Blending Ensemble Learning Model for 12-Lead Electrocardiogram-Based Arrhythmia Classification

Abstract

The increasing prevalence of heart diseases has driven the development of automated arrhythmia classification systems using machine learning and electrocardiograms (ECGs). This paper presents a novel ensemble learning method for classifying multiple arrhythmia types using 12-lead ECG signals through a blending technique. The framework employs a predetermined meta-model from foundation models, while the remaining models serve as potential base estimators, ranked by accuracy. Using sequential forward selection and meta-feature augmentation, the system determines an optimal base estimator set and creates a meta-dataset for the meta-model, which is optimized through grid search with k-fold cross-validation. Experiments conducted with seven diverse machine learning algorithms (Adaptive Boosting, Extreme Gradient Boosting, Decision Trees, k-Nearest Neighbors, Logistic Regression, Random Forest, and Support Vector Machine) demonstrate that the proposed blending solution, utilizing an LR meta-model with three optimal base models, achieves a superior classification accuracy of 96.48%, offering an effective tool for clinical decision support.