CC BY
30
Научни трудове на Съюза на учените в България-Пловдив, Серия Г. Медицина, фармация и дентална медицина т. XIX. ISSN 1311-9427 юни 2016. Scientific works of the Union of Scientists in Bulgaria-Plovdiv, series G. Medicine, Pharmacy and Dental medicine, Vol. XIX, ISSN 1311-9427 Medicine and Dental medicine June 2016.
ИЗСЛЕДВАНЕ ПРИГОДНОСТТА НА ЛИНЕЙНИТЕ КЕПСТРАЛНИ
КОЕФИЦИЕНТИ ПРИ АВТОМАТИЧНО РАЗПОЗНАВАНЕ НА НЕГАТИВНИ ЕМОЦИОНАЛНИ СЪСТОЯНИЯ ОТ ЕЕГ СИГНАЛИ
Фирган Ферадов Технически университет - Варна
STUDY OF THE QUALITY OF LINEAR FREQUENCY CEPSTRAL COEFFICIENTS FOR AUTOMATED RECOGNITION OF NEGATIVE EMOTIONAL STATES FROM EEG SIGNALS Firgan Feradov Technical University - Varna
Abstract: In the following paper a study on the qualities of Linear Frequency Cepstral Coefficients (LFCC) as a feature for recognition of negative emotions from EEG signals is presented. The experimental evaluation of the descriptors was carried out on EEG data taken from the DEAP database. Multiple tests were performed and filter banks, containing varying numbers of filters, were used to extract the LFCC for different frequency bands from the EEG signal spectrum. The extracted coefficients were used to create person specific SVM models performing classification of negative emotional states and accuracy up to 75.7% was achieved. Based on our study we have concluded that the LFCC can successfully be used as a feature for automated recognition of negative emotional states.
Key words: EEG, Emotion detection, LFCC
INTRODUCTION
Electroencephalography (EEG) is a method for measuring the electrical activity of the brain. It is a traditional and well-established practice in medicine, used for the diagnosis, monitoring and study of neurological diseases. In the last decade a rapid increase in the interest towards automated analysis of EEG signals is observed. There is a vast number of studies, aiming not only to develop and improve the diagnostic capabilities of the EEG measurement, but also to increase its applicability in other fields too.
One of the fields towards which there is pronounced interest is the automated recognition of emotions, where the selection of accurate and descriptive features poses a major challenge. There is a number of different approaches to the feature extraction task, one of the most widely used being the analysis of different frequency components of the EEG signal. Some of the methods include extracting and using specific frequency bands in the frequency domain of the EEG signal [1,2], calculating ratios and derivatives of the extracted frequency bands [3] using wavelet decomposition of the EEG signal [4] and others. Another, less examined approach is the use of feature extraction techniques, traditionally employed in other fields. In [6, 7] Othman et al. studies the qualities of features, based on the Mel Frequency Cepstral coefficients (MFCC) used for emotion recognition from EEG signals. They successfully classify different emotional states and conclude that the MFCC based approach shows potential for detection of basic emotions.
In the presented paper the use of Linear Frequency Cepstral as a feature for classification of negative emotional states is proposed. The coefficients are extracted from short-time frames taken from 32 channel EEG signals.
FEATURE EXTRACTION
All extracted features are computed for short frames of the EEG signal, obtained through a sliding window of 492 milliseconds which moves with a skip rate of 164 milliseconds. Successive frames overlap with 66.7% and the total signal length is 60 seconds. Each frame contains data for all 32 channels of the EEG signal. The total number of successive overlapping frames in a recording with N samples is:
fx 1 (1)
where the operator fix stands for rounding towards the smaller integer number, L is the predefined step size in samples, and K is the frame size, also in samples.
As a next step in the feature extraction process Fast Fourier Transform (FFT) is performed on every extracted frame, and the spectrum of the short-time EEG windows is obtained. A filter bank, containing a number of triangular window functions with 50% overlap between subsequent windows is applied to the spectrum of the signal. After the application of the filters, the spectrum of each filtered frequency band is calculated:
P
S, = log10 [£|S(k)|2.tf.(k)|, i = 1,2,...,M (2)
where S. is the output of the i-th filter, |S(&)|2 is the power spectrum, N is the FFT size, H. is the current filter and M is the total number of filters in the filter bank. Using the obtained S. values the LFCC are calculated by performing decorrelation of the filter bank outputs via discrete cosine transform (DCT):
( ro-as)^ r = o -, (3)
LFCC(r) = |S cos[r(i-°.5)nj,r = 0,1,...R-1
where r is the LFCC index, and R < M is the total number of unique LFCC that can be computed. Statistical standardization of the calculated LFCC is performed, so that their distributions are normalized to zero mean value and unit standard deviation on the base of the training set recordings:
Z = (4)
^p
After the normalization, the LFCC extracted from one frame are used to form a feature vector. All feature vectors are combined into a matrix, which is used for the training of a person-specific SVM classification model. For the creation of the SVM models a radial basis function kernel is used.
EXPERIMENTAL PROTOCOL
3.1 Data
The experimental evaluation of the LFCC features was carried out using the following experimental protocol.
The data for the experimental evaluation of the features is taken from the DEAP [7] database, which is a freely distributed database dedicated to the study of human affective states. It contains a number of physiological signals, including 32 channel EEG, recorded from 32 participants, while
they are watching music videos.
During the experimental evaluation of the features the data of ten participants from the DEAP database is used. Participant data is chosen based on the possibility of even separation of the EEG recordings in two groups - EEG recordings of negative emotional states and EEG recordings of non-negative emotional states. Each participant's data is split into two groups depending on the reported like\dislike ratings given for each song, which range from 1 to 9. During the separation, EEG recordings of songs with like/dislike rating lower than 4 were considered recordings of negative emotional states, and were tagged as "negative". Songs with rating higher than 4 were considered recordings of non-negative emotional states, and were tagged as "other". For cases in which a disparity of the elements in the two groups was observed a 5% tolerance to the separation threshold was applied.
The two sets of "negative" and "other" recordings were further separated into three subsets -training, development, and testing data. The data of the three subsets is distributed in the following ratio: 20% training, 20% development and 60% testing. All of the provided EEG recordings of the chosen participants were used in the experimental evaluation, which totals to 400 EEG recordings.
3.2 Model Generation
The evaluation of the EEG signal descriptors was conducted through the use of a person-specific detector of negative emotional states [8], its operation consisting of three major steps -model creation, threshold estimation and classification.
The first step of the classification process is the creation of a classification model. The "negative" and "other" training data sets, containing the most indicative examples for both classes, were used to create a person-specific SVM classifier with radial basis function kernel.
The generated model was then tuned on the development dataset, so that a person-specific decision threshold Tr can be computed:
1Y" D + — Ym D
Tr = n-m--(5)
2
where Dmg and Dpos are the portions of development data consisting of n recordings with negative tags and m with non-negative tags (neutral or positive). In order to find out the optimal training parameters for every emotion classifier, a series of grid searches were implemented on the development dataset.
During the evaluation of the trained classifiers, the threshold (5) was used to make a decision for each recoding. The person-specific recognition accuracy of the descriptor was evaluated in terms of percentage correct detections:
H
accuracy =-x 100, [%] (6)
Nrec
where H is the number of correctly classified recordings (or "hits") and Nrec is the total number of recordings in the testing datasets of the participant.
RESULTS
Based on the described experimental protocol an evaluation of the descriptive qualities of the LFCC for different number of filters was performed. During the experiments the EEG data of participants numbered 2, 11, 17, 21, 22, 24, 28, 29, 30 and 32 in the DEAP database was used. The achieved experimental results are presented in Table.1.
In the first column of the table, named „Feature", the different features used during the
experimental evaluation are listed. The first feature "Energy" is the energy of the signal, calculated in the frequency domain, without the use of any filters. The rest of the features are the LFCC calculated after the application of a filter bank. The experiments were carried out with filter banks containing 2, 3, 5, 7, 8, 10, 15, 20 and 30 filters. In columns 2 to 11 the results, achieved for the datasets of each participant are presented. In the last column "Mean" the mean accuracy for all participants for the given feature is presented.
Table 1: Classification results achieved during the experimental evaluation of the LFCC
features.
Feature P№02 P№11 P№17 P№21 P№22 P№24 P№28 P№29 P№30 P№32 Mean
Energy 66.7% 59.1% 72.7% 73.9% 81.8% 72.7% 83.3% 69.6% 82.6% 63.6% 72.6%
2 Filt. 66.7% 63.6% 77.3% 78.3% 86.4% 72.7% 75.0% 82.6% 69.6% 63.6% 73.6%
3 Filt. 79.2% 63.6% 81.8% 78.3% 81.8% 63.6% 75.0% 82.6% 78.4% 72.7% 75.7%
5 Filt. 79.2% 68.2% 77.3% 78.3% 81.8% 63.6% 70.8% 86.9% 73.9% 63.6% 74.4%
7 Filt. 62.5% 68.2% 86.4% 82.6% 81.8% 63.6% 75.0% 82.6% 78.4% 59.1% 74.0%
8 Filt. 62.5% 63.6% 72.7% 78.3% 90.9% 63.6% 70.8% 86.9% 78.4% 63.6% 73.2%
10 Filt. 62.5% 63.6% 77.3% 78.3% 81.8% 63.6% 79.2% 86.9% 78.4% 63.6% 73.5%
15 Filt 66.7% 68.2% 77.3% 78.3% 81.8% 63.6% 79.2% 82.6% 78.4% 63.6% 73.9%
20 Filt. 58.3% 68.2% 77.3% 65.2% 81.8% 40.9% 75.0% 78.3% 78.4% 59.1% 68.2%
30 Filt. 58.3% 68.2% 68.2% 69.6% 81.8% 50.0% 66.7% 78.3% 73.9% 63.6% 67.9%
From the results, presented in the table it can be seen that the Linear Frequency Cepstral coefficients provide a relatively high recognition accuracy - up to 75.7% in the case where a filter bank containing 3 filters was employed. It was observed that the achieved mean accuracy decreases with the increase of the number of filters. This effect occurs due to the narrowing of the frequency band covered by one filter. In addition it can be added that the EEG recordings are highly person-specific and the performance of different features often varies significantly between data of different participants.
CONCLLUSION
We have presented a study on the qualities of the Linear Frequency Cepstral coefficients as a feature for the recognition of negative emotional states. The performance of the features was evaluated using filter banks containing varying number of filters. Based on the conducted experiments we have concluded that the LFCC can be used successfully as descriptors for EEG based emotion recognition tasks.
REFERENCES
[1] G. Chanel, K. Ansari-Asl, and T. Pun, "Valence-arousal evaluation using physiological signals in an emotion recall paradigm," IEEE International Conference on Systems, Man and Cybernetics, Montreal, Que., Oct. 2007, pp. 2662-2667.
[2] T. Pun, T. Alecu, G. Chanel, J. Krongegg, and S. Voloshynovskiy, "Brain-computer interaction research at the computer vision and multimedia laboratory, university of Geneva," IEEE Trans. On Neural Systems and Rehabilitation Engineering, vol. 14, no. 2, June 2006.
[3] Bos, Danny Oude. "EEG-based emotion recognition." The Influence of Visual and Auditory Stimuli (2006): 1-17
[4] M. Murugappan, N. Ramachandran, and Y. Sazali. "Classification of human emotion from EEG using discrete wavelet transform." Journal of Biomedical Science and Engineering 3.04 (2010): 390.
[5] M. Othman, A. Wahab and R. Khosrowabadi, "MFCC for robust emotion detection using EEG," Communications (MICC), 2009 IEEE 9th Malaysia International Conference on, Kuala Lumpur, 2009, pp. 98-101.
[6] M. Othman, et al. "EEG emotion recognition based on the dimensional models of emotions." Procedia-Social and Behavioral Sciences 97 (2013): 30-37.
[7] Koelstra, S., Muhl, C., Soleymani, M., Lee, J.-S., Yazdani, A., Ebrahimi, T., Pun, T., Nijholt, A., Patras, I. DEAP: A Database for Emotion Analysis using Physiological Signals, IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18-31, 2012.
[8] Feradov, F., Ganchev, T. "Detection of Negative Emotional States from Electroencephalographic (EEG) signals", Annual Journal of Electronics, vol. 8, 2014, pp. 66-69.
