An Optimized Stacking Ensemble Framework With Synthetic Data Augmentation and Gradient-Boosted Feature Selection For Heart Disease Classification

Abstract

According to the World Health Organization, heart disease causes about 17.9 million deaths every year and is one of the leading causes of death worldwide. Early diagnosis improves patient outcomes and lowers healthcare costs, but conventional diagnostic methods are costly, time-consuming and require specialist medical skills. This study develops an optimized machine learning framework based on a stacking ensemble for accurate heart disease prediction and early clinical diagnosis. Five benchmark heart disease datasets (Cleveland, Hungarian, Switzerland, Long Beach and Statlog) were combined to increase the diversity and reliability of the data. Synthetic data augmentation, based on bootstrap sampling and controlled noise injection, was used to compensate for the small sample size and to enlarge the dataset to about 30,000 observations. Gradient Boosted Decision Trees (GBDT) were used to identify the most important predictors. The stacking ensemble consisted of four base learners (Random Forest, Decision Tree, K-Nearest Neighbours and Logistic Regression) and a Logistic Regression meta-learner. Model performance was assessed with accuracy, precision, recall, F1-score and the area under the receiver operating characteristic curve (ROC-AUC) using five-fold cross-validation and a held-out test set. The proposed stacking ensemble achieved an accuracy of 99.97%, with precision, recall and F1-score close to 100% for both classes and a ROC-AUC of 1.000. The combination of several data sources, synthetic data augmentation and ensemble learning was central to this performance, and the optimized stacking ensemble shows promise for the early diagnosis of heart disease in clinical decision support.