A deep learning architecture for forecasting daily emergency department visits with acuity levels

Abstract

Accurate forecasting of Emergency Department (ED) visits is important for decision-making purposes in hospitals. It helps to form tactical and operational level plans, which facilitates staff and resource allocations in advance. A dataset recording the daily visits of patients at the ED of a regional hospital over a 3-year period is used in this study. Patients are triaged into 3 acuity levels: P1, P2 and P3, with P1 being patients with severe or life threatening conditions, whereas P3 being patients with minor injuries requiring less urgent attention. A novel deep learning forecasting structure, which has advantages of both Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), is being developed and applied to forecasting daily visits (up to 56 days into the future) for the different acuity levels. The features included in this study are calendar days, public holidays, Pollution Standard Index (PSI) readings, rainfall and daily average temperature. The effectiveness of our newly developed model, in terms of forecasting accuracy, is demonstrated and compared with other deep learning models. Our model achieves mean absolute percentage errors (MAPEs) of 17.37%, 7.19%, 6.11% and 4.50% in forecasting P1, P2, P3 and total visits respectively, and has demonstrated superior performance when evaluated against state-of-the-art studies in the literature. This study illustrates that utilization of our hybrid model comprising LSTM with CNN layers can provide a significant improvement over these existing deep learning models for ED daily visits forecasting.