Advancing breast cancer risk prediction in national cohorts: the role of mammogram-based deep learning

To date, significant efforts devoted to developing breast cancer risk prediction models that produce modest discriminatory accuracy in the range of 59-68% and underperform in Black women compared to non-Hispanic white (NHW) women, limiting their clinical impact on equitable precision-based/personalized breast cancer prevention and screening. Emerging data suggest that deep learning (DL) models based on mammographic images outperform traditional breast cancer risk prediction models based on clinical and risk factors and breast density alone. Yet there remain major gaps to resolve before widespread roll out of DL methods in order to enhance outcomes and reduce health disparities in breast cancer. These models have been developed in clinical or screening cohorts and require further validation in “real world” settings. Equally unknown is how performance of MDL risk models may differ between Black and NHW women, as few previous studies included sufficient numbers of Black women for meaningful interpretation. Finally, the combination of polygenic risk scores (PRS) and MDL risk scores could significantly improve clinical utility of these risk stratification tools, yet MDL and PRS clinical rollout are largely being evaluated separately, primarily due to lack of availability of both mammographic and genetic data in the same study population. We propose to leverage existing unparalleled resources from two complementary national cohorts, the Sister Study and the Black Women’s Health Study, with digital screening mammograms, genomic data, and extensive epidemiologic and clinical data to address these evidence gaps (ncases, 971; ncontrols, 8793). In Aim 1, we will validate the performance and clinical risk stratification of Mirai, a published MDL model with the highest accuracy in screening cohorts to date, in our epidemiologic cohorts. In addition to considering breast cancer risk overall, we will evaluate performance of this MDL model for early and advanced stage and for ER+ and ER– breast cancer, outcomes with significant clinical implications for prognosis and treatment. In Aim 2, we will evaluate the addition of PRS to MDL risk scores in improving risk stratification, and determine net reclassification. In Aim 3, we will determine whether MDL models perform equally well and impacts risk stratification within strata of breast density, as defined in widespread implementation of breast density notification legislation (dense vs. non-dense). These results may provide crucial data for guiding follow up and surveillance of women with and without dense breasts. To ensure equitable application of results, we will examine differences in results between Black and NHW women in all aims. MDL models hold great promise as breast cancer risk assessment tools, and offer clinical efficiency and optimization by eliminating the need to collect detailed family history and risk factor data. Validation in “real world” epidemiologic cohorts will advance the field substantially. The research proposed here has great potential to identify women at both high and low risk of breast cancer, inform personalized screening and risk reduction strategies, and even to reduce racial disparities in breast cancer outcomes.