Machine Learning Approach to Predict Life-threatening Outcomes with Admit Electrocardiograms of Hospitalized Patients with COVID-19

Abstract

COVID-19 has been straining the health care systems worldwide due to its high fatality rate and high infection rate. Recent studies found that COVID-19 can cause life-threatening cardiovascular complications in patients. Therefore, there is a need for early risk-stratified tools. In this thesis, I proposed two machine learning algorithms for predicting the probabilities of mortality or developing cardiovascular complications for hospitalized patients with confirmed COVID-19. Machine learning is a technique that combines statistics and algorithms which enable the machine to extract underlying features and draw inferences from a huge amount of data. The models in this work only use a standard 10-second 12-lead intake electrocardiogram (ECG) as input data. The models were trained and evaluated with a database containing 1270 intake ECGs which were split into a training set, a validation set, and a testing set in a ratio of 8:1:1. The prediction results yield an average sensitivity of 0.65 and an average specificity of 0.52 for all-cause mortality. For predicting life-threatening cardiovascular outcomes, it reaches an average sensitivity of 0.63 and an average specificity of 0.44.