Graph structure learning based DL model for ECG anomaly prediction

Abstract

Electrocardiography (ECG) is widely used in cardiography as a non-invasive diagnostic tool for providing a graphical representation of the electrical activity in the heart over a duration of time. It captures the electrical impulses generated by cardiac muscles and is used to detect several types of cardiac conditions, such as hypertrophy and arrhythmia. In this study we work on developing a deep learning model which can effectively classify abnormalities from 12 lead ECG data. We use the PTB-XL dataset, the largest publicly available dataset for 12 lead ECGs.In order to harness the inter-relationship from the data from the 12 leads, we model them asa graph, with the graph structure being learned and design a deep learning model consistingof a graph convolution network (GCN) and present a comprehensive quantitative evaluation, demonstrating comparable performance when compared to existing state of the art works.