利用行政健康数据进行乳腺癌的个性化风险评估

Breast cancer screening programs are one-size-fits-most approaches with suboptimal participation rates. Population-level administrative health databases provide a unique opportunity to build scalable, data-driven risk assessment tools capable of identifying women who may benefit from more personalized screening strategies. We assembled nearly two decades of longitudinal health data, including mammographic screening history, medication use, physician visits, and hospital discharge abstracts, for 1.74 million women in British Columbia, among whom 39,211 incident breast cancers were diagnosed. Our team is developing new breast cancer risk assessment models to predict each woman's individual time until Breast Cancer Onset (BCo) using administrative health data from Canada's publicly funded healthcare system. We are applying machine learning Individual Survival Distribution (ISD) models, which identify each subject x with a distribution S (t | x), showing the probability that x's time until BCo is at least t more years, for all t > 0. We can then use these models to estimate each woman's expected time until BCo, as well as her risk score. In preliminary models using 25 features with known/suspected links to breast cancer, random survival forest (RSF) achieved the highest concordance index (CI = 58.9%), while multitask logistic regression (MTLR) achieved a competitive 5-year Brier score (BS = 0.0068) and a low mean absolute error (MAE = 30.4 months). These early results demonstrate the feasibility of leveraging administrative health data for personalized breast cancer risk prediction. Ongoing work will substantially expand the feature sets to improve model discrimination.