Ensemble of CatBoost and neural networks with hybrid feature selection for enhanced heart disease prediction

Heart disease remains one of the leading causes of mortality globally, necessitating accurate and efficient prediction models. This paper presents a novel ensemble model combining CatBoost and neural networks (ECNN) to improve heart disease prediction accuracy. The proposed methodology incorporates two key innovations: the sequential SHAP and RFE hybrid optimization (SSHO) technique for feature selection and the dynamic gradient-sharing mechanism (DGSM) to facilitate efficient interaction between CatBoost and neural networks. SSHO dynamically selects relevant features based on SHAP values, while DGSM shares gradient information to optimize learning. The ECNN model was trained using the personal key indicators of heart disease dataset, addressing class imbalance with SMOTE. The experimental results demonstrate the model’s superior performance with an accuracy of 91%, precision of 94%, and F1-score of 92%. These findings surpass previous studies’ results and highlight the ECNN model’s novelty in improving prediction accuracy and interpretability. The integration of SSHO and DGSM offers a scalable approach to heart disease prediction, making it a valuable contribution to clinical decision support systems.