ST-DGCN: A Novel Spatial-Temporal Dynamic Graph Convolutional Network for Cardiovascular Diseases Diagnosis

Recently, Deep Learning (DL) technology has made significant progress in the field of automatic diagnosis of cardiovascular diseases (CVDs) based on electrocardiograms (ECGs). Multi-lead ECG signals are physiological signals obtained through a series of lead systems based on potential differences between electrodes placed on the limbs and chest. However, most DL models treat them as one-dimensional signals distributed in Euclidean space, often focusing only on features along the temporal dimension and neglecting the spatial relationships between different leads. Different studies indicate that these spatial relationships are physiologically significant for the diagnosis of CVDs, as they represent the activity of different regions of the heart. Given the advantages of Graph Convolutional Networks (GCNs) in analyzing non-Euclidean data, this study proposes a novel method for CVD diagnosis. The method begins by segmenting ECG signals into multiple single-lead segments and converting them into the nodes of a graph. Subsequently, these nodes are interconnected through spatial-temporal connections based on their relationships of physiological structures. The proposed model utilizes a dynamic graph convolutional network to capture the spatial-temporal features of the ECG signals and employs hierarchical pooling techniques to mitigate issues of oversmoothing and overfitting. Compared to other state-of-the-art models (SOTAs), this model achieved at least a 6.5% and 9.3% increase in F1 scores on the Chapman and PTB-XL databases, respectively, such significant performance advantages highlight the effectiveness and reliability of the model in classifying ECG signals, providing a powerful tool for the diagnosis of cardiovascular diseases.