Machine learning models for predicting severe COVID-19 outcomes in hospitals

Abstract

Objectives The aim of this observational retrospective study is to improve early risk stratification of hospitalized Covid-19 patients by predicting in-hospital mortality, transfer to intensive care unit (ICU) and mechanical ventilation from electronic health record data of the first 24 hours after admission. Methods and Results Our machine learning model predicts in-hospital mortality (AUC=0.918), transfer to ICU (AUC=0.821) and the need for mechanical ventilation (AUC=0.654) from a few laboratory data of the first 24 hours after admission. Models based on dichotomous features indicating whether a laboratory values exceeds or falls below a threshold perform nearly as good as models based on numerical features. Conclusions We devise completely data-driven and interpretable machine-learning models for the prediction of in-hospital mortality, transfer to ICU and mechanical ventilation for hospitalized Covid-19 patients within 24 hours after admission. Numerical values of CRP and blood sugar and dichotomous indicators for increased partial thromboplastin time (PTT) and glutamic oxaloacetic transaminase (GOT) are amongst the best predictors.