Multiperceptive Region of Spatial-Temporal Graph Convolutional Shrinkage Network for Arrhythmia Recognition

In the realm of intelligent arrhythmia recognition, deep learning (DL) methods have dominated in recent years. However, these methods often struggle to explicitly capture relationships between signals. Graph convolutional networks (GCNs) are designed to efficiently mine data relationships in graph-structured data with a topological structure. Nevertheless, existing GCN methods have notable limitations. First, one limitation of GCNs is their fixed perceptive region, which restricts their ability to effectively capture and represent complex features. Second, noise is inevitably introduced into input data during the collection and feature generation of electrocardiogram (ECG) sequences. To tackle these concerns, combined with a shrinkage block, a multiperceptive region of spatial-temporal graph convolutional shrinkage network (MPR-STSGCN) is proposed for effective intelligent arrhythmia recognition. In MPR-STSGCN, features from different perceptive regions are not only learned but also fused through a shrinkage block that incorporates channel attention and a soft thresholding module. The shrinkage block is specifically designed to reduce unimportant features by dynamically learning thresholds. Moreover, weighted graphs are employed to represent data samples and signify dissimilarities in their relationships. The effectiveness of the proposed method is validated on the MIT-BIH Arrhythmia dataset. The proposed MPR-STSGCN makes progress compared with the state-of-the-art methods.