Risk stratification for COVID-19 hospitalization: a multivariable model based on gradient-boosting decision trees.

Area of Science:

• Epidemiology
• Health Informatics
• Biostatistics

Background:

• The COVID-19 pandemic strained healthcare systems globally.
• Accurate prediction of severe outcomes like hospitalization is crucial for resource allocation.
• Existing risk stratification methods may not fully capture patient complexity.

Purpose of the Study:

• To develop and validate a multivariable model for predicting hospitalization risk in SARS-CoV-2 infected patients.
• To leverage routinely collected health administrative data for robust risk prediction.
• To compare the model's performance against established empirical risk rules.

Main Methods:

• Utilized a large cohort of adult patients in Ontario, Canada, who tested positive for SARS-CoV-2 via PCR.
• Employed Extreme Gradient Boosting (XGBoost) algorithm for multivariable risk modeling.
• Validated the model using a held-out dataset and assessed performance using discrimination (AUC) and calibration metrics.

Main Results:

• The XGBoost model demonstrated high predictive accuracy with an AUC of 0.852 in the development cohort and 0.8475 in the validation cohort.
• Hospitalized patients were significantly older, more likely male, and had more comorbidities than non-hospitalized patients.
• The top 10% of patients identified by the model accounted for 47.4% of hospitalizations, and the top 30% accounted for 80.6%.

Conclusions:

• A validated, accurate risk stratification model for COVID-19 hospitalization was developed using readily available health administrative data.
• This model can support population health management strategies for COVID-19.
• The findings highlight the potential of machine learning in public health surveillance and resource management during pandemics.