A microsleep is a brief and involuntary sleep-related loss of consciousness of up to 15 s during an active and attention-demanding task. Such episodes of unresponsiveness are of particularly high importance in people who perform high-risk and monotonous activities requiring extended- attention and unimpaired visuomotor performance, such as car and truck drivers, train drivers, pilots, and air-traffic controllers, where microsleeps can, and do, result in catastrophic accidents and fatalities. Microsleep-related accidents can potentially be avoided and thereby lives saved, if microsleeps are noninvasively and accurately predicted.

The aim of this study was to explore various inter-channel relationships in the electroencephalogram (EEG) for detection/prediction of microsleeps. In addition to feature-level and decision-level data fusion techniques, ensemble classification techniques were investigated to improve microsleep detection/prediction accuracies.

The data used in this research were from a previous study in which 15 healthy non-sleep- deprived participants performed two 1-h sessions of 1-D continuous tracking task. Eight subjects were included for the analyses, all of whom experienced at least one microsleep during the two sessions of the task. A gold-standard was formed by integrating two independent measures of the video-rating from a human expert and tracking performance. An average duration of microsleep was 2.16 min/h.

Three univariate feature sets, namely variance, wavelet spectral power, and entropy features were extracted from individual channels of EEG. Seven bivariate and symmetric feature sets were directly extracted from pairs of EEG channels. Three of the 7 feature sets were non-normalized (i.e., covariance, wavelet cross-spectral power, joint entropy) and 4 were normalized (i.e., Pearson’s correlation coefficient, wavelet coherence, mutual information, phase synchronization index (PSI)). Eleven feature sets were extracted from coefficients of a multivariate autoregressive (MVAR) model, fitted to the EEG time series. Five of the 11 feature sets were non-causal (i.e., cross-spectral power, coherence, imaginary part of the coherency (iCOH), partial coherence (pCOH), PSI), whereas 6 were causal (i.e., partial directed coherence (PDC), generalized PDC (GPDC), directed transfer function (DTF), normalized directed transfer function (nDTF), direct DTF (dDTF), full frequency DTF (ffDTF)). All features were extracted from a window length of 5 s of EEG . Feature sets extracted from EEG time series of each session were demeaned with respect to mean of their first 2 min features.

Two supervised feature-selection techniques were used to select discriminatory and relevant features from each feature set. The first technique was classifier-dependent feature-selection, which comprised Pearson’s correlation coefficient-based filter followed by Fisher’s score-based ranking and a linear discriminant analysis (LDA)-based wrapper. The second technique was classifier-independent, which was a combination of relevance-based ranking (i.e., mutual information between a feature and class labels) and relevance-based sequential forward selection (SFS) method.

Two linear classifiers, namely LDA and linear support vector machine (LSVM), were used to validate the efficacy of a feature set at detecting/predicting microsleeps.

To improve detection/prediction accuracies of microsleeps, all 21 feature sets were fused at feature-level and decision-level. Two-stage feature-selection was performed in feature-level fusion and the homogeneous (same type) classifiers were used in decision-level fusion.

The data of each subject was assumed to be an individual cluster. Three ensemble classification techniques were proposed and, in addition to soft majority voting, compared with single classifier. In one ensemble classification technique, individual clusters, using divergence, were combined to form overlapping clusters.

Non-normalized features performed better than normalized and causal features. Baseline correction (demeaning) of non-normalized features substantially improved the performance metrics on microsleep states and onsets. For most of the feature types, performances of classifier- dependent feature-selection was better than classifier-independent. With classifier-dependent feature-selection, LDA classifier, compared to LSVM, gave slightly higher performance.

Overall, joint entropy was the best single feature set. Data fusion was no better than best performing single feature set. All of the ensemble techniques resulted in overall improved performances over single classifier.

The overall highest detection performance metrics (phi = 0.48, area under the curve of precision recall (AUCPR) = 0.51, area under the curve of receiver operating characteristic (AUCROC) = 0.95) and (phi = 0.11, AUCPR =0.09, AUCROC = 0.91) on microsleep states and onsets, respectively using single feature sets were achieved with joint entropy features.

Compared to the best performing single feature set, feature-level fusion resulted in similar detection performance metrics for microsleep states of (phi = 0.49, AUCPRR = 0.49, AUCROC = 0.96) and for microsleep onsets of (phi = 0.10, AUCPR = 0.09, AUCROC = 0.92).

Overlapping clusters resulted in the overall highest detection performance metrics of (phi = 0.50, AUCPR = 0.52, AUCROC = 0.96) and (phi = 0.12, AUCPR =0.10, AUCROC = 0.93) on microsleep states and onsets, respectively.

We report the overall highest microsleep (state and onset) detection and prediction perfor- mance metrics achieved on this data set. Our results are a way forward in using EEG inter-channel relationships as features at detecting/predicting microsleeps. Contrary to the literature on other classification tasks, our results favour non-normalized inter-channel features at classifying microsleeps from responsiveness. Our results do not support data fusion of single modality (e.g., EEG) data in classification of microsleeps. All of the proposed ensemble techniques improved detection/prediction accuracies of microsleeps, using EEG, over a single classifier.

Notwithstanding our achievements, the performances remain too low in general for real-life applications.