Forecasting emergency department overcrowding using the CEDOCS score - predictable, but not actionable?

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Health Informatics

Background:

• Emergency department (ED) overcrowding is a significant challenge impacting patient care and operational efficiency.
• The Clinical Conditions and Emergency Department Operations Control System (CEDOCS) score is a metric used to assess ED crowding.
• Predictive modeling offers a potential solution for anticipating and mitigating ED overcrowding.

Purpose of the Study:

• To develop an explainable machine learning (ML) model for predicting ED overcrowding using the CEDOCS score.
• To evaluate the actionability and clinical utility of the developed ML model in a real-world setting.

Main Methods:

• Retrospective analysis of 300,000 ED visits across two German hospitals (2019-2023).
• Utilized electronic health records (EHR), calendar, and ambulance data to predict CEDOCS scores.
• Employed SHAP values for model explainability and conducted a six-month intervention study to assess clinical impact.

Main Results:

• The ML models achieved strong predictive performance, with low Root Mean Square Errors (RMSE) for one-, two-, and three-hour forecasts.
• Key predictors of overcrowding included the current CEDOCS score, time of day, and intensive care unit (ICU) ambulance diversions.
• Clinician stress was moderately associated with CEDOCS, but actions taken were not directly linked to predictions or model error, rather to perceived tool accuracy.

Conclusions:

• The developed ML model shows robust predictive capabilities for ED overcrowding based on the CEDOCS score.
• Explainability analysis (SHAP) confirmed known drivers of overcrowding.
• Accurate CEDOCS prediction alone may not suffice for effective clinical decision support; operational demands beyond occupancy metrics are crucial for actionable insights.