SE-ECGNet: Multi-scale SE-Net for Multi-lead ECG Data

Abstract

Cardiovascular disease is a life-threatening condition, and more than 20 million people die from heart disease. Therefore, developing an objective and efficient computer-aided tool for diagnosis of heart disease has become a promising research topic. In this paper, we design a multi-scale shared convolution kernel model. In this model, two paths are designed to extract the features of electrocardiogram(ECG). The two paths have different convolution kernel sizes, which are 3×1 and 5×1, respectively. Such multi-scale design enables the network to obtain different receptive fields and capture information at different scales, which significantly improves the classification effect. And squeeze-and-excitation networks (SE-Net) are added to every path of the model. The attention mechanism of SE-Net learns feature weights according to loss, which makes the effective feature maps have large weights and the ineffective or low-effect feature maps have small weights. Our team name is CQUPT_ECG. Our approach achieved a challenge validation score of 0.640, and full test score of 0.411, placing us 8 out of 41 in the official ranking.