Predicting Frequent Emergency Department Visits Among Children With Asthma Using Electronic Health Record Data

Background: Asthma is one of the most common chronic conditions among children and is the third leading cause of pediatric hospitalization among children under age 15. Asthma-related emergency department (ED) visits are common and expensive to the health system. One of the proposed cost savings measures is for hospitals to hire care managers who can coordinate care for individuals with chronic diseases such as asthma, for example by educating families about general preventative care practices, reconciling medications, and answering questions. It would be valuable to prospectively identify children likely to be frequent ED users, in order to enroll in care management programs. It is unclear if electronic health record (EHR) data can be used to predict which patients will frequently use the ED for asthma. Objective: To explore the predictability of frequent ED use among pediatric asthma patients in New York City using data from an EHR from one medical center. Methods: We performed a literature review and interviewed 5 physicians affiliated with Weill Cornell Medical Center to generate a list of potential predictors of ED use. We operationalized a subset of these predictors from an EHR system. We performed bivariate statistics to examine the unadjusted relationship between each variable and frequent ED use. Then we evaluated and compared the performance of several machine-learning algorithms to predict which children with asthma will use the ED two or more times in the next year. The algorithms we used were: logistic regression best subsets, Least Absolute Shrinkage and Selection Operator (LASSO) regression, Random Forests, Classification and Regression Trees (CART), and Support Vector Machines (SVM). We evaluated model performance based on Area Under the Curve (AUC), positive predictive value (PPV), sensitivity, calibration, and classification error. Results: We operationalized 52 predictors. Bivariate analysis showed significant associations between many of the clinical and demographic predictors and frequent ED use. All the predictive algorithms performed similarly, with very good area under the curve (AUC) values, but poor positive predictive value and sensitivity. We selected a two variable model as our final model based on the predictors that appeared significant across the algorithms: number of ED visits in the previous year and type of insurance. Publicly insured patients with asthma who used the ED four or more times in the baseline year have a 50% or greater probability of being a frequent ED user in the following year. The same utilization pattern is seen among privately insured patients who have six or more ED visits in the baseline year. Conclusions: Children who are currently frequent users of the ED are likely to continue to do so. The threshold for identifying these children is lower among children with public insurance compared to those with private insurance. A two variable (prior ED visits and insurance status) model to predict which children with asthma will be future ED users is as accurate as predictions from several machine learning algorithms. These observations can be used to identify children with asthma who may benefit from enrollment in a care management program, using data from an EHR