Automatic Concurrent Arrhythmia Classification Using Deep Residual Neural Networks

Abstract

This paper addresses the PhysioNetlComputing in Cardiology Challenge 2020. The challenge presents a problem to classify 26 types of arrhythmias and normal sinus rhythm using 12-lead electrocardiogram data. We were able to successfully perform the classification task using an eight layer deep residual neural network (ResNet). The skip connections present in the ResNet allowed the model to train faster and produce better challenge score. We also investigated sixteen other models that included convolution and recurrent neural network based models along with interpretability based attention mechanism as all of them are well suited for time series classification problems. The results depicted that the 8 layer ResNet model outperformed other models in terms of challenge score consuming significantly less time during the training phase. We preferred batch wise training to avoid having all the data in memory during training thereby alleviating the problem of memory choking. Our team, deepzx987, obtained a challenge score of 0.305 on validation data, -0.035 on the full test set, and ranked 35th in this year's challenge.