ECG signal classification with binarized convolutional neural network

Arrhythmias are a group of common conditions associated with irregular heart rhythms. Some of these conditions, for instance, atrial fibrillation (AF), might develop into serious syndromes if not treated in time. Therefore, for high-risk patients, early detection of arrhythmias is crucial. In this study, we propose employing deep convolutional neural network (CNN)-based algorithms for real-time arrhythmia detection. We first build a full-precision deep convolutional network model. With our proposed construction, we are able to achieve state-of-the-art level performance on the PhysioNet/CinC AF Classification Challenge 2017 dataset with our full-precision model. It is desirable to employ models with low computing resource requirements. It has been shown that a binarized model requires much less computing power and memory space than a full-precision model. We proceed to verify the feasibility of binarization in our neural network model. Network binarization can cause significant model performance degradation. Therefore, we propose employing a full-precision model as the teacher to regularize the training of the binarized model through knowledge distillation. With our proposed approach, we observe that network binarization only causes a small performance loss (the F1 score decreases from 0.88 to 0.87 for the validation set). Given that binarized convolutional networks can achieve favorable model performance while dramatically reducing computing cost, they are ideal for deployment on long-term cardiac condition monitoring devices. (Source code is available at https://github.com/yangfansun/bnn-ecg).