Motivation ECoG-based overnight sleep scoring ...

1 downloads 0 Views 1MB Size Report
ECoG-based overnight sleep scoring. Modeling interictal variability in seizure prediction. Seizure prediction algorithms (SPAs) try to differentiate ictal or preictal ...

Scoring Sleep in Interictal Electrocorticographic Recordings Farid Yaghouby1, Pradeep Modur2, and Sridhar Sunderam1* 1 – Department of Biomedical Engineering, University of Kentucky 2 – Department of Neurology, University of Texas Southwestern-Austin Program

Motivation

We acquired simultaneous PSG/ECoG data from patients with refractory epilepsy to: A. Develop ways to score and track sleep directly from ECoG. B. Test the effect of vigilance state on ECoG-derived seizure prediction variables during interictal and preictal periods. C. Explore ways to correct SPA output for confounding effects of vigilance changes.

Derivation of sleep variables from ECoG We have previously shown [4] how to automatically score sleep into W, N1, N2, N3, and R states using hidden Markov models (HMMs) of sleep variables derived from the EEG. The sleep variables used were EEG band power ratios that emphasize differences between the vigilance states, estimated in 30s windows: S1 = [delta: 0.5-4 Hz] : [theta: 4-9 Hz], S2 = [alpha+beta:9-30 Hz] : [delta+theta: 0.5-9 Hz], and S3 = [low delta: 0.5-2 Hz] : [sigma: 12-16 Hz].

We examined different ECoG grid derivations (bipolar, common average, etc.) in five epilepsy patients being evaluated for surgery to see which, if any, correlated best with simultaneous EEG. While local bipolar derivations are most commonly used in ECoG analysis, distant bipolar (DB) derivations were found to correlate significantly better with EEG in terms of the computed sleep variables. We therefore used DB-ECoG signals for the rest of our analysis.

Interictal trends in dynamical variables used in the seizure prediction literature [1], such as the linear cross-correlation peak (P1) and mean Hilbert phase coherence (P2), could reflect normal vigilance changes but be mistaken for seizure precursors. We fitted linear models of P1 and P2 computed between neighboring DB-ECoG pairs to gauge the contribution of sleep variables S1, S2 and S3 to their interictal dynamics. Such models could be used to correct SPAs for the confounding effects of sleep-related transitions. ECoG sleep variables

2

S

1

S

2

S

0

3

-2

ECoG seizure prediction variables

We scored sleep in these recordings using an unsupervised four-state HMM fitted to the time series of sleep features S1, S2 and S3, but separately for the DB-ECoG and EEG signals. In the 12 hour overnight sample shown below, trends in ECoG sleep variables mirror those of the EEG; four or five prominent cycles related to sleep are apparent in both. W/N1

ECoG

R

P1

0.3

P2

Objectives

We compared S1, S2 and S3, and PC1 (the first principal component of 7 signal band power fractions) for EEG (frontalcentral or central-occipital derivations) and DB-ECoG signals from outside the seizure focus for 15 overnight recordings (without seizures) in five patients. Of these, five recordings were excluded due to intermittent lapses in surface EEG quality. Results showed strong correlation with means above 60% (n=10) for all variables. Hence vigilance dynamics can be inferred from ECoG without the need for EEG or auxiliary measurements (EMG, EOG).

EEG-based sleep scoring ECoG-based sleep scoring

Seizure prediction algorithms (SPAs) try to differentiate ictal or preictal states from the interictal baseline electrocorticogram (ECoG) [1]. But rarely do they discriminate between states of vigilance within that baseline. Vigilance state affects not only SPA accuracy [2], but seizure likelihood as well [3]; and poor sleep can precipitate seizures. It is important to observe and track dynamical changes in interictal vigilance state to understand how they influence seizures. This insight may then be used to improve SPA accuracy. While most SPAs are developed from invasive ECoG recordings, sleep scoring requires polysomnography (PSG)i.e., a surface electroencephalogram (EEG) with muscle tone, ocular activity, heart rate, and other variableswhich is not routinely done during ECoG monitoring. In this study, we prospectively acquired combined scalp and intracranial electrographic recordings to develop algorithms for tracking sleep directly from ECoG in patients with refractory epilepsy.

Modeling interictal variability in seizure prediction

ECoG-based overnight sleep scoring

0.2

0.1 0.2 0.15 0.1 0.05 0

N2

N3 S3 (ECoG) S2 (ECoG)

data model

data model

2

4

6 Time (h)

8

10

12

The seizure prediction variables appeared to follow interictal trends in the sleep variables. Linear model predictions of P1 and P2 were evaluated on out-ofsample data on the basis of the mean absolute scaled error (MASE) in prediction. MASE is normalized by the mean absolute fluctuation in the true values and was significantly lower than chance in 12 of 15 recordings (p

Suggest Documents