Enhanced cardiovascular disease risk prediction using explainable machine learning and data balancing

Cardiovascular disease (CVD) remains a leading cause of mortality worldwide, making accurate and early risk prediction essential for effective preventive care. However, traditional risk assessment models often fail to capture complex nonlinear clinical patterns, suffer from class imbalance, and provide limited interpretability, restricting their clinical adoption. To address these challenges, this study proposes an explainable and balanced ensemble machine learning (ML) framework for robust CVD risk prediction. Four ensemble classifiers–Gradient Boosting (GB), Extreme Gradient Boosting (XGB), CatBoost (CB), and Extra Trees (ET)–were evaluated on two publicly available benchmark datasets: Heart Disease Classification (HDC) and Cardiovascular Disease (CD). Class imbalance was mitigated using the Synthetic Minority Over-sampling Technique (SMOTE), and model performance was optimized through systematic hyperparameter tuning. A comprehensive evaluation was conducted using 5-fold cross-validation with multiple metrics, including accuracy, recall, F1-score, and ROC-AUC. To enhance transparency, explainable artificial intelligence (XAI) techniques–SHAP, LIME, and permutation-based feature importance–were employed to analyze both global and local feature contributions. Experimental results demonstrate that CatBoost consistently outperformed competing models, achieving an accuracy of 98.88% (HDC) and 99.44% (CD), with corresponding ROC-AUC scores of 99.90% and 99.97%. SMOTE significantly improved minority-class sensitivity, while XAI analysis identified clinically relevant predictors such as troponin, kcm, glucose levels, slope, and exercise-induced angina, aligning with established cardiovascular risk factors. Overall, the proposed SMOTE–CatBoost framework delivers accurate, fair, and interpretable CVD risk prediction, offering strong potential for clinical decision support and precision medicine. Future work will focus on validating the framework using multimodal and real-world longitudinal patient data to further enhance generalizability.