Predicting Future Frequent Users of Emergency Departments in California State

A large percentage of emergency department (ED) visits originates from a small percentage of patients who keep returning to the ED. Being able to flag these frequent users in advance can help clinicians to take appropriate interventions to reduce the number of ED visits, thereby reducing cost and improving quality of care. In this paper we present machine learning models that can predict future ED utilization of individual patients, using only information from the present and the past. We train decision trees (DT), boosted decision trees (AdaBoost) and logistic regression (LR) models on discharge records from California-licensed hospitals from the years 2009 and 2010, and evaluate their predictive accuracy for the years 2011-2013. We also study the impact of including different groups of demographic, frequency of ED visits, distance to emergency department, and clinical features on the accuracy of our predictive models. Overall there are three key findings of this study. First, all three techniques (LR, DT and AdaBoost) have a strong predictive ability to discriminate frequent ED users (number of visits ≥ 5). Second, our models show consistent outcomes across all three test years in our dataset, which is a desired property when a predictive tool that is stable and consistent year over year is required. Third, least and most frequent ED users are comparatively easier to predict when compared to moderate ED users (with higher sensitivity and AUC scores).