Title: A Dual-Stream Gated Cross-Attention 1-D ResNet with Tree-Ensemble Meta-Learning for Robust Inter-Patient ECG Arrhythmia Classification
Authors: Nguyen Phu Hai Anh, Nguyen Minh Duc
Volume: 10
Issue: 9
Pages: 20-32
Publication Date: 2026/09/28
Abstract:
Automated classification of cardiac arrhythmias from single-lead electrocardiograms supports screening and wearable monitoring but remains challenged by inter-patient variability, class imbalance, and false alarms. This study proposes a Dual-Stream Gated Cross-Attention 1D ResNet integrated with a tree-ensemble meta-learning model. Evaluation followed the inter-patient de Chazal protocol and AAMI EC57 recommendations using the MIT-BIH Arrhythmia Database. A patient-independent split comprised 18 training patients, 4 validation patients, and 22 held-out test patients containing 49,684 beats. The framework combines physiology-guided augmentation, RR-interval dynamics, morphological representations, gated cross-attention, and regularized logistic-regression stacking of neural-network and gradient-boosted-tree probabilities. On the held-out test set, the stacking ensemble achieved 95.66% positive predictivity, 76.68% sensitivity, and an F1 score of 85.12% for ventricular ectopic beats, reducing the ventricular false-alarm rate to 4.34%. For supraventricular ectopic beats, sensitivity increased from 16.55% with the strongest tree baseline to 75.78%. A soft-voting variant achieved the highest overall accuracy of 92.64% and ventricular sensitivity of 92.95%, providing an alternative operating point. Grad-CAM analysis indicated attention to the P-wave region for supraventricular beats and the widened QRS complex for ventricular beats. With 226,000 parameters, a 223 KB INT8 footprint, and per-beat latency below 1.5 ms, the model shows potential for embedded arrhythmia monitoring.