A novel deep convolutional neural network for arrhythmia classification

The electrocardiogram (ECG) is a key standard for monitoring the activity regularity of heart. Many cardiac abnormalities will be demonstrated from ECG including arrhythmia, which may be a life-threatening disease. Automatic classification of arrhythmias is one of the most valuable topics in medicine and bioinformatics. we develop a baseline network (network A) and a multi-scale fusion CNN architecture (network B) based on network A to automatically identify five different types of heartbeats in ECG signals. We also design multi-scale signal, including the signals of 250 sample (set A) and 360 sample (set B), which is fed into set B. Our experiment is conducted in set A and set B derived from a publicly available database MIT-BIH arrhythmia database (MITAB). An average accuracy of 92.81%, a sensitivity and specificity of 95.84%, 93.92% respectively is obtained using network A. Meanwhile, an average accuracy of 95.48%, a sensitivity and specificity of 96.53%, 87.74% respectively is obtained using network B. The results demonstrate that multi-scale fusion CNN architecture (network B) is slightly better than network A due to adding a convolution block consist of three convolution layers, which aggregate the feature from all the convolution branches. To the best of our knowledge, the proposed model can be leveraged as a rapid tool to identify different types of arrhythmias.