A Heterogeneous Stacking Ensemble for Cardiac Arrhythmia Detection Combining Deep Learning and Heart Rate Variability Features

Abstract

One of the most clinically crucial problems in biomedical signal processing is automated detection of cardiac arrhythmias from electrocardiogram (ECG) signals. Previous studies have either used deep learning models which were applied to raw waveforms or classical machine learning algorithms which are built on handcrafted features, or a combination of both the methods. The present study proposes a five-model stacked ensemble  that merges three deep learning architectures with two gradient-boosted tree classifiers operating on heart rate variability (HRV) features. This work proposes a hybrid model that integrates a CNN-Transformer hybrid, a multi-scale one-dimensional convolutional neural network (CNN),  XGBoost, a bidirectional long short-term memory network with self-attention (BiLSTM) and LightGBM. A final decision was made by a logistic regression meta-learner combining with their probability outputs. In total 33,109 annotated beats were evaluated from 44 MIT-BIH Arrhythmia database records through five class AAMI EC57 taxonomy. The model achieves an overall accuracy of 97.72% with weighted F1- score of 0.98 and a averaged ROC-AUC of 0.96 respectively. In case of critical ventricular ectopic beat (VEB) class, the model achieves a recall of 0.97 with F1-score of 0.90. A study by nine-configuration ablation support that each component contributes to final performance, with the full ensemble outperforming all subsets. A cross-dataset generalisation experiment on 370 PTB-XL records reveals perfect classification of normal rhythms alongside expected degradation for minority ectopic classes, attributable to structural differences between the two databases. The results demonstrate that heterogeneous model fusion, combining morphological and temporal ECG analysis with HRV-derived statistics, offers a robust and interpretable approach to automated arrhythmia screening.