A New Emotional Classification Method Based on the Combined Characteristics of EEG Signals

The electroencephalogram (EEG) signals has been widely used in emotion recognition. The entropy features are often used in the emotional recognition using EEG. However, its recognition accuracy remains to be improved. In this study, a new EEG feature combining the frontal asymmetry and differential entropy features is used to classify negative and positive emotion. The result of our research shows that the combined feature is better than the entropy features in emotion recognition. The average classification accuracies of the frontal asymmetry combined with differential entropy features and entropy feature in our study are 72.1% and 67.7% respectively. This result indicated that this combination feature is more suited for emotional classification.


Introduction
Emotion is vital to humans. It not only helps people communicate, but also plays a key role in rational and intelligent behavior. There are many ways for emotion recognition. Facial expressions as a recognition method are often used [1]. In recent years, other methods have been continuously improved for emotion recognition. Some physiological signals are widely used, such as electrocardiogram (ECG), electromyogram (EMG), electroencephalogram (EEG). Among these, EEG-based emotion recognition plays a very important role, which has received the attention of many researchers [2,3].
The EEG is a direct measurement of neural activity. Researchers have done many works to identify emotions based on the EEG signals. Among them, how to classify human's emotional states is a prime issue.
There are two models describing emotional states, the discrete and the dimensional model. The discrete model includes six basic emotional states, such as anger, disgust, fear, sadness, happiness, and surprise [4]. The dimensional model comes up with the valence-arousal space [5]. The valence and arousal scales represent the valence arousal space for all emotional states. In recent years, based on DEAP database [6], many studies have used the database to classify valence and arousal [7,8].
Davidson used brain asymmetry to analyze EEG [9], which achieved good results and was used widely by other researchers [10,11,12]. The entropy feature like approximate entropy (ApEn) and sample entropy (SampEn), as an important feature of emotion recognition, are capable to extract emotional information from complex EEG signals. Seyyed Abed Hosseini uses the approximate and wavelet entropy to analysis emotion state [13] .Yong Zhang et al use the method of empirical mode decomposition (EMD) and sample entropy for emotional recognition [14].
In our research, we combine the advantages of asymmetry with entropy information, and put the positive and negative emotions as feature vectors into the support vector machines for classification The results show that the recognition rate of a single subject reaches 92.1% using the combined features, and the average is about 5% higher than the use of the different entropy features. In future studies, the combination of frontal asymmetric and different entropy features will have better prospects in emotional recognition.

Data acquisition
The DEAP database recorded 32 participants' EEG and peripheral physiological signals, in which each participant watched 40 one-minute videos. Then the participant performed a self-assessment(SAM) to quantify emotional response to the video after watching each video. The valence model is considered in this study. Valence measures the pleasantness of the stimulus and varies from unhappy or sad to happy or joyful. We only want to study the effects of EEG on emotions, so we extract features from EEG channels.

Data selection
Each video also has a specific value. First, we mark each trail with a positive negative flag according to the valence score. As shown in figure 1(a).The valence score of negative is less than 5, and if greater than 5,is considered as positive. We selected the data of the SAM valence score of the test and the valence of the original data to make the data more reliable. In addition, because the level of the dominance score determines whether the emotion of the subject is successfully evoked. In general, if the dominance score is greater than 3, the emotion of the subject is aroused [15], so we select out the trial with a dominance score greater than 3.
The purpose of this study is to prove the role of the frontal for the emotion, so we chose F3 and F4 channels, as is shown in figure 1 Alpha waves are more associated to a aware mental state, and are more visible over the parietal and occipital lobes. Beta waves are related to an active state of mind, more prominent in the frontal cortex and over other areas during intense focused mental activity [16].So we choose alpha and beta waves because they are more related to the emotion.
As shown in figure 2, we have compiled a summary of all the data selected by 32 participants, and clearly showed the relationship between the valence and the positive and negative emotions.

Frontal asymmetry
To quantify the asymmetry, Davidson uses the asymmetry index DI = (L -R) / (L + R), where L and R are the powers of the specific bands in the left and right hemispheres, respectively. One type of study suggests that the left frontal is associated with positive emotions, and the right frontal is associated with negative emotions [17,18]. In fact, this theory has been widely accepted and used for research [19,20].
In our study, the F3 and F4 channels of the alpha and beta bands in the frontal were used. Based on the above asymmetry calculation, the logarithm was taken to obtain the asymmetry index, as Fas.
In order to remove the influence of individual differences, we removed the baseline of the resting state for each subject, and then put this into the SVM as a feature vector.
(2) Determine the distance as follows (3) Calculate the number of [ ( ), ( )] < for all of ,and its ratio to Where ( > 0) is a threshold.
(4) Take the logarithm for ( ),and then calculate its average for all of ,defined as ( ).
(5)Increase the dimension to + 1,and repeat steps of (1)-(4),we can get a value +1 ( ). (6) Finally define ApEn as: Thus the ApEn is determined by the parameters and . We choose = 2 and = 0.2 * in our research, the is the standard deviation of the signal.

SVM
SVM is a machine learning method which is based on statistical learning theory of Vapnik-Chervonenkis dimension theory [21], mainly solves the small sample problem. For two-class problem, given training set{ , } =1 , input set ∈ ,output set ∈ and ∈ {−1,1}, we use the following linear classifier, The nonlinear SVM classifier takes the following form, Where is the Lagrangian multiplier, ( , ) is Kernel function. In our study, we use nonlinear SVM as a classifier, we use the RBF kernel function, which is expressed as follows, Where controls the width of the kernel function.

Result
We apply a 128-point window on the pre-processed EEG signal of each channel and each ban to calculate the Fas, ApEn and SanpEn. As shown in Figure 3, the accuracy of the classification of each subject is obtained by comparing the combined feature with the different entropy. For each subject, the accuracy of using the combined feature is higher than the accuracy using the entropy feature. And It can be seen from the figure that the classification accuracy of the 23rd participant is 92.1% using the combined feature, and the average classification rates of the 32 subjects using the two features are 72.1% and 67.7%, respectively.

Conclusion
In this paper, a new combined feature is proposed and compared with different entropy. According to the result, the combined feature is more suitable for emotion classification than different entropy. And this new feature increases the stability. At the same time, we can conclude that the frontal asymmetry is a key feature in emotional recognition.

Discussion
This study used the combined features to identify EEG emotions. And it's also shown that the result of recognition accuracy is improved after adding asymmetric features. The characteristic of entropy is a nonlinear parameter that only to analyze the complexity information of a single channel in a certain band [22,23], but the asymmetry reflects the difference of the emotional changes to the 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  left and right hemispheres [24]. So we combine the two features mentioned above, made a result of 72.1% for valence with 32 participants. Additionally, for the DEAP database, there are several works that use different features to classify valence. Zheng [25] compared with multiple features, and the result showed that the different entropy as the feature can make for the highest accuracy 69.67% with 32 participants; Chung's [26] result is 66.6% .In this study, all bands were used to compute the spectral powers of the 93-channel that generated by the 32 traditional channels and 61virtual channels using bipolar montage with 32 participants, Kumar's [27] result is 61.17% that derived the features of bispectrum for the Fp1 and Fp2 channels of the EEG signals. The comparison with the above study shows that we use the combined feature of frontal asymmetry and entropy feature have a higher recognition rate, which proves that the asymmetry has a important relationship with emotion, because it reflects the common information of the left and right hemispheres.
Therefore, it is also affirmed that the frontal asymmetry is an important feature of emotion recognition. In recent years, prefrontal asymmetry has also been used to analyze mood regulation [28], which will place more emphasis on asymmetry in our future studies of emotional regulation.