Predicting Emergency Department Trolley Numbers Using Interpretable Deep Learning Approaches

Emergency Department (ED) overcrowding remains a critical challenge in Ireland. University Hospital Limerick consistently reports the nation’s highest trolley counts, contributing to prolonged patient wait times, staff burden, and reduced patient flow.To address these pressures, the overarching objective of this research is to develop a suite of machine learning models capable of short-term (e.g. fortnightly) forecasting of ED trolley numbers. Four machine learning models were chosen, each selected for the unique interpretability it contributes to the forecasting process.• SARIMAX is interpretable through explicit parameters such as trend, seasonality, autoregressive and moving-average terms. The coefficients directly show how past trolley counts, seasonal cycles, and exogenous variables (e.g., surge beds, weather) influence predictions, allowing stakeholders to clearly understand why the model forecasts an increase or decrease. Its advantages include a transparent, explainable structure and the ability to easily communicate causal-like relationships.• LSTM brings interpretability by capturing long-term dependencies, enabling the model to learn how earlier patterns influence future trolley counts.• Transformer model, attention scores reveal which past time steps or covariates the modelfocuses on when making predictions, providing insights into the temporal importance of theinput features.• The Temporal Fusion Transformer provides multiple built-in interpretability mechanisms, including a Variable Selection Network that highlights the most influential covariates, Temporal Attention that shows which past days the model relies on, and quantile forecasts that communicate uncertainty in predictions. Its advantages include clear visibility into which features drive predictions and uncertainty estimates that are valuable for operationalplanning.