Deep Learning-Based Automatic Classification of Cardiac Arrhythmia Using Multi-Spectral Attention and BiLSTM Networks

Abstract

Cardiac arrhythmia is a condition charactWerized by irregular heartbeats that may lead to serious health complications such as stroke or heart failure. Electrocardiogram (ECG) is a widely used, noninvasive method to monitor and detect arrhythmias. However, due to signal variability and class imbalance in datasets, the accurate and automated classification of arrhythmias remains a challenge. This study presents a deep learning framework combining Residual Networks (ResNet), Squeeze-and-Excitation (SE) blocks, and Bidirectional Long Short-Term Memory (BiLSTM) layers enhanced with Multi-Spectral attention and synthetic minority oversampling (SMOTE). The proposed model is trained and validated on the MIT-BIH, AFDB, and CinC databases. Results show improved classification accuracy, especially in minority arrhythmia classes, and high generalization ability across datasets, surpassing conventional machine learning and deep learning baselines.