An Automated Cardiac Arrhythmia Classification Network for 45 Arrhythmia Classes Using 12-Lead Electrocardiogram

Abstract

Electrocardiogram is a non-invasive, inexpensive, and widely used diagnostic tool for arrhythmia diagnosis in clinics. Deep learning techniques have shown great promise in electrocardiogram signal analysis, enabling automatic and accurate detection of various cardiac arrhythmia. This paper proposes an automated multi-label cardiac arrhythmia classification network based on a convolutional neural network. The network aims to detect and classify 45 different cardiac arrhythmia classes using 12-lead electrocardiogram data. Unlike previous studies, our approach incorporates both the residual structure and channel attention mechanism. Thus, we developed two key schemes to improve classification performance: the Global Channel Attention Block and the Short Residual Block. The Global Channel Attention Block incorporates dilated convolutions to preserve overall features. It focuses on the important characteristics of each arrhythmia class from the original electrocardiogram data during the training process. The Short Residual Block employs a residual structure to enhance classification accuracy. The network’s performance is evaluated using a large-scale 12-lead electrocardiogram database for arrhythmia study on PhysioNet and the 2018 China Physiological Signal Challenge dataset. In particular, the proposed classification network shows the highest scores in average precision, recall, F1 score, area under the receiver operating characteristic, and accuracy compared to existing convolutional neural network-based arrhythmia classification networks in a large-scale 12-lead electrocardiogram database for arrhythmia study on PhysioNet.

Keywords

• 12-lead electrocardiogram
• deep learning
• convolutional neural network
• multi-label classification
• cardiac arrhythmia classification