3. DATA ANALYSES
3.2. EEG
3.2.6. Independent modulator (IM) decomposition
Here, we present first results of a new method for decomposing fluctuations in the selected independent component (IC) power spectra into independent modulators. The method was applied to the selected IC activity spectra from each subject. Mean logarithmic power at each frequency was subtracted from each single window power spectral estimate. The resulting time series of logarithmic spectral deviations were then concatenated, giving a matrix of size (f c, t), where f the number of frequency bins, c the number of subject ICs, and t is the number of time windows. For each subject, this matrix was reduced to its first 10 principal dimensions by PCA. The dimension-reduced log spectral data were then decomposed by infomax ICA to find independent modulators of log spectral power across subsets of time windows within all ICs. Infomax ICA finds a matrix, W, that linearly unmixed the IC spectral activations, x, into a sum of maximally temporally independent, and spatially fixed modulators, u, such that u = Wx. The rows of the resulting ‘activation’ matrix, u, are the independent modulator activations, and its columns, the time points of the input data.
Columns of the inverse matrix, W-1, give the relative projection weights from each independent modulator to each frequency bin of each component. The projection weights, which are the frequent patterns, of the modulator provide information about the modulating frequencies of the modulator.
Fig.3-11. shows the relation between single subject comodulation analysis and component clusters. Different subject has different components. For example, S4 in this figure has only component of frontal. Then we fill in a 'V' marks in the grid of the frontal scalp map.
It shows that the mean cluster maps are inserted in the 1st row. Then we take the spectral activates of the selected components to do single subject comodulation analysis. The procedure of the single subject comodulation analysis is shown in Fig. 3-12.
Figure 3-11: The relation between single subject comodulation analysis and component clusters. Different subject has different components. For example, S4 in this figure has only component of frontal. Then we fill in a 'V' marks in the grid of the frontal scalp map. It shows that the mean cluster maps are inserted in the 1st row. Then we take the spectral activates of the selected components to do single subject comodulation analysis. The procedure of the single subject comodulation analysis is shown in Fig. 3-12.
Figure 3-12: The independent modulator decomposition procedure. The independent modulator decomposition procedure can be divided into five steps. First, the selected dipolar ICs is divided into 1s windows FFT window power spectra then transformed to log power.
Each colored trace represents the power spectrum for a single 1s window. The thick black line is the mean power spectrum of all windows. Third, the mean is removed from each power spectrum. Forth, these data is converted into matrix format. We concatenate this tall matrix from one IC with the same sort of information from all the other selected components, and then we have a matrix of dimensions 'spectra (or frequencies)' by 'spectral windows'. This matrix was then submitted to PCA for dimension reduction to 10 dimensions and then to ICA to find independent spectral modulators from the mean across these selected components.
Each row is an IC and each column in this illustration is an IM. What come out of this decomposition are projection weights of 10 independent modulators. The projection weights, which are the frequent patterns, of the modulator provide information about the modulating frequencies of the modulator.
Fig. 3-13 shows how to plot the frequency weights of the components from each subject for 1 of the modulator. As described in Fig.3-11, it’s needed to fill up the table. In this figure 'V' marks means the frequency weight of each component and the mean frequency weight is plotted by red thick line.
Figure 3- 13: Method of plotting the frequency weights of the components from each subject for the theta-beta modulator. In this figure 'V' marks means the frequency weight of each component and the mean frequency weight is plotted by red thick line.
3.2.7. LDE-Sorted IM activation analysis and statistical analysis
In order to find the inter-subject relationships between the IM activations and the alertness level, the LDE-sorted analysis method was applied to the IM activation across subjects. The method sorts the smoothed IM activations according to the LDE index to assess the brain dynamics corresponding to the transition from lower LDE to larger LDE. For group analysis, we assumed the alertness levels of all subjects in the lowest LDE states were the same and the difference of the lowest LDE values corresponding to different subjects are caused by the individual reaction speed. To compare the modulator power for the high and slow local driving error across subjects, a paired-sample Wilcoxon signed rank test (signrank, Matlab statistical toolbox, Mathworks) was applied. The significant onset of the alpha increase was earlier than that of the theta increase. All statistical comparisons in this study, a significant level was set at p <0.05.
Figure 3-14: An example of the sorted spectral analysis. The left subplot of Fig 3-6 is a subject’s original LDE trajectory (the blue line) and the corresponding modulator power changes (the red line). The right subplot sorts the LDE values in ascending order and shows the transient modulator powers corresponding to the sorted LDE values. It can be found that the modulator power is increasing at the beginning and will decrease at the latter when LDE values are ascending.
4. Results
4.1. Behavior performance
All subjects’ LDE were ranged from 0 to 65 units, indicated that subjects got drowsiness in our experimental paradigm. Fig. 4-1 shows the plots of the sorted trials by response time in each subject. The performance was changes almost continuously, and we can analyze the EEG power changes accompany with LDE variations continuously.
Figure 4-1: The plots of sorted trials by response time of each subject.
Table 4-1: Subject list
4.2. Component spectral fluctuations related to performance changes
The grand results show that the trends of alpha and theta power changes from good performance to poor performance were similar between different brain regions. The performance changes are shown in Fig. 4-2. From the above result, we compared the EEG fluctuations in time series between different components of intra-subject to confirm that the drowsiness related alpha and theta rhythm in these components may be modulated by the same nucleus or synchronized by cortical-cortical interaction. ICA power fluctuations of parietal, occipital, frontal and central components in time series were shown in the Fig. 4-2.
The result shows that the alpha power fluctuation of parietal component was highly correlated with the fluctuation of occipital component. Additionally, the theta power fluctuations of frontal, parietal and central component were highly correlated with each other.
Figure 4-2: The single subject results of performance change accompanying the component power changes. It shows that the trends of alpha and theta power changes from good performance to poor performance were similar between different brain regions. Hence, we compared the EEG fluctuations in time series between different components of intra-subject to confirm that the drowsiness related alpha and theta rhythm in these components may be modulated by the same nucleus or synchronized by cortical-cortical interaction. Scalp maps on the bottom right show the ICA power fluctuations of parietal, occipital, frontal and central components in time series. The result shows that the alpha power fluctuation of parietal component was highly correlated with the fluctuation of occipital component. Additionally, the theta power fluctuations of frontal, parietal and central component were highly correlated with each other.
4.3. Component clustering results
We clustered all components of 17 subjects into 7 groups, and showed the remarkably meaningful and consistent 5 clusters from Fig. 4-3 to Fig.4-7, with the averaged scalp maps, each scalp map and the dipole source locations in the clusters. Where Fig. 4-3 shows frontal cluster, Fig.4-4 shows occipital cluster, Fig. 4-5 shows right motor cluster, Fig. 4-6 shows left motor cluster, and Fig. 4-7 shows parietal cluster. The results of clustering analysis and dipole fitting displayed that most of the brain areas involved in the lane-keeping driving task. These cluster are more remarkable and stable between different participants in the lane-keeping driving task. Hence, only the components that include in these cluster clusters were for further analysis.
In each figure, equivalent dipole source location, spectra and scalp maps for independent component clusters are shown. Scalp maps are shown on the top. Dipole source location is shown on the bottom right. Spectra are shown on the bottom left.
Figure 4-3: Equivalent dipole source location,
spectra
and scalp maps for independent component clusters of frontal cluster.Figure 4-4: Equivalent dipole source location,
spectra
and scalp maps for independent component clusters of occipital cluster.Figure 4-5: Equivalent dipole source location,
spectra
and scalp maps for independent component clusters of right motor cluster.Figure 4-6: Equivalent dipole source location, spectra and scalp maps for independent component clusters of left motor cluster.
Figure 4-7: Equivalent dipole source location, spectra and scalp maps for independent component clusters of parietal cluster.
4.4. Single Subject Independent modulator (IM) decomposition Results
In the section, we discuss the decomposition results of independent modulator. In Fig.4-8., representative independent modulation patterns from one subject. IM window weights and frequency characteristics derived by multiplying the series of spectral deviations from the mean in each 1-s overlapping time window with the PCA/ICA unmixing matrix.Each row represents a computed IC, and each column an IM. The figure shows that IM1 modulated all ICs in alpha-band, and IM2 modulated all ICs in theta-beta-band. The leftmost row shows the IC scalp maps, and top column, the IM window weight histograms. The numbers in the bottom are the correlation coefficient between IM activation and LDE.
Figure 4-8: Representative independent modulation patterns from one subject. IM window weights and frequency characteristics. It is derived by multiplying the series of spectral deviations from the mean in each 1-s overlapping time window with the PCA/ICA unmixing matrix. Each row represents a computed IC, and each column an IM. The figure shows that IM1 modulated all ICs in alpha-band, and IM2 modulated all ICs in theta-beta-band. The leftmost row shows the IC scalp maps, and top column, the IM window weight histograms. The numbers in the bottom are the correlation coefficient between IM activation and LDE.
4.5. Frequency characteristics of the independent modulators
Fig. 4-9 shows the normalized frequency patterns of two stable modulators in each cluster. Top panels in Fig. 4-9 show the averaged scalp maps of the clusters, which we obtained in component clustering. Middle panels in Fig. 4-9 display frequency characteristics of the one stable modulator, and bottom panels in Fig. 4-9 exhibit another one. The modulated frequency patterns of corresponding ICs were derived from the column of the inverse matrix, W-1, during the procedure of single subject Independent modulator (IM) decomposition. We normalized the frequency characteristics to observe the common characteristic of each IM.
The thin red lines in Fig. 4-9 indicate the modulating frequency of the IM to each component, and the thick red lines in Fig. 4-9 indicate the averaged frequency characteristics of the IM in the cluster. The IMs in middle panels of Fig. 4-9 modulate the theta band, and reveal a peak near 15 Hz in the patterns. This theta dominant modulator affected all areas in our experiment design. The IMs in bottom panels of Fig. 4-9 demonstrate the alpha band modulated in very wide areas.
Figure 4-9: The normalized frequency patterns of two stable modulators related to alertness changes in each cluster. Top panels show the averaged scalp maps of the clusters. Middle panels display frequency characteristics of the one stable modulator, and bottom panels exhibit another one. The modulated frequency patterns of corresponding ICs were derived from the column of the inverse matrix, W-1 , during the procedure of single subject Independent modulator (IM) decomposition. We normalized the frequency characteristics to observe the common characteristic of each IM. The thin red lines indicate the modulating frequency of the IM to each component, and the thick red lines indicate the averaged frequency characteristics of the IM in the cluster. The IM in middle panels modulate the theta band, and reveal a peak near 15 Hz in the patterns. This theta dominant modulator affected all areas in our experiment design. The IM in bottom panels demonstrate the alpha band modulated in very wide areas.
4.6. Independent modulator activities accompanying performance changes
Fig. 4-10 to Fig.4-16 show the two IM activities related to performance changes. First, we compared the intra-subject fluctuations of two IMs which may be modulated by the nucleus or synchronized by cortical-cortical interaction in time series during different alertness levels to confirm that the alpha and theta-beta modulator related drowsiness level.Letter a-c shows the fluctuations of the LDE, theta-beta modulator, and alpha modulator for two subjects in time series. In Fig. 4-10 to Fig.4-15, (a) exhibits the changes of the LDE during whole experiment, and (b) is the fluctuations of the theta-beta modulator, and (c) shows the fluctuations of the alpha modulator. The theta-beta modulator fluctuates very little during the low LDE periods, and increases monotonically from low LDE to high LDE.
Different with the theta band, the alpha modulator fluctuates very large during the low LDE periods. The correlation results between the LDE and two modulators of each subject were
Figure 4-10: The performance changes related to two IM activities of subject 1.
Figure 4- 11: The performance changes related to two IM activities of subject 2.
Figure 4-12: The performance changes related to two IM activities of subject 4.
Figure 4-13: The performance changes related to two IM activities of subject 14.
Figure 4-14: The performance changes related to two IM activities of subject 15.
Figure 4-15: The performance changes related to two IM activities of subject 16.
Figure 4-16: The LDE-Sorted IM activation on two IM activations of all subjects. It showed the mean and SD of the two LDE-sorted IM activations. Fig. 4-16a displays the result of the LDE-Sorted theta-band dominant IM activation, and LDE-Sorted theta-beta modulator activity was significantly increases monotonically from low LDE to high LDE, and the LDE-Sorted alpha modulator, which shows in Fig. 4-16b, was remarkably increases and then sustains from low LDE to high LDE. Additionally, we observed the variances of the two IM activities in the same value of LDE across 17 subjects. The variations were very small in all participants, revealing that the results of the LDE-sorted have inter-subject consistence.
Table 4-2: The correlation coefficients between two modulators and the LDE SUBJECT No. Theta-Beta Modulator Alpha Modulator
SUBJECT 1 0.7861 0.5713
Mean Value 0.83246 0.43497
Standard Deviation 0.06812 0.22137
5. Discussion
5.1 The component spectral fluctuations related to performance
The theta and alpha power changes of several components from good performance to poor performance were similar between different brain regions. It is consistent with past studies (Lin et al., 2006; Huang et al., 2008). Hence, we compared the component fluctuations in time series between different components of intra-subject to confirm that the drowsiness related alpha and theta rhythm in these components may be modulated by the same nucleus or synchronized by cortical-cortical interaction.
5.2 The modulatory model of the neural system
The modulatory model underlying this analysis was illustrated schematically in Fig. 1-1.
Independent component analysis (ICA) was applied to EEG data to identifies temporally distinct (independent) signals generated by partial synchronization of local field potentials within cortical patches (b) and summing in different linear combinations at each electrode depending on the distance and orientation of each cortical patch from the scalp (a) and reference electrode (c). The spectra of resulting cortical independent components (ICs) monotonically decrease with frequency, on average, but exhibit large and frequent variations across time. These spectral modulations may be modeled as exponentially weighted influences of several near-independent modulator (IM) processes (d) that independently modulate the activity spectra of one or more independent component (IC) signals. On converting the IC spectra to log power, combined IM influences on IC spectra are converted to log-linear weighted sums of IM influences, allowing a second linear ICA decomposition, applied to the IC log power spectra, to separate the effects of the individual IM processes (d) across EEG frequencies and IC sources (b).
investigate the modulation model. They have examined the correlation between occipital EEG alpha rhythm, selectively extracted using ICA, and fluctuations in the BOLD effect during an open versus closed eyes and an auditory stimulation versus silence condition. Occipital alpha amplitude is consistent with metabolic changes occurring simultaneously, sites in the medial thalamus and in the anterior midbrain, with about 2.5 sec lag. Goncalves et al.(2005) used simultaneous recording of electroencephalogram/functional magnetic resonance images (EEG/fMRI) to identify blood oxygenation level-dependent (BOLD) changes associated with spontaneous variations of the alpha rhythm, which is considered the hallmark of the brain resting state. They also found the BOLD signal was positively correlated with the alpha power in small thalamic areas, supporting the cortical and subcortical modulation of electroencephalographic alpha rhythm.
A combined PET/EEG study by Schreckenberger et al. also supported the hypothesis of a close functional relationship between thalamic activity and alpha rhythm in humans mediated by corticothalamic loops. (Schreckenberger et al., 2004) This functional corticothalamic loop supports the concept of cortical control of thalamic activity in humans for modulating cortical EEG activity.
Physiological processes that may produce these patterns include several brain systems regulating brain and behavioral arousal and/or valuation judgments of stimulus and other events via the release by midbrain or brainstem neurons of modulatory neurotransmitters – dopamine (DA), acetylcholine (ACh), norepinephrine (NE), serotonin, etc. – through their extensive cortical and thalamic projections (Robbins, 1997; Bardo, 1998).
5.3 The modulator fluctuations from alertness to drowsiness
In this study, we employed independent modulator decomposition to the spectral fluctuations of independent components, and two independent modulators related to performance changes were found in all subjects.
The alpha modulators were very sensitive to the performance changes, so the alpha modulator fluctuates very large during the low LDE periods. Therefore, the alpha modulator in this study might be partially contributed by cortical idling, decreased attentiveness and decline of movements. In the earlier finding (Lee et al., 1999), the frequency of the 8-14 Hz synchronizes in the thalamocortical system during quiet sleep. The EEG fluctuations from low error to high error were similarly during frontal, central-parietal and occipital lobe in the experiment. So that the alpha modulator maybe also modulated by thalamus during drowsiness.
The theta-beta modulators increase monotonically from low LDE to high LDE. Theta rhythm is the EEG characteristic of sleep stage 1 and microsleep (Bear et al., 2001; Gennaro et al., 2001; Thomas et al., 2003). Past studies also reported the fluctuations in the modulation of the beta-wave amplitude related to an indirect measurement of drowsiness (Poupard et al., 2001). Jung et al. (1997) also reported theta and beta band power were high correlated with performance changes.
The correlation coefficient between modulator power change and LDE is shown in Table 4-2. The standard deviation is large because the inter-subject behavior states are varied. Some subjects were involved more period of stage 1 sleep, others included more time of cortical idling, so that results of the independent modulator decomposition depend on drowsiness state.
6. Conclusions
In this study, we model spectral fluctuations of independent components from EEG activations as the actions of independently modulator processes, and report two main classes
In this study, we model spectral fluctuations of independent components from EEG activations as the actions of independently modulator processes, and report two main classes