Abstracts

Rationale: People with epilepsy who frequently use the emergency department (ED) may have either poor disease control or poor access to care, and thus may benefit from intensive case management services. However, they may also seek care at multiple hospitals, complicating efforts to coordinate their care. Health information exchange (HIE) is technology that connects electronic data from clinical information systems at different hospitals. Connecting data from multiple sites may make it easier to prospectively identify frequent ED users, but the predictive accuracy of this approach is unknown. Here, we report the performance of a statistical model to predict frequent ED users based on HIE data.Methods: We conducted a retrospective cohort study to create and evaluate a predictive model to identify frequent ED users among people with epilepsy. We used data from an HIE organization linking seven academic hospitals in Manhattan, NY. We identified people with epilepsy via the consensus definition of probable epilepsy : a single ICD9 code of 345.x (epilepsy), or two ICD9 codes of 780.9 (convulsion) on different days. The dataset contained two consecutive years of history for 8198 people with epilepsy. For predictors, we included seven utilization variables (number of ED visits, inpatient visits, outpatient visits, days with any radiology studies, head CTs, brain MRIs, and other brain imaging studies), three demographics variables (home zip code [10 categories], age [9 categories], and gender), and 33 binary comorbidity variables (Jette 2011). For the outcome, we defined the binary variable, frequent ED use, as four or more visits in a year. We measured the predictor variables over one year, and the outcome variable in the subsequent year. We used logistic regression as the prediction algorithm. We performed model selection with the lasso technique. We used 10-fold cross validation to select the model that yielded the largest mean AUC (area under the ROC curve). We examined the accuracy of the model by assessing ED use (in year two) of 100 individuals chosen by the model to be the most likely frequent ED users.Results: 491 of the cohort of 8198 people with epilepsy used the emergency room four or more times in year two. The lasso technique selected eight variables for the model: ED visits, inpatient visits, outpatient visits, living in Manhattan (i.e. geographically close to the hospitals), alcohol abuse, depression, drug use, and pulmonary disease. With this model, the AUC was 0.89, indicating very good predictive accuracy. Among the 100 individuals proposed by the model to be the mostly likely frequent ED users in year two, 76 were correctly identified (four or more visits), 8 used the ED one to three times, and 16 did not use the ED at all.Conclusions: Health information exchange has the potential to support prospective identification of future frequent ED users, among people with epilepsy, with very good predictive accuracy. Care coordination efforts based on these predictions would target individuals at high risk for chronic ED use.