The flowchart of the utilized EEG data analysis procedure is showing in Fig 3-1. The EEG data were preprocessed using a simple low-pass filter and a high-pass filter with cut-off frequency above 50 Hz and below 0.5 Hz, respectively, to remove 60Hz line noise, high-frequency artifacts and electrogalvanic signals before further analysis. Firstly, independent component analysis (ICA) was applied to decompose EEG signals into various temporally statistical independent activations (ICA components) and calculated the moving-averaged log power spectra of the resultant ICA components. Then we clustered the components of all volunteers to find the stable and inter-subject consistency components.
Finally, we used the sorted spectral analysis to investigate the EEG dynamic changes from alertness to drowsiness. Details of the utilized analysis methods mentioned in above are introduced as flowers.
3.1 Independent Component Analysis
ICA is a signal processing technique that separates multi-channel observation data into temporally independent stationary sources by the obtained un-mixing matrix after training [26]. By using ICA, we attempt to remove EEG artifacts and extract EEG sources in different brain areas that involve in driving or alertness level changes.
ICA methods have been extensively applied to the blind source separation problem since the 1990s [27-30]. Subsequent technical reports [31-37] demonstrated that ICA was a suitable solution to the problem of EEG source segregation, identification, and localization. In this study, we used an extended version of the infomax algorithm of Bell and Sejnowski [38] that can separate sources with either super- or sub-Gaussian distributions, to decompose distinct brain activities.
The ICA is a statistical “latent variables” model with generative form:
)
where A is a linear transform called a mixing matrix and the s are statistically mutually i independent. The ICA model describes how the observed data are generated by a process of mixing the components s . The independent components i s (often abbreviated as ICs) are i latent variables, meaning that they cannot be directly observed. Also the mixing matrix A is assumed to be unknown. All we observed are the random variables x , and we must estimate i both the mixing matrix and the IC’s s using the i x . i
Therefore, given time series of the observed data x(t)=
[
x1(t) x2(t) Λ xN(t)]
T in N-dimension, ICA will find a linear mapping W such that the unmixed signals u(t) are statically independent.u(t)=W x(t). (2) After ICA training, we can obtain 30 ICA components u(t) decomposed from the measured 30-channel EEG data x(t) (2 of the 32 channels recorded by the left and right mastoid electrodes were the reference).
).
Fig 3-2 shows an example of the scalp topographies of ICA weighting matrix W corresponding to each ICA component by projecting each wi,j onto the surface of the scalp, which provides spatial information about the contribution of each ICA component (brain source) to the EEG channels.
Fig 3-3 shows the time course signals, scalp maps and power spectra of some typical independent components representing different types of artifacts and EEG sources. Fig 3-3 (A) shows the eye blink component which had some large peaks and its physiological origin is
from far frontal site. Fig 3-3 (B) shows the horizontal eye movement component which had large fluctuations and the physiological origin is also from far frontal site. Figs 3-3 (C) and (D) show the temporal muscle component and the channel noise component that also had peaky activations and without spread scalp maps. In addition, there were spectral peaks above 20 Hz for temporal muscle component shown in Fig 3-3 (C). Fig 3-3 (E) shows the EEG source whose scalp map spreads smoothly. Hence we removed the artifact components including eye blink components, eye movement components, temporal muscle components, and channel noise components in our experiment through independent component analysis [39-41]
3.2 Smoothed Power Spectral Analysis
Moving-averaged spectral analysis of the EEG data of the extracted ICA components was first accomplished using a 750-point Hanning window with 250-point overlap. Windowed 750-point epochs were further subdivided into several 125-point subwindows using the Hanning window again with 25-point step. Each 125-point frame was extended to 256 points by zero-padding to calculate its power spectrum by using a 256-point fast Fourier transform (FFT), resulting in power-spectrum density estimation with a frequency resolution near 1 Hz.
A moving median filter was then used to average and minimize the presence of artifacts in the EEG records of all sub-windows. Previous studies [42,43] show that the transient amplitudes of EEG power spectrum involved in wake-sleep regulation are very different. The cortex produces low amplitude and fast oscillations during waking, and generates high-amplitude, slow cortical oscillations during the onset of sleep. Their reports also showed that the EEG spectral amplitudes correlated with the wake-sleep transition more linearly in the logarithmic scale than in the linear scale. Thus, the ICA power spectra were further converted into a logarithmic scale. The resultant time series of ICA log power spectra for each session consisted of the power spectra of 30 ICA components across 40 frequencies (from 1 to 40 Hz)
stepping at 2-second (500-point, an epoch) time intervals [18]. Fig. 3-4 shows the smoothed spectral analysis procedure.
Since alertness level fluctuates with cycle lengths longer than 4 minutes [11, 44], we smoothed the ICA power spectra and the driving performance time series by using a causal 90-second square moving-averaged filter to eliminate variances at cycle lengths shorter than 1–2 minutes. The smoothed driving performance was called “local driving error (LDE)” as shown in Fig 3-5. The LDE is an indirect index of the alertness level and we will assess the relationships between subject’s local driving error and his/her smoothed ICA log power spectra to investigate human’s EEG spectral changes from alertness to drowsiness in driving.
3.3 Independent Component Clustering
In order to find the stable and inter-subject consistency sources related to alertness changes, we clustered the EEG sources of all volunteers. The components of all volunteer were clustered semi-automatically based on the gradients values:
[
Gx
i,Gy
i]
(4) of the component scalp maps [41]. K-mean algorithm [45] was utilized for clustering. The K-mean clustering is to classify or to group objects based on attributes/features into K number of groups. K is a positive integer number. The grouping is done by minimizing the sum of squares of distances between the data and the corresponding cluster centroid as:∑
− 2= i i k
k
x y
e
(5) where e represent the square error, k x andi y represent the data point and cluster centers, k respectively. Fig. 3-6 shows the diagram of component clustering analysis.For each ICA activation map, we perform an EEG source localization procedure to locate its single dipole. By localizing multiple dipoles independently, we substantially reduce our search complexity and increase the likelihood of efficiently converging on the correct solution.
The independent EEG processes and their equivalent dipole source locations were obtained by using the EEGLAB toolbox (Makeig, et al., 2004).
3.4 Sorted Spectral Analysis
Since the LDE is an indirect index of the alertness level, we propose the sorted spectral analysis method that sorts the smoothed ICA log power spectra according to the LDE index to assess the brain dynamics corresponding to the transition from alertness (lower LDE values) to drowsiness (larger LDE values) in driving. Fig. 3-7 shows an example of the sorted spectral analysis. The left subplot of Fig. 3-7 is a subject’s original LDE trajectory (the blue line) and the corresponding alpha power changes (the red line). The right subplot sorts the LDE values in ascending order and shows the transient alpha powers corresponding to the sorted LDE values. It can be found that the alpha power is increasing at the beginning and will decrease at the latter when LDE values are ascending. According to our experimental results presented in Chapter 4, the power changes of some ICA component clusters accompanying with the LDE increasing can be obviously observed. It is noted that 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.