Prediction of patient flow in the emergency department using explainable artificial intelligence

Abstract

Introduction Overcrowding in emergency departments (ED) is a significant problem affecting patient outcomes, hospital length of stay, and staff job satisfaction. This issue often stems from unpredictable patient flow and suboptimal resource allocation. Objectives This study aims to develop two machine learning (ML) models to assist in early and accurate resource allocation in EDs. The first model predicts patient admission at the time of triage, while the second predicts the specialty of care needed indicated by the initial ward transfer. Methods The study leverages the Medical Information Mart for Intensive Care (MIMIC-IV) database with 425,000 ED visits including basic vital signs, medications, presentation information, diagnoses, and demographic information about the patients. Ensemble tree classifiers are employed for model development, and model's explainability is assessed by investigating feature importance. The best model is selected based on the balance between performance and explainability. Features’ importances are calculated and presented using SHapley Additive exPlanations and models’ intrinsic feature importance. Results The best-balanced admission prediction model in terms of classification performance and explainability achieved an accuracy of 0.775 and an area under the receiver operating curve (AUROC) of 0.779 by using eXtreme Gradient Boosting (XGBoost). The resource allocation prediction model, using a one-vs-rest approach, attained an AUROC of 0.783 again by using XGBoost. The models shared acuity and age in the three most important features, whereas admission ratio and gender were the additional features for admission prediction and resource prediction, respectively. Conclusion The study successfully demonstrates the potential of ML models in predicting patient admission and required specialty care at the ED triage stage. While the admission prediction model shows moderate performance compared to existing studies, the resource prediction model exhibits superior performance compared to related works. The research highlights the importance of explainability in ML models, suggesting the need for further practical implementation to refine and validate these models in real-world settings.