Ascertaining cause-specific emergency department demand using forecast combinations

Emergency department (ED) overcrowding is a recurrent public health concern, which may be alleviated by forecasting ED visits. Numerous ED forecasting models exist, making it challenging for decision makers to select appropriate forecasts and plan under model and forecast uncertainty. It remains unclear whether incorporating environmental covariates as model predictors or aggregating cause-specific ED visit forecasts to obtain total ED visit forecasts can each contribute to improving forecast accuracy. To address this gap, we developed a framework to generate accurate probabilistic forecasts using forecast combination schemes. We developed probabilistic forecast combinations schemes which can directly combine predictive distributions/quantiles in a linear/non-linear fashion, using static/dynamic weighting schemes. These schemes were tested in probabilistically forecasting cause-specific ED visits in Singapore, an equatorial city state. We incorporated a high-dimensional set of predictors to further augment model performance. We documented the forecast performance of all forecasting models across 1- to 12-week ahead horizons. Our study showed that aggregating cause-specific forecasts to provide all-cause ED visit forecasts can enhance overall ED visit forecasting performance—particularly for KNN, XGBoost, and Elastic Net. Forecast combinations, in particular, the linear opinion pool can lead to excellent and stable forecasts over all individual forecasts in 164 out of the 180 cause–horizon combinations examined. Within the considered forecast horizons, longer horizons led to more scenarios where forecast combinations outperformed individual models significantly (p < 0.05). Forecast combinations exhibited stable performance across forecast horizons, whereas individual models showed greater variability—some performed well at shorter horizons but deteriorated at longer ones, and vice versa. We also found that quantile forecast combinations can generate confident forecasts while maintaining good accuracy. Forecasting cause-specific ED visits can provide fine-scale forward guidance on resource optimization and ED crowding preparedness. Probabilistic forecast combinations can characterize the uncertainty of forecasts and hedge against model selection uncertainty in a robust manner. However, performance during COVID-19 was not assessed, which may limit generalizability under structural breaks.