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

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Infectious Disease Epidemiology

Background:

• Hospitalized COVID-19 patients face unpredictable clinical trajectories.
• Early identification of high-risk patients is crucial for timely intervention and resource allocation.
• Existing risk stratification methods may not fully leverage early electronic health record data.

Purpose of the Study:

• To develop and validate a machine learning model for predicting in-hospital mortality, intensive care unit (ICU) transfer, and mechanical ventilation in COVID-19 patients.
• To utilize laboratory data from the first 24 hours post-admission for early risk stratification.
• To create interpretable and data-driven predictive models.

Main Methods:

• Retrospective observational study design.
• Development of machine learning models using electronic health record data, focusing on the first 24 hours of admission.
• Evaluation of model performance using Area Under the Curve (AUC) for predicting mortality, ICU transfer, and mechanical ventilation.
• Comparison of models using numerical versus dichotomized laboratory features.

Main Results:

• The machine learning model achieved high predictive accuracy for in-hospital mortality (AUC = 0.918) and ICU transfer (AUC = 0.821).
• The model also predicted the need for mechanical ventilation with moderate accuracy (AUC = 0.654).
• Dichotomous features (threshold-based) performed comparably to numerical features, simplifying model input.

Conclusions:

• Machine learning models can effectively stratify risk for hospitalized COVID-19 patients within 24 hours of admission.
• Key predictors include C-reactive protein (CRP), blood glucose, partial thromboplastin time (PTT), and glutamic oxaloacetic transaminase (GOT).
• These interpretable models offer a valuable tool for early clinical decision-making and resource management in COVID-19 care.