DRGCN-BiLSTM: An Electrocardiogram Heartbeat Classification Using Dynamic Spatial-Temporal Graph Convolutional and Bidirectional Long-Short Term Memory Technique

Abstract

An automated cardiac rhythm classification using electrocardiograms is crucial for accurate and timely diagnosis of cardiovascular disease. Recent advances in deep learning have facilitated automated arrhythmias recognition, surpassing traditional ECG methods that depend on manual feature extraction. Despite significant progress in existing arrhythmias classification techniques, the current method fails to utilize spatial-temporal interaction and task-specific features include temporal dependencies among observations and a significantly imbalanced class distribution in the dataset. Thus, the accuracy of network-based ECG heart-beat classification still requires improvements. To address this issue, an effective classification algorithm combined with synthetic minority oversampling technique (SMOTE)-Tomek, dynamic perceptive region spatial-temporal graph convolutional network with bidirectional long-short term memory (DRGCN-BiLSTM) is proposed for effective intelligent arrhythmia recognition. This $ \mathrm {DRGCN\_BiLSTM}$ model employs a trainable weighted $\epsilon $ -neighborhood graph to capture the pattern of time series within ECG segments. The SMOTE-Tomek technique addresses data imbalance, while BiLSTM captures temporal features from the graph network. The presented technique was validated on the MIT-BIH arrhythmia database, comprising a total of 109,253 ECG beats and the experimental results demonstrate the average accuracy of 99.92% respectively. The proposed method achieves superior performance as compared to state-of-the-art techniques, which results in better diagnosis of heart-related problems.