Prediction-Driven Staffing for Emergency Departments: What to Predict and How to Predict

We conduct a comparative study of prediction-driven strategies for staffing hospital emergency departments (ED). We evaluate three approaches: (i) a machine learning (ML) approach that relies on census forecasts and applies a straightforward patient-to-nurse ratio to determine staffing; (ii) a staffing-level informed machine learning (SIML) approach that models the mapping from staffing levels to congestion outcomes and chooses the staffing plan that minimizes the associated cost; and (iii) a queueing-informed (QI) approach that leverages a calibrated queueing model to guide staffing decisions. We evaluate the three approaches using real ED arrival patterns. ML, which overlooks the endogeneity of queueing dynamics, can suffer from varying degrees of delayed feedback. SIML performs well when training and evaluation conditions align, but can be sensitive to distribution shifts. QI typically achieves the best results under correct model specification, though it is vulnerable to misspecification, for which we provide a diagnostic tool. Finally, we offer practical guidance to help hospitals select the most suitable approach given their data and modeling expertise.