hide
Free keywords:
Resting state fMRI, EEG, EEG-fMRI, Sleep, Classification
Abstract:
A growing body of literature suggests that changes in consciousness are
reflected in specific connectivity patterns of the brain as obtained
from resting state fMRI (rs-fMRI). As simultaneous
electroencephalography (EEG) is often unavailable, decoding of
potentially confounding sleep patterns from rs-fMRI itself might be
useful and improve data interpretation. Linear support vector machine
classifiers were trained on combined rs-fMRI/EEG recordings from 25
subjects to separate wakefulness (S0) from non-rapid eye movement (NREM)
sleep stages 1 (S1), 2 (S2), slow wave sleep (SW) and all three sleep
stages combined (SX). Classifier performance was quantified by a
leave-one-subject-out cross-validation (LOSO-CV) and on an independent
validation dataset comprising 19 subjects. Results demonstrated
excellent performance with areas under the receiver operating
characteristics curve (AUCs) close to 1.0 for the discrimination of
sleep from wakefulness (S0 vertical bar SX), S0 vertical bar S1, S0
vertical bar S2 and S0|SW, and good to excellent performance for the
classification between sleep stages (S1 vertical bar S2: similar to 0.9;
S1 vertical bar SW:similar to 1.0; S2 vertical bar SW:similar to 0.8).
Application windows of fMRI data from about 70 s were found as minimum
to provide reliable classifications. Discrimination patterns pointed to
subcortical-cortical connectivity and within-occipital lobe
reorganization of connectivity as strongest carriers of discriminative
information. In conclusion, we report that functional connectivity
analysis allows valid classification of NREM sleep stages. (C) 2015
Elsevier Inc. All rights reserved.
