Features Extraction

The features were the EEG data in Frontal and Motor component. Before we analyzed the extracted component data by SOM, the feature had to be processed.

We wanted the features clearer and less variation. The flowchart of features processing was showed in the Fig. 3-2. We proposed some methods to process our extracted data. The detail steps and meaning of each method would be described in detail.

Downsizing

We designed the stimulus onset asynchrony (SOA) experiment, especially the different time interval of the dual tasks. The extracted features were the epoch-based.

There was huge amount of information in our original EEG data. The data were epoch-based and included the information of timing. There were five thousand time points (1 second is the baseline and the other 4 seconds is phasic) in one epoch. In this research, the feature with combined with the EEG signals in Frontal and Motor

components to provide the more clear phenomenon than the data in each single component. Although the time points of EEG epochs in the two components equaled to five seconds, the dimensions of combined features were twice than the original EEG epochs in single component. We wanted to reduce the dimensions of the combined features and not lose any information about time and the perturbation of frequencies in each epoch. We could see the Fig. 3-3 to understand the detail information. Each epoch was equally divided into ten intervals. The length of phasic part in each epoch was 4 seconds so one interval was 400 milliseconds. We applied Fast Fourier Transform (FFT) for each interval to transform the signal from time domain to frequency domain.

.

Fig. 3-2: The flowchart of steps about EEG data processing and analyzing.

The two parts are feature processing and then using the processed data to train the maps by SOM algorithm. We applied these steps to process the EEG signals (as showed in the left part of this figure). Then the power spectra were the input data for

After applying the FFT, The main difference in power spectra among five cases could be observed about 5~14 Hz in Fontal component and 8~25 Hz in Motor component [17]. But there were 50 frequencies for each time point in the original EEG epoch. To reduce the dimensions of each interval, these values in the active bands by dealing with tasks were reserved. We just preserved the 1~20 Hz of Frontal component and 1~30Hz of Motor component in each interval. Then the data for Frontal (Motor) component in each interval were 20 (30) Hz and there were 10 intervals in each epoch. The features of Frontal or Motor components were reduced from 4000 time dimensions to 200 or 300 frequency and time dimensions, respectively. After this step, we got fewer dimensions and preserved the timing and frequency information in each interval.

Removing Baseline

There were four designed sessions in one complete experiment and the EEG signal was collected during one hour. There were many epochs in each session, and Fig. 3-3: The figure showed ten intervals and the baseline.

The length of each epoch from event-onset to event-offset during the experiment was 4000 milliseconds (4 seconds). One trial was divided to ten intervals so the length of all intervals was 400 milliseconds. The frequencies less than 20 (30) Hz in Frontal (Motor) component were reserved in this step. There were 50 points in each intervals and one epoch would be reduced to 500 points.

the events were presented to the subjects randomly in order to prevent anticipative [38]. Since we had designed different cases with the combination of the driving and the mathematic tasks, thus the EEG response related to different cases should be extracted from the analyzed EEG signals. To investigate the changing on brain activities between single- and dual- task conditions in a virtual environment, we just analyzed the EEG signals from onset of the event to the end of that epoch. The baseline was the mean of the EEG signal one second before the event onset. In order to investigate the changes in spectral power and the perturbations in the oscillatory dynamics of ongoing EEG, the baseline of each EEG epoch was removed by a dividing method. The unit of EEG signal is decibel (dB), and the dB is a logarithmic unit of measurement that expresses the magnitude of a physical quantity relative to a specified or implied reference level.

Because the FFT was applied to all EEG epochs, there were 50 frequency points originally. However, after the step of downsizing was processed, there were just 20/30 frequency points (1~20Hz/1~30Hz) for Frontal/Motor component. The length of baseline was 1000 milliseconds in Fig. 3-3. In other words, there were 1000 time points and 20 or 30 frequency points in each time point. The baseline was averaging all same frequencies which located at this time interval. Because dB is a logarithmic unit, each particular frequency in the EEG epoch was divided by that frequency during the baseline. After this step, we can ensure the EEG signals were main caused by the responding the tasks, excluded of reasons by the “state of mind”.

There were four sessions in one complete experiment. Each session of all experiments was set in the same circumstance, and the subjects were asked to keep the same psychological and physical situation during the experiment. However different people might not have the same phenomena for the same task; in other words,

among people. We wanted to analysis the influence of distraction instead of the difference among all subjects. In order to decrease the diversity in people and keep the variation among all five cases, we proposed this method of subtracting mean vector.

Reducing Variation

All epochs of five cases in the same subject were extracted from the data set.

There were five hundred points in each epoch which was contained two hundred points from Frontal component and three hundred points from Motor component.

Each dimension of all epoch extracted before from case 1 to case 5 was averaged to get one mean value. There was one mean number for one dimension from these extracted EEG epochs. We called this vector mean vector. The dimensions of mean vector and processed EEG signals were the same. The method of computing was showed in Fig. 3-4.

We subtracted this mean vector from each EEG epochs in that subject. If this step Fig. 3-4: The method of calculating the mean vector.

All EEG epochs of five conditions in the same subject were extracted from the data set. The mean value would be calculated for each dimension of these epochs by averaging all number in the current dimension. There were five hundred dimensions in each epoch and we would get the same numbers of mean value. These mean values were called mean vector.

of subtracting mean vector was not performed, the variation among subjects would be presented by the SOM map. In practice, the performance of the maps with subtracting the mean vector would be better. We would discuss this issue into details in Section 5-1.

Normalizing

Although we subtracted the mean vector for all epochs in each subject, the variation among the trials was still in our data. For example, someone performed two tasks A and B. However these two tasks were in the same condition, they were not happened in sequence. Many events in other condition would be appeared during the interval between A and B. by reason of events random occurred, the level of excited for these two tasks might be different. Before running the Self-Organizing Map with the multiple high dimension data, it is important to reduce the variation among different epochs. We carried out a normalization method like Z-Score to remove this abnormality. This algorithm was applied for each particular case by the order number of subjects. There were n trials in each case of one subject. First the mean value (Smean) of those trials was computed by the following equation (1):

(1)

Where X is the input space, n is the total number of epochs in each case, i is the index of epoch number, and d is the dimensions of the input space.

In the second step the standard deviation (Sstd) was calculated by the same data by the following equation (2):

. (2)

Then we took these two values to normalize all epochs in that case. Each frequency in every trial was subtracted by the mean value (Smean) and divided by the standard deviation (Sstd). The normalization of trial T was processed by the following equation