Abstract 7407: Generalizability of an AI-based mammogram risk score (MRS) for breast cancer among racially and ethnically diverse women

Abstract Breast cancer remains the most common cancer diagnosed among women worldwide. While risk prediction models hold great promise for enhancing preventive breast cancer management, a major challenge lies in accurately identifying women at elevated risk across diverse racial and ethnic groups. Mammogram risk scores (MRS), an AI-driven feature summarizing mammographic texture, have shown potential to accurately classify 5-year breast cancer risk in racially diverse populations.1 However, similar prediction performance across groups does not necessarily ensure the model's generalizability. To address this, we analyzed data from two large North American cohorts—Emory and British Columbia—encompassing over 226, 000 racially diverse women, including Non-Hispanic White (NHW), Non-Hispanic Black (NHB), East Asian, South Asian, and Indigenous women. We assessed the distribution of MRS by race and ethnicity and observed little variation in median MRS between the two cohorts overall, among race and ethnicity groups within each cohort, or between cohorts within race and ethnicity (where such comparisons were feasible). We further examined the association between MRS and breast cancer risk both overall and stratified by race/ethnicity. Statistically significant associations were found, with hazard ratios (HR) per standard deviation increase in MRS ranging from 2.03 (95% CI: 2.00-2.06) to 2.15 (95% CI: 1.93-2.40) after adjusting for age. This significant association was consistent across racial/ethnic groups: NHB (2.12, 95% CI: 1.92-2.51), East Asian (2.10, 95% CI: 2.02-2.18), South Asian (2.09, 95% CI: 1.72-2.49), and Indigenous women (2.19, 95% CI: 1.65-2.79). Prediction accuracy, as assessed using the 5-year AUC, also showed no statistically significant differences by race/ethnicity within each cohort. Our findings indicate no clinically meaningful differences in MRS performance across these groups, suggesting broad applicability. The MRS could be effectively used in clinical practice to estimate age-specific breast cancer incidence, including for individuals with mixed ancestry or self-reported identities commonly encountered in real-world settings. 1. Jiang S, Debbie B, Rosner B, Tamimi R, Colditz G. Development and validation of dynamic 5-year breast cancer risk model using repeated mammograms. JCO Clinical Cancer Informatics, In press. 2024. Citation Format: Shu JIang, Debbie L. Bennett, Peter Kraft, Graham A. Colditz. Generalizability of an AI-based mammogram risk score (MRS) for breast cancer among racially and ethnically diverse women [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7407.