A Robust Hybrid Ensemble Approach for Cardiovascular Disease Prediction with Feature Optimization

Abstract

Cardiovascular diseases (CVDs) continue to be the world's leading cause of death, accounting
for 17.9 million deaths annually. For effective prevention and intervention, accurate early
prediction systems are crucial. The inability of conventional statistical risk models to capture
the complex, nonlinear interactions among clinical risk factors frequently limits their
predictive accuracy. To address these problems, this study evaluates optimized machine
learning methods for cardiovascular disease prediction using the Framingham Heart Study
dataset, a reliable longitudinal clinical resource. A comprehensive preprocessing pipeline that
included clinically guided feature selection, the removal of incomplete records, feature
standardization, and class imbalance correction using Borderline-SMOTE2 was implemented
to enhance minority class detection. Four extreme-tuned classifiers—Random Forest,
AdaBoost, Support Vector Machine (SVM), and Decision Tree—were systematically trained
and compared using a number of evaluation metrics, including accuracy, ROC-AUC,
sensitivity, specificity, F1-score, and Cohen's Kappa. Experimental results demonstrate that
the optimized AdaBoost and Random Forest models achieved the best predictive performance,
with accuracies of 94.12%. AdaBoost had the best overall discrimination, with an ROC-AUC
of 0.9775, F1-score of 0.9406, and Cohen's Kappa of 0.8824. With a ROC-AUC of 0.9764,
F1-score of 0.9402, and Kappa of 0.8823, Random Forest ranked second. SVM provided
strong sensitivity (0.9419) but lower specificity, while the Decision Tree performed
comparatively poorly. These findings confirm that ensemble-based methods provide
improved stability, balanced classification, and better generalization for cardiovascular risk
prediction. The proposed framework offers a clinically reliable and understandable decision-
support solution for early detection of cardiovascular disease.