A hybrid artificial intelligence approach for cardiovascular disease prediction using ensemble learning, neural networks, and interpretable models

This study presents a stacked ensemble learning framework for accurate prediction of cardiovascular disease (CVD) risk using the Framingham Heart Disease dataset. Five highly differentiable base classifiers—XGBoost, LightGBM, CatBoost, Gradient Boosting, and AdaBoost—have been integrated through a Multilayer Perceptron (MLP) meta-learner. Preprocessing entailed Hampel filtering for outlier elimination, mean imputation for handling missing data, Min-Max normalization, PCA dimension reduction on the basis of nine components, and SMOTE for class balance restoration. Stacked ensemble model produced 96.06% accuracy, 96.12% F1-score, and 99.31% ROC-AUC, significantly superior to stand-alone classifiers. In a bid to ensure interpretability, feature importance was explored and revealed that components relating to blood pressure, smoking, cholesterol, and age played most critical roles in risk estimation. Correspondingly, these features possessed complex, non-linear effects demonstrating threshold-like behavior reflecting model’s decision-making. Correlation analyses corroborated good model alignment, where CatBoost and XGBoost revealed highest agreement of feature importance with the ensemble. This work illustrates the merit of uniting comprehensive learners with explainable AI for reliable, interpretable, and highly scalable CVD risk classification, making the architecture deployable in the clinic for early detection and personalized preventive strategies.