Effectiveness of Machine Learning for COVID-19 Patient Mortality Prediction Using WEKA

Abstract

Timely detection of patients with a high mortality risk in coronavirus disease 2019 (COVID-19) can substantially improve triage, bed allocation, time reduction, and potential outcomes. A potential solution is using machine learning (ML) algorithms to predict mortality in COVID-19 hospitalized patients. The study's objective was to create and verify individual risk assessments for mortality using anonymous demographic, clinical, and laboratory findings at admission, as well as to assess the possibility of death using machine learning. We used a standardized format and electronic medical records. Data from 2,313 patients were collected from two Muhammadiyah hospitals from January 2020 to July 2022. Utilizing each patient's clinical manifestation state at admission and laboratory parameters, 24 demographic, clinical, and laboratory results were studied. The algorithms analyzed were AdaBoost, logistic regression, random forest, support vector machine, naïve Bayes, and decision tree, which were applied through WEKA version 3.8.6. Random forest performed better than the other machine learning techniques, with precision, sensitivity, receiver operating characteristic (ROC), and accuracy of 78.6%, 78.7%, 85%, and 78.65%, respectively. The three top predictors were septic shock (OR=21.518, 95% CI=4.933–93.853), respiratory failure (OR=15.503, 95% CI=8.507–28.254), and D-dimer (OR=3.288, 95% CI=2.510–4.306). Machine learning–based predictive models, especially the random forest algorithm, may make it easier to identify patients at high risk of death and guide physicians' appropriate interventions.