Predicting coronary artery disease with ensemble-based feature selection and grid search-tuned SVM under class imbalance conditions

Abstract Coronary artery disease (CAD) is a common heart condition that leads to numerous deaths. The standard diagnostic tool for CAD, coronary angiography, is expensive, time-consuming, and carries harmful side effects. To overcome these limitations, computer-aided decision-making through machine learning algorithms offers an effective alternative for early CAD detection. This study uses the Z-Alizadeh Sani dataset, comprising 303 samples with 54 features. Feature selection was performed using a combination of XGBoost and Random Forest algorithms to identify the most significant predictors. To mitigate class imbalance, three oversampling techniques were applied, including Conditional Tabular Generative Adversarial Network (CT-GAN), Adaptive Synthetic Sampling (ADASYN), and Synthetic Minority Over-sampling Technique (SMOTE). ADASYN outperformed CT-GAN and SMOTE in improving classification outcomes. Five classifiers including Multi-Layer Perceptron (MLP), Random Forest (RF), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) were tested and optimized using both grid search and random search. The highest balanced accuracy was achieved by the ensemble feature selection combined with ADASYN oversampling and support vector machine (SVM) with grid search resulting in 96.86% and an AUC of 1.00. This approach highlights the potential for integration into clinical decision support systems to assist in early CAD diagnosis.