Western Journal of Emergency Medicine: Integrating Emergency Care with Population Health

Introduction: Emergency department (ED) crowding harms patient care and strains hospital operations. Forecasting daily patient arrivals can help ED medical directors, department chairs, and hospital administrators plan staffing and bed capacity in advance, rather than reacting after a surge has already occurred. This study compared five computer-based forecasting methods, including three machine learning models (eXtreme gradient boosting [XGBoost], light gradient boosting machine [LGBM], and long short-term memory [LSTM]) and two traditional statistical models (autoregressive integrated moving average [ARIMA] and seasonal ARIMA [SARIMA]), to determine which most accurately predicts ED patient volume seven days in advance, using eight years of daily visit data from an academic freestanding ED.

Methods: This retrospective study analyzed 3,202 consecutive days of daily patient volume (March 27, 2017, to December 31, 2025) from an academic freestanding ED with approximately 31,102 annual visits. Each model used the preceding 365 days of data to forecast the subsequent 7 days, repeated across the study period to reflect the weekly cycle by which most EDs plan staffing. Our primary measure of accuracy was mean absolute error (MAE), defined as the average number of patients per day by which a forecast missed the actual volume. Secondary measures included root mean squared error (RMSE), mean absolute percentage error (MAPE), the coefficient of determination (R², a score from 0 to 1 reflecting how well predictions tracked actual day-to-day volume), and the percentage of forecasts accurate within 5, 10, 15, and 20 patients, thresholds selected for their relevance to staffing decisions. The best-performing model on initial metrics, XGBoost, served as the reference model for statistical comparison using the Diebold-Mariano (DM) test, which determines whether an observed difference in accuracy between two models is statistically meaningful or attributable to chance.

Results: Among 453 rolling weekly forecast windows, SARIMA, XGBoost, and LGBM demonstrated the lowest absolute errors and best forecasting of patient volume, with MAE values of 9.60 (95% CI, 9.29–9.91), 9.61 (95% CI, 9.31–9.91), and 9.64 (95% CI, 9.33–9.94) patients/day, respectively; these three models therefore missed the actual daily patient count by approximately 9 to 10 patients, on average. Models ARIMA and LSTM were less accurate, with mean MAE values of 10.54 (95% CI, 10.21–10.87) and 12.01 (95% CI, 11.63–12.39) patients per day. Relative to the XGBoost reference (MAPE 13.82%), ARIMA (MAPE 14.93%) and LSTM (MAPE 17.27%) showed a marked decrease in predictive accuracy. For operational planning, the top-performing models (SARIMA, XGBoost, and LGBM) correctly predicted daily volume within a margin of 20 patients in 90.3% to 90.6% of instances, while LSTM performed less well and reached this threshold in only 83.3% of forecasts. Statistical testing confirmed these differences were unlikely due to chance, with XGBoost significantly outperforming ARIMA (DM = −6.09, P < .001) and LSTM (DM = −6.96, P < .001).

Conclusion: Models SARIMA, XGBoost, and LGBM performed comparably well, and all three would provide reliable, operationally useful forecasts for ED staffing and resource allocation decisions. A traditional statistical method, SARIMA, which is simpler and less resource-intensive to implement, performed as well as the newer machine learning methods evaluated in this study. This finding suggests that hospitals do not necessarily need to invest in more complex artificial intelligence tools to obtain accurate ED volume forecasts.