A Recall-Oriented Comparative Evaluation of Machine Learning Models for Heart Disease Detection on the Cleveland Heart Disease Dataset

Heart disease screening models are usually tuned for accuracy, although a missed diagnosis costs far more than a false alarm. This study compared six classifiers (Decision Tree, Random Forest, Support Vector Machine, XGBoost, Logistic Regression and a Multilayer Perceptron) on the 303-patient Cleveland Heart Disease dataset, in which 139 patients had angiographic disease. A widely shared copy of the dataset was found to carry a reversed outcome label, so all experiments used the original repository file. Hyperparameters were tuned by repeated nested cross-validation, and each model's decision threshold was fixed on training data to reach a recall of at least 0.90. Across 50 outer test folds, every model met the recall target (0.89 to 0.91). Logistic Regression achieved the highest mean specificity (0.68; 95% CI 0.60 to 0.77) and precision-recall AUC (0.91), although its advantage over Random Forest and Support Vector Machine was not statistically significant. The Decision Tree ruled out far fewer disease-free patients (specificity 0.43). Restricting the inputs to five pre-test variables reduced Logistic Regression specificity to 0.57. Logistic Regression is recommended as an interpretable, recall-oriented baseline, subject to external validation in unselected primary-care populations.