Graph-enhanced deep learning for ECG arrhythmia detection: An integration of CNN-GNN-BiLSTM approach

Abstract

Early and accurate detection of cardiac arrhythmias is crucial for preventing severe cardiovascular events. This study proposes a CNN–GNN–BiLSTM integrated framework for automated ECG arrhythmia classification, combining spatial, relational, and temporal learning to achieve enhanced predictive accuracy. Convolutional Neural Networks (CNNs) serve as feature extractors from ECG spectrograms, while Graph Attention Networks (GATs) capture inter-beat relationships through graph-based modeling. In parallel, Bidirectional Long Short-Term Memory (BiLSTM) networks refine temporal dependencies, ensuring robust sequential representation. Outputs from GAT and BiLSTM modules are concatenated to form a unified feature representation, which is passed through a fully connected classifier for final prediction. The model is evaluated on three benchmark ECG datasets—MIT-BIH, PTB, and Chapman-Shaoxing—as well as a combined 11-class dataset, demonstrating superior generalization. Results indicate significant performance improvement over conventional deep learning approaches, achieving 96.0% overall accuracy and up to 99.89% accuracy on MIT-BIH. The proposed framework effectively mitigates misclassification errors and offers a scalable, real-time solution for AI-driven cardiac monitoring systems.