Abstracts

Rationale: Frequent ED use by children with epilepsy indicates poor access to care and/or poor seizure control. Accurate predictions of future ED use can identify individuals for intensive outpatient services, such as care management. We developed and validated a predictive algorithm to identify children at risk for 4 or more ED visits per year.Methods: Using a retrospective cohort design, we used data from one year to identify children at risk for 4+ ED visits in the subsequent year. We developed the algorithm using electronic health records from a single center (Nationwide Children’s Hospital; NCH) from 2013 - 2014 (""EHR""). We assembled 100+ variables, reviewed results from ten regression variants and machine learning tools, and selected a model that balanced parsimony, interpretability, and accuracy. We selected a threshold for the model that would identify approximately 200 patients for enrollment in a care management intervention at NCH, based on the operational capacity of this program. We validated the final model, with this threshold, on two additional data sources: health information exchange data from New York City 2009 - 2010 (""HIE"") and a pediatric Accountable Care Organization in Central Ohio 2010 - 2011 (""ACO"").Results: The EHR data contained 2063 children. Bivariate analyses found multiple potential predictors of ED use: younger age, male, Medicaid, frequent use of health services, medical complexity, several comorbidities and AEDs, and home zip code characteristics. We selected a two variable model: the PEER (Pediatric Epilepsy Emergency Room) Score is the sum of the number of ED visits and the number of head CTs in one year. In the EHR data, the PEER Score predicted frequent ED use with similar performance to other regression variants and machine learning algorithms. A PEER score of 3 identified an appropriate size cohort (slightly fewer than 200 patients). (Table 1). Among 2063 children in the EHR data, 181 (8.7%) had a PEER score ≥ 3 in the first year. In the following year, 138 of these 181 (i.e. 76%) visited the ED at least once, and 53 (29%) visited 4 or more times. We validated the algorithm in groups of 2312 children in the HIE data and 1996 children in the ACO data. A PEER score of ≥ 3 identified 12-21% of children in these two cohorts. Among identified children, 59-60% used the ED at least once in the subsequent year, and 15-21% used the ED 4 or more times. (Table 2)Conclusions: Using data from an EHR, the PEER score identifies children at high risk for future frequent ED use, and performs as well as several machine learning algorithms. The PEER score successfully identifies at risk children in other datasets, though with reduced accuracy. Ongoing work is focused on improving predictive accuracy by building models tailored to each dataset.