Forecasting daily patient arrivals at an Emergency Department of Specified Academic Hospital

Abstract

The hospital Emergency Department (ED) has become the main point of entry for patients in modern hospitals, resulting in frequent overcrowding; as a result, hospital management is increasingly paying attention to the ED to provide better quality service to patients. This study seeks to build time series (Autoregressive Integrated Moving Average) and machine learning (XGBoost, Gradient Boosting Regressor and Voting Regressor) regressor models, evaluate the performance of each and use the best model to forecast daily attendance. A comprehensive analysis of data related to patient arrivals at a hospital, focusing on different times of day is performed. The study was conducted in the Emergency Department of a specified South African public hospital. A dataset of patient arrivals from May 2019 to November 2021 has been collected, with a total of 47 461 observations used for the analysis. A time series model and three machine learning regressor models were investigated. Detailed statistical and exploratory analyses, time series plots, model training, and model validation efforts are carried out. The study delves into various aspects such as stationarity testing, normality testing, and the use of different transformation methods to achieve stationarity. Machine Learning algorithms are employed, with a hyperparameter tuning phase to obtain optimal coefficients. The evaluation matrices Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE) and Mean Percentage Difference (MPD). Lastly, the chosen model is used to forecast Normal Hours and After Hours. The Voting Regressor emerged as the most reliable, showing consistent performance across both training and test datasets, whereas models like ARIMA and XGBoost struggled with autocorrelation issues and peak predictions, respectively. Overall, while the Gradient Boosting Regressor performed well on training data, it exhibited potential overfitting, suggesting the Voting Regressor as the preferable model for handling the complex patterns of patient arrivals.