Abstracts

Rationale: There is an established relationship between sleep and seizure states in animal and human models of epilepsy. There has also been a significant effort to develop seizure prediction algorithms both as a warning tool for patients and as part of closed-loop control systems. To date, published methods and corresponding performance have been difficult to implement and reproduce. Although it has been noted that false-prediction rates follow a diurnal variation (Schelter 2006), the state of vigilance (SOV) has not been used as a seizure prediction feature. We therefore examined the likelihood of seizure onset as a function of SOV in the chronic tetanus toxin model of temporal lobe epilepsy. We then quantify prediction performances utilizing SOV. Methods: Long-Evans rats were focally injected in the ventral hippocampus with tetanus toxin to induce chronic spontaneous seizures. We use existing algorithms to classify SOV (Sunderam 2007) and identify seizures from 380 hours of recordings in 6 rats. SOV was classified in 10 s windows as either wake, rapid-eye-movement sleep (REM) or non-REM sleep (NREM). We utilize a forecast seizure probability as the predictor output, and quantify its performance with a Brier score (Jachan 2009). We consider three predictor models - one based on overall seizure rate, one on seizure rate and SOV, and one on SOV, time of day (TOD) and seizure rate. We considered the prediction formalism presented in (Schelter 2006b), and optimized performance with respect to seizure prediction horizon and seizure onset period. We separated the data into training and testing sets. Means and variances were then computed from this process. Results: We find for this model that the likelihood of REM prior to seizure dominates all other states up to 180 s ahead of seizure onset. Visual verification of data suggests that seizures are arising from REM and not the transition out of REM. A predictor conditioned on SOV performs significantly better than the unconditioned seizure predictor. However, the addition of TOD to a predictor already conditioned on SOV does not further improve predictive power as indicated by the Brier score. Conclusions: Our observations are in contrast to findings that REM is an effectively anti-epileptic state (Shouse 2000), and challenge prevailing hypothesized mechanisms of sleep-related epileptogenesis (Shouse 2000, Colom 2006). Furthermore, they support the notion that SOV is a primary biologically relevant feature for prediction of seizure. We further conjecture that it may overwhelm in significance other seizure prediction features, although that remains to be demonstrated. Schelter B et al. Epilepsia 47(12):2057-70, 2006. Sunderam S et al. J Neurosci Meth (163):373, 2007. Jachan M et al. IFMBE Proceedings, 22(11):1701-1705, 2009. Schelter B et al. Chaos 16, 013108, 2006b. Shouse MN et al. Clin Neurophys 111(S2):9-18, 2000. Colom LV et al. J Neurophys 95(6):3645-53, 2006.