Predicting Heart Disease Using Machine Learning and Ensemble Models: A Comparative Study

Overview & Synopsis

Heart disease (CVD) is still the leading cause of death globally, with nearly 32% of humans dying from heart disease each year. Furthermore, traditional diagnostic methods (ECG, angiography, stress testing, etc.) are invasive, expensive, and struggle to make adequate predictions. This study examines a diverse machine learning system predicting heart disease utilizing both standalone classifiers and ensemble techniques. The combined dataset consisted of 2,091 patients with 15 demographic and clinical characteristics, with balanced cases (having heart disease and not) for each patient. Following significant rigorous preprocessing requirements (normalization, encoding, dimensionality reductions (PCA, UMAP, autoencoders), and stratified cross-validation) our work tests multiple individual supervised, unsupervised, and ensemble based learning methods. The best performing models were stacking, LightGBT, and the voting classifier which achieved testing accuracies of 97.37% and produces nearly perfect ROC-AUC values of 0.995. Other models including ensemble, but more weak classifiers, such as AdaBoost, produced terrible performance too. The precision and recall analyzed upon each of the methodologies, within confusion matrices, confirmed these ensemble-based models maintained a strong sensitivity and specificity while minimizing the mis-classification rates. Our findings suggest machine learning pipelines can support heart disease prediction which is accurate and reliable, while supporting patient detection, and support clinical decision.