Abstracts

Rationale: Numerous measures computed from the EEG have been proposed for detecting preictal states in epilepsy, but few perform better than chance when tested on prospective data. Multivariate measures of coherence across brain areas have shown some seizure-predictive ability, but with poor specificity (i.e., high false positive rate). Sleep-wake state is one factor known to bias the number of false positives generated by seizure prediction algorithms (SPAs). It is to be expected that switching between different normal states of vigilance (SOV) is accompanied by changes in coherence; however, little is known about the effect of normal vigilance cycles on the dynamics of SPA measures, and on seizure likelihood. We have previously studied the effect of SOV on seizures in a rodent model of temporal lobe epilepsy. Here we use this model to test the effect of SOV changes on measures of EEG coherence, and compare values to the preseizure period.Methods: Tetanus toxin implanted in the hippocampus of rats generated spontaneous, intermittent seizures for weeks after a latency of 2-4 days. Seizures were detected and verified using video-EEG. Univariate EEG power in different frequency bands, along with head acceleration, was computed from cortex and hippocampus and used to discriminate SOV i.e., NREM sleep, REM sleep, quiet wake (QW) or active wake (AW) in sequential 10s epochs. Multivariate measures of coherence between pairs of EEG channels, specifically the linear cross-correlation peak r² and the broadband Hilbert phase coherence R, were computed in 10s epochs and averaged over all channel combinations. Distributions of r² and R were estimated for different SOV in the post-implant latent period prior to any seizures, and compared with a 5 min period prior to each seizure after the development of spontaneous seizures.Results: Measures of EEG coherence were found to vary with SOV in epileptic rats. The most significant difference (p<0.01) was a surge in value during AW behaviors such as exploratory motion. Differences between REM, NREM, and QW were not as distinctive for the chosen measures; furthermore, the preseizure period (n = 30 seizures) could not be distinguished from REM, NREM, and QW. Finally, the SOV during the preseizure period was predominantly scored as REM or NREM sleep.Conclusions: Not only do measures of EEG coherence offer poor contrast between the preseizure period and normal vigilance cycles, but changes in coherence may frequently track transitions to some normal but more seizure-permissive state (e.g., slow wave sleep or arousal) rather than a unique preictal state. Neither of the above facilitates the development of accurate seizure prediction algorithms. However, it is possible that specific choices made in this study for instance, averaging of measures over all available channels or the use of broadband instead of band-limited coherence may have contributed to the lack of specificity. Nevertheless, it is clear that, at a minimum, SPAs must incorporate adaptive state-dependent corrections to achieve the level of performance desirable in clinical use.