Content uploaded by Lina María Sepúlveda-Cano
Author content
All content in this area was uploaded by Lina María Sepúlveda-Cano on Oct 23, 2017
Content may be subject to copyright.
Characterization Framework for Ex-combatants Based on EEG and Behavioral
Features
Andr´
es Quintero-Zea1, Lina M. Sep´
ulveda-Cano2, M´
onica Rodr´
ıguez Calvache1, Sandra Trujillo
Orrego3, Natalia Trujillo Orrego3,4and Jos´
e D. L´
opez1
1SISTEMIC, Engineering Faculty, Universidad de Antioquia UDEA, Calle 70 No. 52 - 21, Medell´
ın, Colombia
2ARKADIUS, Engineering Faculty, Universidad de Medell´
ın UDEM, Cra. 87 No. 30 - 65, Medell´
ın, Colombia
3Mental Health Group, School of Public Health, Universidad de Antioquia UDEA, Cra. 53 No. 61-30, Medell´
ın, Colombia
4Neuroscience Group, Faculty of Medicine, Universidad de Antioquia UDEA, Medell´
ın, Colombia
Abstract— This paper presents a framework to characterize
the emotional processing of Colombian ex-combatants from ille-
gal groups. The classification process is performed using EEG-
ERP data and behavioral features from psychological tests. The
results show that ex-combatant and civilian populations can be
automatically separated using supervised techniques. With this,
we can provide a decision support system for psychologists to
improve current interventions aimed to help ex-combatants to
make a successful reintegration to civilian life.
Keywords— Emotional Processing, Emotional Recognition
Task, ERP, Ex-combatants, Supervised Learning.
I INTRODUCTION
Ex-combatants from illegal groups in Colombia mani-
fest an increased expression of aggression. A previous work
demonstrated that they present alterations on their emotional
processing [1]. However, this study did not analyze early
stages of physiological processing. Veterans mainly mani-
fest differences on the processing of unpleasant or violent
emotions [2]. On ex-combatants, we hypothesize an atypical
functioning on similar mechanisms. In order to improve the
current interventions, we need biological markers to charac-
terize atypical functioning of emotional processing.
In this line, Electroencephalography (EEG) is widely used
as an index of neurophysiological activity associated to
electrical activations in the brain, measured with electrodes
placed on the scalp [3]. Given that event-related potentials
(ERP) capture neural activity related to both sensory and cog-
nitive processes, it is the most used approach to characterize
EEG changes [4].
Several studies have reported relations between ERP com-
ponents and the behavior or personality of healthy sub-
jects. Results in [4] show an overview of the different ERP
waveforms and the major findings in various psychiatric
conditions. Relations between impulsivity and P3 ampli-
tude/latency were studied in [5] with subjects suffering of
high anxiety. While [2] showed that combat veterans with
PSTD exhibit greater ERPs.
This paper presents a framework based on ERP compo-
nents and psychological tests for automatic characterization
of ex-combatants, in order to design strategies that are useful
for their reintegration into civilian life. For feature selection
and extraction, Partial Least Squares (PLS) method is used as
a supervised projection, aiming to preserve the components
that are maximally related with the labels. Results show that
we indeed can characterize ex-combatants from civilians. The
proposed methodology can be used as a decision support sys-
tem to develop efficient intervention protocols.
II MATERI AL S
A Participants
The participants were 30 Colombian ex-combatants (two
female) from Agencia Colombiana para la Reintegraci´
on
(ACR) reintegration program, and 20 Colombian individu-
als (paired by sex, age and school level). All subjects partic-
ipated voluntarily and signed an informed consent in agree-
ment with the Helsinki declaration. Demographic informa-
tion is provided in Table 1.
Table 1: Demographic information.
Ex-combatants Civilians
n= 30 n= 20 p
Gender 2:28 2:18 0.678
Age (years) M = 37.50 M = 36.15 0.589
SD = 8.22 SD = 9.17
Educational M = 10.33 M = 11.05 0.373
level (years) SD = 3.10 SD = 2.14
B Emotional Recognition Task Procedure
For this study, we implement a modification of the Dual
Valence Association Task (DVAT) [6]. A two-alternative,
forced-choice task, in which participants are asked to classify
words or faces displayed on a computer screen according to
their valence, into one of two categories (positive or negative)
as quickly as possible [6]. Our modifications consisted in the
inclusion of Neutral as a third valence level, and the removal
of the simultaneous stimuli block described in [6].
Fig. 1 shows the pipeline of a single trial. Block trials are
presented one-by-one with strict alternation between words
and faces, and no more than two consecutive stimuli with the
same valence.
Fig. 1: Experimental design. Trials start with a fixation cross, followed by
a target stimulus: single stimulus face or word. Feedback is provided only in
error trials. Time between trial completion and onset of the subsequent trial
(ISI) varies between 700 and 1000 ms.
C ERP Recordings
EEG signals were sampled at 1000 Hz from a 64-channel
Neuroscan SynAmps2 amplifier. They were band-pass fil-
tered between 0.1 and 30 Hz. We selected 60 electrodes for
this study (HEO, VEO, CB1 and CB2 were excluded as they
do not record neural activity). Acquired EEG signals were
re-referenced off-line to average electrodes, and downsam-
pled to 500 Hz. Continuous EEG data were segmented from
200 ms prior to the stimulus to 800 ms after. All segments
contaminated with eye movement were removed from fur-
ther analysis using Independent Component Analysis (ICA)
and visual inspection. Artifact-free segments were averaged
to obtain ERPs. The preprocessing stage was performed us-
ing EEGLAB Toolbox [7]. Unless DVT presents faces and
words, only word stimuli were considered for this work.
D Psychological Tests
Both ex-combatants and civilians completed a neuropsy-
chological evaluation. It included two psychological tests: the
Social Ability Scale of Gismero (EHS) [8], and the Reactive-
Proactive Aggression Questionnaire (RPQ) [9].
The EHS measures the assertion of individual’s everyday
social interaction. It consists of 33 items, 28 redacted in neg-
ative sense (absence or lack of abilities), and 5 in positive
sense. It explores 6 dimensions of the social skills. This scale
allows obtaining a global score and punctuations for each of
the six dimensions [8].
The theoretical foundation of the RPQ suggests that the
exploration of reactive and proactive aggression has to in-
clude the motives associated with behavior, context and type
of reaction; such as physical and verbal aggression [9]. It has
23 items divided in two dimensions rated as never,sometimes
and often for frequency of occurrence [9].
III METHO DS
A Features Extraction
The main ERP components and their properties are sum-
marized in [10]. In accordance with the nature of the DVAT,
two peaks are identified as dominants: (i) N170, a a mem-
ber of the N2 family with latency between 156 and 189 ms,
which reflects expert object recognition; and (ii) P300 (or
P3), which occurs in response to an unexpected stimulus type
approximately 300 ms after the stimulus onset.
In this work, we use the following procedure to identify the
lag and amplitude of the N170 and P300 ERP components:
i) a multilevel 1-D non-decimated Haar wavelet decomposi-
tion is calculated for the ERP signal, ii) the minimum (N170)
and maximum (P300) values of the first approximation coef-
ficient are found between 150 and 350 ms, iii) the occurrence
time of both peaks is brought to the original ERP signal, and
iv) the peak amplitudes are calculated. Fig. 2 shows a ran-
dom ERP signal, its first approximation coefficient, and the
detected peaks.
-200 0 200 400 600 800
Time [ms]
-120
-100
-80
-60
-40
-20
0
20
40
60
Amplitude [µ v]
ERP
Approximation coefficient
Fig. 2: ERP component detection. The peaks N170 and P300 are detected
from the first approximation coefficient of the non-decimated Haar wavelet
decomposition of the ERP. The peaks amplitudes are then obtained from the
original ERP signal.
Finally, we complete the feature vector with the psycho-
logical scores. Specifically, for each participant we take the
global score from the EHS, the punctuations in each dimen-
sion (reactive and proactive) of the RPQ, and the total RPQ
punctuation.
B Partial Least Square Regression
The main idea of PLS regression is to find a linear model
to describe some predicted variables Xin terms of other ob-
servable variables Y. For a Lclassification problem as in [11],
with (xi,yi)∈X×{C1,...,CL},x∈Rp, with Qthe number of
features and nthe number of observations. The sample vec-
tors Xand response Ymatrices are given by:
X=
x11 . . . x1Q
.
.
.....
.
.
xn1. . . xnQ
Y=
1 0 . . . 0
.
.
..
.
.....
.
.
0 0 . . . 1
(1)
where each row of Ycontains ones in positions denoting
class labels. PLS regression searches for a set of components
(called latent vectors) that perform a simultaneous decompo-
sition of Xand Yas in Eq. (2). These components explain as
much as possible of the covariance between Tand U[12].
(X=TP>+E
Y=UQ>+F(2)
Here, matrices Eand Fare the error terms, assumed to be
i.i.d. normal, Pand Qare orthogonal loading matrices, and
Tand Uare the projections of Xand Y. The algorithm used
in this work for PLS implementation is the well–know Non-
Linear Iterative Partial Least Squares Algorithm (NIPALS).
C k-Nearest Neighbors Classifier
The k-Nearest Neighbors (kNN) algorithm starts from a
training set Xe=x1,...,xm⊂X,m<n, labeled with a class
label yj∈Y. Its objective is to classify an unknown sample
χ. According to [13], for each xi∈Xethe distance between
χand xicould be calculated as follows:
d{χ,xi}=∑
q∈Q
wqδ(χq,xiq),(3)
where δis any distance metric, i.e., Mahalanobis, Euclidean,
Minkowski, Hamming, among others; and wqis the dis-
tance weighting function. The knearest neighbors are se-
lected based on this distance metric. The class of χcould be
assigned to the majority class among the nearest neighbors;
nevertheless, there are several ways to select the χlabel, ac-
cording to the weighting factor.
D Support Vector Machine Classifier
Support vector machines (SVM) are linear classifiers de-
veloped on statistical learning theory (SLT) by Vapnik [14].
SVM are supervised learning models that aim to find an opti-
mal hyperplane to separate the points of two classes. C-SVM
is a widely used type of SVM (see [15] for implementation
details).
For linearly separable data, the maximum margin hyper-
plane is determined by the constrained optimization problem
minimize
α
1
2
m
∑
i=1
m
∑
j=1
yiyjαiαjκ(xi,xj)−
m
∑
j
αj
subject to
m
∑
i=1
αiyi=0,
0≤αi≤C,i=1,...,m,
(4)
where κ(xi,xj)represents a kernel function, and Cis the
penalty factor that controls the complexity of the SVM. The
decision function is
f(x) = sgn m
∑
i
αiyiκ(xi,x) + b!(5)
For the C-SVM model specification, the Gaussian (or ra-
dial basis) kernel is defined as
κ(x,y) = exp −kx−yk2
2σ2!(6)
The penalty factor Cand the parameter σof the kernel must
be tuned by minimizing the estimated generalization error.
IV RESULTS AND DISCUSSION
We used the classification accuracy to tune the parame-
ters of both the number of PLS components, and the number
of neighbors for the kNN classifier. The highest results were
obtained with two components and three neighbors. The C-
SVM model selection was made by grid search. The best per-
formance was reached with C=1.6 and σ=1.0.
Results show that although both classifiers achieved over
75% of accuracy (see Table 2), the SVM-based classifier
presented better performance, achieving 80.00% of accuracy
with a confidence interval of (63.87 - 90.88)%. Despite that
evidence points out the existence of alterations in the emo-
tional processing of ex-combatants, the confidence intervals
are wide due to the limited size of the sample, and because
some of the features found also reflect non-emotional cog-
nitive processes that may be common to both populations.
The sample size of our study was modest because neuropsy-
chological assessment combined with ERPs in illegal ex-
combatants is difficult and uncommon.
Table 2: Accuracy, sensitivity and specificity reached with the proposed
methodology.
Method Accuracy (%) Sensitivity (%) Specificity (%)
(Conf. Interval) (Conf. Interval) (Conf. Interval)
KNN 77.50 85.00 70.00
(61.14 - 89.03) (69.48 - 94.39) (53.29 - 83.21)
SVM 80.00 85.00 75.00
(63.87 - 90.88) (69.48 - 94.39) (58.48 - 87.14)
Regarding to the population grouping, Fig. 3 shows that
ex-combatants conform a well defined group, while civilians
tend to share some features with the other group, i.e. some
civilian samples mix with the ex-combatants group of sam-
ples. This complicates their identification as a group.
(a) (b)
Fig. 3: Classification results for (a) kNN and (b) SVM. Note that civilians
are spread, while ex-combatants are more compact. This drive to lower
specificity values.
V CONCLUSION
This paper introduced a framework to characterize ex-
combatants from civilian people, as a first step to design in-
terventions to treat their rage issues. This study showed that
supervised techniques with EEG and behavioral features may
differentiate emotional processing from ex-combatants with
high accuracy and sensitivity.
ACK NOWL EDGEM ENTS
The authors appreciate the assistance of Agencia Colom-
biana para la Reintegraci´
on. This work was partially
supported by Colciencias Grants [122266140116 and
111556933399], CODI-UDEA INV518-16, doctoral fellow-
ship call 647 (year 2014), and research project 762 (Univer-
sidad de Medell´
ın and Neurocentro de Pereira).
CON FLICT OF IN TE REST
The authors declare that they have no conflict of interest.
REFERENCES
1. Tob´
on Carlos, Iba˜
nez Agust´
ın, Velilla Lina, et al. Emotional process-
ing in Colombian ex-combatants and its relationship with empathy and
executive functions Social Neuroscience. 2015;10:153-165. PMID:
25302548.
2. Shu I-Wei, Onton Julie A., Prabhakar Nitin, O’Connell Ryan M., Sim-
mons Alan N., Matthews Scott C.. Combat veterans with PTSD af-
ter mild TBI exhibit greater ERPs from posterior-medial cortical ar-
eas while appraising facial features Journal of Affective Disorders.
2014;155:234 - 240.
3. Light Gregory A, Williams Lisa E, Minow Falk, et al. Elec-
troencephalography (EEG) and Event-Related Potentials (ERPs)
with Human Participants Current protocols in neuroscience.
2010;CHAPTER:Unit–6.2524.
4. Sur Shravani, Sinha VK, others . Event-related potential: An overview
Industrial psychiatry journal. 2009;18:70.
5. Pascalis Vilfredo De, Strippoli Enrica, Riccardi Patrizia, Vergari Fabi-
ola. Personality, event-related potential (ERP) and heart rate (HR)
in emotional word processing Personality and Individual Differences.
2004;36:873 - 891.
6. Ib´
a˜
nez Agust´
ın, Urquina Hugo, Petroni Agust´
ın, et al. Neural Process-
ing of Emotional Facial and Semantic Expressions in Euthymic Bipolar
Disorder (BD) and Its Association with Theory of Mind (ToM) PLoS
ONE. 2012;7:1-12.
7. Delorme Arnaud, Makeig Scott. EEGLAB: an open source toolbox for
analysis of single-trial EEG dynamics including independent compo-
nent analysis Journal of Neuroscience Methods. 2004;134:9 - 21.
8. Gismero E. EHS Escala de habilidades sociales Madrid: TEA Publica-
ciones de Psicolog´
ıa Aplicada. 2000.
9. Raine Adrian, Dodge Kenneth, Loeber Rolf, et al. The Reactive–
Proactive Aggression Questionnaire: Differential Correlates of Reac-
tive and Proactive Aggression in Adolescent Boys Aggressive Behavior.
2006;32:159–171.
10. Key Alexandra P Fonaryova, Dove Guy O., Maguire Mandy J.. Linking
brainwaves to the brain: An ERP primer Developmental Neuropsychol-
ogy. 2005;27:183–215.
11. Blaszczyk Pawel. Semi-supervised Feature Extraction Method Using
Partial Least Squares and Gaussian Mixture Model Lecture Notes in
Engineering and Computer Science.. 2015.
12. Abdi Herv. Partial least squares regression and projection on latent
structure regression (PLS Regression) Wiley Interdisciplinary Reviews:
Computational Statistics. 2010;2:97–106.
13. Cunningham Padraig, Delany Sarah Jane. k-Nearest neighbour classi-
fiers Multiple Classifier Systems. 2007:1–17.
14. Vapnik Vladimir N. The Nature of Statistical Learning The-
orych. Methods of Pattern Recognition, :138–155. New York, NY:
Springer–Verlag2nd ed. 1995.
15. Shi Guo-You, Liu Shuang. Model selection of RBF kernel for C-SVM
based on genetic algorithm and multithreading in 2012 International
Conference on Machine Learning and Cybernetics;1:382-386 2012.
Address of the corresponding author:
Author: Andr´
es Quintero Zea
Institute: Universidad de Antioquia
Street: Calle 70 No. 52 - 21
City: Medell´
ın
Country: Colombia
Email: andres.quintero@udea.edu.co
