Abstracts

Rationale: The ability to anticipate or predict seizures with reasonable accuracy would benefit patients with intractable epilepsy. Many of the seizure prediction algorithms (SPAs) proposed to date extract dynamical features ( prediction variables ) of the electrocorticogram (ECoG) and consider abnormal trends in advance of seizures as evidence of a preictal state (i.e., an abnormal state especially conducive to seizures) [Brain 2007; 130:314-33]. But sleep-wake state changes can limit the performance of some SPAs [Epilepsia 2006; 47:2058-70]. Vigilance state (R=REM, N=NREM, W=Wake) is already known to bias seizure likelihood [Epilepsia 1997; 38:56-62] in addition to circadian phase, sleep debt and stress, but a detailed investigation of its effects on seizure predictability is lacking. One limitation is that polysomnography (PSG) is needed to determine vigilance state, but PSG is not commonly performed in conjunction with ECoG [Sleep 2000; 23:231-4], which is the source of most (if not all) data used to develop and test SPAs. The purpose of this study is to determine how prediction variables derived from ECoG are influenced by vigilance state. This is addressed by analyzing ECoG and PSG data recorded simultaneously from patients with refractory epilepsy.Methods: We analyzed prospectively acquired ECoG and PSG (C3-O1 or C4-O2, EOG, submental EMG) data from two patients undergoing intracranial monitoring as part of presurgical evaluation at UTSMC. Electrodes for PSG were placed contralateral to the craniotomy. Vigilance state was scored from overnight recordings using a hidden Markov model (HMM) that labeled each 30 s epoch as W, N1/2, N3, or R. Signal features used in the model were EEG / and ( + )/( + ) power ratios respectively and root-mean-squared EMG and EOG. This algorithm gave moderate-to-high agreement (Cohen s kappa of 0.6-0.8) with expert scores on a Physionet PSG database [IEEE-BME 2000; 47:1185-94] for six subjects. Next, three prediction variables linear cross-correlation peak, mean phase coherence, and sample entropy were computed from differential ECoG near the seizure focus. Mean values of prediction variables for the four HMM-scored vigilance states were compared using analysis of variance.Results: All of the prediction variables derived from ECoG varied significantly with vigilance state (p < 0.001). Although post hoc pair-wise comparisons showed no consistent patterns of difference between vigilance states across prediction variables or patients, some differentiation was apparent between sleep and wake states.Conclusions: Univariate and bivariate ECoG features typical of those incorporated in SPAs were found to vary significantly with vigilance state as determined by PSG. Our preliminary analysis of simultaneous ECoG/PSG recordings from two patients suggests that such algorithms should correct for normal changes in vigilance state in order to minimize false predictions. Collection and analysis of ECoG-PSG recordings from more patients undergoing intracranial evaluation is ongoing and expected to yield more definitive ways of correcting for state and improving SPA performance.