Objective: The ongoing emergence of novel severe acute respiratory syndrome coronavirus 2 strains such as the Omicron variant amplifies the need for precision in predicting severe COVID-19 outcomes. This study presents a machine learning model, tailored to the evolving COVID-19 landscape, emphasizing novel risk factors and refining the definition of severe outcomes to predict the risk of a patient experiencing severe disease more accurately.
Methods: Utilizing electronic health records from the Healthjump database, this retrospective study examined over 1 million US COVID-19 diagnoses from March 2020 to September 2022. Our model predicts severe outcomes, including acute respiratory failure, intensive care unit admission, or ventilator use, circumventing biases associated with hospitalization, which exhibited ∼4× geographical variance of the new outcome.
Results: The model exceeded similar predictors with an area under the curve of 0.83 without lab data to predict patient risk. It identifies new risk factors, including acute care history, health care encounters, and distinct medication use. An increase in severe outcomes, typically 2-3× higher than subsequent months, was observed at the onset of each new strain era, followed by a plateau phase, but the risk factors remain consistent across strain eras.
Conclusion: We offer an improved machine learning model and risk score for predicting severe outcomes during changing COVID-19 strain eras. By emphasizing a more clinically precise definition of severe outcomes, the study provides insights for resource allocation and intervention strategies, aiming to better patient outcomes and reduce health care strain. The necessity for regular model updates is highlighted to maintain relevance amidst the rapidly evolving COVID-19 epidemic.
Keywords: COVID-19; Clinical care; Clinical epidemiology; Predictive score; Risk assessment.
Copyright © 2023 Elsevier Inc. All rights reserved.