Abstracts

Rationale: Antiepileptic drug trials in the current era compare seizure prevalence in cohorts of patients with respect to an observed baseline. Because there are often baseline observation periods which establish inclusion into a trial, patients whose seizure frequency is transiently higher than usual may be included and have lower seizure prevalences during the experimental intervention period due to regression to the mean. This may manifest as high placebo response rates. We performed computer simulations to estimate the effect size of mean regression under various assumptions of seizure timing. Methods: We performed computer simulations of drug trials with varying parameters of 1) baseline observation length, 2) intervention length, 3) seizure frequency cutoff for trial inclusion, and 4) seizure rate parameters. Seizures were modeled as points in a Poisson process using an exponential distribution with rate parameters taken as samples from a lognormal distribution. Rates of seizure prevalence were described by median seizure reduction compared to baseline period. The simulations were coded in the programming language Julia (version 1.1) and run on a Linux workstation using an Intel Core i5-4430 CPU. Trials were conducted with patient cohorts of 10,000. Results: Median seizure change was quite variable and ranged from greater than 99% reductions to increases of 400%. The highest values for median seizure reduction were associated with short baseline periods, long intervention periods, and high inclusion criteria cutoffs. Rate parameter magnitude bore a non-monotonic relationship to median seizure changes. Conclusions: Antiepileptic drug trials' effect size estimates are biased due to minimum seizure counts for inclusion and comparison of seizure frequencies to baseline rates. The size of the estimate bias increases with intervention-to-baseline length ratio and with minimum seizure counts. As a function of underlying rate parameters estimate bias may have local optima. This relies on assumptions of independent seizure occurrences (ie, discounting clustering). To reduce bias, AED trials should be designed with longer baseline periods and lower cutoffs for seizure rates. Funding: No funding