Machine Learning for Queue Prioritization: Applications to the Emergency Department

Queue prioritization is a common practice that allocates limited resources to heterogeneous customers to improve operational outcomes and customer satisfaction in service systems such as call centers and emergency departments.The first part of the dissertation studies a queue prioritization problem with two customer types under imperfect information. The service provider uses a binary classification model to estimate the probability of being a high-importance customer upon a customer’s arrival. If the likelihood is above a certain threshold, the customer is provided with priority service, which is faster and not too much more variable than non-priority service. The service provider wants to minimize the average waiting costs by selecting the optimal threshold. The ROC curve shows the performance of the binary classification algorithm in terms of sensitivity and specificity at various thresholds. Changing the threshold usually impacts the classification algorithm's sensitivity and specificity in opposite directions. The traditional threshold selection method tends to optimize a ROC curve-based metric and does not consider the operational externalities. This dissertation analyzes the optimal threshold policy in terms of the ROC curve, i.e., sensitivity and specificity, by considering the operational nature of the service systems. We find that optimal policy trades a loss in specificity for a higher gain in sensitivity. The second part of the dissertation is an empirical study on queue prioritization where customer priority depends on other customers' characteristics and system attributes, focusing on its application to the Emergency Department at the University of Chicago Medicine (UCM). We model the patient prioritization problem using a discrete choice framework and implement a tree-based segmentation algorithm that generates ED system clusters where a similar patient prioritization rule is observed. We find that room type, waiting room census, and time of the day are the most important system-level attributes; acuity and waiting time are the most important patient-level attributes for patient prioritization. High acuity patients are prioritized for the primary service area, while low acuity patients are prioritized for the fast-track area in the ED. The First-Come-First-Served principle is generally followed within the same acuity class. As the waiting room gets crowded and resource utilization increases, the adherence to acuity-based prioritization increases. Finally, we develop a tree-based segmentation algorithm that creates patient clusters and incorporates the cluster membership in the discrete choice model to capture the patient-level nonlinear and interaction effects