|dc.description.abstract||Lapses in responsiveness (‘lapses’), especially behavioural microsleeps (‘microsleeps’) and attention lapses, involve complete disruption of performance from 0.5 to 15 s and can result in injury or death, especially in the transport sector (e.g., pilots, air-traffic controllers, truck and car drivers, etc.). The existence of a real-time monitoring system for detecting lapses could reduce accidents and save lives. The Christchurch Neurotechnology Research Programme (www.neurotech.org.nz) is a leader in lapse research in terms of characterization and EEG-based detection of microsleeps. Despite its achievements, there is still some way to go toward developing a system capable of accurately detecting, let alone predicting, microsleeps in the real world.
Functional magnetic resonance imaging (fMRI) has shown increases and decreases in blood oxygen level dependent (BOLD) activity in certain regions of the brain prior to and during microsleeps. During microsleeps, the BOLD signal (and, hence, neural activity) decreases in the thalamus and posterior cingulate cortex but increases in several cortical regions, including the inferior frontal cortex, posterior parietal cortex, and occipital cortex. Furthermore, the extent of decreases in neural activity in the thalamus increases with the duration of microsleeps. Therefore, identifying regions with increased or decreased activity based on EEG functional imaging would provide more information for current EEG-based microsleep detectors.
A limitation of current EEG-based detectors of microsleeps is that analysis is performed on sensor-space EEG data which does not, and cannot, take advantage of the dipolar pattern of brain source signals. Hence, they cannot estimate signals from specific regions in the brain related to microsleeps, such as the thalamus. Synchronous activation and alignment of neurons produces electrical currents and changes in potentials on the scalp which can be recorded by EEG sensors, and these sources can be modelled via head-modelling techniques as equivalent current dipole sources. The electrical activity of these sources is not distributed uniformly on the scalp and has a dipolar pattern. Therefore, techniques such as spatial filters, which assume a dipolar constraint, can be used to reconstruct EEG source-space signals, i.e., reconstruct the time-series of virtual voxels. In this way, it is possible to reconstruct/enhance the signal of different brain regions and/or identify which brain regions have activations/deactivations associated with microsleeps.
In this project, the aim was to perform EEG-based functional imaging of the brain so as to enhance brain activity and identify brain locations which have changes in activity during microsleeps. A literature review identified potential signal processing techniques, including spatial filters and head-modelling techniques, which might have application in the reconstruction of brain electrical activity in 3D source-space rather than EEG sensor-space.
Once the appropriate techniques from minimum-variance and minimum-norm spatial filters were identified, they were systematically evaluated via simulations. The results of these simulations provided valuable information on how these spatial filters perform for weak and strong sources, their spatial resolution, and their biased behaviour to different locations in source space. After evaluation of the results, the weight-normalized minimum-variance spatial filter was chosen as the most appropriate technique for the study of electrical activity associated with microsleeps, due to its superior spatial resolution, unbiased behaviour in reconstruction from different brain locations, and suitability for weak sources.
A new technique, source-space independent component analysis (source-space ICA) was developed for detection, localization, and time-course reconstruction of multiple electrical sources in the brain. The components identified by source-space ICA also have considerable potential for investigation of 3D functional connectivity. The concept of source-space ICA is to apply a vector spatial filter to a 3D scanning grid of the brain, to reconstruct the time-courses of all of these voxels, and apply singular value decomposition (SVD) and ICA to these time-courses to identify the independent sources.
Another technique developed is Voxel-ICA, which is able to estimate the orientation of brain sources and increase the quality of the reconstructed signal by application of ICA on post vector-spatial filtering to separate the source of interest from other background activity and estimate orientation.
For brain functional imaging and signal enhancement of microsleeps, previous EEG recordings from 5 subjects who had had microsleeps when performing a 50-minute 2D tracking task were explored. The 64-channel 10-20 system had been used for EEG sensor configuration. Microsleeps were identified based on complete disruption of racking performance for 0.5-15 s accompanied by drowsy behaviour and partial or full eye-closure.
Functional imaging of EEG was performed by two methods: source-space spectral analysis and source-space ICA. For the first method, the weight-normalized minimum-variance beamformer and band-pass filtering were applied to the real EEG of subjects containing microsleeps. In both methods, a minimum-variance beamformer was used to project back the EEG (sensor-space) data to source-space. For source-space spectral analysis, the power changes in the source-space due to the microsleeps were measured and presented in the form of tomographic maps. In contrast, in the source-space ICA approach, source signals were separated from each other by ICA of the source-space signals, and those sources which had higher activity during or in the vicinity of microsleeps were identified visually and their corresponding mixing coefficients, found by ICA, were shown as tomographic maps. As full or partial eye-closure during microsleeps produces a dominant posterior alpha activity (8-12 Hz) and an anterior electro-oculogram delta (0-4 Hz) artefact, both functional imaging techniques were also applied to EEG during voluntary eye-closures. This allowed any common power changes during microsleeps and voluntary eye-closures to be identified and discounted.
The advantages of source-space derived signals over sensor-space signals are that (1) the data are analysed in source-space and therefore, can be directly presented as tomographic maps, (2) the spatial filters, such as minimum-variance beamformers, suppress sensor noise and thus do not require band-pass filtering which usually results in loss of desired data, and (3) the reconstructed time-series via a minimum-variance beamformer for a brain location gives an estimate of the brain’s neuroelectrical activity for that location. Since beamformers suppress brain activity from other regions, their application effectively provides enhancement of weak electrical activity from the region of interest.
For source-space spectral analysis, although there were plenty of microsleeps available (94 events), they came from only 3 subjects, as the other two subjects had too few microsleeps processable due to most of their microsleeps having EEG highly contaminated with artefacts. Conversely, for source-space ICA it is possible to analyse a single microsleep. Therefore, we were able to apply source-space ICA to the data from all 5 subjects.
In the case of source-space spectral analysis, after reconstruction of the virtual voxel time-courses, spectral analysis was performed to detect any changes in power in the brain during microsleeps relative to a 4 s period immediately preceding each microsleeps. The results were statistically analysed via Wilcoxon and Mann-Whitney tests, thresholded (α < 0.05), and presented as tomographic maps. To remove the effects of eye-closure from power changes during microsleeps, eye-closure EEGs were analysed separately and removed from the functional images of microsleeps. Significantly increased delta activity was seen in posterior regions of the brain, including the calcarine fissure and cerebellum, during the microsleeps. There are indications that this increased delta activity may be due to K-complexes, generally considered to be associated with stage 2 sleep, but this has still to be confirmed and further explored. Decreased activity was seen in the gamma band from motor cortex of 2 of the 3 subjects. However, contrary to expectations, no change was seen in the thalamus.
In source-space ICA of microsleeps, 29 microsleeps from 5 subjects were analysed to identify components active during or in the vicinity of microsleeps. In all subjects there were 1 or 2 alpha-band components time-locked to some of their microsleeps, all of which were from the hippocampus (bilateral or right-lateralized). In contrast, there were 2 to 6 time-locked alpha band components during voluntary eye-closures in all subjects. However, importantly, none of the voluntary eye-closure components were from hippocampal regions. In addition, the microsleep alpha-band components were bursts of activity from 0.5 s to ∼4 s bursts and are similar to, and may have been, sleeps spindles.
Source-space ICA also identified several theta-band components with strong burst-like activity in the vicinity of microsleeps. However, unlike the alpha-band components, the theta-band components were only loosely time-locked to microsleeps. The theta-band components were mostly found to arise bilaterally from the frontal orbital cortex.
This is the first study to have carried out EEG-based source-space analysis (source-space spectral analysis and source-space ICA) of microsleep neural activity. This has led to improved understanding of the presence, characteristics, and location of neuroelectrical activity in the brain relevant to microsleeps. Brain regions which have consistent changes in activity during microsleeps but not eye-closure may provide valuable new features for EEG-based microsleep detectors. As current detectors do not make use of the 3D spatial information of the brain sources, providing higher weights to such regions (i.e., spatial constraints) has the potential to improve the performance of current microsleep detectors.||