Abstract:The automatic classification of cardiac arrhythmias holds immense clinical value in smart healthcare and mobile health monitoring. The long-term temporal dependencies and complex, variable waveform characteristics of electrocardiogram (ECG) signals, combined with the limited computational resources of portable monitoring devices, have long posed major challenges to the practical implementation of this technology. This study utilizes the Chapman-Shaoxing arrhythmia dataset (macro-level rhythm segments) from Shaoxing People’s Hospital and the MIT-BIH arrhythmia dataset (micro-level single-beat features) from Beth Israel Hospital as its data foundation. The raw signals from both datasets are first denoised using a noise reduction scheme. Based on this, a multi-module heterogeneous model is constructed: CNN extracts local spatial features, ECA dynamically optimizes channel weights, and BiGRU captures global bidirectional temporal information. Experimental results show that the model demonstrates outstanding performance on the Chapman-Shaoxing and MIT-BIH public datasets, achieving overall accuracy rates of 99.08% and 99.74%, respectively. Compared with traditional BiLSTM or heavyweight ResNet models, the proposed model maintains extremely high recognition accuracy while significantly reducing parameter count and computational overhead. This study provides a robust and efficient algorithmic foundation for embedded deployment in portable ECG monitoring devices.