Research Paper
Expert Systems with Applications 38 (2011) 1859–1865
Contents lists available at ScienceDirect
Expert Systems with Applications
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a
Multivariate autoregressive models and kernel learning algorithms for classifying driving mental fatigue based on electroencephalographic
Chunlin Zhao a,b,⇑, Chongxun Zheng a, Min Zhao a, Yaling Tu a, Jianping Liu b a Institute of Biomedical Engineering, Xi’an Jiaotong University, Key Laboratory of Biomedical Information Engineering of Education Ministry, No. 28, Xianning West Road, Xi’an, Shanxi 710049, PR China b Engineering College of Armed Police Force, ShanQiao Road, Xi’an, Shanxi 710086, PR China
a r t i c l e i n f o a b s t r a c t
Keywords: KPCA SVM MVAR EEG Driving mental fatigue
0957-4174/$ - see front matter � 2010 Elsevier Ltd. A doi:10.1016/j.eswa.2010.07.115
⇑ Corresponding author at: Institute of Biomedica University, Key Laboratory of Biomedical Informatio Ministry, No. 28, Xianning West Road, Xi’an, Shanxi 13152097819.
E-mail addresses: zclwj1974@163.com (C. Zhao), Zheng), zhaoclzm@163.com (M. Zhao), tuyaling_0 liujp96@sohu.com (J. Liu).
Long-term driving is a significant cause of fatigue-related accidents. Driving mental fatigue has major implications for transportation system safety. Monitoring physiological signal while driving can provide the possibility to detect the mental fatigue and give the necessary warning. In this paper an EEG-based fatigue countermeasure algorithm is presented to classify the driving mental fatigue. The features of mul- tichannel electroencephalographic (EEG) signals of frontal, central and occipital are extracted by multi- variate autoregressive (MVAR) model. Then kernel principal component analysis (KPCA) and support vector machines (SVM) are employed to identify three-class EEG-based driving mental fatigue. The results show that KPCA–SVM method is able to effectively reduce the dimensionality of the feature vec- tors, speed up the convergence in the training of SVM and achieve higher recognition accuracy (81.64%) of three driving mental fatigue states in 10 subjects. The KPCA–SVM method could be a potential tool for classification of driving mental fatigue.
� 2010 Elsevier Ltd. All rights reserved.
1. Introduction techniques based on physiological phenomena achieving higher
Mental fatigue is a gradual and cumulative process and is thought to be associated with a disinclination for any effort, re- duced efficiency and alertness and impaired mental performance. The major symptoms of mental fatigue is a general sensation of weariness, feeling of inhibition and impaired activity. Driving men- tal fatigue is widely recognized as a core safety issue in the trans- portation. This is four times more likely to be a contributor to workplace impairment than drugs or alcohol. Driving mental fati- gue-related road accidents alone cost around Australian $ 3 billion per year and become a substantial financial burden on the commu- nity (The Parliament of the Commonwealth of Australia, 2000). Developing and establishing an accurate and non-invasive real- time system for monitoring driver’s mental fatigue is quite impor- tant to reduce road accidents and lower social cost in traffic safety.
As a result, numerous field studies and laboratory experiments were conducted to produce the real-time and non-obtrusive means for detecting driving mental fatigue. Among these different driving mental fatigue detection methods, there are two main fields with
ll rights reserved.
l Engineering, Xi’an Jiaotong n Engineering of Education 710049, PR China. Tel.: +86
cxzheng@mail.xjtu.edu.cn (C. 119@yahoo.com.cn (Y. Tu),
detection accuracy. One approach focuses on driver and vehicle physical changes such as the inclination of the driver’s head, sag- ging posture, and decline in gripping force on steering wheel or the open/close state of the eyes or steering angle, vehicle lateral position, vehicle speed and vehicle yaw rates (Hu & Zheng, 2009; Lal, Craig, Boord, Kirkup, & Nguyen, 2003; Sayed & Eskandarian, 2001; Smith, Shan, & da Vitoria Lobo, 2000). But these methods are limited to depending on the vehicle type and driving condi- tions. The other approaches focus on the fields to measure physio- logical changes such as eye-blinking, heart-rate, pulse-rate or skin- electric-potential, particularly, brain waves, as a means of detect- ing a human mental fatigue state. While numerous physiological indicators were available to measure mental fatigue, the EEG is widely regarded as the physiological ‘‘gold standard” for the assessment of mental fatigue (Bouchner, 2006; Lal & Craig, 2001; Lin et al., 2005; Jap, Lal, Fischer, & Bekiaris, 2008). EEG signals con- tain a lot of information of the cognitive states such as alertness and arousal and they have plentiful information related to the dif- ferent physiological states of the brain and can be a very effective medium for understanding the complex dynamical behavior of the brain.
Mental fatigue is a complex phenomena which is relative to nerve-central activity. One must look at activity distributed over the entire scalp in order to detect brain state during mental fatigue. Thus, multichannel EEG must be recorded because single-channel
1860 C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865
brainwaves do not provide enough information (Anderson, Stolz, & Shamsunder, 1998; Franaszczuk, Blinowska, & Kowalczyk, 1985).
In this study, a multivariate autoregressive (MVAR) model is ap- plied to extract EEG features for measuring driving mental fatigue. A newly-developed machine-learning technique – support vector machine (SVM) combined kernel principal component analysis (KPCA) is adopted to differentiate three driving mental fatigue states.
Fig. 1. The scores of self-report (**p < 0.005).
2. Materials and methods
2.1. Subjects
To reduce inter-subject differences, 13 male graduate students (mean age, 23.8 years; range 22–26 years) were recruited from students of Xi’an JiaoTong University to perform the experiments. All participants provided informed consent prior to participating in the study. All subjects did not have any actual driving experi- ence and none of them was able to operate the stick shift car. They were familiar with operating a computer and had the experience of playing video game. All subjects were trained before the experi- ment until they performed the simulative driving system expertly. None of them worked night shifts or used prescription medication and medical contraindications such as severe concomitant disease, alcoholism, drug abuse, and psychological or intellectual problems likely to limit compliance. According to their self-reports, all sub- jects had normal or corrected-to-normal vision and were right- hand dominated.
2.2. Apparatus
The driver simulator equipment consisted of a car frame with an in-built steering wheel, gas and brake pedals, clutch, manual shift and a horn and turn signal. The visual display of the (virtual reality) VR-based driving simulative environment is a 19 inch Li- quid Crystal Display at a distance of 80 cm from the subject’s eyes. The LCD shows the road environment, the current speed and other road stimuli. The system also can provide engine noise and nearby traffic noise. The simulative route and traffic sign are standardized with national traffic law.
2.3. Experiment design
Previous literatures pointed out that driving mental fatigue oc- curred in a monotonous driving environment. Thus, a highway scene was selected in our experiment. Furthermore, we designed the simulative driving track with the following requirements: The route was simple so that the drivers could perform as easily as possible; There were few scenery changes and no moving ob- jects in the three-lane road with no inclination to reduce outside stimuli; A very light curvature was chosen so that drivers should pay their attention to steering all the time. One lap would take about 7 min when the subjects kept the car speed at about 100 km/h. Each driving experiment lasted about 150 min continuously.
This study had the institute’s Human Research Ethics Commit- tee approval, and was conducted in a dimly lit, sound-attenuated, electrically shielded and temperature-controlled laboratory. Train- ing was carried out previously. Participants were asked to sleep adequately the day before the study, refrain from consuming alco- hol caffeine, tea or food as well as smoking approximately 12 h be- fore the study, and reported compliance with these instructions. To avoid the influence of circadian fluctuations on subjects, the exper- iments were conducted approximately at 8:00 AM or 2:30 PM dur- ing the normal work-time. Before the experiment, the subjects
learned the whole procedure to well understand the procedure and the instructions, and the psychological self-report measures of mental fatigue were conducted. Subjects then performed the simulated driving without any break either until 150 min elapsed or until volitional exhaustion occurred. During the driving, subjects were asked to restrict all unnecessary movements as much as pos- sible and to try their best to maintain constant speed and avoid car accidents. There were not any questionnaires and any additional measurements during driving, so as to maintain a monotonous condition. At the end of all experiment sessions, the same psycho- logical self-report measure of fatigue was also carried out.
2.4. Data acquisition
The physiological signals were recorded by a Neruoscan system with international 10–20 lead systems. EEGs were recorded using a 32 channel electrode cap with sintered Ag/AgCl electrodes from scalp positions FP1, FP2, F7, F3, Fz, F4, F8, FT7, FC3, FCz, FC4, FT8, T7, C3, Cz, C4, T8, TP7, CP3, CPz, CP4, TP8, P7, P3, Pz, P4, P8, O1, Oz, O2 (Fig. 1). Vertical electrooculogram (EOG) was recorded using bipolar electrodes placed above and below the left eye. All sites were referenced to linked mastoids. The connecting impedance was kept below 5 kX. All physiological signals were sampled at 500 Hz with 0.05–70 Hz band-pass filter and 50 Hz notched. The EEG larger than +100 uV was rejected as artifact. Eye movement contamination of EEG was firstly removed by adaptive filtering methods.
Meijman (1994) reported that the length of preceding work hours had negative effects on mental fatigue by impairment of the mental performance capacity itself or by negative changes of the willingness to spend mental capacity in order to sustain an adequate performance. The conclusion was consistent with empir- ical evidence. In this study, the driving mental fatigue is classified into three levels: the alert, the medium fatigue and the extreme fa- tigue according to time-on-task. Thus epochs of the experimental beginning (0–15 min), the middle (75–90 min) and the end (135– 150 min) are selected to investigate different mental fatigue states. After artifact detection and ocular correction, 30-s continuous EEG data of each epoch for each subject are selected to be analyzed.
During the whole simulative driving, the mental fatigue signs such as rubbing, yawning and nodding, the driving performances such as car accidents, flameout, and other operating errors are re- corded manually by an observer to validate mental fatigue states.
2.5. EEG features extracted based on MVAR
MVAR model is the extension form of the univariate AR model and can capture data flowing from a number of channels simulta-
C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865 1861
neously. Analysis of the multichannel EEG with this method can be used in researching the synchronization of brain structures, the de- gree of coupling between channels, the estimation of phase delays, and eventually the direction of spreading of brain activity (Ander- son et al., 1998; Franaszczuk et al., 1985; Neumaier & Schneider, 2001)
MVAR model with pth order can be expressed as:
v n ¼ Xp i¼1
Aiv n�i þ en; en ¼ noiseðCÞ ð1Þ
where m-dimensional vector en is the vector of multivariate zero mean uncorrelated white noise process and covariance matrix C e Rm�m, A1, A2, . . . , Ap e R
m�m are the coefficient matrices of the MVAR model. The vector v e Rm is a vector, which consists of sam- pling signal by m channels at n time. Eq. (1) shows that the multi- variable signals at n time can be estimated by their values at past time and the white noise.
An AR model of a sequence of observations may be found by estimating the parameter matrices by way of a least squares proce- dure that minimizes the sum of squared errors. This is the Yule– Walker equations:
�½Rð1ÞRð2Þ � � �RðpÞ� ¼ ½Að1ÞAð2Þ � � �AðpÞ�eR ð2Þ where
eR ¼ Rð0Þ Rð1Þ � � � Rðp � 1Þ RTð1Þ Rð0Þ � � � Rðp � 2Þ
..
. .. . ..
.
RTðp � 1Þ RTðp � 2Þ � � � Rð0Þ
2 66664
3 77775
The solution of these equations is the coefficient matrices of the MVAR model.
Before coefficients calculating, the model order should be deter- mined. The model order can be found by means of criteria derived from information theory. Previous research tested the sensitivity of MVAR performance depending on the model order and demon- strated that small changes of model order do not influence results. The Akaike’s Information Criterion (AIC) was found as the most sat- isfactory for model order determination (Franaszczuk et al., 1985). It is used in this work for MVAR model fitting
AICðkÞ¼ N log½detðbV eÞ�þ 2m2k ð3Þ where N is the number of experimental data points of the sampled signal, m is the number of inputs (or channels in this case), bV e is the estimated covariance matrix of the noise processes. bV e can be determined from the formula:
bV e ¼ Rð0ÞþX k
i¼1 Ai RðiÞ ð4Þ
where Ai, R(i) are the estimated: matrix of coefficients and matrix of covariance.
After the optimal model order p is determined. Let bDk be the combined model coefficient matrix of the kth data segment
bDk ¼ðbAk;1 bAk;2 � � � bAk;pÞ ð5Þ Then, feature vector can be constructed as follows:
~xk ¼ðbDk;1: bDk;2: bDk;3: bDk;4: bDk;5: bDk;6:ÞT ð6Þ where the coefficient vector bDk;1: represents the ith row of the ma- trix bDk. The feature vectors obtained from all data segments will be saved for later analysis. In this study, EEG data of six electrodes (Fp1, Fp2, C3, C4, O1, O2) are selected for analyzing. According to AIC, the order of the MVAR is selected as 3. For a three-order model and six channels, the size of the feature vector is 108.
2.6. Kernel based dimensionality of feature space reduction
Before executing a learning algorithm, additional vector space transformations need to be applied on the initial features for improving classification performance and reducing the dimension- ality of the data. Kernel principal component analysis (KPCA), pro- posed by Scholkopf, Smola, and Muller (1998), is one approach of generalizing linear PCA into nonlinear case using the kernel meth- od. The basic idea is to map the original input vectors into a high- dimensional feature space and then to calculate the linear PCA in this feature space. KPCA as a nonlinear feature extractor leads to better classification than the linear ones. KPCA algorithm is used to reduce the dimensionality of EEG features and the Gaussian function is selected as the kernel function.
2.7. Multiclass support vector machine
Support vector machine (SVM), a novel machine learning algo- rithm, has been recently proven to be a promising tool for both data classification and pattern recognition (Vapnik, 1998). SVM is also a kernel-based classification technique that is based on the margin-maximization principle, which makes SVM have better generalization ability than the other traditional learning machines that are based on the learning principle of empirical risk minimiza- tion. SVM uses the kernel-mapping to map the data in input space to a high-dimensional feature space in which the problem becomes linearly separable. There are many kinds of kernels that can be used, such as the linear, polynomial and radial basis function (RBF) kernels (Zhang, Zhou, & Jiao, 2004). To reduce the search- space of parameter sets, in this study we train all datasets only with the RBF kernel.
The earliest used implementation for SVM multiclass classifica- tion is probably the one-against-all (OA) method (Simard & Vapnik, 1994). It constructs k SVM models where k is the number of clas- ses. The ith SVM is trained with all of the examples in the ith class with positive labels, and all other examples with negative labels. Another major method is called the one-against-one (OO) method (Knerr, Personnaz, & Dreyfus, 1990). Assume training data from the ith and the jth classes. This method constructs k(k � 1)/2 classifiers where each one is trained on data from two classes. If decision function says x is in the ith class, then the vote for the ith class is added by one. Otherwise, the jth is increased by one. Then we pre- dict x is in the class with the largest vote. The decision strategy is called ‘‘Max Wins”.
In this study, two multiclass SVM methods also are adopted to identify the three driving mental fatigue states.
3. Results
3.1. Self-report about driving mental fatigue
The subjective component of fatigue is very important, ques- tionnaire investigations may be important in the study of driver mental fatigue. Questionnaires can provide information about fati- gue such as the feelings when fatigue appeared and factors contrib- uting to fatigue. According to the self-report questionnaires, all the subjects felt tired, bored and drowsy when the driving task was over. They also reported that these feelings became stronger and there were difficulties to concentrate and focus their attention on the driving task as the driving time increased. To keep the monot- onous driving environment, the questionnaires just only were car- ried out at two epochs: pre-driving: before driving task; post- driving: after that task. Fig. 1 shows the psychological self-report measures of mental fatigue according to Li’s subjective fatigue scale (LSFS) (Li, Jiao, Chen, & Wang, 2003).
1862 C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865
The self-report questionnaires revealed subjects as almost not fatigued before the driving task and moderately to extremely fati- gued after driving. Compared with the pre-driving, the subjective scores increased significantly (t = �9; df = 9; p < 0.001) after the end of the driving.
3.2. Some objective indicators of driving mental fatigue
To maintain scientific validity, questionnaires should not be the sole identifier of fatigue symptoms. More objective measures need to accompany them for verification of fatigue. The subject’s man- nerisms such as rubbing, yawning and nodding, the driving perfor- mance details such as car accidents, flameout, and other operating errors and the vertical EOG were combined to validate the different driving mental fatigue status. The EOG was used to identify blink artifact in the EEG data as well as changes in blink types such as the small and slow blinks that characterize fatigues.
Table 1 shows the proportion of subjects categorized according to the mannerisms identified from the manually recorded data as well as the driving performance in each state. There almost are not fatigue physical mannerisms during alert states. Over 60% sub- jects showed fatigue physical mannerisms during medium state. In extreme state, these mannerisms are observed in all subjects. The lapse of driving proportion also increases linearly from alert to ex- treme fatigue state.
In Table 1, each validated mannerisms for subjects should sat- isfy these criterions respectively: 1 Rubbing/5 min; 1 Yawn/ 5 min; 1 Noddings/30s; 1 Flameout/7.5 min; 1 Car accidents/ 7.5 min; 1 Other driving errors/15 min.
The blink frequency can be estimated by EOG. The data of blinks were obtained from EOG by identifying the peak of blink based on wavelet detection method. Fig. 2 shows the blink frequency at the three mental fatigue states.
The blink frequency is increasing from alert to extreme state. The ANOVA results show a highly significant difference in the three
Table 1 The mannerisms of fatigue and lapses in driving performance.
Mannerisms Alert (%)
Medium fatigue (%)
Extreme fatigue (%)
Rubbing 10 60 100 Yawns 0 70 100 Noddings 0 70 100 Flameout 0 30 80 Car accidents
(collisions) 0 60 90
Other driving errors 0 40 70
Fig. 2. The blink frequency of three fatigue states (*p < 0.05; **p < 0.005).
mental fatigue states. (F(2, 18) = 8.876, p = 0.002;). The least differ- ence (LSD) post hoc analysis shows that the alert state is significant different from medium and extreme fatigue states, but there is no statistical difference between medium and extreme fatigue states.
3.3. The classification results
The subjective and objective measures indicate that driving mental fatigue is induced after a long time simulative driving task. In order to distinguish different mental fatigue states, 30 s EEG data of each subject in three epochs are selected to be analyzed. The EEG data is divided into 2 s segment with 0.2 s overlapping. The sample set includes 420 data segments for each subject. As the number of EEG data segment available is limited in this exper- iment, a 27-fold cross-validation test is applied. For each subject, the 20% of this sample set is selected as testing set and the 80% is selected as training set randomly, the classification accuracy is calculated over 27 trials with different random selection of training and testing set.
For a three-order MVAR model and six channels, the size of the EEG feature vector is 108. KPCA is applied to reduce the size of fea- ture vectors and then the lower-dimensional vectors are consid- ered as input of SVM. The test performance of the classifiers can be determined by the computation classification accuracy which is defined as the proportion of the correct decisions number and total cases number.
When KPCA is applied to reduce the dimension of features, Gaussian function is selected as the kernel function. Fig. 3 shows the average accuracy with different kernel parameter r.
From Fig. 3, we also find that the kernel parameter r is a factor which can influence classification accuracy. The max-accuracy reaches 80.8% when r equals to 1.
The average classification accuracies of all subjects under differ- ent numbers of the feature dimensions are illustrated in Fig. 4.
Fig. 4 shows that classification accuracy fluctuates with the number of feature dimension. The max-accuracy achieves 81.64% when the number of features equals 25. The classification results do not improve with the features dimension increasing.
In this study, the classification accuracy is obtained by averag- ing the classifying result of the three-classes. Fig. 5 represents the classifying result about different mental fatigue states over all subjects.
Fig. 5 shows that the max-accuracy about three mental fatigue states has some difference. The alert and extreme fatigues obtain a slightly higher accuracy than medium fatigue state. The max-accu- racy is alert 81.6%, medium 80%, extreme 83.8% respectively.
As a basis for comparison, we observe the accuracy of classifica- tion under the condition of the various extraction features using
Fig. 3. The classification accuracy under different r.
Fig. 4. The classification accuracy under different feature dimensions.
Fig. 5. The classification accuracy of three fatigue states.
Fig. 6. The comparison of different classification algorithms.
C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865 1863
KPCA and linear PCA (LPCA) respectively. And we also compare the results to the original SVM without using KPCA or LPCA, the origi- nal RBF network and KPCA–RBF network classification. In addition,
Table 2 The comparison of the accuracy (%) of different number of feature dimensions.
Algorithm The number of features dimension
15 20 25 30
KPCA–SVM(OA) 78.80 80.83 81.64 81.18 PCA–SVM 55.96 60.54 64.01 68.19 SVM 78.71 78.71 78.71 78.71 KPCA–RBF 70.37 76.26 78.77 79.15 RBF 76.06 76.06 76.06 76.06 KPCA–SVM(OO) 79.72 80.44 80.51 80.70
the accuracy of two kinds of multiclass SVM (OA and OO) is con- trasted when applying KPCA. Fig. 6 shows the classification result of various algorithms under different numbers of the feature dimensions.
From Fig. 6, it can be found that KPCA–SVM shows the best per- formance with the highest accuracy. The maximal classification accuracy achieves 81.64% while the number of feature dimensions equals 25. Although the feature-vector dimensions are reduced by KPCA, the accuracy is better than SVM with high-dimension origi- nal feature data. However, the result of LPCA–SVM is not desired and the classification accuracy is below 72%. Compared with LPCA–SVM, the accurate rate of KPCA–SVM is improved signifi- cantly. For multiclass SVM, the result based on OA method is better than that of OO method on the whole. For the original non-reduc- ing feature dimensions data, the SVM takes the better performance than that of RBF network. Furthermore, KPCA–RBF shows better performance than that of RBF with high-dimension original feature data while feature dimensions is more than 20. Table 2 compares the performance of KPCA–SVM with the other types of classifica- tion algorithm under different number of feature dimensions.
In Table 2, it can be seen that the max-accuracies do not belong to one fixed dimension for different classification algorithms. When the number of features dimension is reduced by KPCA, the fewer dimension combines with the faster speed of classifying, we should balance the best performance and the classification speed.
Compared with the original SVM, the KPCA–SVM can accelerate the classification speed and accuracy of driving mental fatigue effectively, which greatly reduces the dimensionality of input fea- tures. Moreover, the performance of KPCA–SVM is greater or more than that of LPCA–SVM and RBF network. KPCA–SVM is a promis- ing classifier for driving mental fatigue.
4. Discussion
It is well known that there is a strong link between time-on- task and mental fatigue progression. Many studies have demon- strated this validity (Otmani, Pebayle, Roge, & Muzet, 2005; Ting, Hwang, Doong, & Jeng, 2008). The self-report results in our study revealed that subjects were slightly fatigued before and moder- ately or extremely fatigued after the driving test. Lal and Craig (2002) used subject’s mannerisms and EOG signs as independent variables to identify different mental fatigue phases with excellent reliability and satisfying result. In this present experiment, the mannerisms of fatigue and lapses in driving performance are ab- sent at the start of task and increase with time-on-task till almost all subjects showed these signs at the end of the task. According to the statistical results, mental fatigue mannerisms and collisions accidents have occurred across over 60% subjects under medium fatigue status. This should be the notable symbol of increased men- tal fatigue. The time-on-task actually has a negative effect on dri- ver’s performances and behaviors. The result of blink frequency also indicates that they are significantly different among three epochs. The conventional blinks during the alert phase are replaced
The max-accuracy
35 40 45
81.10 81.52 80.96 81.64 70.04 71.21 71.82 71.82 78.71 78.71 78.71 78.71 79.80 79.24 80.23 80.23 76.06 76.06 76.06 76.06 80.10 80.32 80.24 80.70
1864 C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865
by fast rhythmic blinks during mental fatigue. The blink frequency increases positively with the extent of mental fatigue. This result is consistent with that Lal and Craig (2002) reported.
In this experiment, the classification accuracy of medium fati- gue is slightly lower than those of alert and extreme fatigue states. To three-class problem, the lower classification accuracy in one class means this class sample is similar to one of the rest samples. This result suggests that medium state might lie between alert and extreme state.
It has been known for many years that the change in brain arou- sal involves specific changes in oscillatory brain activity and the EEG can reflect the fluctuation of alertness level. The EEG signal may be one of the most predictive and reliable index to assess mental fatigue. However, there are different EEG rhythm changes on different scalp regions during mental fatigue. Jap et al. (2008) also made the same conclusion. It is necessary to look at activity distributed over the entire scalp in order to detect brain state changes. Previous study also indicated that the application of mul- tivariate approach for the determination of the information flow in brain structures brought very rich and important information about the interactions between brain structures (Franaszczuk et al., 1985; Kus, Blinowska, Kaminski, & Basinska-Strarzycka, 2005). In this study, EEG data of six electrodes (Fp1, Fp2, C3, C4, O1, O2) were selected for analyzing. MVAR model is employed to extract model coefficients as EEG features. The model can give us the information on the mutual relationships between relevant structures, particularly on the degree of their synchronization in the frequency domain. The features extracted by MVAR should be sensitive to the change of driving mental fatigue. One can see when looking at the averages across subjects that the MVAR gives the best three-states classification accuracy at 81.64%.
For the high-dimensions of extracted feature by MVAR, KPCA method is applied to reduce the dimensions of feature vectors. KPCA is a generalization of PCA in a feature space by a kernel func- tion that could be nonlinear. Compared with LPCA, KPCA can ex- tract more efficient features that are useful for the classification purpose. That is why the results become better when LPCA is re- placed by KPCA.
The performance of SVM classifier is nonsensitive to the sample size and the dimensionality. In addition, its ability to produce sta- ble and reproducible results makes it a good candidate for solving many classification problems (Burges, 1998). In this paper, the SVM shows more excellent results than RBF network under the same dimensions. The classification abilities of two kinds of SVMs, respectively trained by OA and OO method, are further compared. The experimental result proves that OA method is better in achiev- ing the generalization ability of the SVM for three-class problem.
Combined KPCA with SVM, obtains the best performance with the highest average accuracy over all subjects. The classification re- sults indicate that KPCA–SVM show better performance than that of original SVM with high-dimension original feature data even the dimensions of feature vector are few. KPCA method could sig- nificantly reduce the dimensions of the feature vectors in a high- dimensional feature space by a nonlinear mapping. The low- dimensional feature representation preprocessed by KPCA would accelerate the speed and improve the accuracy of the classifier. The KPCA–SVM can obtain the satisfying accuracy in classifying three-level driving mental fatigue with the lower dimensions of feature space.
5. Conclusion
Driving mental fatigue is a complicated physiological and psy- chological process. This paper presents a framework based on EEG for classifying driving mental fatigue. The MVAR extract EEG
features effectively which are sensitive to the change of driving mental fatigue. Then the kernel-based algorithm KPCA leads to bet- ter classification and faster calculation speed for its nonlinear transformations and feature vector dimensions reduction capacity. The experimental results show that KPCA–SVM algorithm en- hances the generalization ability of the classifier and improve the accuracy of driving mental fatigue states recognition. The classify- ing model could be potential for evaluating driving mental fatigue.
However there are two factors that play important roles in clas- sification, the kernel parameter r and the number of feature vector dimensions. The two factors corresponding to the best accuracy of each subject are different. It is still a further ongoing research issue that is how to choose the optimal kernel parameter r and the num- ber of feature vector dimensions in order to take the higher classi- fication accuracy.
Acknowledgment
The project is supported by National Science Foundation of Chi- na under Grant No. 30670534.
References
Anderson, C. W., Stolz, E. A., & Shamsunder, S. (1998). Multivariate autoregressive models for classification of spontaneous electroencephalographic signals during mental tasks. IEEE Transactions on Biomedical Engineering, 45(3), 277–286.
Bouchner, P. (2006). A complex analysis of the driver behavior from simulated driving focused on fatigue detection classification. WSEAS Transactions on Systems, 5(1), 84–91.
Burges, J. C. (1998). A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery, 2, 121–167.
Franaszczuk, P. J., Blinowska, K. J., & Kowalczyk, M. (1985). The application of parametric multichannel spectral estimates in the study of electrical brain activity. Biological Cybernetics, 51, 239–247.
Hu, S. Y., & Zheng, G. (2009). Driver drowsiness detection with eyelid related parameters by support vector machine. Expert System with Applications, 36, 7651–7658.
Jap, B. T., Lal, S., Fischer, P., & Bekiaris, E. (2008). Using EEG spectral components to assess algorithms for detecting fatigue. Expert Systems with Application, 36, 2352–2359.
Knerr, S., Personnaz, L., & Dreyfus, G. (1990). Single-layer learning revisited: A stepwise procedure for building and training a neural network. In Neurocomputing algorithms, architectures and applications. NATO ASI Series (Vol. F68, pp. 41–50).
Kus, R., Blinowska, K. J., Kaminski, M., & Basinska-Strarzycka, A. (2005). Propagation of EEG activity during continuous attention test. Bulletin of Polish Academy of Sciences Technical Sciences, 53(3), 217–222.
Lal, S. K. L., & Craig, A. (2001). A critical review of the psychophysiology of driver fatigue. Biological Psychology, 55, 173–194.
Lal, S. K. L., & Craig, A. (2002). Driver fatigue, electroencephalography and psychological assessment. Psychophysiology, 39, 313–321.
Lal, S. K. L., Craig, A., Boord, P., Kirkup, L., & Nguyen, H. (2003). Development of an algorithm for an EEG-based driver fatigue countermeasure. Journal of Safety Research, 34, 321–328.
Li, Z. Y., Jiao, K., Chen, M., & Wang, C. T. (2003). Effect of magnitopuncture on sympathetic and parasympathetic nerve activities in healthy drivers assessment by power spectrum analysis of heart rate variability. European Journal of Applied Physiology, 88(4–5), 404–410.
Lin, C.-T., Wu, R.-C., Liang, S.-F., Chao, W.-H., Cheng, Y.-J., & Jung, T.-P. (2005). EEG- based drowsiness estimation for safety driving using independent component analysis. IEEE Transactions on Circuits and Systems, 52(12), 2726–2738.
Meijman, T. F. (1994). Mental fatigue and the efficiency of information processing in relation to work times. International Journal of Industrial Ergonomics, 20, 31–38.
Neumaier, A., & Schneider, T. (2001). Estimation of parameters and eigenmodes of multivariate autoregressive modes. ACM Transactions on Mathematical Software, 27(1), 37–57.
Otmani, S., Pebayle, T., Roge, J., & Muzet, A. (2005). Effect of driving duration and partial sleep deprivation on subsequent alertness and performance of car drivers. Physiology and Behavior, 84(5), 715–724.
Sayed, R., & Eskandarian, A. (2001). Unobtrusive drowsiness detection by neural network learning of driver steering. Proceedings of the Institution of Mechanical Engineers – Part D – Journal of Automobile Engineering, 215(9), 969–975.
Scholkopf, B., Smola, A., & Muller, K. R. (1998). Nonlinear component analysis as a kernel eigenvalue problem. Neural Computation, 10(5), 1299–1319.
Smith, P., Shan, M., & da Vitoria Lobo, N. (2000). Monitoring head/eye motion for driver alertness with one camera. In Proceeding of the international conference on pattern recognition (Vol. 4, pp. 636–642).
C. Zhao et al. / Expert Systems with Applications 38 (2011) 1859–1865 1865
The Parliament of the Commonwealth of Australia. (2000). Beyond the midnight oil. Managing fatigue in transport. House of Representatives Standing Committee on Communication, Transport and the Arts, Australia.
Ting, P.-H., Hwang, J.-R., Doong, J.-L., & Jeng, M.-C. (2008). Driver fatigue and highway driving: A simulator study. Physiology and Behavior, 94(3), 448–453.
Vapnik, V. N. (1998). Statistical learning theory. New York: John Wiley and Sons Inc.. Zhang, L., Zhou, W. D., & Jiao, L. C. (2004). Wavelet support vector machine. IEEE
Transactions on Systems, Man and Cybernetics, 34(1), 34–39.