Deep learning mammography-based breast cancer risk model, its serial change, and breast cancer mortality

Although numerous breast cancer risk prediction models have been developed to categorize individuals by risk, a substantial gap persists in evaluating how well these models predict actual mortality outcomes. This study aimed to investigate the association between Mirai, a deep learning model for risk prediction based on mammography, and breast cancer-specific mortality in a large cohort of Korean women.

This retrospective cohort study examined 124,653 cancer-free women aged ≥ 34 years who underwent mammography screening between 2009-2020. Participants were stratified into tertiles by Mirai risk scores and categorized into four groups based on risk changes over time. Cox proportional hazards regression models were used to evaluate the associations of both baseline Mirai scores and temporal risk changes with breast cancer-specific mortality.

Over 1,075,177 person-years of follow-up, 31 breast cancer-related deaths occurred. The highest Mirai risk tertile showed significantly higher breast cancer-specific mortality than the lowest tertile (hazard ratio [HR], 5.34; 95% confidence interval [CI] 1.17-24.39; p for trend = 0.020). Temporal Mirai score changes were associated with mortality risk: those remaining in the high-risk (HR, 5.92; 95% CI 1.43-24.49) or moving from low to high risk (HR, 5.57; 95% CI 1.31-23.63) had higher mortality rates than those staying in low-risk.

The Mirai model, developed to predict breast cancer incidence, was significantly associated with breast cancer-specific mortality. Changes in Mirai risk scores over time were also linked to breast cancer-specific mortality, supporting AI-based risk models in guiding risk-stratified screening and prevention of breast cancer-related deaths.