Using machine learning for predicting intensive care unit resource use during the COVID-19 pandemic in Denmark

Area of Science:

• Medical Informatics
• Computational Biology
• Epidemiology

Background:

• The COVID-19 pandemic significantly strained hospital resources, necessitating tools for effective resource allocation.
• Accurate forecasting of intensive care unit (ICU) demand is crucial for hospital planning and management during public health crises.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models, specifically Random Forest (RF), in predicting future ICU resource requirements.
• To forecast the risk of ICU admission and mechanical ventilation use for SARS-CoV-2 positive patients.

Main Methods:

• A retrospective analysis of health records from 42,526 SARS-CoV-2 positive patients in Denmark.
• Training RF models to predict the n-day risk (n=1-15) of ICU admission and mechanical ventilation use.
• Assessing model performance using Area Under the Receiver Operator Characteristic Curve (ROC-AUC) and coefficient of determination (R²).

Main Results:

• RF models achieved high accuracy in predicting ICU admission (ROC-AUC: 0.981-0.995) and mechanical ventilation use (ROC-AUC: 0.982-0.997).
• Forecasting models accurately predicted ICU capacity (R²: 0.334-0.989) and ventilation needs (R²: 0.446-0.973).
• Performance decreased with longer forecast horizons (larger n), with 5-day forecasts showing strong predictive power (ICU capacity R²=0.928, ventilation R²=0.854).

Conclusions:

• Random Forest-based ML models are effective for accurate, short-term (n not too large) forecasting of ICU resource requirements.
• These predictive models can support hospital resource allocation and planning during pandemics.
• The study highlights the potential of ML in optimizing healthcare resource management during public health emergencies.