Predicting Patient Waiting Times at an On-Call Emergency Line at a Dutch GP Office

Abstract

A challenge is posed by the prediction of patient waiting times at an emergency call line operated by a GP office in the Netherlands. The office currently employs a prediction approach based on a Discrete Event Simulation model, which has not been compared to different approaches. Previous studies suggest that ARIMA-family models and LSTM might perform well in predicting patient waiting times. We select the most accurate model among ARIMA, SARIMA and SARIMAX models for this particular case and compare it to the current simulation-based model. We also compare the most accurate ARIMA-family model with LSTM in terms of their prediction accuracy of patient waiting time prediction for this particular case. The data used in this thesis were provided by an external company and they span from January 2022 to April 2023. The time series is analyzed using an hourly frequency and the predicted outcome is the average patient waiting time per hour measured in minutes. Box-Jenkins method and external forecast validation are applied in modelling and selection of the most accurate model for forecasting. The most accurate among the ARIMA-family models is the SARIMAX model. This model uses exogenous variables: calendar variables and number of incoming calls per hour to model and predict the average patient waiting time per hour. The SARIMAX model with incoming calls is also more accurate than the current simulation model for short- and long-range predictions. However, LSTM has better prediction accuracy for this case than any ARIMA-family model. The implementation of LSTM is recommended for this case. We also provide tentative results regarding the effect of the number of staff available for patient waiting time predictions, and we suggest that it is investigated in greater detail and with more data available in future work.