Cardiovascular Disease Classification Using Machine Learning: A Comparative Study with SMOTE Class-Imbalance Correction
TL;DR
The pipeline approach ensured no data leakage in cross-validation, and the findings support ensemble ML models with SMOTE as a preprocessing step for imbalanced CVD datasets.
Abstract
Cardiovascular disease (CVD) is the major cause of global mortality, accounting for approximately 17.9 million annual deaths and 32% of all global fatalities. The dataset revealed a real class imbalance of 58% CVD-positive cases (580) and 42% CVD-negative cases (420), with an imbalance ratio of 1.381. Without correction, classifiers trained on this data are biased toward the majority class, inflating accuracy while suppressing sensitivity on the minority class. To correct class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE) on the training set, compare six supervised machine learning classifiers for binary CVD detection, and evaluate their performance on the original test set. Six classifiers Decision Tree, KNN, SVM, Gradient Boosting, Random Forest, and Logistic Regression were evaluated. A stratified 80:20 train-test split was applied, followed by SMOTE only on the training set (336 → 464 samples per class). To prevent data leakage, SMOTE was embedded within an imbalanced-learn Pipeline. Performance on the original test distribution was evaluated using Accuracy, Precision, Recall, F1 Score, MCC, and AUC-ROC. After SMOTE balancing, Logistic Regression and Random Forest achieved the highest F1 score (0.9871), with AUC-ROC values of 0.9982 and 0.9990, respectively. Gradient Boosting achieved the highest AUC-ROC (0.9993) and was the most stable model in cross-validation (99.20% ± 0.98%). Feature importance analysis identified maxheartrate, oldpeak, and noofmajorvessels as the top three discriminative predictors. The pipeline approach ensured no data leakage in cross-validation, and the findings support ensemble ML models with SMOTE as a preprocessing step for imbalanced CVD datasets.