Abstracts

Rationale: For children with epilepsy, past emergency department (ED) use roughly predicts future use. The additive predictive value of insurance coverage and disease severity is underexplored. Methods: Using 2013-2014 electronic records from two centers, we conducted a retrospective cohort study to predict ED use. We used logistic regression, benchmarked against machine learning algorithms. For robustness, we fit models on half the data, and evaluated performance on the other half, 100 times. We tested generalizability with statewide administrative data from California and New York from 2010-11. We estimated the break-even cost of an intervention to reduce ED visits by 10% among the 10% highest risk individuals, using median ED and inpatient reimbursements from one center. Results: A three-concept prediction rule (prior ED use, insurance status, number of anti-epileptic drugs) predicted frequent ED use with performance similar to machine learning (center1, N=2730: median AUC 0.82 [interquartile-interval 0.8?"0.82] vs. best machine learning algorithm 0.84 [0.83?"0.85]; center2, N=786: 0.75 [0.73-0.76] vs. 0.73 [0.72-0.75]). A two-concept model (prior ED use, insurance status) also performed well (center1: 0.81 [0.79?"0.81], center2: 0.74 [0.72?"0.75], NY: N=11,711, AUC=0.72; CA: N=12,384, AUC=0.70). (Table). Estimated yearly per-patient break-even intervention costs ranged from $51-246 if only ED discharges are preventable, and $734-3543 if ED-to-inpatient admissions are also preventable. For small potential cohorts (20-50) at one center, the three-concept model meaningfully outperformed the two-concept model. (Figure). Conclusions: Prior ED use and insurance status accurately predict future use among children with epilepsy in multiple datasets. Consideration of disease severity may meaningfully improve predictions in some circumstances. Funding: This work was funded by Pediatric Epilepsy Research Foundation (PERF) and Empire Clinical Research Investigators Program (ECRIP)