Machine Learning Approaches to Predict Patient’s Length of Stay in Emergency Department

Abstract

As the COVID-19 pandemic has afflicted the globe, health systems worldwide have also been significantly affected. This pandemic has impacted many sectors, including health in the Kingdom of Jordan. Crises that put heavy pressure on the health systems’ shoulders include the emergency departments (ED), the most demanded hospital resources during normal conditions, and critical during crises. However, managing the health systems efficiently and achieving the best planning and allocation of their EDs’ resources becomes crucial to improve their capabilities to accommodate the crisis’s impact. Knowing critical factors affecting the patient length of stay prediction is critical to reducing the risks of prolonged waiting and clustering inside EDs. That is, by focusing on these factors and analyzing the effect of each. This research aims to determine the critical factors that predict the outcome: the length of stay, i.e., the predictor variables. Therefore, patients’ length of stay in EDs across waiting time duration is categorized as (low, medium, and high) using supervised machine learning (ML) approaches. Unsupervised algorithms have been applied to classify the patient’s length of stay in local EDs in the Kingdom of Jordan. The Arab Medical Centre Hospital is selected as a case study to justify the performance of the proposed ML model. Data that spans a time interval of 22 months, covering the period before and after COVID-19, is used to train the proposed feedforward network. The proposed model is compared with other ML approaches to justify its superiority. Also, comparative and correlation analyses are conducted on the considered attributes (inputs) to help classify the LOS and the patient’s length of stay in the ED. The best algorithms to be used are the trees such as the decision stump, REB tree, and Random Forest and the multilayer perceptron (with batch sizes of 50 and 0.001 learning rate) for this specific problem. Results showed better performance in terms of accuracy and easiness of implementation.