abstract
- This retrospective single center study evaluated the performance of Mirai, a mammography-based deep learning model for predicting 1¿5-year breast cancer risk in Mexican women using longitudinal screening data collected between January 2014, and December 2019. All available mammographic examinations acquired during the study period were included to generate individualized risk scores, rather than limiting the analysis to a single baseline mammogram. Pathology reports and longitudinal imaging follow-up served as the reference standard. Model performance was assessed across the full cohort using the area under the receiver operating characteristic curve (AUC) and the concordance index (C-index). In a total of 41,883 mammographic examinations, Mirai achieved a mean AUC of 0.617 (95% Confidence Interval (CI): 0.571¿0.633) and a C-index of 0.610 (95% CI: 0.551¿0.659) for predicting future breast cancer risk over the 1¿5-years. These findings suggests that Mirai shows moderate discriminatory performance when applied to longitudinal mammography data in a real-world Mexican screening population, highlighting the challenges of transferring mammography-based risk models across populations and imaging settings.