A deep neural network based on multi-model and multi-scale for arrhythmia classification

Abstract

Automatic arrhythmia classification has been a cross research hot spot of artificial intelligence applied on biomedical engineering. In this study, a deep neural network based on multi-model and multi-scale net (MMnet) was proposed to classify arrhythmia. Firstly, the raw one-dimensional (1-D) ECG recordings from the public 2018 China Physiological Signal Challenge database (CPSC 2018 DB) was preprocessed to remove noise by a wavelet transform algorithm based on db6, and the clean 1-D recordings was transformed into their corresponding two-dimensional (2-D) time–frequency diagrams by a modified frequency slice wavelet transform (MFSWT). Then, the clean 1-D ECG recordings and the corresponding 2-D diagrams were used as model 1 and model 2 of the multi-model input of MMnet to feed into the multi-scale net of the MMnet which included a backbone module for extracting initially rough features and a dilated convolution module for further fine features. So, the more comprehensive and accurate information within the 1-D recordings and the 2-D diagrams was acquired for classifying all 9 arrhythmia categories (i.e., N, AF, I-AVB, RBBB, PAC, PVC, STD, LBBB and STE). Finally, the results for 9-class classification were precision 84.91 %, recall 82.64 % and F1 83.52 %. The results indicate the MMnet can obtain relatively more comprehensive information within the ECG recordings.