EEG data analyses

The fig. 2-6 shows the flowchart of EEG data analysis. The detail procedures were described as the followings.

2.6.1 EEG Data pre-processing

EEG signals were first filtered with a simple low-passed and a high-passed filter

with the cut-off frequencies at 50 Hz and 0.5 Hz to remove the line noise, other high-frequency noises and electrogalvanic signals. The filtered EEG signals were screened and rejected grossly data contaminated by other artifacts including muscle contractions the body movement artifacts, and bad channels before the further EEG analysis (fig. 2-7).

2.6.2 Independent Component Analysis

Independent component analysis (ICA) method (fig. 2-8) has extensively applied to blind source separation problem since 1990s [36]-[44]. Subsequent technical reports [45]-[52] demonstrated that ICA is a suitable solution to the problem of EEG source segregation, identification, and localization. Applying ICA algorithm on EEG raw data in drowsiness experiment could remove the EEG artifacts including the eye blinking as well as the eye movements and could further extract EEG sources associated with human drowsiness. In this study, 43-independent components (ICs) were calculated from applying ICA algorithm on 43-channel EEG data. The fig. 2-9 shows the 43 ICs of a single subject (subject 1). The IC-1, -2,-3, -4, -6, -8, -11 were identified as EEG sources and the IC-5, -10, -13, -14, -16, -17, -18, -19, 20, -21, -23, -24, -25, -27, -28, -31, -33, -34, -35, -36, -37, -38, -39, -40, -41, -42, -43 were identified as noise components. The frontal central midline (FCM), occipital-midline (OM) and bi-lateral occipital (BLO) components were selected by visual inspection of scalp topography and then the component signals were used for advancing analysis. In order to extract EEG source related to drowsiness only from forehead, we also applied the ICA algorithm on forehead 15 channels. Although these forehead EEG channel signals were similar on time domain and their topographical locations were very close, ICs related to drowsiness could still be decomposed from these similar channel

signals after applying the ICA algorithm.

2.6.3 Dipole source localization

In order to find the stable and inter-subject consistent sources related to the drowsiness, the FCM and OM ICs were selected from all volunteers. For each IC activation map, we performed an EEG source localization procedure to locate its single dipole. By localizing multiple dipoles independently, we substantially reduced our search complexity and increased 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 [53].

2.6.4 Moving-averaged power spectra analysis

The fig. 2-10 shows the moving-averaged spectral analysis of the extracted 43 ICs/channel data. Signals were first accomplished by using a 750-point Hanning window overlapped with 250-point. The 750-point epochs were further subdivided into several 125-point sub-windows using the Hanning window again advanced by 25-point step. Each 125-point sub-window was zero padded into 256 points and applied by a 256-point fast Fourier Transform (FFT). A moving median filter was then used to minimize the presence of artifacts in all sub-windows. The moving-averaged power spectra were further converted into a logarithmic scale for spectral correlation and driving performance estimation. Each session were consisted of 43 ICs logged power spectra were then estimated across 50 frequencies (from 1 to 50 Hz) stepping at a 2-second (500-point, an epoch) time interval. The fig. 2-12A shows the temporal changes of the theta-band power, which is smoothed using a

causal 90-second square moving-averaged filter advancing at 2-second steps, and the LDE.

2.6.5 Correlation analysis

To select drowsiness-related components, we correlated the smoothed IC power spectrum with the LDE. Since the alertness level was fluctuated with cycle lengths longer than 4 minutes, we smoothed the temporal profile of the power spectra and driving performance with a causal 90-second square moving-averaged filter. The Pearson’s correlations coefficients between changes in the ICs log power spectrum

and driving performance at each EEG frequencies are expressed as .

The ICs with the highest correlation coefficients among several frequency bands located in FCM and OM were selected for drowsiness estimation. The same correlation estimating processes were also applied on 15 forehead components.

2.6.6 Feature extraction and drowsiness estimation

In this study, we used a multivariate linear regression model [54] to estimate the subject’s LDE based on the power spectra of several signals (FCM, OM, forehead ICs, and forehead channel). The 5-Hz band power with the highest correlation coefficient was selected from each subject as inputs to train the individual linear regression model. For each subject, the ICA unmixing matrix obtained from the training session was applied on testing session which the data were collected in different day. The estimated LDE traces were obtained by multiplying the selected features obtained from the testing session and the linear regression model. The fig. 2-11 shows that the

estimation results of a single subject (subject1). The correlation coefficients between the estimated and actual LDE traces were 0.85 and 0.9 for the within-session testing and the cross-session testing respectively.

2.6.7 LDE sorted spectral analysis

The power changes in theta and alpha bands were sorted according to LDE. The LDEs from 5 to 85 were evenly divided into 80 bins and the corresponded alpha- and theta-band power in each bin was averaged. The fig. 2-12B shows the average LDE sorted theta band power.

2.6.8 Coherence analysis

The EEG coherence is a normalized measurement of the coupling between two signals at any given frequencies [55][56]. The coherence value of each frequency bin was calculated by

, which is the extension of Pearson’s correlation coefficient to complex number pairs.

In this equation, f denotes the spectral estimate of two EEG signals x and y for a given frequency bin (λ). The numerator contains the cross-spectrum for x and y (fxy), while the denominator contains the respective autospectra for x (fxx) and y (fyy). For each frequency bin (λ), the coherence value (Cohxy) is obtained by squaring the magnitude of the complex correlation coefficient R. This procedure yields a real number between 0 (no coherence) and 1(maximal coherence). In this study, the coherence of FCM and OM IC signals were evaluated from 20 sessions across 10 subjects.