Electrocardiogram Arrhythmia Classification with SE-ResNet: Effects of Signal Windowing and Hybrid Class Balancing

Automated electrocardiogram (ECG) arrhythmia classification requires models that can preserve diagnostically relevant waveform morphology while remaining robust to severe class imbalance. This study evaluates a One-Dimensional Convolutional Neural Network (1D-CNN) integrating Squeeze-and-Excitation Residual Network (SE-ResNet) on the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) Arrhythmia Database, with particular emphasis on signal window length and data-balancing strategy. Two input windows, 180 and 360 samples, were compared together with Tomek-Links, Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), Tomek-SMOTE, and Tomek-ADASYN. The dataset was stratified into 80% training and 20% testing subsets, with an internal validation split during training. All balancing procedures were applied only to the training data, and performance was assessed across the five Association for the Advancement of Medical Instrumentation (AAMI) heartbeat classes using Accuracy, Precision, Recall, F1-score, Area Under the Curve (AUC), and computational time. The 360-sample window consistently produced stronger classification results than the 180-sample window, indicating that the longer segment retained more complete ECG morphology. Among the evaluated configurations, the 360-sample SE-ResNet with Tomek-SMOTE achieved the highest overall performance, with 98.77% accuracy, 98.76% weighted F1-score, and an AUC of 0.9962. However, this configuration required 921.96 s of training, whereas Tomek-Links alone achieved 98.63% accuracy with substantially lower computational cost. These results show that signal windowing and class-balancing design jointly influence both predictive performance and computational demand, providing a practical basis for selecting ECG classification configurations under different resource constraints.