Performance Comparison of Tree-Based Models for Heart Disease Prediction Using Feature Selection and SMOTE

Abstract

Heart disease remains the leading cause of mortality worldwide, highlighting the need for accurate early prediction models. This study proposes a machine learning framework for heart disease prediction using the BRFSS 2015 Heart Disease Health Indicators dataset, which contains 253,680 records and 22 attributes. The proposed approach integrates Synthetic Minority Oversampling Technique (SMOTE) for class imbalance handling, mutual information-based SelectKBest feature selection (k = 15), and three tree-based classifiers: Decision Tree, Random Forest, and XGBoost. A leakage-free preprocessing pipeline was implemented to ensure that SMOTE was applied only to the training data, and classification threshold optimization was performed to improve minority class detection. Model performance was evaluated using Accuracy, Precision, Recall, F1-score, and ROC-AUC metrics. Experimental results show that XGBoost achieved the best performance with a cross-validation ROC-AUC of 0.9815 and a test ROC-AUC of 0.8444 at an optimized threshold of 0.20. The findings demonstrate that the proposed integration of oversampling, feature selection, and threshold optimization can improve predictive performance for imbalanced cardiovascular risk data, providing a practical foundation for machine learning–based decision support in early heart disease risk screening.