Abstract PD14-02: PD14-02 Ten-year follow-up of an image-based AI risk model for breast cancer

Abstract

Abstract Background: Artificial intelligence based, image-derived short-term risk models for breast cancer have shown high discriminatory performance compared to traditional lifestyle familial-based risk models. However, the long-term performance has not yet been investigated. Methds: In this study, we investigated the long-term performance for predicting breast cancer throughout 10 years using an image-based risk model and compared the results to a traditional lifestyle familial-based risk model. We performed a nested case-control study based on a mammography screening cohort conducted since 2010 in Sweden for women aged 40-74. Mammograms, age, lifestyle and familial risk factors were collected at study entry. In the breast cancer register update in 2022; 2,028 incident breast cancers were included together with 8,398 controls that were matched to the cases on year of prior baseline mammogram. The image-based model extracted mammographic features (density, microcalcifications, masses, left-right breast asymmetries of these features) and age from the baseline mammograms. Tyrer-Cuzick risk model used self-reported lifestyle and familial risk factors to estimate risk at study-entry. Absolute risks were estimated using the risk models. We estimated model performances using Area Under the receiver operating characteristic Curves (AUC) statistics of the absolute risks and, risk ratios of women classified as high-risk and low risk using NICE and USPSTF guidelines. Results: The AUCs of the image-derived risk model ranged from 0.76 (95%CI 0.72-0.81) to 0.66 (95%CI 0.65-0.67) for breast cancers developed 1-10 years after study-entry. The corresponding Tyrer-Cuzick AUCs were 0.68 (95%CI 0.63-0.73) to 0.62 (95%CI 0.60-0.63). For estrogen negative and symptomatic cancers, the AUCs for the image-derived model were ≥0.75 during the first 2 years. Women with high and low mammographic density showed similar AUCs. Throughout the 10-years of follow-up, 20% of all women with cancers were deemed high risk at study-entry by the image-derived risk model compared to 6% of all women with cancers identified as high risk by the lifestyle familial-based model (p< 0.01). Conclusion: The image-derived model outperformed the lifestyle familial-based model both for short-term and long-term risk assessment and, could be used for identifying women who possibly could benefit from additional examinations and primary prevention. Citation Format: Mikael Eriksson, Kamila Czene, Emily F. Conant, Per Hall. PD14-02 Ten-year follow-up of an image-based AI risk model for breast cancer [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr PD14-02.