Robust Predictive Model for Heart Disease Diagnosis Using Advanced Machine Learning Techniques

Abstract

This study presents a hybrid ensemble learning framework designed to enhance the predictive accuracy, robustness, and generalizability of heart disease classification models. The framework integrates three base classifiers: Decision Tree (DT), Gaussian Naive Bayes (GNB), and K Nearest Neighbor (KNN), which are combined using a stacking ensemble method with Logistic Regression (LR) as the meta learner. Each classifier contributes a distinct analytical perspective: DT models nonlinear relationships, GNB provides probabilistic reasoning, and KNN captures similarity-based patterns. Logistic Regression aggregates their outputs to produce a unified predictive decision. To mitigate class imbalance commonly observed in clinical datasets, the Synthetic Minority Oversampling Technique (SMOTE) is applied to generate synthetic samples of the minority class, improving the model’s ability to recognize underrepresented cases. Hyperparameter optimization is performed using the Optuna framework, which applies the algorithm to efficiently explore parameter configurations. The proposed model was evaluated on a publicly available heart disease dataset and achieved an accuracy of 99.61%, precision of 99.62%, recall of 99.59%, F1 score of 99.60%, and specificity of 99.58%, corresponding to a false positive rate of only 0.42 percent. These results demonstrate the framework’s strong ability to accurately identify heart disease cases while minimizing misclassification. The integration of SMOTE, stacking, and Optuna optimization contributes to its superior performance and robustness. Consequently, this approach shows strong potential for integration into clinical decision support systems to assist healthcare professionals in reliable and timely diagnosis.