Patients visits forecasting in emergency departments: Self-adapting LSTM models for evolving data distributions

Area of Science:

• Healthcare Operations Research
• Artificial Intelligence in Medicine
• Time Series Forecasting

Background:

• Emergency department (ED) overcrowding is a persistent healthcare challenge.
• Accurate patient visit forecasting is crucial for resource optimization.
• Traditional models struggle with abrupt data distribution shifts, necessitating manual retraining.

Purpose of the Study:

• To develop a deep learning forecasting framework for ED visits.
• To ensure forecasting accuracy under evolving conditions without manual intervention.
• To create a robust system for dynamic healthcare environments.

Main Methods:

• Proposed an adaptive forecasting approach using a Seq2Seq LSTM architecture.
• Incorporated Per-Sequence Scaling (PSS) and Continual Learning (CL) for adaptive updates.
• Evaluated against LSTM, ARIMA, and Prophet using historical ED visit data across different phases (pre-COVID, COVID, post-COVID).

Main Results:

• The adaptive LSTM significantly reduced forecasting errors on short-to-medium horizons (1-day, 1-week) compared to baselines.
• Achieved consistent Mean Absolute Scaled Error (MASE) values below 1.
• Demonstrated robustness by continuously adjusting to new data distributions without full retraining, outperforming other models in post-COVID conditions.

Conclusions:

• The adaptive Seq2Seq LSTM framework provides a robust solution for ED forecasting in dynamic environments.
• It offers operational viability, outperforming Prophet and ARIMA once data distributions stabilize.
• The model's incremental self-adjustment capability is ideal for real-time hospital deployment under privacy and IT constraints.