Abstracts

Rationale: Most seizure detection/prediction algorithms (SDPAs) try to distinguish abnormal states from the interictal baseline in the extraoperative electrocorticogram (ECoG), but do not discriminate between different states of vigilance. Vigilance state affects not only SDPA accuracy, but also the likelihood of seizure generation. It is thus important to examine dynamic changes in the interictal vigilance state to understand how they influence seizures; this insight may then be used to improve SDPA performance. While most SDPAs are developed from ECoG recordings in patients being evaluated for epilepsy surgery, standard guidelines for scoring sleep require polysomnography (PSG), which is not routinely acquired during chronic ECoG monitoring. In this study, we acquired combined scalp and intracranial electrographic recordings to develop algorithms for tracking sleep directly from the ECoG in patients with refractory epilepsy. Methods: Combined ECoG and PSG (EEG, EKG, EOG, and EMG) recordings were prospectively acquired from five patients undergoing presurgical evaluation at the University of Texas Southwestern Medical Center. Fifteen seizure-free overnight interictal segments (8 to 15 h duration; mean 11 h) were extracted. ECoG contacts closest to the scalp EEG contacts were analyzed using a bipolar derivation. Signal power was estimated separately for the chosen EEG and ECoG signals in consecutive 30 s windows in five frequency bands (delta, theta, alpha, sigma, and beta) corresponding to recognized cortical rhythms and compressed into three "sleep" features: i.e., power ratios (e.g., delta/theta) that vary between the predominant vigilance states (NREM 1-3, REM, and Wake). For each segment, the correlation between EEG and ECoG sleep features was tested. Then the feature time series was fitted separately for EEG and ECoG to a hidden Markov model (HMM) with five states. Both HMMs were used to predict the sequence of vigilance states in each recording, and their agreement assessed by Cohen's kappa. The intent was to determine whether sleep tracking based on ECoG is feasible. Results: The three sleep features derived from the ECoG were strongly correlated with the corresponding features computed from the EEG with mean correlation coefficients r of 60±3, 64±7 and 62±6%, respectively (n = 15). When five recordings with poor EEG quality were excluded, r increased to 67±3, 73±4 and 73±4%, respectively (n = 10). Categorical agreement for the two HMMs was moderate (mean kappa 44±3%, n = 15) and improved when poor EEG recordings were excluded (mean kappa 49±4%, n = 10). Conclusions: The results demonstrate that ECoG and EEG signals, derived from anatomically similar locations on the cortex and scalp respectively, convey analogous information regarding vigilance states during sleep in patients with epilepsy. The ECoG can therefore be used to infer changes in baseline vigilance state and to draw correlations with ictogenesis and putative preictal changes. Such correlations might improve the specificity of SDPAs. This is of special relevance given the emergence of devices for ambulatory recording and responsive neurostimulation.