Predicting Emergency Department Patient Arrivals at Hospitals Using Machine Learning Techniques

Area of Science:

• Healthcare Operations Research
• Machine Learning Applications
• Time Series Forecasting

Background:

• Emergency Departments (EDs) experience significant operational challenges due to unpredictable patient volumes and overcrowding.
• Accurate short-term demand forecasting is essential for efficient ED staffing and resource allocation.
• Minimizing patient entry delays requires precise hourly arrival predictions.

Purpose of the Study:

• To develop and evaluate a forecasting framework for hourly ED patient arrivals.
• To compare the performance of six different forecasting models, including machine learning approaches.
• To assess the impact of engineered temporal features on prediction accuracy.

Main Methods:

• Utilized de-identified hourly patient arrival data from an ED in Madinah, Saudi Arabia (January-November 2024).
• Engineered 183 features including time encodings, holiday indicators, autoregressive lags, and volatility measures.
• Compared Seasonal Naive, ETS, Ridge Regression, LightGBM, TCN, and LSTM models using Bayesian optimization and an asymmetric loss function.

Main Results:

• Ridge Regression yielded the lowest Mean Absolute Error (MAE) at 3.75 (R² = 0.52).
• Hybrid TCN and LSTM models showed comparable performance (MAE 3.80 and 3.85).
• All evaluated machine learning models significantly outperformed baseline methods (ETS, Seasonal Naive).

Conclusions:

• Combining engineered temporal features with linear or sequence models provides accurate hourly ED arrival forecasts.
• The operational sufficiency of forecast accuracy for staffing requires site-specific clinical validation.
• Further research is needed to integrate these forecasting tools into real-world ED operations.