Predicting Mortality and ICUs Transfer in Hospitalized COVID-19 Patients Using Random Forest Model

Abstract

Background: The objective of the present study was to identify prognostic factors associated with mortality and transfer to intensive care units (ICUs) in hospitalized COVID-19 patients using random forest (RF). Also, its performance was compared with logistic regression (LR).
 Methods: In this retrospective cohort study, information of 329 COVID-19 patients were analyzed. These patients were hospitalized in Besat hospital in Hamadan province, the west of Iran. The RF and LR models were used for predicting mortality and transfer to ICUs. These models' performance was assessed using area under the receiver operating characteristic curve (AUC) and accuracy.
 Results: Of the 329 COVID-19 patients, 57 (15.5%) patients died and 106 (32.2%) patients were transferred to ICUs. Based on multiple LR model, there was a significant association between age (OR=1.02; 95% CI=1.00-1.05), cough (OR=0.24; 95% CI=0.10-0.56), and ICUs (OR=7.20; 95% CI=3.30-15.69) with death. Also, a significant association was found between kidney disease (OR=3.90; 95% CI=1.04-14.63), decreased sense of smell (OR=0.28; 95% CI=0.10-0.73), Kaletra (OR=2.53; 95% CI=1.39-4.59), and intubation (OR=8.32; 95% CI=3.80-18.24) with transfer to ICUs. RF showed that the order of variable importance has belonged to age, ICUs, and cough for predicting mortality; and age, intubation, and Kaletra for predicting transfer to ICUs.
 Conclusion: This study showed that the performance of RF provided better results compared to LR for predicting mortality and ICUs transfer in hospitalized COVID-19 patients.