MST-BiLSTM: An Improved Bi-LSTM Method Based on Multi-Scale Spatio-Temporal Feature Fusion and Attention Mechanism for ECG Recognition

Electrocardiography (ECG) plays a crucial role in assessing heart health and detecting cardiac abnormalities. The development of wearable single-lead ECG devices has transformed long-term remote monitoring, offering unparalleled convenience. However, capturing the dynamic nature of ECG signals, along with local nuances and effective discriminative features, remains a challenge in automatic ECG classification. To address this, we propose MST-BiLSTM, a deep learning model that integrates multi-scale Spatio-Temporal feature fusion with Bi-LSTM. By employing a multi-scale structure and attention mechanism, the model effectively extracts and merges spatiotemporal features from single-lead ECG signals, emphasizing critical distinctions for arrhythmia classification. The model is validated on five heartbeat types using the MIT-BIH Arrhythmia Database, achieving impressive results with an average accuracy of 99.68 %, precision of 99.69 %, sensitivity of 99.67 %, and specificity of 99.68 %. These results highlight MST-BiLSTM's superior capability in capturing local ECG features and generating more discriminative feature representations than traditional methods. The proposed approach offers a robust solution for ECG classification, particularly in conditions of limited data, contributing to improved cardiac health monitoring and management.