An Artificial Intelligence-Based Framework for Predicting Emergency Department Overcrowding: Development and...

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Operations Research

Background:

• Emergency department (ED) overcrowding is a persistent challenge impacting patient care and hospital efficiency.
• Current reactive management strategies are insufficient for effective patient flow.
• Machine learning (ML) offers predictive capabilities for proactive interventions.

Purpose of the Study:

• Develop ML models for predicting ED waiting room occupancy (waiting count) at hourly and daily resolutions.
• Enable proactive resource allocation and mitigation of ED overcrowding.
• Forecast waiting counts 6 hours ahead (hourly) and average daily waiting counts.

Main Methods:

• Utilized integrated internal and external data from a southeastern US hospital ED.
• Trained and evaluated eleven ML algorithms, including traditional and deep learning approaches.
• Optimized feature combinations and assessed model accuracy under various conditions.

Main Results:

• Time series vision transformer plus (TSiTPlus) achieved the best hourly prediction (MAE 4.19).
• Explainable convolutional neural network plus (XCMPlus) yielded the best daily prediction (MAE 2.00).
• Both models outperformed traditional forecasting methods.

Conclusions:

• Developed effective ML models for forecasting ED waiting counts at hourly and daily intervals.
• Demonstrated the value of diverse data integration and advanced modeling for proactive hospital management.
• These tools can improve patient flow and reduce ED overcrowding.