MACHINE LEARNING FOR RESOURCE ALLOCATION IN EMERGENCY DEPARTMENT

Abstract

Emergency Departments (EDs) face constant pressure from unpredictable patient arrivals, limited beds, and staffing challenges. These issues often lead to longer ED stays, higher costs, and clinician burnout. While machine learning (ML) is widely used in healthcare, it has rarely been applied to real-time staffing and resource allocation in the ED.The objective of this research is to (1) identify key factors that predict prolonged ED length of stay (EDLOS >6 hours), (2) simulate nurse staffing needs under different patient demand scenarios, and (3) forecast daily admissions with time-series models to support short-term planning. In this study almost 40,000 ED patient visits were analyzed, including demographics, vitals, acuity, primary diagnosis, and staffing levels. Most patients were between 50 to 80 years of age, 40% were African American and retired, and more than half were on Medicare. Common risks included obesity and smoking with higher acuity (2 & 3). Top admitting diagnoses were related to trauma, gastrointestinal and neurological, and top admitting units were internal medicine and surgery. Three supervised Machine Learning models, Logistic Regression, Support Vector Machine and CHAID Decision Tree were tested on 4 sets of data cohort to predict which patients would stay longer than 6 hours. Logistic regression gave balanced but moderate accuracy, SVM achieved very high accuracy (over 95%) on training set but moderate on largest data set, while CHAID decision trees provided easier-to-interpret rules with reasonable performance. Together, these models show the trade-off between predictive power and interpretability. Acuity and time series variables showed important predictors for longer EDLOS. We also performed what-if simulations on different staffing scenarios and analyzed EDLOS in different scenarios such as reducing staff, high patient volume, high acuity, and above baseline RN staffing. A time series ARIMA forecasting model was also built to project daily admissions to 30-day prediction. When compared with actual data, it showed reliable predictions with 5% margin of error. The findings of this study offer a practical decision-support framework by combining prediction, simulation, and forecasting. The implications are broadly valuable such as hospitals can cut costs by reducing overtime & inefficient staffing, HR can improve nurse retention by balancing workloads, and patients benefit from shorter wait times and smoother care transitions. Beyond the ED, this approach is adaptable to other hospital units, making it a scalable solution that supports efficiency, staff well-being, and better clinical outcomes.