Emotional Valence Detection based on a Novel Wavelet Feature

Extraction Strategy using EEG Signals

Hao Zhang

, Shin’ichi Warisawa

1,2

and Ichiro Yamada

1,2

School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, Japan

Graduate School of Frontier Science, The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa, Chiba, Japan

Keywords:

Emotion Assessment, Valence Detection, Brain-Computer Interface (BCI), EEG, Wavelet Feature, Probabilis-

tic Neural Network (PNN), Genetic Algorithm (GA).

Abstract:

This paper presents a novel feature extraction strategy in the time-frequency domain using discrete wavelet

transform (DWT) for valence level detection using electroencephalography (EEG) signals. Signals from dif-

ferent EEG electrodes are considered independently for the ﬁrst time in order to ﬁnd an optimum combination

through different levels of wavelet coefﬁcients based on the genetic algorithm (GA). Thus, we take into consid-

eration useful information obtained from different frequency bands of brain activity along the scalp in valence

level detection, and we introduce a new set of features named the cross-level wavelet feature group (CLWF).

The effectiveness of this approach is strongly supported by the analytical results of experiments in which EEG

signals with valence level labels were collected from 50 healthy subjects. High accuracy was achieved for both

2-level (98%) and 3-level valence detection (90%) by applying leave-one-out cross validation using a proba-

bilistic neural network (PNN). In addition, light-weighted sets with less than half EEG recording electrodes are

proposed, which can achieve a high accuracy (86% for 3-level valence detection) with offering convenience

of users and reducing computational complexity.

1 INTRODUCTION

Emotion assessment is essential for gaining a deeper

understanding of human beings and preventing

lifestyle diseases. A large body of literature indicates

that negative emotions have harmful effects on peo-

ple’s health (Sirois, 2003), whereas positive emotions

foster good physical health (Salovey and Rothman,

2000).

Several emotion modeling approaches have been

proposed in the existing literature, including discrete

models (Darwin, 1872), appraisal models (Arnold,

1950), and dimensional models (Wundt, 1905). We

adopt a category of dimensional models that consider

multidimensional space, where each dimension repre-

sents a fundamental property shared among all emo-

tions in this study. Many researchers (Constantinou

et al., 2013) agree that emotions have at least two

indexes, including valence and arousal, and emotion

can therefore be represented through complex inter-

actions related to valence and arousal. In the emotion

model proposed by Russell (Russell, 1980), arousal

represents a quantity that varies from calm to ex-

citement, whereas valence represents a quantity that

varies from positive to negative.

In previous studies, the voice signal has been con-

sidered to be an effective signal for emotion related

analysis. An emotional valence detection system was

accuracy (Tahon et al., 2012). A large number of

researchers have adopted physiological signals for

evaluating different levels of valence. Picard et al.

proposed a method for positive and negative classi-

ﬁcation with 87% accuracy using electromyography

(EMG), galvanic skin response (GSR), respiration,

blood volume pulse (BVP), and electrocardiogram

(ECG) signals (Picard et al., 2001). In this research,

we study the valence dimension of the emotion model

proposed by Russell, which is related to positive ver-

sus negative affect.

The interest in using electroencephalography

(EEG) for valence detection has increased recently

because of the latest signiﬁcant developments in

brain-computer-interface (BCI) research. The fact is

that ignoring a user’s emotional state with any inter-

face dramatically impedes the performance and leads

to a bad user experience. BCI systems are now be-

ing developed for a wide range of applications in ar-

Zhang H., Warisawa S. and Yamada I..

Emotional Valence Detection based on a Novel Wavelet Feature Extraction Strategy using EEG Signals.

DOI: 10.5220/0004764600520059

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2014), pages 52-59

ISBN: 978-989-758-010-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

eas such as healthcare, gaming, security, and mar-

keting. These applications can beneﬁt largely from

knowing and adapting their operations according to

users’ emotional states. This emotion information can

then be utilized to provide a better user experience by

controlling the BCI through affect. It has been shown

that a correlation exists between emotions and brain

activity (Nardi, 2011), especially in the prefrontal

cortex and the amygdala (Adolphs et al., 2003). A

small amount of research on emotion related classi-

ﬁcation is based on EEG, and most of this research

has not been able to obtain good results compared

with the performance using other physiological sig-

nals. Based on the asymmetry represented by EEG

signals, the mean accuracy for three valence classes

(negative, neutral, and positive) of picture stimula-

tions was reported to be 44% based on frequency do-

main features and 49% based on combining time and

frequency domain EEG features(Schaaff and Schultz,

2009). By combining EEG and some physiological

signals (heart rate and pulse), Takahashi and Tsuk-

aguchi achieved a recognition rate of 62.3% with a

neural network (NN) classiﬁer and 59.7% with a sup-

port vector machine (SVM) for classifying pleasure

and displeasure (Takahashi and Tsukaguchi, 2003).

Hosseini adopted entropy analysis of EEG signals and

achieved pleasure and displeasure classiﬁcation with

an accuracy of 72.35% using SVM (Hosseini, 2011).

The objective of this study was to ﬁnd effective

features and to formulate a robust method that can

distinguish between different valence levels with high

accuracy from EEG signals. Affective pictures were

adopted in this research to elicit different valence lev-

els. In section II, the experimental design and record-

ing protocol are introduced. In section III and IV, the

analytical model including feature extraction and op-

timization is described. The results are discussed in

section V, and ﬁnally, in section VI, conclusions that

have been drawn based on the proposed research and

perspectives are discussed.

2 EXPERIMENTAL DESIGN

2.1 Valence Elicitation

Three levels of valence have been deﬁned based

on arousal-valence space based on the emotion

model proposed by Russell (Russell, 1980). The

most widely adopted emotion elicitation technique

is to use pictures to evoke various emotion states

based on valence-arousal space (Katsis et al., 2011).

We adopted a very popular picture-based, emotion-

evoking database called the International Affective

Picture System (IAPS) (Lang et al., 2008), which

contains pictures labeled with values of valence and

arousal. We chose pictures with a small range of va-

lence values but a large variance of arousal values in

order to collect data that would ensure a robust model

of valence detection. The IAPS labels respective pic-

tures for males and females, so we selected differ-

ent pictures for male and female subjects in order to

achieve the same valence level elicitation. The posi-

tioning in terms of valence and arousal of the pictures

selected for the three levels of valence is illustrated

in Fig. 1 against all pictures for males and females,

respectively.

Valence

Arousal

Group 1

Group 2

Figure 1: Two groups of pictures selected from IAPS for va-

lence stimulation for Females (Group 1) and Males (Group

2). The red dots represent the pictures selected from IAPS

among all pictures, shown as gray dots.

2.2 Experimental Protocols

The experiments were designed to stimulate a certain

level of valence from viewing multiple pictures from

the IAPS database. The subjects were required to sit

in a dark room and look at the pictures that appeared

on a screen. The protocol of the stimulation proce-

dure used with each subject is illustrated in Fig. 2.

We collected EEG signals from 50 healthy Japanese

subjects (35 males and 15 females). The ages of the

subjects ranged from the 20s to the 70s. The EEG

signals were recorded by a Nihon kohden EEG-1200

using electrodes placed according to the international

10-20 system; the sampling rate of EEG signal acqui-

sition was 1000 Hz. This experiment was conducted

with the permission from Research Ethics and Safety

committee of The University of Tokyo.

2.3 EEG Dataset

EEG signals are affected by noise such as pulses, line

noise, and artifacts. The noise generated from eye

blinks is the most difﬁcult type to deal with by using

signal processing techniques such as wavelet analy-

sis. For this reason, the subjects were asked to refrain

from blinking their eyes during the experiments. A ﬁl-

ter was applied to delete the line noise at 50 Hz. Previ-

ous research (Murugappan et al., 2010) shows that the

EmotionalValenceDetectionbasedonaNovelWaveletFeatureExtractionStrategyusingEEGSignals

Pic1

Pic2

Pic10Pic3

10 s per picture

Pic11

Pic12

Pic20Pic13

10 s per picture

Pic21

Pic22

Pic30Pic23

10 s per picture

3 min. rest

Level 1

s!mula!on

Level 2

s!mula!on

Level 3

s!mula!on

Figure 2: Valence stimulation procedure.

emotions are not only correlated to the brain activity

recorded in the frontal area but also to that recorded

in other areas, and therefore, higher accuracy was ob-

tained by using EEG electrodes along all the scalp.

We used 16 channels (Fp1, Fp2, F3, F4, C3, C4, P3,

P4, O1, O2, F7, F8, T3, T4, T5, T6) for analysis based

on the International 10-20 system. On the basis of

the correlation theory of brain activity asymmetry and

emotions (Alves et al., 2008), our approach is to com-

pare the frontal area, the posterior area, and the en-

tire area of brain activity with electrodes on both left

and right brain hemispheres, and this approach gains

a better understanding of how emotion states can be

interpreted using EEG signals in different areas. The

test cases using different combinations of EEG elec-

trodes are illustrated in Fig. 3.

Frontal Area (FA) Posterior Area (PA) Entire Area (EA)

Figure 3: Test cases of brain activity in different areas.

3 METHODS

3.1 Wavelet Analysis

We adopted discrete wavelet decomposition (Morlet,

1984) to better understand the frequency and location

information of the EEG signals. The discrete wavelet

decomposition deﬁnition is expressed as

m,n

∞

−∞

f (t)ψ

m,n

(t)dt, (1)

m,n

(t) = a

−m/2

ψ(a

−m

t − nb

), (2)

where ψ

m,n

(t) is a wavelet mother function, the in-

teger m controls wavelet dilation, and n controls trans-

lation. Here, a

is a speciﬁed ﬁxed dilation setup

parameter set at a value greater than 1, and b

the location parameter, which must be greater than 0.

Also, T

m,n

are the discrete wavelet values given on a

scale-location grid of index m,n, which are known as

wavelet coefﬁcients or detail coefﬁcients.

Wavelet analysis has been successfully used for

feature extraction and for many classiﬁcation tasks

based on EEG (Sherwood, 2009). No objective

method for mother wavelet selection has been devel-

oped so far for classiﬁcation, and the methods are

mostly empirical. We selected Daubechies 5 wavelet

(Db5) since it has been adopted successfully for mod-

eling complex signals such as EEG signals (Liu and

Xu, 2010), and we applied seven levels of wavelet de-

composition, which is determined by taking the EEG

frequency bands into consideration.

3.2 Statistical Parameters

Statistical parameters shows effectiveness on demon-

strating affective states from physiological signals

(Van den Broek and Westerink, 2009). Several statis-

tical parameters including mean, standard deviation

(std), skewness, and kurtosis are used in this study to

demonstrate the characteristics of transformed EEG

signals using discrete wavelet decomposition.

3.3 Genetic Algorithm (GA)

We adopted a genetic algorithm (Holland, 1975) for

feature selection. The GA mimics the process of nat-

ural evolution to ﬁnd beneﬁcial adaptations to a com-

plex environment. A chromosome in GA is an en-

coding that represents the decision variable of an op-

timization problem. A ﬁnite set of chromosomes in

GA is called a population. Each chromosome is rated

by the ﬁtness function on its ‘ﬁtness’, which deter-

mines how good it is in solving the optimization prob-

lem. Crossover refers to the generation of two new

offspring by mating two parental chromosomes. A

mutation simply ﬂips randomly the binary value of

one or more bits. Crossover and mutation provide op-

portunities for chromosomes that have higher ﬁtness

values to evolve.

The reasons why we adopt GA are shown in the

following. According to previous studies, the corre-

lation theory of brain activities asymmetry and emo-

tions were proved by researchers (Alves et al., 2008).

GA allows us to realize our proposal to select ap-

propriate information from each EEG electrode and

eventually obtain the overall pattern of brain activ-

ities monitored by multiple EEG electrodes. Other

data-driven feature selection approaches cannot give

attention to the overall pattern of brain activities. The

HEALTHINF2014-InternationalConferenceonHealthInformatics

selected features can be only extracted from one or

several EEG electrodes so that it fails to demonstrate

the whole brain activities asymmetry since there is

no priori biological knowledge applied on those al-

gorithms. These considerations are supported by a re-

cent paper proposed in a similar application concern-

ing neuroimaging (Chu et al., 2012), as they have two

ﬁndings related to feature selection including data-

driven feature selection was no better than adopting

whole data and a priori biological knowledge was ef-

fective to guide feature selection. As for the imple-

mentation of GA with proposed features, we general-

ized our problem of wavelet level selection according

to different EEG positions to an optimization issue of

minimizing the ﬁtness function.

3.4 Principal Component Analysis

(PCA)

PCA (Jolliffe, 2002) is a common visualization

method that is applied through dimensionality reduc-

tion by performing a covariance analysis between fac-

tors. Technically, a principal component can be de-

ﬁned as a linear combination of optimally weighted

observed variables, the ﬁrst k principal components

capture the greatest variance in the data among all k-

dimensional orthonormal linear combinations of the

original variables. In our case, PCA was used for fea-

ture space visualization to observe the separability of

different valence levels by speciﬁc features.

3.5 Probabilistic Neural Network (PNN)

Artiﬁcial neural networks are effective when used

with such signals as EEG since they are very robust in

dealing with non-linear and complex signals. More-

over, the fault tolerance of artiﬁcial neural networks

is necessary in order to reduce the inﬂuence of noise.

PNN (Specht, 1990) is one kind of artiﬁcial neural

network that has been proven to be suitable for clas-

siﬁcation tasks by many researchers (Sivakumar and

.Kannan, 2009). The operations are designed into a

multi-layered feed-forward network with four layers.

In this work, PNN is adopted considering its charac-

teristics of fast training process and additional train-

ing samples can be added without extensive retrain-

ing, which make it practical for model improvement

when we have large training database. The inputs of

PNN were the optimized features, and the output was

the predicted valence level.

4 IMPLEMENTATION

4.1 Cross-level Wavelet Feature

Extraction

The statistical parameters have been shown to be use-

ful for extracting features from raw or processed EEG

signals. Although time or frequency domain features

from EEG signals failed to provide good accuracy

for valence detection (Schaaff and Schultz, 2009),

(Takahashi and Tsukaguchi, 2003), (Hosseini, 2011),

we explore new possibilities in time-frequency do-

main features. Murugappan et al. conducted a study

on clustering different emotions by wavelet features

from EEG signals using 24 and 63 channels (Muru-

gappan et al., 2007) and represented the potential of

time-frequency features for identifying emotions. The

theory of wavelet decomposition can roughly takes

into consideration EEG frequency bands. The fre-

quency information of each decomposition level is

listed in Table 1.

Table 1: Frequency range of different wavelet decomposi-

tion levels with EEG frequency band information.

Decomposition Frequency EEG frequency

level range (Hz) band

A7 0-4 Delta

D7 4-8 Theta

D6 8-16 Alpha

D5 16-32 Beta

D4 32-63 Low gamma

D3 63-125 High gamma

D2 125-250 —

D1 250-500 —

Our strategy is to formulate a novel cross-level

wavelet feature group (CLWF) for valence detection

based on GA, as opposed to the strategy of extracting

a mono-level wavelet feature group (MLWF). EEG

consists of a multi-channel system that can record a

great deal of information with a lot of noise. Much

research has focused on how to omit a number of

electrodes and select the most useful ones in order

to achieve a moderate level of performance (Arvaneh

et al., 2011). The studies that have adopted wavelet

features have only considered a single decomposition

level or have combined levels from all EEG electrodes

(Sherwood, 2009). However, those studies ignore the

fact that useful information for classiﬁcation may be

represented in different frequency ranges among dif-

ferent EEG electrodes. If we implement this idea

EmotionalValenceDetectionbasedonaNovelWaveletFeatureExtractionStrategyusingEEGSignals

EEG signals

Fp1

Wavelet

Decomposed signals

GA selected levels

Cross-level signals

D2 of EEG signal in O2

D1 of EEG signal in F3

D3 of EEG signal in Fp1

Statistical

measurement

Cross-level features

Figure 4: Schematic of cross-level wavelet feature estimation from EEG signals.

by a brute-force search, the order of magnitude for

the searching cases will be n

, where m is the num-

ber of EEG electrodes, and n represents the levels of

wavelet decomposition. For instance, there will be

152,587,890,625 cases with 16 EEG electrodes and 5

levels of wavelet decomposition, which makes it ex-

tremely difﬁcult to ﬁnd the optimized feature combi-

nations. Thus, GA is introduced in this step to solve

this optimization issue. The ﬁtness function is de-

signed as Eq. 3.

Fitness = 1 −Accuracy (3)

Accuracy is calculated using the leave-one-out cross

validation (LOOCV) method based on PNN. The in-

put feature for GA is a matrix with n levels of wavelet

decomposition for all EEG electrodes. A chromo-

some is an array of numbers that represent levels of

wavelet coefﬁcients related to EEG electrodes; the

GA will select one level from each electrode to formu-

late the optimized CLWF. GA parameters are given in

Table 2.

Table 2: GA parameters.

Population size 10 chromosomes

Elite count 2

Mutation rate 0.01

Crossover rate 0.8

A relatively limited population size and high mu-

tation rate as well as the LOOCV design in the ﬁt-

ness function are for preventing overﬁtting in GA. A

schematic of the entire process of feature estimation is

illustrated in Fig. 4. In this ﬁgure, EEG signals from

positions O2, F3, and Fp1 are used as examples to il-

lustrate the feature extraction process. The best com-

bination of wavelet coefﬁcients (those indicated in red

in Fig. 4) are selected based on GA from a huge num-

ber of combinations, and the statistical parameters are

calculated for a later classiﬁcation process.

4.2 Reduced Set of EEG Electrodes

We explore different sets of EEG electrodes to fur-

ther understand the functional area for emotional ac-

tivities, and reduce the number of EEG electrodes for

the convenience of users and computational simplic-

ity. Besides two comparison sets of frontal and poste-

rior areas, the same GA approach is performed to ﬁnd

another optimized set of EEG electrodes placement

for each subject. The input for GA is a matrix with

n levels of wavelet decomposition for all EEG elec-

trodes; the GA will select less than a certain number

of electrodes for the reduced set. Multiple electrodes

are needed to understand the difference of brain activ-

ities in different brain areas, we consider 6 electrodes

are appropriate based on preliminary attempts and for

comparisons with the other two sets using 6 electrodes

mentioned in Fig. 3.

4.3 Classiﬁcation

Since each EEG record corresponds to a picture stim-

ulation from IAPS, it is easy to give labels to the clas-

siﬁcation tasks. The approach is to consider two test-

ing scenarios for each individual including two va-

lence levels (level 1: displeasure, level 3: pleasure)

and three valence levels (level 1: displeasure, level 2:

neutral, level 3: pleasure) for model validation by cal-

culating the accuracy based on LOOCV using PNN.

5 RESULTS AND DISCUSSION

Our goal was to formulate robust features for valence

detection using EEG signals, so the discernibility of

different valence levels achieved by the proposed fea-

tures was visualized using PCA. From the PCA re-

sults shown in Fig. 5, we cannot observe any dis-

tinct structures in feature space by using statistical

HEALTHINF2014-InternationalConferenceonHealthInformatics

−1

−0.5

0.5

1.5

−0.5

0.5

1.5

−0.8

−0.6

−0.4

−0.2

0.2

0.4

0.6

0.8

1st princ

ipal co

mponent

3rd principal component

2nd principal c

omponent

1st princ

ipal co

mponent

3rd principal component

−1

−0.5

0.5

1.5

−1.5

−1

−0.5

0.5

−0.4

−0.2

0.2

0.4

0.6

0.8

−1

−0.5

0.5

−0.6

−0.4

−0.2

0.2

0.4

−0.3

−0.25

−0.2

−0.15

−0.1

−0.05

0.05

0.1

0.15

Valence level 1

Valence level 2

Valence level 3

2nd principal component

1st princ

ipal co

mponent

3rd principal component

2nd principal component

1st princ

ipal co

mponent

3rd principal component

a b c

Figure 5: Visualization of feature space using PCA. a, First 3 principal components of features (std) calculated from raw

EEG signals. b, First 3 principal components of MLWF, which were the std calculated using the best performance level

after wavelet decomposition. c, First 3 principal components of CLWF, which were the std calculated using the cross-level

decomposed signals from discrete wavelet decomposition selected using GA.









1 2 3

****

*****

Frontal

sterior

tire

Average accuracy (%)







1 2

*****

Average accuracy (%)

3-level

2-level













0 1 2 3 4 5 6 7 8 9 10 11 12 13

Entire

Posterior

Frontal

Raw

D1-D5

MLWF

CLWF

Average accuracy (%)

Feature groups

EEG sets

No. of levels

Figure 6: Average results for valence detection on 50

participants using different feature groups and EEG sets.

a, Average accuracy for 3-level valence detection using

12 different feature groups (mean ± s.e.m). b, Com-

parisons of effectiveness for 3-level valence detection us-

ing EEG signals in different brain areas by proposed fea-

tures (CLWF); ****p<0.00001, *****p<0.000001 by ana-

ysis of variance (ANOVA) plus Tukey’s Honestly Signiﬁ-

cant Difference (HSD) test. c, Comparison of 2-level (L1:

98.4% and L3: 97.8% respectively) and 3-level (L1: 94.0%,

L2: 86.6%, and L3: 90.2% respectively) overall valence

classiﬁcation accuracy using proposed features (CLWF);

*****p<0.000001 by paired t-test.

measurements such as mean, kurtosis, and skewness

based on raw EEG signals, but a clearer structure

can be observed using statistical measurement std, al-

though there are also overlaps between different va-

lence levels (Fig. 5a). We conﬁrmed from the test

0 5 10 15 20 25 30 35 40 45 50

EEG electrode no.

Subject no.

Frontal electrodes 65.7%

Posterior electrodes 34.3%









1 2 3 4

****

Frontal

Posterior

Entire

GA-reduced

Average accuracy (%)

b c

Figure 7: GA based EEG electrodes reduction. a, GA-

reduced EEG electrodes distribution; no. 1-16: Fp1,

Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4,

T5, T6. b, Comparison results of GA-reduced EEG sets;

****p<0.00001 by ANOVA plus Tukey’s HSD throughout

the ﬁgure. c, Selected electrodes’ brain area summary.

results on the simplest two-level valence detection,

80% accuracy was obtained by using std, while near

or slightly better than chance accuracy (ranging from

50% to 60%) was obtained by using the other three

statistical measurements.

Based on these facts, std was selected for fur-

ther analysis in order to develop our new strategy

for feature extraction. After applying wavelet de-

composition, we tested the performance using PNN

by LOOCV and selected the optimum performance

wavelet decomposed level. Then the features were vi-

sualized using PCA (Fig. 5b), which indicates that the

discernibility was further improved compared to the

original features extracted based on raw EEG signals.

EmotionalValenceDetectionbasedonaNovelWaveletFeatureExtractionStrategyusingEEGSignals

Fig. 5c illustrates the proposed cross-level wavelet

features in this work. We obtained clearer clusters

by plotting them using the ﬁrst three principal com-

ponents, which proves the robustness of the proposed

strategy to extract features. Fig. 6a shows the aver-

age accuracy by adopting different groups of features

for three valence levels classiﬁcation. To study the re-

sults illustrated by Fig. 6a, we ﬁrstly apply analysis of

variance (ANOVA) to check if the means representing

average accuracy are unequal. It turns out that at lease

one mean is different by p<0.000001, and then we

apply Tukey’s Honestly Signiﬁcant Difference (HSD)

test to ﬁnd which means are unequal. We found that

every pair of means including proposed feature group

are signiﬁcantly different by p<0.000001. Some pairs

of means are not signiﬁcantly different by the condi-

tion of p>0.05, they are A7 and D6, Raw and D6,

Raw and A7, pairs between D1, D2, D3, D4, and D5.

Generally, the EEG signals from the frontal area

are more effective for valence detection than the ones

from the posterior area. However, just as in the re-

search results demonstrated by Murugappan et al.

(Murugappan et al., 2007), (Murugappan et al., 2010),

our classiﬁcation results also indicate that the accu-

racy can be improved by covering the entire area

with more EEG electrodes. Fig. 6b shows compar-

isons using data from frontal area, posterior area, and

entire area of EEG signals with p<0.00001 for all

pairwise comparisons by ANOVA plus Tukey’s HSD.

Our ﬁndings are also consistent with the emotion the-

ory proposed by Heller, which argues that the frontal

and parieto-temporal regions are involved in emotion

(Heller, 1993).

Our proposed cross-level wavelet feature group

(CLWF) that is searched from D1 − D5 (higher EEG

frequency bands that can produce much better av-

erage accuracy from different EEG electrodes based

on GA can largely improve the classiﬁcation perfor-

mance to an accuracy of 98% with two-level and 90%

with three-level valence detection. Fig. 6c illus-

trates the comparison of 2-level and 3-level valence

detection using proposed features with p<0.000001

by paired t-test. In contrast, the accuracy can be im-

proved very much comparing to simply combining

all the levels from D1 − D5 shown in Fig. 6a. The

results demonstrate the importance and practicability

of the cross-level strategy for extracting features in

the time-frequency domain for valence level detection

using EEG signals. Our analytical results that pro-

vide higher accuracy by using EEG signals collected

from the entire area are consistent with an important

statement in Borod’s emotion model (Borod, 1992),

in which emotions are represented in cortico-limbic

networks rather than in particular areas of the brain.

Electrodes selected by GA for each subject are il-

lustrated in Fig. 7a in which a blue dot indicates “a se-

lected electrode” and the results using the reduced set

of EEG electrodes for 3-level emotional valence de-

tection are illustrated in Fig. 7b for comparisons with

other set of electrodes. As we can see, a better accu-

racy (86.1%) can be achieved using the reduced set of

EEG electrodes compared with frontal and posterior

area EEG electrodes. Another ﬁnding shown in Fig.

7c from statistical summarization of the electrodes se-

lected from 50 subjects shows that more frontal area

electrodes (65.7%) are selected compared with pos-

terior area electrodes (34.3%), which is also consis-

tent with the viewpoint that frontal area brain activi-

ties can better demonstrate emotion (Heller, 1993).

6 CONCLUSIONS AND FUTURE

WORK

We proposed a new strategy to extract time-frequency

domain features from EEG signals in cross levels

of wavelet decomposition coefﬁcients with different

EEG electrodes for valence level detection. The pro-

posed features (CLWF) substantially increase the ac-

curacy compared to conventional features extracted

directly from EEG signals or from transformed sig-

nals in the time-frequency domain. We achieved 83%

accuracy using 6 EEG electrodes and 90% accuracy

using 16 EEG electrodes for three-level valence de-

tection, and 97% accuracy using 6 electrodes and 98%

accuracy using 16 electrodes for two-level valence de-

tection by using the proposed features.

The results show the importance of taking into

consideration information in different frequency

bands with different EEG electrodes, as it results in

higher accuracy than that achieved when only con-

sidering a single level or combined levels of wavelet

decomposed signals from different EEG electrodes.

A GA-reduced light-weight set of EEG electrodes is

also proposed with competitive accuracy (86%) for

3-level valence detection for increasing usability and

further reducing computational complexity. Future

work on interpreting the physiological meanings will

be necessary and a user-independent model will be

proposed based on EEG for valence detection.

The results achieved in this research would be in-

terests of practitioners in a number of related ﬁeld

such as health informatics and BCI. This work gives

hints of many paths for future model development.

We emphasize a work with more comprehensive

study of applying different machine learning tech-

nologies on our new proposed features.

HEALTHINF2014-InternationalConferenceonHealthInformatics

REFERENCES

Adolphs, R., Tranel, D., and Damasio, A. R. (2003). Disso-

ciable neural systems for recognizing emotions. Brain

and Cognition, 52(1):61–69.

Alves, N. T., Fukusima, S. S., and Aznar-Casanova, J. A.

(2008). Models of brain asymmetry in emotional pro-

cessing. Psychology & Neuroscience, 1(1):63–66.

Arnold, M. (1950). An Excitatory Theory of Emotion.

McGraw-Hill, New York, xxiii edition.

Arvaneh, M., Guan, C., Ang, K. K., and Quek, C. (2011).

Optimizing the channel selection and classiﬁcation

accuracy in eeg-based bci. IEEE Transactions on

Biomedical Engineering, 58(6):1865–1873.

Borod, J. C. (1992). Interhemispheric and intrahemispheric

control of emotion: a focus on unilateral brain dam-

age. J Consult Clin Psychol, 60(3):339–348.

Chu, C., Hsu, A.-l., Chou, K.-h., Bandettini, P., and Lin, C.

(2012). NeuroImage Does feature selection improve

classiﬁcation accuracy? Impact of sample size and

feature selection on classiﬁcation using anatomical

magnetic resonance images. NeuroImage, 60(1):59–

70.

Constantinou, E., Bogaerts, K., Diest, L. V., and den Bergh,

O. V. (2013). Inducing symptoms in high symptom re-

porters via emotional pictures: The interactive effects

of valence and arousal. Journal of Psychosomatic Re-

search, 74(3):191–196.

Darwin, C. (1872). The Expression of the Emotions in Man

and Animals, with an Introduction, Afterword, and

Commentaries by Paul Ekman. Oxford Univ Press,

Oxford, 3rd edition.

Heller, W. (1993). Neuropsychological mechanisms of

individual differences in emotion, personality, and

arousal. Neuropsychology, 7,:476–489.

Holland, J. H. (1975). Adaptation in Neural and Artiﬁcial

Systems. University of Michigan Press, Ann Arbor,

Michigan, 1st edition.

Hosseini, S. A. (2011). Emotion recognition method using

entropy analysis of eeg signals. International Journal

of Images, Graphics, and Signal Processing, 3(5):30–

36.

Jolliffe, I. T. (2002). Principal Component Analysis.

Springer,, 2nd edition.

Katsis, C., Katertsidis, N., and Fotiadis, D. (2011). An in-

tegrated system based on physiological signals for the

assessment of affective states in patients with anxiety

disorders. Biomedical Signal Processing and Control,

6:261–268.

Lang, P. J., Bradley, M. M., and Cuthbert, B. N. (2008). In-

ternational affective picture system (iaps): Technical

manual and affective ratings. Technical Report A-8.

Liu, J. and Xu, J. (2010). Compared study of the analyzing

methods for eeg data. In Proc. of 3rd IEEE Interna-

tional Conference on Computer Science and Informa-

tion Technology, volume 9, pages 445–448.

Morlet, J. (1984). Decomposition of hardy functions into

square integrable wavelets of constant shape. SIAM.

Math. Anal, 15(4):723–736.

Murugappan, M., Ramachandran, N., and Sazali, Y. (2010).

Classiﬁcation of human emotion from eeg using dis-

crete wavelet transform. Journal of Biomedical Sci-

ence and Engineering, 3:390–396.

Murugappan, M., Rizon, M., Nagarajan, R., Yaacob, S., Zu-

naidi, I., and Hazry, D. (2007). Eeg feature extraction

for classifying emotions using fcm and fkm. Inter-

national Journal of Computers and Communications,

1(2):299–304.

Nardi, D. (2011). Neuroscience of Personality: Brain Savvy

Insights for All Types of People. Radiance House, 1.0

edition.

Picard, R. W., Vyzas, E., and Healey, J. (2001). Toward

machine emotional intelligence: analysis of affective

physiological state. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 23(1):1175–1191.

Russell, J. (1980). A circumplex model of affect. Journal of

Personality and Social Psychology, 39(6):1161–1178.

Salovey, P. and Rothman, A. J. (2000). Emotional states

and physical health. The American psychologist,

55(1):110–121.

Schaaff, K. and Schultz, T. (2009). Towards emotion recog-

nition from electroencephalography signals. In 3rd

International Conference on Affective Computing and

Intelligent Interaction. Institute of Electrical and Elec-

tronics Engineers.

Sherwood, J. (2009). On classiﬁability of wavelet fea-

tures for eeg-based brain-computer interfaces. In Pro-

ceedings of International Joint Conference on Neural

Networks. Institute of Electrical and Electronics Engi-

neers.

Sirois, B. C. (2003.). Negative emotion and coronary heart

disease. Behavior Modiﬁcation, 27(1):83–102.

Sivakumar, A. and .Kannan, K. (2009). A novel feature

selection technique for number classiﬁcation problem

using pnn - a plausible scheme for boiler ﬂue gas anal-

ysis. Sensors and Actuators B: Chemical, 139:280–

286.

Specht, D. F. (1990). Probabilistic neural networks. Neural

Networks, 3:110–118.

Tahon, M., Degottex, G., and Devillers, L. (2012). Usual

voice quality features and glottal features for emo-

tional valence detection. In the 6th International Con-

ference of Speech Prosody. Tongji University Press.

Takahashi, K. and Tsukaguchi, A. (2003). Remarks on emo-

tion recognition from multi-modal bio-potential sig-

nals. In Proc. of IEEE International Conference on

Systems, Man, and Cybernetics. Institute of Electrical

and Electronics Engineers.

Van den Broek, E. L. and Westerink, J. H. D. M. (2009).

Considerations for emotion-aware consumer products.

Applied ergonomics, 40(6):1055–64.

Wundt, W. (1905). Grundriss der Psychologie [Fundamen-

tals of Psychology]. Engelman, Leipzig, 7th edition.

EmotionalValenceDetectionbasedonaNovelWaveletFeatureExtractionStrategyusingEEGSignals