See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262288467

Detecting Emotion from EEG Signals Using the Emotive Epoc Device

Conference Paper · December 2012

DOI: 10.1007/978-3-642-35139-6_17

CITATIONS READS

60 7,289

2 authors:

Rafael Ramirez Zacharias Vamvakousis

University Pompeu Fabra University Pompeu Fabra
102 PUBLICATIONS   568 CITATIONS    5 PUBLICATIONS   97 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

TELMI Technology Enhanced Learning of Musical Instrument Performance View project

TELMI Technology Enhanced Learning of Musical Instrument Performance View project

All content following this page was uploaded by Rafael Ramirez on 18 February 2015.

The user has requested enhancement of the downloaded file.

 Detecting Emotion from EEG Signals
Using the Emotive Epoc Device

Rafael Ramirez and Zacharias Vamvakousis

Department of Information and Communication Technologies

Universitat Pompeu Fabra
Roc Boronat 138, 08018 Barcelona, Spain
{rafael.ramirez,zacharias.vamvakousis}@upf.edu

Abstract. The study of emotions in human-computer interaction has

increased in recent years in an attempt to address new user needs. At the
same time, it is possible to record brain activity in real-time and discover
patterns to relate it to emotional states. This paper describes a machine
learning approach to detect emotion from brain activity, recorded as elec-
troencephalograph (EEG) with the Emotic Epoc device, during auditory
stimulation. First, we extract features from the EEG signals in order to
characterize states of mind in the arousal-valence 2D emotion model. Us-
ing these features we apply machine learning techniques to classify EEG
signals into high/low arousal and positive/negative valence emotional
states. The obtained classifiers may be used to categorize emotions such
as happiness, anger, sadness, and calm based on EEG data.

1 Introduction

The study of emotions in human-computer interaction has increased in recent

years. This is due to the growing need for computer applications capable of de-
tecting the emotional state of users [17]. Motivated by every day interaction
among humans, a great part of the research in this area has explored detecting
emotions from facial and voice information. Under controlled situations, cur-
rent emotion-detection computer systems based on such information are able
to classify emotions with considerable accuracy [18]. However, emotions are not
always manifested by means of facial expressions and voice information. Psy-
chologists distinguish between physiological arousal, behavioral expression, and
the conscious experience of emotions. Facial and voice information is related
only to behavioral expression which can be consciously controlled and modified,
and which interpretation is often subjective. Thus, other approaches to detect
emotion have been proposed which focus on different physiological information
such as heart rate, skin conductance, and pupil dilation [18,16]. A still relatively
new field of research in affective brain-computer interaction attempts to de-
tect emotion using electroencephalograms (EEGs) [2,3]. There have been several
approaches to EEG-based emotion detection, but there is still little consensus
about definite conclusions.

F.M. Zanzotto et al. (Eds.): BI 2012, LNCS 7670, pp. 175–184, 2012.

c Springer-Verlag Berlin Heidelberg 2012
176 R. Ramirez and Z. Vamvakousis

In this paper we describe an approach to detecting emotion from electroen-

cephalogram signals measured with a (low-cost) Emotiv EPOC headset. We
present to subjects auditory stimuli from a library of emotion-annotated sounds
and record their response EEG activity. We then filter and process the signal
in order to extract emotion-related features and apply machine learning tech-
niques to classify emotional states into high/low arousal and positive/negative
valence (e.g. happiness is a state with high arousal and positive valence, whereas
sadness is a state with low arousal and negative valence). Our approach differs
from previous works in that we do not rely in subject self-reported emotional
states during stimuli presentation. Instead, we use a library of emotion-annotated
sounds publicly available for emotional research. Figure 1 illustrates the different
steps of our approach.

Fig. 1. Schematic view of the system

The rest of the paper is organized as follows: Section 2 presents background

to this research. Section 3 described the data collection process and methods.
In Section 4 we describe our approach to EEG-based emotion classification and
report the results, and finally Section 5 presents some conclusions and further
research.

2 Background
The firing of neurons in the brain trigger voltage changes. The electrical activity
measured by the electrodes in an EEG headset corresponds to the field poten-
tials resulting from the combined activity of many individual neuronal cells in the
 Detecting Emotion from EEG Signals Using the Emotive Epoc Device 177

brain cortex. However, the measured cortical activity is distorted by the tissue
and skull between the electrodes and the neurons. This introduces noise and re-
duce the intensity of the recorded signals. In despite of this, EEG measurements
still provide important insight into the electrical activity of the cortex.
The frequency of EEG measurements ranges from 1 to 80Hz, with amplitudes
of 10 to 100 microvolts [9]. Signal frequencies have been divided into different
bands, since specific frequency waves are normally more prominent in particular
states of mind. The two most important frequency waves are the alpha waves
(8-12Hz) and the beta waves (12-30Hz). Alpha waves predominantly originate
during wakeful relaxation mental states, and are most visible over the parietal
and occipital lobes. Intense alpha wave activity have also been correlated to brain
inactivation. Beta wave activity, on the other hand, is related to an active state
of mind, most prominent in the frontal cortex during intense focused mental
activity [9].
Alpha and beta wave activity may be used in different ways for detecting
emotional (arousal and valence) states of mind in humans (more details later).
Choppin [3] propose to use EEG signals for classifying six emotions using neural
networks. Choppin’s approach is based on emotional valence and arousal by
characterizing valence, arousal and dominance from EEG signals. He characterize
positive emotions by a high frontal coherence in alpha, and high right parietal
beta power. Higher arousal (excitation) is characterized by a higher beta power
and coherence in the parietal lobe, plus lower alpha activity, while dominance
(strength) of an emotion is characterized as an increase in the beta / alpha
activity ratio in the frontal lobe, plus an increase in beta activity at the parietal
lobe.
Oude [14] describes an approach to recognize emotion from EEG signals mea-
sured with the BraInquiry EEG PET device. He uses a limited number of elec-
trodes and trains a linear classifier based on Fishers discriminant analysis. He
considers audio, visual and audiovisual stimuli and trains classifies for posi-
tive/negative, aroused/calm and audio/visual/audiovisual.
Takahashi [18] use a headband of three dry electrodes to classify five emotions
(joy, anger, sadness, fear, and relaxation) based on multiple bio-potential signals
(EEG, pulse, and skin conductance). He trains classifiers using support vector
machines and reports the resulting classifying accuracy both using the whole set
of bio-potential signals, and solely based on EEG signals.
Lin at al. [10] apply machine-learning techniques to categorize EEG signals ac-
cording to subject self-reported emotional states during music listening. They pro-
pose a framework for systematically seeking emotion-specific EEG features and
exploring the accuracy of the classifiers. In particular, they apply support vector
machines to classify four emotional states: joy, anger, sadness, and pleasure.

3 Data Collection

EEG data in this study were collected from 6 healthy subjects (3 males and 3
females) with average age of 30.16 during listening to emotion-annotated sounds.
178 R. Ramirez and Z. Vamvakousis

For collecting the data we used the Emotiv EPOC headset, recently released by
the Emotiv Company [7]. This headset consists of 14 data-collecting electrodes
and 2 reference electrodes, located and labeled according to the international
10-20 system [12]. Following the international standard, the available locations
are: AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8 and AF4. Figure
2 shows the 14 Emotiv EPOC headset electrode positions. The EEG signals are
transmitted wirelessly to a laptop computer.

Fig. 2. Emotiv EPOC headset electrode positioning

Subjects were instructed to look at a cross in the a computer screen and

to remain seated during the experiment. Subjects listened to selected sounds
from the IADS library of emotion-annotated sounds [1] which is available for
emotion research. Based on the annotations provided by the stimuli databases,
we selected 12 sound stimuli situated in the extremes on the arousal-valence
emotion plane: three positive/aroused, three positive/calm, three negative/calm,
and three negative/aroused. The stimuli were selected to be as much as possible
on the extremes of the two-dimensional emotion plane and as unanimous as
possible, since we do not consider self-reporting information to cater for person-
dependent deviations.
Initially, the subjects are informed about the experiment procedure and in-
structed to follow the usual guidelines during stimuli presentation (e.g. do not
blink or move). Once this was done, 12 sound stimuli are randomly presented
each one for five seconds and a 10 second silent rest is inserted between stimuli.
 Detecting Emotion from EEG Signals Using the Emotive Epoc Device 179

The purpose of the 10 second silent rests is to set a neutral emotional state of
mind in between stimuli.

4 Data Classification
4.1 Feature Extraction
As mentioned in the previous section, in EEG signals the alpha (8-12Hz) and beta
(12-30Hz) bands are particular bands of interest in emotion research for both
valence and arousal [13]. The presence of EOG artifacts (eye movement/blinking)
is most dominant below 4Hz, ECG (heart) artifacts around 1.2Hz, and EMG
(muscle) artifacts above 30Hz. Non physiological artifacts caused by power lines
are normally present above 50Hz [6,5].
Thus, fortunately a byproduct of extracting the alpha and beta frequencies is
that much of the noise present in EEG signals is considerably reduced. We apply
bandpass filtering for extracting alpha and beta frequency bands. Using Fourier
frequency analysis, the original signal is split up in frequencies in order to remove
specific frequencies, before transforming back the signal with only the frequen-
cies of interest. For this research, we apply the bandpass filter implementation
provided by the OpenVibe software [15].
From the EEG signal of a person, we determine the level of arousal, i.e. how
relaxed or excited the person is, by computing the ratio of the beta and alpha
brainwaves as recorded by the EEG. We measure the EEG signal in four locations
(i.e. electrodes) in the prefrontal cortex: AF3, AF4, F3 and F4 (see Figure 2).
As mentioned before, beta waves are associated with an alert or excited state of
mind, whereas alpha waves are more dominant in a relaxed state. Alpha activity
has also been associated to brain inactivation. Thus, the beta/alpha ratio is a
reasonable indicator of the arousal state of a person.
In order to determine the valence level, i.e. negative or positive state of mind,
we compare the activation levels of the two cortical hemispheres. This is moti-
vated by psychophysiological research which has shown the importance of the
difference in activation between the cortical hemispheres. Left frontal inactiva-
tion is an indicator of a withdrawal response, which is often linked to a negative
emotion. On the other hand, right frontal inactivation may be associated to an
approach response, or positive emotion.
As mentioned before, high alpha activity is an indication of low brain activity,
and vice versa. Thus, an increase in alpha activity together with a decrease in
beta waves may be associated with cortical inactivation [13]. F3 and F4 are the
most used positions for looking at this alpha activity, as they are located in the
prefrontal lobe which plays a crucial role in emotion regulation and conscious
experience.
Although previous research suggests that hemispherical differences are not
an indication of affective valence (feeling a positive or negative emotion), it has
been suggested that it is an indication of motivational direction (approach or
withdrawal behavior to the stimulus) [8]. In general, however, affective valence is
related to motivational direction. Therefore, comparing hemispherical activation
180 R. Ramirez and Z. Vamvakousis

seems to be a reasonable method to detect valence. Thus, we estimate the valence

value in a person by computing and comparing the alpha power a and beta power
b in channels F3 and F4. Specifically,

valence = aF 4 /bF 4 − aF 3 /bF 3 .

4.2 Learning Task

In this section we describe our approach to training and evaluating classifiers for
the task of detecting the emotional state of mind of a person given the person’s
observed EEG data. We approach this problem as a two 2-class classification
problem. In particular, we apply machine learning techniques to classify high/low
arousal and positive/negative valence emotional states. The obtained classifiers
can be used to classify emotions such as happiness, anger, sadness, and calm.
Figure 3 shows these emotions in the arousal/valence plane.

Fig. 3. Emotional states and their positions in the valence/arousal plane

We are interested in inducing two classifiers of the following forms:

ArousalClassif ier(EEGdata([t, t + c])) → {high, low}
and
V alenceClassif ier(EEGdata([t, t + c])) → {positive, negative}
where EEGdata([t, t + c]) is the EEG data observed at time interval [t, t + c]
and {high, low} and {positive, negative} are the sets of emotional states to be
discriminated. The results reported in this paper are obtained with c=1s and
with increments of t of 0.0625s. For each subject in the EEG data sets we train
a separate classifier.
 Detecting Emotion from EEG Signals Using the Emotive Epoc Device 181

4.3 Algorithms
In this paper we evaluate two classifiers, Linear Discriminant Analysis (LDA)
[11] and Support Vector Machines (SVM) [4], for classifying an emotion state for
each EEG segment. Linear discriminant analysis and the related Fisher’s linear
discriminant are methods used in statistics, pattern recognition and machine
learning to find a linear combination of features which characterizes or separates
two or more classes of objects or events. The resulting combination may be
used as a linear classifier. LDA is closely related to regression analysis, which
also attempt to express one dependent variable as a linear combination of other
features. In regression analysis however, the dependent variable is a numerical
quantity, while for LDA it is a categorical variable (i.e. the class label).
On the other hand, SVM is one of the most popular supervised learning algo-
rithms for solving classification problems. The basic idea in SVM is to project
input data onto a higher dimensional feature space via a kernel transfer func-
tion, which is easier to be separated than that in the original feature space.
Depending on input data, the iterative learning process of SVM would even-
tually converge into optimal hyperplanes with maximal margins between each
class. These hyperplanes would be the decision boundaries for distinguishing dif-
ferent data clusters. Here, we use linear and radial basis function (RBF) kernel
to map data onto a higher dimension space. The results reported are obtained
using the LDA and SVM implementations in the OpenVibe software [15].
We evaluated each induced classifier by performing the standard 10-fold cross
validation in which 10% of the training set is held out in turn as test data while
the remaining 90% is used as training data. When performing the 10-fold cross
validation, we leave out the same number of examples per class. In the data sets,
the number of examples is the same for each class considered, thus by leaving
out the same number of examples per class we maintain a balanced training set.

4.4 Results
Given that we are dealing with 2-class classification tasks and that the number
of instances in each class is the same, the expected classification accuracy of the
default classifier (one which chooses the most common class) is 50% (measured in
correctly classified instances percentage). For the high-versus-low arousal, and
the positive-versus-negative valence classifiers the average accuracies obtained
for SVM with radial basis function kernel classifier were 77.82%, and 80.11%,
respectively. For these classifiers the best subject’s accuracies were 83,35%, and
86.33%, respectively. The correctly classified instances percentage for each sub-
ject and each learning method is presented in Figures 4 and 5.

4.5 Discussion
The difference between the results obtained and the accuracy of a baseline classi-
fier, i.e. a classifier guessing at random confirms that the EEG data contains suffi-
cient information to distinguish between high/low arousal and positive/negative
182 R. Ramirez and Z. Vamvakousis

Fig. 4. Classifiers (LDA, SVM with linear kernel, and SVM with radial basis function
kernel) accuracies for high-versus-low arousal for all subjects

Fig. 5. Classifiers (LDA, SVM with linear kernel, and SVM with radial basis function
kernel) accuracies for positive-versus-negative valence for all subjects

valence states, and that machine learning methods are capable of learning the
EGG patterns that distinguish these states. It is worth noting that both learning
algorithm investigated (LDA and SVM) produced better than random classifi-
cation accuracies. This supports our statement about the feasibility of training
classifiers using the Emotiv Epoc for the tasks reported.
The accuracy of the classifiers for the same task for different subjects varies
significantly, even using the same learning method. Subjects producing high
 Detecting Emotion from EEG Signals Using the Emotive Epoc Device 183

accuracies with one learning method tend to produce high accuracies with the
other learning methods. These uneven accuracies among subjects may be due to
different degrees of emotional response between different individuals, or to the
amount of noise for different subjects. In any case, it has been reported that
there exists considerable variation in EEG responses among different subjects.
It is worth mentioning that in all the experiments performed we provided
no self-assessment information about the emotional states by the subjects. This
contrasts with other approaches (e.g. [10]) where EEG data is categorized ac-
cording to subject self-reported emotional states. Incorporating self-assessment
information would very likely improve the accuracies of the classifiers.

5 Conclusion

We have explored and compared two machine learning techniques for the problem
of classifying the emotional state of a person based on EEG data using the
Emotiv Epoc headset. We considered two machine learning techniques: linear
discriminant analysis and support vector machines. We presented the results
of the induced classifiers which are able to discriminate between high-versus-
low arousal and positive-versus-negative valence. Our results indicate that EEG
data obtained with the Emotiv Epoc device contains sufficient information to
distinguish these emotional states, and that machine learning techniques are
capable of learning the patterns that distinguish these states. Furthermore, we
proved that it is possible to train successful classifiers with no to self-assessment
of information about the emotional states by the subjects. As future work, we
are particularly interested in systematically exploring different feature extraction
methods and learning methods in order to improve the accuracy of the induced
classifiers.

Acknowledgments. This work is supported by the Spanish TIN project DRIMS

(TIN2009-14274-C02-01).

References
1. Bradley, M.M., Lang, P.J.: International Affective Digitized Sounds (IADS): Stim-
uli, Instruction Manual and Affective Ratings. The Center for Research in Psy-
chophysiology, University of Florida, Gainesville, FL, USA (1999)
2. Chanel, G., Kronegg, J., Grandjean, D., Pun, T.: Emotion Assessment: Arousal
Evaluation Using EEG’s and Peripheral Physiological Signals. In: Gunsel, B., Jain,
A.K., Tekalp, A.M., Sankur, B. (eds.) MRCS 2006. LNCS, vol. 4105, pp. 530–537.
Springer, Heidelberg (2006)
3. Choppin, A.: Eeg-based human interface for disabled individuals: Emotion expres-
sion with neural networks. Masters thesis, Tokyo Institute of Technology, Yoko-
hama, Japan (2000)
4. Cristianini, N., Shawe-Taylor, J.: An Introduction to Support Vector Machines.
Cambridge University Press (2000)
 184 R. Ramirez and Z. Vamvakousis

5. Coburn, K., Moreno, M.: Facts and artifacts in brain electrical activity mapping.
Brain Topography 1(1), 37–45 (1988)
6. Fatourechi, M., Bashashati, A., Ward, R.K., Birch, G.E.: EMG and EOG artifacts
in brain computer interface systems: A survey. Clininical Neurophysiology (118),
480–494 (2007)
7. Emotiv Systems Inc. Researchers, http://www.emotiv.com/researchers/
8. Harmon-Jones, E.: Clarifying the emotive functions of asymmetrical frontal cortical
activity. Psychophysiology 40(6), 838–848 (2003)
9. Kandel, E.R., Schwartz, J.H., Jessell, T.M.: Principles of Neural Science. Mc Graw
Hill (2000)
10. Lin, Y.-P., Wang, C.-H., Jung, T.-P., Wu, T.-L., Jeng, S.-K., Duann, J.-R., Chen,
J.-H.: EEG-Based Emotion Recognition in Music Listening. IEEE Transactions on
Biomedical Engineering 57(7) (2010)
11. Mika, S., et al.: Fisher Discriminant Analysis with Kernels. In: IEEE Conference
on Neural Networks for Signal Processing IX, pp. 41–48 (1999)
12. Niedermeyer, E., da Silva, F.L.: Electroencephalography, Basic Principles, Clinical
Applications, and Related Fields, p. 140. Lippincott Williams & Wilkins (2004)
13. Niemic, C.P.: Studies of emotion: A theoretical and empirical review of psychophys-
iological studies of emotion. Journal of Undergraduate Research 1, 15–18 (2002)
14. Bos, D.O.: EEG-based Emotion Recognition: The Influence of Visual and Auditory
Stimuli
15. OpenViBE: An Open-Source Software Platform to Design, Test, and Use Brain-
Computer Interfaces in Real and Virtual Environments. MIT Press Journal Pres-
ence’ 19(1), 35–53 (2010)
16. Partala, T., Jokiniemi, M., Surakka, V.: Pupillary responses to emotionally
provocative stimuli. In: ETRA 2000: Proceedings of the 2000 Symposium on Eye
Tracking Research & Applications, pp. 123–129. ACM Press, New York (2000)
17. Picard, R.W., Klein, J.: Toward computers that recognize and respond to user
emotion: Theoretical and practical implications. Interacting with Computers 14(2),
141–169 (2002)
18. Takahashi, K.: Remarks on emotion recognition from bio-potential signals. In: 2nd
International Conference on Autonomous Robots and Agents, pp. 186–191 (2004)

View publication stats