Abstract
Heart disease continues to rank as the foremost contributor to premature death across the globe, and developing reliable automated screening tools has become a pressing priority in clinical informatics. This study presents a comparative analysis of three machine learning classifiers — Logistic Regression (LR), Random Forest (RF), and Gradient Boosting (GB) — for binary heart disease prediction using the UCI Cleveland Heart Disease dataset. The dataset comprises 297 patient records with 13 clinical features following removal of missing values. Models were evaluated using an 80/20 stratified train-test split with Min-Max normalisation. Logistic Regression achieved the highest AUC-ROC of 0.9554 and accuracy of 83.33%, followed by Random Forest (AUC-ROC = 0.9375, accuracy = 85.00%) and Gradient Boosting (AUC-ROC = 0.8828, accuracy = 76.67%). Feature importance analysis identified chest pain type (cp), thalassemia (thal), and maximum heart rate (thalach) as the most influential predictors. Results indicate that well-calibrated linear classifiers can match or exceed ensemble methods in structured clinical settings, a practically significant finding for risk-stratification tools where probability estimates must be trusted.