Improving breast cancer risk assessment with neighborhood indicators using a case-control study: insights from All of Us

Background Accurate breast cancer risk assessment is needed to reduce disparities. We assessed whether neighborhood indicators could replace race and ethnicity without reducing predictive performance. Methods We conducted a matched case–control study utilizing All of Us Research Program data (May 2017–June 2022) among participants assigned female at birth aged ≥18. We created six subgroups: four retrospective, one prospective, and one survey-only. We matched cases with up to three controls, by age (±2 years) and follow-up. We developed three conditional logistic regression models: (1) Model 0 (clinical factors, race, ethnicity), (2) Model 1 (Model 0, neighborhood indicators), and (3) Model 2 (clinical factors, neighborhood indicators). Model performance was assessed using discrimination (AUC), risk stratification (top 20%), and decile-based calibration (logarithmic observed/expected ratios). Findings Among 413,457 participants, 46,427 were matched (34,810 controls; 11,617 cases): 29,884 (64.4%) Non-Hispanic (NH) White, 7280 (15.7%) NH Black, and 6308 (13.6%) Hispanic, other groups not reported due to small sample size. Across analytic subgroups, models demonstrated similar but minimal discrimination (pooled AUCs 0.54–0.62), risk stratification (52–234 cases [26.7–37.8%] in the top 20% of risk), and calibration. Replacing race and ethnicity with neighborhood indicators changed AUC by 0.06 and cases in the top 20% of predicted risk by less than 4 percentage points. Interpretation Neighborhood indicators may replace race and ethnicity in risk models while preserving performance, despite limited discrimination. Their use may better identify social and structural factors relevant to resource and place-based interventions. Future research should consider race-conscious structural measures to refine prediction. Funding NIH, NCI, Breast Cancer Research Foundation.