided informed consent, where applicable, and participated under protocols approved by ethics committees. The corresponding ethics approvals are as follows: MMHS—approved at Mayo Clinic under 08-000847; KARMA— approved by the Swedish Ethical Review Authority under 2010/958-31/1; NHS—approved at Mass General Brigham under 1999P002112.

Author contributions

M.G.-C., P.P.C. and C.V.M. conceived of and designed the study. M.G.-C. and P.P.C. contributed equally in supervising the work. G.L.G. also contributed to the supervision. C.V.M., X.Y., Y.H.J., C.G.S., C.G. and M.E. performed data analysis. A.N.H. assisted with analyses. C.V.M., P.P.C. and M.G.-C. produced the initial draft and final version of the manuscript. All authors reviewed the manuscript and approved the final version.

Funding

Open access funding provided by the National Institutes of Health.

Data availability

The datasets used in the current analysis will not be made publicly available due to restraints imposed by the ethics committees of individual studies; requests for individual-level data for all the participants in the full cohort of any study can be made to the individual studies.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41523-025-00813-z.