3-D VCG Loop-Based Myocardial Infarction Identification With Multiscale Spatial Residual Attention

Early diagnosis of myocardial infarction (MI) is crucial for timely treatment and prognosis of patients. Currently, most automatic classification methods for MI rely on features extracted from 12-lead electrocardiograms (ECGs) and three-lead vectorcardiograms (VCGs). However, information loss during spatial mapping and the lack of spatial morphology details collectively hinder the effectiveness of MI identification. To address these challenges, a novel multiscale spatial residual attention network (MS-SRAN) is proposed for automated feature extraction of 3-D VCG loop within 3-D spatial domains. The spatial representation of 3-D VCG loop from three-lead VCG signals is reconstructed to reflect the spatial distribution characteristics of cardiac tissue electrical activity. By embedding a multiscale feature fusion module and a spatial attention module within 3-D ResNet, MS-SRAN captures continuous spatiotemporal variations and intricate local spatial features of 3-D VCG loops. The experiments on the publicly available PhysioNet/PTBDB diagnostic database have demonstrated that our approach can achieve 99.96% and 97.33% of accuracy for intrapatient and interpatient identification of MI events, respectively, which outperforms existing approaches.