Abstracts

Rationale: Clinical trials and physicians evaluate therapies by comparing seizure frequencies before and after an intervention begins. One key question currently remains unanswered: how much intrinsic variability in seizure frequency is expected? Methods: We studied three independently collected datasets: patient reported data (SeizureTracker.com or ST), physician curated patient reported data (Human Epilepsy Project or HEP), and chronically implanted intracranial data (NeuroVista, or NV). The NV data was subdivided further into clinically reported seizures, electrically equivalent to the clinical but unreported, and electrographic only seizures. A nonlinear model was fit retrospectively to the overall standard deviation given an overall mean for each of the 3 datasets. To test the ability of the model to predict variability prospectively, each dataset was segmented into trial sized blocks of 6 months (2 for baseline, 1 for titration, 3 for test). Patients contributed as many blocks as available in their diaries. Three methods of predicting the variability were computed: (A) the model (B) the baseline variability (C) "fixed" variability - 50-150% of baseline being assumed. A 7.5-month additional dataset was exported from ST after the models were tested for validation. Results: After applying criteria for inclusion of complete data, ST had 3016 patients, HEP had 274, and NV had 15. The nonlinear model was able to predict an individual patient's overall standard deviation using overall mean, with high accuracy across each dataset (see figure). The model fits showed R^2 of 0.827 for ST, 0.901 for HEP, and 0.911,0.935, 0.971 for the 3 types of NV data. The sequential predictions were correct 62-98% for method A, 77-98% for method B and 31-43% for method C, across datasets. In the ST validation set (N=1820), the predictions were correct 87%, 87% and 59% of the time for methods A, B and C. Conclusions: Using data from very different sources, 2 models for predicting seizure frequency variability can be employed that outperform the fixed variability model implicit in typical clinical trials, and these were validated in a dataset external to the original testing that produced the models. Of note, each of these datasets included anti-seizure medication changes at times, and despite these instabilities, both prediction methods A and B still showed much higher accuracy than the fixed approach C. These predictions may be of use in clinical trial design, as well as in outpatient management of drug-resistant epilepsy. Funding: Support for this study was provided by the National Institutes of Neurological Disease and Stroke Intramural Research Program