ABSTRACT
COVID-19 disease can be classified into various stages depending on the severity of the patient. Patients in severe stages of COVID-19 need immediate treatment and should be placed in a medical-ready environment because they are at high risk of death. Thus, hospitals need a fast and efficient method to screen large numbers of patients. The enormous amount of medical data in public repositories allows researchers to gain information and predict possible outcomes. In this study, we use a publicly available dataset from Springer Nature repository to discuss the performance of three machine learning techniques for prediction of severity of COVID-19: Random Forest (RF), Naive Bayes (NB) and Gradient Boosting (GB). These techniques were selected for their good performance in medical predictive analytics. We measured the performance of the machine learning techniques using six measurements (accuracy, precision, recall, F1-score, sensitivity and specificity) in predicting COVID-19 severity. We found that RF generates the highest performance score, which is 78.4, compared with NB and GB. We also conducted experiments with RF to establish the critical symptoms in predicting COVID-19 severity, and the findings suggested that seven symptoms are substantial. Overall, the performance of various machine learning techniques to predict severity of COVID-19 using electronic health records indicates that machine learning can be successfully applied to determine specific treatment and effective triage.