Machine-Learning Based Risk Prediction of Outcomes in Patients Hospitalised With COVID-19 in Australia: The AUS-COVID Score.

Abstract

Background: We aimed to develop a machine-learning based risk score to predict in-hospital mortality, intubation and adverse cardiovascular events in patients hospitalised with COVID-19. Method(s): This Australian, multicentre, prospective study included 1,714 consecutive adult patients hospitalised with COVID-19. Data were separated into training (80%) and test sets (20%). Eight supervised machine-learning methods were used: LASSO, ridge, elastic net (EN), decision tree, support vector machine, random forest, AdaBoost and gradient boosting. Included variables were established through a feature selection method and considered in groups of 5/10/15/20/all. The final models were selected by balancing the optimal area under the curve (AUC) score with interpretability, through the number of variables. Result(s): 181 (10.6%) patients died in-hospital, 148 (8.6%) patients required intubation and 90 (5.3%) patients had adverse cardiovascular events. The LASSO model performed best (AUC 0.852) for in-hospital mortality with 5 variables: age, respiratory rate, features of COVID-19 on chest X-ray (CXR), troponin elevation and COVID-19 vaccination (>=1 dose). For intubation, the EN model demonstrated optimal performance (AUC 0.752) with five variables: pre-existing cardiovascular disease, gender, COVID-19 vaccination (>=1 dose), CXR and initial oxygen saturation on room air. The EN model also performed best (AUC 0.636) for adverse cardiovascular events with five variables: smoking status, creatinine, pre-existing cardiovascular disease, CXR and troponin elevation. To facilitate real-world use, we built a user-friendly web application which provides a risk score as a percentage. Conclusion(s): The AUS-COVID Score is a robust, pragmatic machine-learning based risk score to predict in-hospital mortality, intubation and adverse cardiovascular events in patients hospitalised with COVID-19.Copyright � 2024