Optimizing heart disease diagnostics: An evaluation of feature selection frameworks and ensemble learning models

Background Cardiovascular disease (CVD) is still one of the leading causes of death worldwide, which highlights the need for accurate and automated decision-support systems. Objective This study presents a non-invasive computational approach to diagnosing CVD. Methods This study combines careful data preprocessing, a step-by-step feature selection process, and advanced ensemble learning models. We used an open-source clinical dataset with 1190 patient records. To select the most important features, we applied three feature selection methods: Mutual Information, Sequential Forward Selection, and LASSO. LASSO performed best, achieving 94.03% accuracy. We tested 14 machine learning (ML) and ensemble classifiers and evaluated them using confusion matrix metrics, F1-scores, and the stability of the Area Under the ROC Curve (AUC) across different classification thresholds. Results The Extra Trees classifier achieved the highest diagnostic accuracy (98.80%). We also used independent t-tests and Chi-square analyses to compare three groups of attributes—behavioral habits, physical symptoms, and clinical or physiological data—between male and female patients. The results showed that men and women were broadly similar in these key risk factors. Conclusion Our high-precision approach provides a cost-effective, reliable, and easy-to-understand decision-support tool that can improve early screening and risk assessment before hospitalization.