Coronary artery disease prediction using Bayesian-optimized support vector machine with feature selection

Area of Science:

• Cardiovascular research
• Medical informatics
• Machine learning in healthcare

Background:

• Coronary Artery Disease (CAD) is a major global health concern.
• Current diagnostic methods like invasive angiography are costly and carry risks.
• There is a need for non-invasive, accurate, and interpretable CAD prediction tools.

Purpose of the Study:

• To develop and validate a non-invasive, interpretable framework for Coronary Artery Disease (CAD) prediction.
• To leverage machine learning for enhanced CAD risk stratification.
• To identify clinically relevant features for CAD prediction.

Main Methods:

• Utilized the Z-Alizadeh Sani dataset for model development.
• Employed a hybrid decision tree-AdaBoost for feature selection (30 features).
• Applied SMOTE oversampling within cross-validation folds to prevent data leakage.
• Optimized Support Vector Machine (SVM) using Bayesian hyperparameter tuning.
• Interpreted model predictions using SHapley Additive exPlanations (SHAP).

Main Results:

• The SVM_Bayesian model achieved 97.67% accuracy, 100% sensitivity, and 99% AUC.
• Outperformed logistic regression, random forest, standard SVM, and SLOA-optimized SVM.
• SHAP analysis identified key predictors like Typical Chest Pain and Age.
• Statistical tests and temporal generalization confirmed model robustness.

Conclusions:

• The proposed framework offers a transparent, generalizable, and clinically actionable tool for CAD risk stratification.
• Demonstrates the potential of interpretable machine learning in cardiovascular disease prediction.
• Highlights the importance of interconnected cardiovascular features in systemic disease prediction.