Abstract
Cardiovascular disease (CVD) remains one of the leading causes of death worldwide, thereby creating a need for risk predictions which are accurate and interpretable. Herein, this work proposes a robust hybrid machine learning framework to predict cardiovascular risk. The framework integrates minimum Redundancy Maximum Relevance (mRMR) feature selection, Principal Component Analysis (PCA), and an explainable multi-model learning strategy. We evaluate the architecture on the Framingham Heart Study dataset comprising 4,240 patient records with mixed clinical and demographic information. A total of seven supervised learning classifiers-Extra Trees, Random Forest, Gradient Boosting, CatBoost, XGBoost, Decision Tree, and Multilayer Perceptron-have been evaluated for performance using stratified 10-fold cross-validation and test validation. To address feature redundancy and class imbalance, mRMR, PCA, and SMOTE were employed. The best-performing model, which combines PCA with Extra Trees, achieves 98.56% accuracy, 97.81 % precision, 99.35 % recall, 98.57 % F1-score, and a ROC-AUC score of 0.9988. The use of explainable AI tools has also been on the rise as a means of improving transparency. Both global and local explanations, which employ SHAP and LIME explanations, identify important risk factors for clinical conditions such as age, blood pressure, and blood glucose. Overall, the model has a good balance of predictive performance, robustness, and interpretability, indicating that the model is appropriate for application in clinical practice.
