Smith, Hugh

Image-derived artificial intelligence (AI) risk models have shown promise in short-term risk assessment for improving breast cancer (BC) screening. No image-derived long-term AI risk model for primary prevention has been developed and externally validated. Individuals aged 31 to 94 years were recruited between 2010 and 2020 in two population-based case cohorts (Olmsted County, Minnesota, US, and KARMA, Sweden) and one hospital-based case-control study (EMBED, Atlanta, US). Median follow-up in the case cohorts was 10 years, with BCs diagnosed through June 2022. Additional validation was performed in EMBED with 3-year follow-up. An AI risk model was developed in Sweden, and we report independent validation in Olmsted/KARMA/EMBED. Absolute 10-year risks were estimated at study entry. Time-dependent discriminatory performance [AUC( t )] and expected-to-observed events (E/O) were estimated. Comparisons were performed with clinical risk tools and the image-based Mirai tool. Across Olmsted/KARMA, 8696 individuals (mean age, 54.4±10.6) and 1633 individuals with incident BC (mean age, 57.0±10.6) were included. Average 10-year risks were 3.83 and 3.14%, and E/O ratios were 0.99 and 0.99. Ten-year AUC( t ) for invasive BCs were 0.72 in Olmsted and 0.72 in KARMA. Similar discriminatory results were observed in EMBED. Our AI risk tool performed significantly better than Mirai in Olmsted/KARMA/EMBED. In KARMA, in the top 10% of high-risk individuals, the AI risk tool predicted 33% of BCs compared with 23%/20%/24% predicted by Tyrer-Cuzick-v8/BCSC-v3/Mirai (all P < 0.01). The 10-year image-derived AI risk model showed significantly higher performance than clinical risk models and Mirai in diverse populations, supporting its clinical potential for primary prevention.