Comparative and Explainable Machine Learning for Coronary Artery Disease Classification
Accurate and interpretable classification of angiographically defined coronary artery disease (CAD) remains a relevant benchmark for clinical machine-learning research. This study compares eight supervised-learning models using the Cleveland Heart Disease dataset. After complete-case exclusion, 297 records with 13 clinical and diagnostic predictors were analyzed. Nominal predictors were one-hot encoded, binary predictors were retained in their original coding, and continuous predictors were standardized for scale-sensitive models. Preprocessing parameters were estimated using training data only and re-estimated within each cross-validation fold. Logistic Regression, a Decision Tree, k-Nearest Neighbors, a Multilayer Perceptron, Random Forest, Gradient Boosting, a Support Vector Machine, and an equal-weight Soft-Voting Ensemble were assessed using a stratified 80/20 train–test split and stratified 10-fold cross-validation. Test performance was evaluated using accuracy, precision, sensitivity, specificity, F1-score, confusion matrices, and the area under the receiver operating characteristic curve (AUC). Under stratified 10-fold cross-validation, Logistic Regression achieved the highest mean accuracy (83.83%), followed by the Multilayer Perceptron (83.16%), while Logistic Regression also achieved the highest mean ROC AUC (0.9093). On the complementary 60-record held-out test partition, the Support Vector Machine and Multilayer Perceptron achieved the highest observed accuracy (86.67%), whereas the Support Vector Machine achieved the highest ROC AUC (0.9565). Tree-based feature-importance analyses and SHapley Additive exPlanations (SHAP) analysis of the Gradient-Boosting model showed relatively large model-attribution values for the number of major vessels visualized by fluoroscopy, the asymptomatic chest-pain category, age, ST-segment depression, cholesterol, and specific thallium stress-test categories. More complex models showed no consistent performance advantage. Excluding the fluoroscopy-derived predictor ca reduced mean cross-validation accuracy and ROC AUC for all eight classifiers. In the more restrictive seven-predictor analysis, which additionally excluded thallium and exercise-test variables, the highest mean cross-validation accuracy was 78.15% (Logistic Regression), while the highest ROC AUC was 0.8315 (Random Forest). Given the small, historical, single-center sample, the inclusion of target-proximal diagnostic predictors, the absence of calibration assessment, and the lack of external validation, the findings should be interpreted as an internal comparative benchmark rather than evidence of clinical applicability or readiness for clinical deployment.
Authors
- Theofanis I. Aravanis (ORCID: https://orcid.org/0000-0003-0329-3200)
- Athanasios Kanavos (ORCID: https://orcid.org/0000-0001-9330-1627)
- Konstantinos Ioannou (ORCID: https://orcid.org/0009-0008-4195-2338)
Institutions
- University of Peloponnese (GR)
Publication Details
- Journal
- AI in Medicine
- Published
- 2026-11-13
- DOI
- https://doi.org/10.3390/aimed1040031
- Primary Topic
- Artificial Intelligence in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00