Forecasting Pediatric Emergency Department Arrivals: Evaluating the Role of Exogenous Variables Using Deep Learning Models
Abstract
Background Forecasting pediatric emergency department (ED) demand remains a critical challenge in healthcare operations. This study aimed to identify exogenous variables influencing pediatric ED visits and evaluate the performance of different forecasting models. Method Using a retrospective observational design, we analyzed 192,347 pediatric ED visits across nine hospitals in Southeast Michigan between 2017 and 2019. Patient data were aggregated into daily arrival counts and enriched with exogenous variables such as weather, air quality, pollen, calendar, Google search trends, and chief complaints. Feature selection was performed using XGBoost and SHapley Additive exPlanations to identify the most influential predictors. Three forecasting models were developed: a Naïve baseline, Long Short-Term Memory (LSTM), and an attention-based neural network. The models were evaluated across 1-day, 7-day, and 14-day forecasting horizons using mean absolute percentage error (MAPE) and R2 metrics. Results LSTM and attention-based model significantly outperformed the Naïve baseline across all horizons. The LSTM model incorporating calendar data achieved the best 1-day forecast (MAPE: 8.71 %, R2: 0.67). For 7-day forecasts, the attention-based model using chief complaint data performed best (MAPE: 9.18 %, R2: 0.57). At 14 days, the attention-based model without exogenous inputs outperformed most LSTM variants, reflecting superior performance in long-range forecasting. Among exogenous variables, calendar and chief complaint data added the most predictive value, while Google Trends and pollen data introduced noise and diminished model performance. Conclusion Combining deep learning architectures with selected external data improves pediatric ED arrival forecasting. From an operational perspective, such forecasts can support more efficient staffing, reduce wait times, and mitigate ED crowding.
