ST-based Deep Learning Analysis of COVID-19 Patients

Abstract

The number of deaths worldwide caused by COVID-19 continues to increase and the variants of the virus whose process we do not yet master are aggravating this situation. To deal with this global pandemic, early diagnosis has become important. New investigation methods are needed to improve diagnostic performance. A very large number of patients with COVID-19 have with cardiac arrhythmias often with ST segment elevation or depression on an electrocardiogram. Can ST-segment changes contribute to automatic diagnosis of COVID-19? In this article, we have tried to answer this question. We propose in this work a method for the automatic identification of COVID patients which exploits in particular the modifications of the ST segment observed on recordings of the ECG signal. Two sources of data allowed the development of the database for this study: 300 ECGs from the "physioNet" database with prior measurement of the ST segments, and 100 paper ECGs of patients from the cardiology department of the hospital X in Tunis registered on (non-covid) topics and covid topics. Four learning algorithms (ANN, CNN-LSTM, Xgboost, Random forest) were then applied on this database. The evaluation results show that CNN-LSTM and Xgboost present better accuracy in terms of classifying covid and non-covid patients with an accuracy rate of 87% and 88.7% respectively.