ST-ReGE: A Novel Spatial-Temporal Residual Graph Convolutional Network for CVD.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning (DL) models have advanced electrocardiogram (ECG)-based cardiovascular disease (CVD) diagnosis.
• Conventional DL models often overlook spatial relationships between ECG leads, focusing primarily on temporal features.
• These spatial relationships are physiologically significant for accurate CVD diagnosis.

Purpose of the Study:

• To propose a novel spatial-temporal residual graph convolutional network (GCN) for enhanced CVD diagnosis using multi-lead ECG signals.
• To incorporate non-Euclidean data analysis to better represent the nature of multi-lead ECG signals.
• To improve the accuracy and efficiency of automated CVD diagnosis systems.

Main Methods:

• ECG signals were divided into single-channel patches and transformed into nodes for GCN analysis.
• Spatial-temporal connections were established between nodes to capture lead relationships.
• Residual GCN blocks and feed-forward networks were employed to mitigate over-smoothing and over-fitting.

Main Results:

• The proposed spatial-temporal residual GCN model demonstrated superior performance in CVD diagnosis.
• The model achieved significant increases in F1-score (5.85% and 6.80%) over state-of-the-art algorithms on PTB-XL and Chapman databases.
• The approach effectively captures both global and detailed spatial-temporal features.

Conclusions:

• The novel GCN model offers a promising approach for intelligent CVD diagnosis.
• This method effectively utilizes the spatial information inherent in multi-lead ECGs.
• The proposed model provides an efficient solution for automated CVD diagnosis, even with limited computational resources.