ECG-based Multi-Class Arrhythmia Detection Using 13-branch Convolutional Neural Network

Jing Zhang1, Junyuan Jing1, Deng Liang1, Xun Chen1, Min Gao2
1University of Science and Technology of China, 2The First Affiliated Hospital of University of Science and Technology of China (Anhui Provincial Hospital)


In the PhysioNet/ Computing in Cardiology Challenge (CinC) 2020, our team USTC-ECG proposes an end-to-end deep learning method to detect multi-class arrhythmia using 12-lead electrocardiogram (ECG) records. 12-lead ECG has two unique attributes: diversity and integrity. From this point, we design a 13-branch convolution neural network (CNN) consisting of 12 branches for learning the diversity and 1 branch for learning the integrity. The 13 branches are independent of each other. Specifically, the fist 12 branches are used to extract diverse features from different leads by taking different leads as input. The last branch is used to extract integrate features by taking 12-lead ECG as input. Finally, we concatenate the features extracted by 13 branches and fed them into a classifier composed of two fully connected layers. At the unofficial phase of CinC 2020 challenge, the proposed method achieves the F_2 score of 0.821, the G_2 score of 0.624 and the Geometric Mean of 0.716, ranking 4 at the leaderborad. It is noted that the K-fold cross validation ensemble is not adopted by the submitted method.