Predicting treatment-related cardiovascular risks in breast cancer patients: development and validation of an interpretable machine learning model.

Purpose: This study aimed to develop and validate an interpretable machine learning model to predict the 1- to 3-year risk of cardiovascular events in breast cancer patients by integrating baseline and treatment variables, while preliminarily investigating the potential association between short-term cardiac function decline and long-term adverse cardiovascular events. Methods: We analyzed electronic medical records from 31,878 breast cancer patients. A composite cardiovascular event outcome was used. Predictors were selected via a two-step process: removing highly correlated variables (|r|≥0.7) and applying LASSO regression with 10-fold cross-validation, which refined 62 initial variables down to 18. Five models were built and compared using the area under the receiver operating characteristic curve (AUC-ROC). The optimal model was interpreted using SHapley Additive exPlanations (SHAP). Results: Among 31,878 breast cancer patients, 3,960 (12.4%) experienced cardiovascular events. The XGBoost model demonstrated the best overall discriminative performance (AUC = 0.790). SHAP analysis identified endocrine therapy, anemia management therapy, and history of cerebrovascular disease as the top three predictors. Crucially, short-term decline in cardiac function was also selected as a significant predictor, supporting its role as a precursor to long-term events. Model robustness was confirmed via sensitivity analysis.