The following is a summary of “Predicting daily emergency department visits using machine learning could increase accuracy,” published in the March 2023 issue of Emergency Medicine by Pappas, et al.
Emergency department visits have been predicted for many years by administrators and physicians alike. Predicting or “forecasting” the amount of ED visits can make it possible to allocate resources more effectively, increasing or downsizing staff, altering OR scheduling, and anticipating the need for considerable resources. For a study, researchers sought to use machine learning to investigate combinations of variables in order to improve prediction accuracy and identify the elements that most closely predict overall ED visits. They expected that machine learning models would be more accurate in predicting St. Joseph Mercy Ann Arbor’s patient visit load for the emergency department (ED) than a straightforward univariate time series model.
As a benchmark for comparison, univariate time series models for daily ED visits, such as ARIMA, Exponential Smoothing (ETS), and Facebook Inc.’s prophet algorithm, were calculated. Data from 2017 and 2018 were used to train machine learning models, such as random forests and gradient-boosted machines (GBM). To assess how effectively the final models predicted real ED patient volumes in data not used during the model fitting process, they were applied to the 2019 data after being produced. Out-of-sample predictive accuracy was used to gauge the accuracy of the machine learning and time series models, and root mean squared error (RMSE) was used to compare them.
The results revealed that the random forest model was the most accurate at predicting daily ED visits in the 2019 test set, followed by the GBM model, using root mean squared error (RMSE) to measure the out-of-sample predictive accuracy of the models. These just marginally outperformed the predictions of the straightforward exponential smoothing model. In contrast, the ARIMA model fared badly. The day of the week was shown to be the most significant predictor of patient volumes (presumably reflecting variations between weekdays and weekends). Some of the seasonality tendencies connected to variations in patient numbers appeared to be captured by weather-related characteristics like maximum temperature and SFC pressure.
Though not significantly, machine learning models outperformed straightforward univariate time series models in forecasting daily patient volumes. The sparse preliminary findings can be supported by more studies. Additional feature engineering and data collection might be helpful for training the models and perhaps enhancing forecast accuracy.