BREST-AI: Three-year breast cancer risk prediction from screening mammograms

Abstract Accurate assessment of breast cancer risk is essential for risk-stratified screening strategies. We developed BREST-AI, a deep-learning model that predicts three-year breast cancer risk from negative screening mammograms. Model development used a curriculum-learning strategy and data from 65,419 women; the model was trained on 58,591 women (9020 cancer, 1633 benign, and 47,938 three-year cancer-free). In the external UK case–control validation (1899 cases and 5697 matched controls), BREST-AI achieved an AUC of 0.751 (95% CI 0.738–0.764), outperforming Mirai (0.700, 95% CI 0.687–0.714; p < 0.001). In a US external cohort (97 cases and 291 matched controls), AUCs were 0.670 (95% CI 0.609–0.731) for BREST-AI and 0.633 (95% CI 0.570–0.696) for Mirai ( p = 0.199); this analysis is underpowered to detect modest differences and should be interpreted as exploratory. In a UK cross-vendor external cohort acquired on Siemens systems (400 cases and 1200 matched controls), cross-vendor domain shift modestly reduced discrimination, but BREST-AI still exceeded Mirai (AUC 0.719 versus 0.685; p = 0.019). When projecting case–control results to the English screening prevalence (8 cancers per 1000 examinations), BREST-AI yielded an estimated positive predictive value of 8.0% and a relative risk of 10.1 at a high-risk operating point defined by 99% specificity. These prevalence-adjusted estimates suggest that BREST-AI could help to identify a small subgroup of women at high three-year breast cancer risk within the NHS Breast Screening Programme.