Machine learning identifies ICU outcome predictors in a multicenter COVID-19 cohort

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Infectious Disease Epidemiology

Background:

• Intensive Care Units (ICUs) faced significant strain during the COVID-19 pandemic.
• Existing methods for stratifying SARS-CoV-2 patient risk and predicting ICU outcomes were insufficient.
• Accurate prediction models are crucial for managing critical COVID-19 cases.

Purpose of the Study:

• To develop and validate a machine learning model for predicting clinical outcomes in ICU patients with COVID-19.
• To stratify patient risk upon ICU admission.
• To predict ICU survival and specific outcomes like need for ECMO or renal replacement therapy.

Main Methods:

• Established a Germany-wide electronic registry for SARS-CoV-2 ICU patients, collecting retrospective and prospective data.
• Evaluated various machine learning approaches for predictive accuracy and interpretability.
• Selected the Explainable Boosting Machine (EBM) as the optimal method, reporting individual parameter functions and interactions.

Main Results:

• Included 1039 patients (596 retrospective, 443 prospective) in the EBM model.
• The EBM model reliably predicted patient survival, with age, inflammatory markers, thrombotic activity, and ARDS severity identified as key predictors.
• Predictors for ECMO therapy included patient age, pulmonary dysfunction, and external transfer; predictors for renal replacement therapy involved age, creatinine, and SOFA score interactions.

Conclusions:

• Explainable Boosting Machine analysis confirmed known and identified novel predictors for outcomes in critically ill COVID-19 patients.
• This machine learning strategy effectively models COVID-19 ICU patient outcomes, surpassing limitations of linear regression.
• Trial registration: NCT04455451.