Abstract
Microsleeps are sudden, short involuntary episodes of sleep (0-15 s) during which a person is unconscious and unresponsive. They are associated with behavioural cues like partial eye-closure, head nods, and loss of muscle control. A person experiencing microsleeps is often unaware of them. This makes the microsleeps perilous, as the temporary loss of responsiveness can result in fatal accidents. Hence, it is of importance that microsleeps are detected and, if possible, predicted.
This work aimed at identifying effective-connectivity-based-brain network characteristic features of EEG which can demarcate a microsleep from preceding responsive states, and evaluating their effectiveness for microsleep detection/prediction in machine-learning-based classifiers. This study used previously recorded behavioural and EEG data from a study comprised of a 1-D tracking task performed by 15 non-sleep-deprived and healthy subjects for two 1-hour sessions. Only those eight subjects who had a minimum of one definite microsleep in at least one 1-hour session were included for analysis. Both intra-band and inter-band brain networks were investigated. In the case of synchronous networks, the EEG signals, after reference electrode standardization technique (REST) rereferencing, were processed to derive time-varying auto-regressive (TVAR) parameters via a General Linear Kalman Filter (GLKF). Then, a time-varying effective connectivity (EC) measure – orthogonal partial directed coherence (OPDC) – was obtained. The effective EC matrices formed using these OPDC measures, with the scalp electrodes as nodes, constituted the synchronous brain networks, which were further processed through graph theory concepts. Synchronous brain circuits were obtained for theta, alpha, and beta bands and analyzed in both P-A (posterior-anterior) and A-P (anterior-posterior) directions.
Inter-band or coupling-based brain networks were obtained through non-linear analysis of the EEG by higher-order spectral analysis (HOSA). Based on the cross frequency coupling (CFC) measure of phase amplitude coupling (PAC) via bi-coherence, the brain networks were constructed for multiple inter-band couplings. Intra-band and CFC-based analyses were also carried out in source-space. Twenty-six Brodmann areas served as the nodes for the brain networks. Source-space reconstruction was also modified according to the simplified source-space reconstruction for Brain Computer Interface data, by reducing the number of dipoles involved based on Brodmann areas combined with weighted minimum norm (WMN), to reduce the implementation time. Global, nodal, and community-based measures were investigated. The non-parametric Wilcoxon signed rank test was used for group comparisons, with Cohen-type effect size. Potential features were identified for four different scenarios: sensor-space intra-band, sensor-space CFC, source-space intra-band, source-space CFC, using group-wise and event-wise consistency analysis. The features were then ranked based on the minimum classification error criterion, exploiting the cross-correlation among them. The probability distribution of the communities during microsleeps and during responsives were calculated from a characterization perspective. The communities present during microsleeps but not responsives, and vice-versa, were also identified.
To explore the features, from a machine-learning perspective, the data were divided into eight training sets by the leave-one-subject-out (LOSO) method. Group and event analyses were done for each of the training data sets, and performance was computed on the corresponding test
data set. Performances of the fusion of features from the four scenarios were also investigated. Lastly, the fusion of EC features with non-EC features were analyzed. Irrespective of the scenario, among the global measures, dynamic modularity showed significant differences in at least one band-direction or coupling-direction, especially mostly in A-P direction. In all four scenarios, the nodal measures of flexibility, integration, core-percentage, node cohesion, node-betweenness, and global efficiency always decreased, while recruitment increased invariably during microsleeps, irrespective of the bands/couplings-direction and electrode/cortical regions, that showed significant changes. Clustering coefficient increased invariably in sensor and source space intra-band, irrespective of the bands/couplings-directions and electrode/cortical regions, that showed significant changes. Promiscuity exhibited decreasing trends in source-space CFC, irrespective of the couplings-directions and cortical regions, that showed significant changes.
From a region point of view, irrespective of measures and bands, parietal, occipital and primary motor cortex areas demonstrated changes during microsleeps in both sensor-space intra-band and CFC. The hand motor area, posterior parietal cortex, and occipital cortex showed consistent changes in both source-space intra-band. These indicates consistency in the regions affected during microsleep in sensor and source-spaces. The bands and couplings that showed changes during microsleeps conform to the functions associated with regions that showed differences during microsleeps. The parietal and occipital regions play a vital role in working memory/attention, motor task performance, and visual perception, which is in line with alpha and theta changes. Theta-alpha, theta-gamma, and alpha-beta are also linked with working memory and auditory, visual and motor tasks, for which the above mentioned regions are instrumental. Moreover, community-evolution measures dominate the top informative features for microsleep state detection. Also, flexibility, recruitment, and node global efficiency were common among the top informative features, for all scenario.
Modularity decreased in microsleep states compared to responsive states, mostly in the anterior to posterior direction (A-P). Despite the change in direction, A-P indicates sleep onset (N2 stage), decrease in modularity associates microsleep with drowsiness and/or presleep (N1). Decrease in global efficiency, and increase in clustering coefficient also associate microsleeps with NREM sleep stages, excluding N1. Also, decrease in flexibility, which is a measure of cognitive flexibility, associates microsleeps with stage 2 NREM sleep. Changes in delta-theta and beta-gamma couplings associate microsleeps with shift to induced unconsciousness state and N2 sleep stage respectively. The absence of changes in global measures and significant changes in nodal measures, thus brings out the similarity between microsleeps and N1 (presleep), a transition from sleep to wakefulness. Also, the flow decreased from A-P (feedback flow) compared to P-A (feed forward) in directional connectivity, unlike sleep, but similar to induced consciousness. Increase in modularity with depth of sleep is also associated with increased implicit learning capability, needed for memory and procedural skill consolidation occurring during the REM and NREM (except N1) sleep stages. Also, there is a need for coordinated activity among brain areas to facilitate the memory and procedural skill consolidation during sleep. But, microsleeps are associated with decrease in modularity, flexibility, integration, global efficiency, and increase in recruitment, irrespective of the four different scenarios. This decrease in measures of learning and integration and increase in measures of segregation like recruitment, not only indicate less learning capability of the brain during microsleeps but reflects disorganized manner of working of the brain during microsleeps. This clearly indicates lack of memory and procedural skill consolidation during microsleeps, unlike the other sleep stages. Also, findings from the regions and bands indicates disruptions in the visuomotor, perceptual, and cognitive skills as in case of sleep, despite very short duration. Hence, from all these findings it can be said that microsleeps are associated with a forced shifts to unconscious states, in the struggle to stay awake, unlike voluntary sleep which involves natural and slow transitions, and share traits of both N1 and other NREM sleep stages, but lack the advantages associated with sleep.
Of the four EC scenarios, the highest microsleep state detection performance was in sensor-space intra-band (phi= 0.35), with an average sensitivity and precision of 0.66 and 0.27 respectively. For source-space intra-band, the state detection performance was essentially the same (phi= 0.34), with an average sensitivity and precision of 0.57 and 0.33 respectively. Fusion of all 4 EC scenarios yielded a phi of 0.36, with an average sensitivity and precision of 0.80 and 0.28, respectively.
Although many EC features changed during microsleeps, these changes are either not exclusive to microsleeps or the variability, due to low signal-to-noise-ratio, is too high to produce robust microsleep features. This is reflected in a high rate of false positives, leading to a substantial drop in performance. Fusion with non-EC log-spectral features gave a phi of 0.44 with source-space CFC features. Fusion with non-EC joint entropy features gave aphi of 0.49 with sensor-space CFC features. The log-spectra + sensor space CFC performed better compared to the the LS baseline (p=0.0078), but JE + source CFC performed no better than the JE baseline performance.The performance obtained for fusion with non-EC indicates the presence of orthogonal information between the log-spectra and the CFC based features. However, the EC features rank below several non-EC features, from prior work, in terms of detection of the microsleep state.