Abstracts

Rationale: Statistical models of seizure frequency can provide insight into epilepsy mechanisms and treatment, and are useful for clinical trial design and analysis. Past work in this field has been limited by the availability of longitudinal data. Using data from SeizureTracker.com, a seizure diary tool with over 1 million recorded seizures across 8 years, we assessed several statistical distributions to identify optimal tools for modeling seizure frequency. Methods: We selected seizure diary excerpts of patients not reporting nonfocal epilepsy or status epilepticus. For each patient, daily seizure counts during first 6-week period in which at least 4 seizures occurred was identified, and the immediately following 12 weeks, were analyzed. The following models were fit to the data: Poisson (PS), Negative Binomial (NB), Negative Binomial with Covariates (NBC), which considers the effects of demographics and etiology, and Negative Binomial with Autocorrelation (NBA), which considers the effect of seizures occurring on the previous two days. Two NBA models, using either presence/absence of seizures (NBAP), or number of seizures (NBAN), were assessed. For the NBC models, all available demographic and etiology covariates were assessed, and those that were significant at the p < 0.05 level (Wald test) were selected. Data for adults (age ?- 18) and children were analyzed separately. Strength of the models was evaluated using Bayesian Information Criteria (BIC), and significance was evaluated by performing ANOVA on the models. Results: We analyzed 133223 seizures from 1049 adult patients and 180594 seizures from 1422 pediatric patients. For both sets of patients, the NB model was significantly superior to the PS model (p < 0.001, p < 0.001). The NBC model for adult data was slightly inferior to NB (p < 0.01) with Tuberous sclerosis complex (TSC), Lennox-Gastaut, and unspecified congenital disorders (UC) having significant effects on seizure frequency. The NBC model for children was slightly superior to NB (p < 0.001) with TSC, injury resulting in hematoma, and age affecting daily seizure frequency. For both children and adults, the NBAP model improved fit over the NB model (p< 0.001, p < 0.001). In both cases, the NBAP model was significantly superior to the NBAN model (p < 0.001,p < 0.001). Conclusions: Overdispersion and autocorrelation are observed in the daily seizure counts of children and adults. Presence/absence of seizures on previous days is a better predictor of seizure frequency than number of seizures on previous days. For adults, inclusion of demographic information and etiology resulted in worse BIC, so increasing model complexity may not be warranted. The NBC model was superior to NB for children, but only slightly so. NBAP best modeled the data in both children and adults. The use of more accurate statistical models of seizure frequency will allow the construction of more relevant clinical trial simulations, and may help plan more effective clinical trials. Funding: Supported by the National Institutes of Neurological Diseases and Stroke Intramural Research Program