Previous studies showed that the strongest response related to human drowsiness was mainly found in the occipital lobe with alpha power increases. However, in order to improve the acquisition procedure in the future, brain activities were not only collected with traditional 32 channel electrode cap but also from subjects’ forehead.
We first applied ICA on the collected EEG signals (including the 15 forehead channels and the 28 traditional channels) to isolate brain sources. Then, coherence analysis was used to find out the coupling relationship between FCM and OM component.
Two ICA components, FCM and OM, were selected to compare with the 15 forehead components and the 15 forehead channels. The ICA sub-band power of FCM, OM, forehead components and forehead channels were extracted from the first session experiment of each subject. The sub-band powers were then used as features to estimate the driver’s drowsiness state with linear regression model.
In order to investigate the relationship between power changes and LDE, the LDE were then sorted to find out the power change from alert to drowsy. The power changes of FCM, OM, and forehead components were then characterized by comparing their power responses. The following paragraphs showed detailed results.
3.1 OM and FCM ICs clusters
Plenty of brain sources involved in the drowsiness experiments. The ICs were extracted by applying ICA algorithm to 43-channel EEG artifact-free data. According to previous studies, the drowsy-related brain activities were found in the occipital lobe, thus we first focus on the brain activities in this region.
In order to discuss the variance of component in each session, topography map,
power spectrum, dipole location of OM in all sessions were grouped in one figure. Fig.
3-1 shows the power spectrum, dipole location and scalp topographies of ICA back-projection matrix W-1 of OM components in 20 sessions. Twenty OM components provide strong evidence that EEG source occur on occipital lobe is stable among all subjects during drowsy experiment, and correlation coefficient between each scalp map and mean scalp map ICA back projection matrix W-1 is 0.87±0.2. The spectrum of each OM component and pair correlation coefficient is 0.904±0.6 (panel B). Small scalp map of panel C shows that the OM components appearing in all the 20 sessions, and the bigger scalp topography in the left-upper corner is the scalp map of average ICA back-projection matrix W-1. Note that dipole of OM components mainly located on occipital lobe, but there are four dual-dipoles that belong to bilateral occipital component in Fig. 3-1.
In addition to OM component that activated from alert to drowsy, FCM components were found stable appearing in our twenty sessions (Fig. 3-2). Fig. 3-2C shows the average FCM scalp map and FCM scalp maps appearing in each session, and the correlation coefficient matrix of W-1 is 0.94±0.06. Fig. 3-2B shows the spectrum of each FCM component and pair correlation coefficient is 0.917±0.25.
Spectrum profile of OM component had a peak in alpha band, but the spectrum characteristics in FCM had a peak in theta band and 15 Hz. In addition to spectrum and scalp topographies, fig. 3-2A shows the different view point of dipole source, dipoles of FCM component mainly located at the midline of frontal lobe, but some dipoles located near thalamus due to indirect channel locations or experimental setup error. Overall, OM and FCM components were found and having stable profile in dipole location, spectrum and topography during the drowsiness experiment.
According to our results, OM and FCM components can be used as EEG features to estimate driver’s drowsiness.
3.2 Coherence between FCM and OM component
In order to figure out the synchronization between FCM and OM ICs, coherence analysis was applied to these two ICs. The coherence value was counted between the one-hour signals from FCM&OM component in each session. Fig. 3-3 illustrates the mean coherence in 1~30 Hz between FCM & OM from 20 sessions, the coherence value donates the coupling degree between two time series. The coherence value in alpha band (around 11Hz) is lower than other frequency band, but higher in theta band (4~7Hz), and the peak appears at 15 Hz. Coherence result shows that phase synchronization in theta band and 15 Hz is higher than alpha band between FCM and OM ICs. Phenomenon appearing in frontal and occipital lobe is similar in theta band.
3.3 Relationship between FCM Component and Forehead Component
In order to collect more signal from frontal lobe, forehead patch were developed.
The 15 forehead components were extracted from the 15 forehead EEG signals.
Although these forehead EEG channel signals were similar on time domain and their topographical locations were very close, but ICs related to drowsiness could still be decomposed from these similar channel signals after applying the ICA algorithm.
The 15 forehead components were decomposed from 15 forehead channel EEG signals, and the forehead component with highest correlation coefficient between spectrum and LDE was selected and then back-projected to the 15 forehead channels with matrix W-1. The result of back projection with the forehead component was shown in Fig. 3-4. These 15 color blocks present the ICA back-projection matrix W-1 value of a forehead component, and figure shows that the component weight is higher
in superior way (away from eye) and symmetry in the vertical direction (channel 13).
Furthermore, correlation coefficients between forehead and FCM component signals on the time course were calculated (0.64±0.17). In addition to FCM component, correlation coefficient between forehead signal and OM component time series is 0.06±0.04.
The forehead components were proved highly correlated to the FCM components, while the correlation coefficients between forehead components and the OM components are very low. These results suggest that there is a brain source located in frontal lobe response to drowsiness and it can be easily collected from forehead EEG signals.
3.4 LDE Estimation
In last section, we have proved that the forehead components were highly correlated to FCM components and thus can also be used as features to estimate subjects’ cognitive state.
In order to monitor subject’s cognitive state, LDE is an indirect index to measure subject’s cognitive stage. The car in the VR scene was designed to drift from the middle of driving lane to test the subjects’ cognitive response. The distance between the middle of the car and the middle of cruising lane was defined as local driving error, which can be used to study the subjects’ cognitive state. For example, if the subject is alert, the car-drifting can be detected and LDE can be minimized by the subject with the steering wheel. In contrast, LDE will become very large if the subject is drowsy.
We use a least-square multivariate linear regression model to estimate subject’s LDE according to the information obtained from the sub-band power spectra analysis of ICs and EEG channel. The OM and FCM components decomposed from 43-ch
EEG signals, the 15 forehead channels and the 15 forehead components were used as drowsiness-related brain signals in this study. For each signal source, the optimal frequency bands were selected according to the correlation coefficients between ICA power spectrum and LDE in the training session. The single-subject model was trained with the EEG features that extracted from the first session experiment for each subject and then used to estimate subject’s driving performance with the EEG features in second session. The results shown in Table 1 are the correlation coefficients between the estimated driving error and the real LDE acquired in the second session.
In reverse, the EEG signals and LDE collected in the second session were used to train the model and the features extracted from the first session signals were used to estimate the driving error. (i.e. S2 est S1)
In Table 1, the averaged correlation coefficient between the estimated driving error and the real driving error is highest (0.89±0.04) when the features were extracted from the OM components, and it’s 0.83±0.07 when FCM features were used.
The result suggests that there is another drowsiness-related brain source located in the frontal lobe. Besides, the forehead signals yield better estimation accuracy than FCM component, which means even without complicated application of electrode cap, forehead signals can be used to estimate the subjects’ cognitive state. The correlation coefficients were very similar when the forehead components and the forehead channels were used, the application of ICA thus become unnecessary for this study. In order to investigate the difference between results of four groups, we first applied the Kolmogorov-Smirnov test (KS test) to each group, all groups are not normal distribution, and second we applied the Wilcoxon Signed-Rank test between the groups. The difference between the OM and the other three groups are significant (p<0.05). Furthermore, there is no significant difference between FCM ICs, forehead ICs, forehead channel through the Wilcoxon Signed-Rank test.
According to our results, the OM lobe is the area that most correlated to drowsiness, but frontal lobe features are also good for the estimation of driving performance. These results suggest that we can detect drowsiness by using a single forehead channel and thus save lots of preparation comparing to wearing the electrode cap.
3.5 LDE sorted spectral analysis of ICA components
In order to investigate the brain dynamics corresponding to the transition from alertness (lower LDE) to drowsiness (larger LDE) in the experiments, the ICA log power spectra were sorted according to their LDE. The smoothed results were given in Fig. 3-5.
The sorted LDE values were shown in ascending order in x-axis and the transient frequency band mean power and standard deviation power corresponding to the sorted driving performance values were shown in y-axis. It can be found that the alpha power increased sharply at lower LDE (around 30) and starting decrease latter as ascending driving performance in OM and forehead component. Oppositely, alpha power of FCM increased 2 dB from LDE 5 to 85. Changes in theta band of all components increase from LDE 5 to 85, but OM component have higher increasing power. Power in theta band increased 6 dB from LDE 5 to 85, and this phenomenon can be used to detect driver’s cognitive state from alert to drowsy. In addition, theta band is a better index than alpha band.