Machine Learning for Intensive Care Unit Length-of-Stay Prediction: A Simulation-Based Approach to Bed Capacity...

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Operations Research

Background:

• Machine learning (ML) models are increasingly used for healthcare predictions, but their operational impact on resource management, like intensive care unit (ICU) bed capacity, is under-researched.
• Traditional ML evaluation metrics often lack practical insights for healthcare decision-makers regarding resource allocation.
• Understanding the real-world implications of ML predictions for operational efficiency is crucial for healthcare management.

Purpose of the Study:

• To evaluate the impact of ML-driven length-of-stay (LOS) predictions on ICU bed capacity management.
• To compare the performance of different ML models (logistic regression and XGBoost) in a simulated ICU environment.
• To assess the practical utility of ML models beyond traditional performance measures for clinical decision support.

Main Methods:

• A simulation study was conducted using the HiRID dataset, comprising high-frequency data from over 33,000 patients.
• Two classification models, logistic regression (LR) and extreme gradient boosting (XGB), were applied to predict ICU LOS.
• ML model predictions were integrated into a simulation framework replicating real-world ICU bed management to assess practical implications.

Main Results:

• Both ML models improved ICU capacity control compared to baseline scenarios.
• XGBoost demonstrated superior performance over LR in the simulation, leading to slight underoccupancy.
• LR resulted in slight overoccupancy, highlighting the nuanced impact of different ML models on bed management.

Conclusions:

• Evaluating ML models within the context of specific healthcare operations, like ICU capacity management, is essential for practical application.
• Simulation-based approaches provide more relevant insights for healthcare practitioners than traditional performance metrics.
• This study bridges the gap between ML predictive accuracy and actionable clinical decision support for efficient resource management.