Arrhythmia Classification from Single Lead ECG by Multi-Scale Convolutional Neural Networks.

Abstract

Arrhythmia refers to any abnormal change from the normal electrical impulses of the heart. Some arrhythmias are manifested as abnormal heartbeat. Effective heartbeat classification is helpful for computer aided diagnosis. Conventional heartbeat classification methods work on information of multiple leads, and need heuristic or hand-crafted feature extraction. In this paper, we propose a new heartbeat classification approach based on a recent deep learning architecture called multi-scale convolutional neural networks (MCNN). A unique feature of our work is that we take single lead ECG as input, rhythm information is not taken into consideration. Such a single-lead setting, although more challenging than multi-lead cases, is often faced in medical practice due to advancements in mobile ECG devices and hence much needed. We exploit the power of convolutional neural networks for find discriminative features in heartbeat time series. The algorithm was tested on public datasets. The overall accuracy is 0.8866, the accuracy on supraventricular ectopic beat is 0.9600, and accuracy on ventricular ectopic beat is 0.9250. The performance is comparable with conventional method using features hand crafted by human experts.