Can big data help us close an epilepsy care gap?

Epilepsy has been termed an “ambulatory care sensitive condition,” meaning that high-quality outpatient epilepsy care can reduce unnecessary emergency department (ED) visits and hospitalizations. Using this framework, high ED use is a marker of poor disease control or limited access to care. Do all people with epilepsy (PWE) with poorly controlled disease or limited access frequent the ED? Are other factors at play? In this changing health care landscape, with accountable care organizations and a need to reduce costs, can we predict PWE who are at high risk of frequent ED use? In this issue of Neurology®, Grinspan et al. predicted frequent ED use in PWE through an innovative use of big data. Using a retrospective cohort analysis of administrative data from the New York Clinical Information Exchange (NYCLIX), a health information exchange network in New York City, the authors analyzed the predictors of frequent ED use in 8,041 PWE over 2 years. In bivariate analysis, adults (vs children and seniors), men, Manhattan residents (vs non-Manhattan), frequent users of health services, users of multiple health systems, and those with comorbidities in year 1 were more likely to have frequent ED use (41 visits) in year 2 of the study. Despite examining numerous statistical models to predict frequent ED use in year 2, the authors found that a simple one-variable logistic regression model, based on the number of ED visits in year 1, performed comparably. Simply put, baseline ED use predicted follow-up ED use. PWE with heavy ED use (111 visits) in year 1 had a $50% probability of frequent ED use in year 2. The authors predicted future frequent ED visits with high accuracy (area under the receiver operating curve .0.85, positive predictive value .70%) but with poor sensitivity (approximately 20%). Although easy to identify a group of PWE at high risk of frequent ED use, it is still difficult to predict ED use for most PWE. Innovatively, the authors used big data to identify PWE at high risk of frequent ED use. By linking patient records from 7 of 13 Manhattan emergency departments, the authors were able to evaluate twothirds of Manhattan’s yearly ED visits. By connecting unaffiliated institutions, the authors determined that PWE who used multiple health systems (vs one system) in year 1 were more frequent ED users in year 2. Because PWE can have seizures at the most inopportune times, it is not surprising that some PWE use multiple different unaffiliated EDs. Although comprehensively examining numerous predictive models, a simple one-variable logistic regression model was the most parsimonious. Unfortunately, big datasets often come with missing information. The authors identified PWE using a “probable epilepsy” algorithm defined by an ICD-9 code of 345.x (epilepsy) or 2 codes of 780.39 (convulsions) on separate days. Different validated epilepsy algorithms, including both ICD-9 codes plus antiseizure medications, have been used to analyze veterans administrative data. The authors were not able to validate their definition, making it likely that a proportion of their cohort included some people without epilepsy (e.g., patients with provoked seizures only, or nonepileptic seizures). Such codes also lack reliable information about disease severity and epilepsy type. Because of limitations in the NYCLIX database, a number of other factors—such as race, ethnicity, socioeconomic status, insurance type, or whether the person had a regular epilepsy source of care and with whom (primary care, neurologist, epileptologist)—were not included in the prediction models. In the Houston/New York City health care use study, black and Hispanic patients with epilepsy had higher rates of generalist visits, emergency room (ER) care, and hospitalizations, and lower rates of specialist visits; and over time, patients with a low socioeconomic status had higher use of the ED. Furthermore, such NYCLIX data do not measure factors such as whether a seizure took place in public, stigma, or utilization of self-management strategies. While Grinspan et al. demonstrated that high ED use predicts future ED use, we are still left not fully knowing whether, and the extent to which, health disparities, a