Multi-label Classification of Abnormalities in 12-Lead ECG Using Deep Learning

Abstract

Identifying arrhythmias from electrocardiogram(ECG) signals remains an intractable challenge. This study aims to develop an effective and non-invasive approach to realize the recognition of arrhythmias based on 12-lead ECG for the PhysioNet/Computing in Cardiology Challenge2020. To this end, we propose a deep learning-based diagnosis approach, called EASTNet which captures the characteristics of cardiac abnormalities and correlation between heartbeats sampled randomly from 12-lead ECG records by a 34-layer 1D-deep squeeze-and-excitation network. Experimenting in the multi-label arrhythmia classification task, our team, EASTBLUE, was unable to rank and score in the hidden validation and test sets, but achieved diagnostic performance with 0.7030 ± 0.0090 metric score using 5-fold cross-validation on the training set. We also investigate the effect of beat sampling on diagnostic performance, and find that the beat sampling plays a role in data augmentation that effectively alleviates network overfitting. These results demonstrate that our approach has good potential application prospects in clinical practice, especially in the auxiliary diagnosis of abnormalities.