Predicting the Unpredictable: Comparing Statistical Forecasting and Deep Learning Models for Forecasting Emergency Department Arrivals

Abstract

The hospital emergency department must be continuously staffed and ready to care for patients. Staffing and demand could be more easily matched if hospitals were able to better predict surges in emergency department visits. In this thesis, we examine methods for the development of forecasting models including feature engineering and model selection. Deep learning-based forecasting methods are compared to traditional statistical methods to predict emergency room arrivals using emergency room time series data over 48 months period. Specifically, we examine RNNs and LSTM architectures and compare them to ARIMA models. Each of the two approaches has tradeoffs in terms of complexity, flexibility and the ability to model external events. We discuss these tradeoffs as they pertain to predicting demand and how they affect the operationalization of a model in an emergency room setting.