Predicting Microsleep States Using EEG Inter-Channel Relationships

• Published in: IEEE transactions on neural systems and rehabilitation engineering 2018-12, Vol.26 (12), p.2260-2269
• Main Authors: Buriro, Abdul Baseer, Shoorangiz, Reza, Weddell, Stephen J., Jones, Richard D.
• Format: Article
• Language: English
• Subjects: class imbalance; Classifiers; Coherence; Correlation analysis; Correlation coefficient; Correlation coefficients; Covariance; Discriminant analysis; EEG; Electroencephalography; Entropy; Entropy (Information theory); Feature extraction; Indexes; inter-channel relationships; LDA; LSVM; microsleep; Neurotechnology; Performance prediction; Sensorimotor integration; Sleep; Support vector machines; Synchronism; Synchronization; Task analysis; Theta rhythms; Training; Wavelet analysis
• ISSN: 1534-4320; 1558-0210
• DOI: 10.1109/TNSRE.2018.2878587

A microsleep is a brief and an involuntary sleep-related loss of consciousness of up to 15 s. We investigated the performances of seven pairwise inter-channel relationships-covariance, Pearson's correlation coefficient, wavelet cross-spectral power, wavelet coherence, joint entropy, mutual information, and phase synchronization index-in continuous prediction of microsleep states from EEG. These relationships were used as the feature sets of a linear discriminant analysis (LDA) and a linear support vector machine classifiers. Priors for both classifiers were incorporated to address the class imbalance in the training data sets. Each feature set was extracted from a 5-s window of EEG with the step of 0.25 s and was demeaned with respect to the mean of first 2 min. The sequential forward selection (SFS) method, based on a serial combination of the correlation coefficient, Fisher score-based filter, and an LDA-based wrapper, was used to select features from each training set. The comparison was based on 16-channel EEG data from eight subjects who had performed a 1-D visuomotor task for two 1-h sessions. The prediction performances were evaluated using leave-one-subject-out cross-validation. For both classifiers, non-normalized feature sets were found to perform better than normalized feature sets. Furthermore, demeaning the non-normalized features considerably improved the prediction performance. Overall, the LDA classifier with joint entropy features resulted in the best average prediction performances (phi, AUCPR, and AUCROC) of (0.47, 0.50, and 0.95). Joint entropy between O1 and O2 from theta frequency band was the most informative feature.