Participants
A total of ten adult volunteers (8 males and 2 females, age range: 18–28, mean age: 23 ± 3.4 SD) participated in current study. All participants were right-handed except one. They did not suffer from any psychological or neurological disorders and they had normal vision. All subjects were informed about the task prior to experiments. The study is conformed by the ethical guidelines of PAARAND specialized center.
Experimental task
The simulator technique allows subjects to interact directly with driving task without risk of operating on actual machine. For this reason participants performed a driving task in a simulator environment. They were asked to drive along a pathway with indicators of turning to left and right. The path was like a butterfly wings (Figure 1) with no obstacles. Driving was comprised of four laps with four turns to right and four to left in each lap. Therefore each subject had experience of 32 turnings, 16 turnings to left and 16 to right. The task lasted 5 minutes and participants drove on cruise control mode with constant speed of 30 (km/h). Visual driving environment was implemented on the 17-inch LCD with 1280*800 resolutions (Figure 2). The specification for computer is described as follows: Pentium 4, 2.8 GHZ CPU, 1GB RAM, Windows XP professional and driving simulator program. The driving simulator was 3D-Driving School Simulator (copyright by BESIER 3D-EDUTAINMENT 2003) based on virtual environment (Figure 3).
EEG recording system
Volunteers were fitted with a 19-channel electrode cap and prepared for EEG recording according to standard techniques. Recorded channels (FP1, FP2, F3, F4, C3, C4, P3, P4, F7, F8, T3, T4, T5, T6, FZ, CZ, PZ, O1, and O2) were selected from the international 10–20 set of electrode positions, with linked-ears montage (Miller et al. 1991). (The MCN system (Modified Combinatorial Nomenclature) renames four points of the 10–20 system T3, T4, T5 and T6 as T7, T8, P7 and P8, respectively). Subjects performed the experiment in a sound-dampened, electrically shielded booth. EEG signals were amplified with MITSAR hardware, and then sent through an analog- to- digital converter. Signals recorded at 500 Hz on a PC running digitize.
Preprocessing
Due to the most dominant frequency bands of brain, (Delta (1–3.5 Hz), Theta (4–7.5 Hz), Alpha (8-14 Hz) and Beta (15-30 Hz)), a band-pass filter of 1-30 Hz transmitted over signals throughout (Michail et al. 2008). A 1-30 Hz phase-shift free Butterworth band-pass filter (12 dB/Octave) was used. Moreover two amplitude thresholds were considered; slow waves up to maximum 50 microvolts and ultimate 30 microvolts for fast waves. (Threshold values were chosen based on Alpha and Beta brain waves normal amplitudes (Sanei and Chambers 2007)). In order to correct detailed artifacts, independent component analysis (ICA) method was implemented. The “Infomax” algorithm was implemented in WinEEG software to analysis raw EEG signals (Delorme et al. 2007). Eye blink artifacts and some other artifacts were corrected using ICA method even if the EOG (Electrooculography) signal was not recorded (Hori and Cao 2011). This method is based on blind source separation procedure from multi-channel EEG data and spatial filtering of some components of EEG signal. After the decomposition of multi-channel signal, the components of signal related to artifacts were selected manually analyzing topographies and waveforms of components. In general the main components are horizontal and vertical eye movements besides temporal muscular activity, which all were predefined to the WinEEG software. The noisy components were selected and ICA algorithm was applied to the whole EEG data. In this study we had nineteen ICA components. However, actually up to two or three components were used maximally related to individuals.
Feature extraction
Quantitative EEG analysis (QEEG) refers to extract features from EEG signal. Multi-channel EEG is digitized further adjusted to remove extra cerebral artifact, and subjected to spectral analysis using the fast Fourier transform (FFT). Extraction of features such as amount of absolute power at each electrode for each frequency band or as a function of frequency is carried out for individuals. All features used in this article were extracted using NeuroGuide; a software of quantitative EEG analyzer.
Feature selection
In order to reduce the number of features easily, all the decompositions were normalized and reshaped to a row vector. For selecting essential and proper features for classification, scalar feature selection method was implemented using T-test criteria which it ranked all features. T-test criterion returns the significance level (p-value) of the test. The p-value is the probability, under the null hypothesis, of observing a value as extreme as or more extreme than the test statistic.
Absolute value of the criterion was used to rank features. Absolute value means that how much a feature is significant to separate two classes. Features with high absolute value were chosen and others were rejected. The scalar feature selection method considers feature ranking between the two classes.
Hopfield neural network
Hopfield is a network with fully connected N artificial neurons which update their activation values. The update of a neuron depends on the other neurons of the network and on itself. A neuron i will be influenced by another neuron j with a certain weight w
ij
, and a threshold value (Hopfield 1984). There is a weight w
ji
associated to input i. The connection weight from neuron i to neuron j, is w
ij
. In general always two conditions are imposed on the weight matrix: symmetry (w
ij
= w
ji
) and no self-connections (w
ii
= 0). The network also has an output. The state of the output is maintained until the neuron is updated. The training method consists of a single calculation for each weight. In Hopfield network instead of ones and zeros, which it is used in the other networks so far, the inputs are −1 and +1 (the neuron threshold is zero). This has to be true for the network to work correctly. Each weight is labeled by giving it a subscript showing which input it’s coming from and which neuron it’s going too. w
ij
comes from input i and is going to neuron j. The training method is to multiply the value of each feature in each pattern corresponding to the index of the weight, so for w
ij
the value of feature i and feature j were multiplied together in each of the patterns. Then the result is added up.
The new activation value (state) of a neuron is computed, in discrete time, by the function (1):
(1)
Where X is the activation value of the n neurons, W is the weight matrix and T is the threshold of each neuron:
(4)
The sign function is defined as:
(6)
Hopfield network converges to a local state. The energy function of a Hopfield network in a certain state is (7):
(7)
E1 is a general energy function. More often, E2 is used which is equivalent to E1.
The way Hopfield networks act, as a pattern is entered to the network, the Hopfield subject to a number of iterations updating all or part of the nodes to a specific value and stopped. The network neurons are then read out to see which pattern is in the network. The idea behind the Hopfield network is that patterns are stored in the weight matrix. The input must contain part of these patterns. The dynamics of the network then retrieve the patterns stored in the weight matrix. This is called Content Addressable Memory (CAM). The network can also be used for auto-association. The stored patterns in the network are divided in two parts: cue and association. By entering the cue into the network, the entire pattern, which is stored in the weight matrix, is retrieved.