Spectral Features in iEEG for Automated Behavioral State Classification
Abstract number :
1.147
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2016
Submission ID :
191436
Source :
www.aesnet.org
Presentation date :
12/3/2016 12:00:00 AM
Published date :
Nov 21, 2016, 18:00 PM
Authors :
Vaclav Kremen, Mayo Clinic, Rochester, MN, USA ; CIIRC, Czech Technical University in Prague, Czech Republic., Rochester, Minnesota; Juliano J. Duque, Mayo Clinic, Rochester, MN, USA ; FFCLRP, University of São Paulo, Ribeirão Preto SP, Brazil, South Amer
Rationale: There has been relatively little investigation of the feasibility of sleep staging from intracranial EEG (iEEG) recordings. We explored the feasibility of automated wake and slow wave sleep (SWS) classification using wide bandwidth iEEG (0.01 ?" 600 Hz) spectral power features from normal and epileptic brain regions. Methods: Intracranial EEG data were recorded from 7 subjects with drug resistant medial temporal lobe epilepsy. All subjects were implanted with intracranial depth electrodes and had simultaneous scalp EEG recordings over multiple days and nights. Seizure onset zones (SOZ) were determined by visual assessment of electrographic seizure discharges by a board certified epileptologist. Visual sleep scoring was performed in accordance with standard scalp montage according to American Academy of Sleep Medicine 2007 methods. All original signals were filtered into six frequency bands as follows: 0.1-4Hz, 4-25Hz, 25-55Hz, 65-100Hz, 100-250Hz, 250-600Hz. The absolute and relative spectral power levels of frequency bands were calculated for all electrodes across all subjects for each 30 second sleep staging epoch. Ten minutes of wake and SWS iEEG data from each subject were used to train a support vector machine binary classifier for each electrode and subject. Prediction accuracy of selected classifiers was subsequently validated using twenty minutes of out of sample awake and SWS data. Results: We performed comparison of classifiers performance using all spectral features in all electrodes in each subject for both SOZ and non-SOZ. Classification accuracy was higher (p < 0.01) using data from non-SOZ electrodes (average accuracy 98.4%) compared to SOZ (average accuracy 94.6%). Conclusions: Our iEEG spectral power feature analyses demonstrated accurate automated SWS and wake state classification for both epileptogenic and non-epileptogenic cortex. Such iEEG-based behavioral state classifiers could feasibly be incorporated into future implantable devices that quantify patient sleep patterns, administer behavioral state specific therapies, and adjust seizure forecasting classifiers. Funding: Mayo Clinic Discovery Translation Grant, National Institutes of Health (R01-NS063039, R01-NS078136), institutional resources for research by Czech Technical University in Prague, Czech Republic, ALISI ?" NPU (LO1212), VES15 II ?" LH15047, and S㯠Paulo Research Foundation 2014/01587-8.
Neurophysiology