Noise Segmentation and Selection of Optimal Frequency Bands

will create huge muscle movement, eye movement, and blink artifacts to the non-invasive measurement of brain potentials. The other noise includes the movement of the 6-DOF motion platform and line noise. Assume these noises of muscle activity, eye, and, cardiac signals are not time locked to the EEG activity, i.e., they are temporal independent; it is very suitable to use ICA to separate the EEG signals from the “mixed” recordings other than using low pass filter. In this chapter, after ICA training, we can obtain 33 ICA components u(t) decomposed from the measured 33-channel EEG data x(t). Fig. 4-15 shows the scalp topographies of ICA weighting matrix corresponding to each ICA component by spreading each wi,j into the plane of the scalp, which provides information about the location of the sources, e.g., eye activity was projected mainly to frontal sites, and the drowsiness-related potential is on the parietal lobe to occipital lobe, etc. We can observe that the most artifacts and channel noises included in EEG recordings are effectively separated into ICA components 1, 2, and 3 as shown in Fig. 4-15. The ICA components 8, 11, and 17 may be considered as effective “sources” related to drowsiness, which will be examined by the correlation analysis.

It is more conservative estimation to just remove possible artifact components than choosing

“sources” components only in avoidance of making erroneous judgments. Thus, the

“corrected” EEG signals can be obtained by re-projection from the ICA components after removing possible “artifact” components using Eq. (4-6) as follows:

)

Besides, the ICA can be used to locate possible positions of the “drowsiness” sources.

Fig. 4-16 shows the resulting correlation spectra of subject 2 in 33 ICA components. The

horizon axis indexes frequency bands between 1 and 40 Hz and the vertical axis indexes the ICA components. The correlation spectra shows a strong evidence between fluctuations in ICA bandpower of frequency bands within 4 to 25 Hz and driving performance with high positive correlations in ICA components 8 and 17. As driving error increases, so does ICA bandpower. Fig. 4-17 show the spatial distributions in scalp topographies of weighting matrices for dominant ICA component 8 that was centered near CPz (22^th) channel and ICA component 17 that was centered on Pz (28^th) /Oz (32^th) channels. The correlations are particularly strong at central and posterior areas, which are similar to the results of previous studies in the driving experiments [32, 34]. For practice and routine application, EEG-based cognitive assessment systems should use as fewer EEG sensors as possible to reduce the preparation time for device wiring and computational cost for continuous alertness level estimation in near real time. According to the analysis shown in Figs. 4-16 and 4-17, the relatively high correlation coefficients with driving performance suggest that it is adequate to use the EEG signals at center position of dominant ICA components to assess the alertness level of subjects continuously.

Figure 4-15. Scalp topography of ICA weighting matrix w_i,j by spreading each w_i,j into the plane of the scalp corresponding to the jth ICA components based on International 10-20 system.

8-12 Hz @ Com. 8 & 17

Frequency (Hz)

Correlation of ICA Power and Driving Performance

ICA Components Index

8-12 Hz @ Com. 8 & 17

Frequency (Hz)

Correlation of ICA Power and Driving Performance

ICA Components Index

Figure 4-16. Correlation spectra between smoothed driving performance and log power spectra of 33 ICA components of Subject-2. It is observed that the bandpower spectra between frequency bands 8~12Hz have highest positive correlation with driving performance in both 8^th and 17^th ICA components.

CPz Pz

CPz Pz

Figure 4-17. Scalp topographies of ICA weighting matrices for dominant components 8 and 17. Note that the CPz channel and Pz channels are at the center position of these two ICA components, respectively.

In this section, we compared the correlation between log subband power spectra and driving error for each frequency bands and individual ICA component to find the optimal subbands and localizations of electrodes according to the scalp topographies of ICA weighting matrices. Previous studies [40, 49-52] showed that it is not applicable to use full EEG frequency bands to accurately estimate individual changes in vigilance and driving error because of the artifacts and individual variability in EEG dynamics accompanying loss of alertness. Even though information about alertness may be distributed over the entire EEG spectrum. Table 4-2 shows the correlation coefficients between different frequency bands of the ICA component 11 or 13 and the driving error of subject-3 in different experimental sessions. The ICA weighting matrices after training were held and used in the testing sessions on different days. The results show the better frequency bands of ICA components 11 and 13 are from 10 to 14 Hz with the correlation rate up to 0.94. Table 4-3 lists the correlation results

and their scalp topographies shown in Figs. 4-14 (c) and (d) demonstrated that the most alpha waves with positive correlation related to micro-sleep could be observed at occipital and central sites. Table 4-4 shows the optimal 2 ICA components and frequency bands ranges corresponding to different subjects according to the higher correlation coefficients between the log subband power spectra and the driving performance. The best frequency bands are 5-9 Hz both in 17^th and 28^th ICA components for subject 1, and 8-12 Hz both in 17^th and 8^th components for subject 2, etc. Table 4-4 demonstrated that the better frequency bands and the ICA components are not the same for different subjects.

The above analyses provide strong and converging evidence that changes in subject alertness level indexed by driving error during a driving task are strongly correlate with the changes in the ICA power spectrum at several frequencies located at central and posterior sites. This relationship is stable over time in different sessions of the same subject, but relatively variable between subjects. These results are consistent with the findings from a simple auditory target detection task reported in [18, 83]. These findings suggest that maximal accuracy the estimation algorithm should be capable of adapting to individual differences in the mapping between EEG and alertness..

Table 4-2.

The correlation coefficients between the log subband power spectra and the driving error of subject 3 corresponding to different frequency bands from 8 to 15 Hz of ICA component 11 and 13 in the training and testing sessions using the same ICA weighting matrices obtained from the training session.

Band ICA

Component Index

8 Hz 9 Hz 10Hz 11Hz 12Hz 13Hz 14Hz 15Hz

Training 0.82 0.89 0.92 0.92 0.92 0.92 0.89 0.87 Testing-1 0.86 0.88 0.88 0.88 0.87 0.86 0.83 0.82 Testing-2 0.79 0.87 0.90 0.92 0.91 0.91 0.86 0.78 Com 11

Testing-3 0.78 0.90 0.93 0.93 0.93 0.94 0.94 0.91 Training 0.77 0.88 0.90 0.91 0.92 0.91 0.90 0.86 Testing-1 0.87 0.90 0.90 0.89 0.88 0.87 0.84 0.80 Testing-2 0.75 0.87 0.87 0.90 0.90 0.88 0.85 0.79 Com 13

Testing-3 0.76 0.89 0.91 0.92 0.93 0.92 0.92 0.89

Table 4-3.

The correlation coefficients between log subband power spectra and the driving error of subject 3 using five best frequency bands (from 10 to14 Hz) corresponding to different single ICA component. The same ICA weighting matrices obtained from the training session were used for testing session performed in the other day.

ICA component 5 11 13 24 26 29 31

Training 0.84 0.93 0.92 0.82 0.89 0.82 0.79

Testing 0.80 0.92 0.91 0.82 0.88 0.78 0.78

Table 4-4.

The optimal 2 ICA components and frequency band ranges corresponding to different subjects according to the higher correlation coefficients between log subband power spectra and the driving performance.

Subject Subject 1 Subject 2 Subject 3 Subject 4 Subject 5

ICA Components 17, 28 17, 8 11, 13 4, 5 22, 25

Bands 5-9 Hz 8-12 Hz 10-14 Hz 4-8 Hz 8-12 Hz

4.6.5. Drowsiness Estimation Based on Log Bandpower of ICA

在文檔中 基於腦波之駕駛員認知反應估測及其在安全駕駛的應用 (頁 109-115)