Daad M. Alhassan, Hala M. Alsuabeyl, Rana A. Almashari, Rand A. Altareefi, Layan A. AlRushaid and Fahima Hajjej, Princess Nourah bint Abdulrahman University (PNU), Saudi Arabia

Emergency department (ED) overcrowding is a persistent operational challenge that degrades patient care quality, lengthens waiting times, and strains hospital resources. Accurate forecasting of patient arrivals enables proactive staffing, bed allocation, and capacity planning. This paper presents a comparative study of machine learning and deep learning models for forecasting daily ED visit volume, developed within a modified CRISP-DM framework. Using de-identified records from the MIMIC-IV-ED (v2.2) database enriched with historical weather observations and calendar features, we engineer temporal predictors and evaluate five models: a Long Short-Term Memory (LSTM) network, a Random Forest Regressor (RFR), Extreme Gradient Boosting (XGBoost), Neural Basis Expansion Analysis for Time Series (N-BEATS), and the Temporal Fusion Transformer (TFT). Models are assessed with Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) on a chronologically held-out test set. The TFT substantially outperformed all baselines, achieving an MAE of 4.46, an RMSE of 5.02, and a MAPE of 41.5%, compared with MAE values of 12.8-14.2 for the remaining models. To promote transparency, SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP), and the TFT attention weights are used to interpret model behaviour. The best model is embedded in an interactive forecasting dashboard, "Marsad," that produces daily and short-horizon hourly forecasts to support proactive ED operations in alignment with Saudi Vision 2030.

Deep Learning, Emergency Department Forecasting, Explainable AI, Healthcare Operations, Machine Learning, Temporal Fusion Transformer, Time-series Prediction