Abstract
<jats:p>cardiovascular disease (CVD) is a leading cause of global mortality and may progress without obvious symptoms. This study proposes CardioSafeAI, an explainable federated learning framework for predicting ten-year coronary heart disease risk from a public dataset. The workflow includes imputation, encoding, Min--Max normalization and SMOTE. A hybrid feature selection method combines Pearson correlation, Information Gain and Gain Ratio. Eleven individual models and a Stacking ensemble were evaluated. The Stacking classifier achieved the highest centralized accuracy of 91.1%. In a five-client simulation, the federated Stacking model achieved 90.80% accuracy and an ROC-AUC of 0.9486. SHAP and LIME provided global and patient-level explanations. CardioSafeAI therefore provides a promising basis for interpretable cardiovascular risk assessment, although independent and prospective clinical validation is required before use in clinical decision making.</jats:p>