DEVELOPMENT OF A LOW-COST SVM-BASED

SPONTANEOUS BRAIN-COMPUTER INTERFACE

Fernando Flórez, José M. Azorín, Eduardo Iáñez, Andrés Úbeda and Eduardo Fernández

Biomedical Neuroengineering Group (NBIO), Miguel Hernández University of Elche, Elche, Spain

Keywords: Brain-computer interface, Electroencephalography, Support vector machines, Emotiv epoc.

Abstract: This paper describes a spontaneous non-invasive Brain-Computer Interface (BCI) using an inexpensive

EEG device. The aim of this work is to determine the feasibility of using the Emotiv Epoc device in a BCI.

BCIs provide a method for interaction with a computer for people with severe communication disabilities.

The EEG signals of five healthy users have been registered and preprocessed. The Fast Fourier Transform

(FFT) has been used to extract the relevant characteristics of the EEG signals. Frequency spectrum between

8 Hz - 30 Hz has been calculated. An offline analysis to the recorded data has been performed using a

Support Vector Machine (SVM) as a classification algorithm in order to differentiate three and four mental

tasks. Results of up to 71% classification accuracy for three tasks and 64% classification accuracy for four

tasks were obtained, showing that the Emotiv Epoc is suitable to be used in a Brain-Computer Interface.

1 INTRODUCTION

A Brain-Computer Interface (BCI) provides a

communication and control system that does not

depend on the brain’s normal neuromuscular output

channels (Wolpaw, et al., 2002). A BCI is based on

the use of the mental activity of a person to generate

control commands in a device (Dornhege, et al.,

2007). For this purpose, the brain activity of the user

has to be registered and processed appropriately in

order to distinguish between different cognitive

processes or “mental tasks”.

BCIs are an alternative to classical methods of

human-machine communication, like a keyboard or

a mouse. For that reason, this kind of interfaces are

very useful for people with severe communication

disabilities. BCIs have been used in different

applications, from the control of a wheelchair

(Galán, et al., 2008) to the control of a smart home

(Guger, et al., 2008).

Brain activity can be registered in several ways,

using invasive and non-invasive techniques. Using

invasive techniques, the activity of a single neuron

or a small group of neurons can be registered with

microelectrodes arrays implanted directly in the

brain. These techniques have been used to determine

the intention of movements in animals (Carmena, et

al., 2003) or to control a cursor on a screen (Serruya,

et al., 2002). Non-invasive techniques use electrodes

on the scalp to measure the EEG signals. For

humans, non-invasive techniques based on EEG

signals are more appropriate due to ethical aspects

and medical risks.

Non-invasive BCIs can be classified as evoked

or spontaneous. In an evoked BCI, the registered

signals evidence the automatic response of the brain

to certain external stimuli (Bayliss, 2003; Sirvent, et

al., 2010). Nevertheless, the need for external stimuli

limits the number of applications. In contrast, in a

spontaneous BCI, the user performs the mental tasks

of his/her own free will (Iáñez, et al., 2010).

Once the EEG signals have been registered, they

have to be processed and filtered, and a

classification has to be done in order to differentiate

between the cognitive processes.

Table 1: Inexpensive EEG devices.

Price Range Device Channels

0€ - 200€

Neurosky MindWave

Modular EEG

OCZ Technology NIA

2-6

200€ - 500€ Emotiv Epoc 14

500€ - 1.000€

Neurobit Optima

Neurobics Pendant EEG

2-4

415

Elgendi M., Vialatte F., Constable M. and Dauwels J..

IMMERSIVE NEUROFEEDBACK - A New Paradigm.

DOI: 10.5220/0003725704650469

In Proceedings of the International Conference on Neural Computation Theory and Applications (Special Session on Challenges in Neuroengineering-

2011), pages 465-469

ISBN: 978-989-8425-84-3

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Sensor placement of the 14 data channels in the Emotiv Epoc, according to the 10/20 International System (left).

Emotiv Epoc (centre). User with the Emotiv Epoc (right).

Different classification algorithms has been used

in BCI experiments (Bashashati, Fatourechi, Ward

and Birch, 2007). Among several of them, like

Linear Discriminant Analysis (LDA) and Neural

Networks (NN), Support Vector Machines (SVM)

provide a powerful method for data classification

(Garrett, Peterson, Anderson and Thaut, 2003). In

this project, a SVM has been used as a classification

algorithm.

Nowadays, several low-cost EEG devices are

available for the consumer. They have a very

affordable price by comparison with the professional

EEG systems, whose prices vary between 10.000€

and 150.000€. Table 1 presents a list of inexpensive

EEG devices. The Emotiv Epoc has been chosen in

this experiment.

In recent years, there has been an increasing

amount of literature on inexpensive BCI. The

Emotiv Epoc device have been used from the control

of a mobile phone (Campbell, et al., 2010) to the

control of a car without using a steering wheel

(AutoNOMOS project Freie Universität Berlin,

2011).

This paper describes a spontaneous non-invasive

EEG-based Brain-Computer Interface using the

Emotiv Epoc device. The BCI developed uses the

Fast Fourier Transform (FFT) to extract the

relevants characteristics of the EEG signals. The

results of five voluntary users will be obtained using

a Support Vector Machine (SVM) as a classification

algorithm to differentiate between mental tasks. The

aim of this experiment is to check the feasibility of

using an inexpensive EEG device in a Brain-

Computer Interface.

The rest of this paper is organized as follows.

Section 2 describes the Brain-Computer Interface

developed and the classifier used is introduced. In

section 3, the experimental results are shown.

Finally, Section 4 contains the conclusions.

2 BRAIN-COMPUTER

INTERFACE

In this section, the Emotiv Epoc device is

introduced. The procedure followed to register the

EEG signals and to extract the relevant features of

the signals is explained. The classifier used in this

experiment is presented.

2.1 EEG Hardware

The EEG signals have been registered using the

Emotiv Epoc headset (Figure 1, centre), released by

the Emotiv Company (Emotiv Systems, 2011).

Emotiv Epoc is a wireless device composed of

14 channels and 2 reference electrodes, located

according to the 10/20 International System

(American Electroencephalographic Society, 1991)

in the positions AF3, F7, F3, FC5, T7, P7, O1, O2,

P8, T8, FC6, F4, F8 and AF4. CMS/DRL reference

electrodes are located in the positions P3/P4 (Figure

1, left). Each electrode has to be moistened with a

saline solution before being used. Once the device is

placed on the scalp (Figure 1, right), the signal

quality of the electrodes has to be checked. It has

been seen that the signal quality decreases when the

electrodes get dry.

The Emotiv Epoc device does not provide the

impedance level of each channel. Instead, the

contact quality of each sensor is represented by a

colour code: Black – No signal; Red – Very poor

signal; Orange – Poor signal; Yellow – Fair signal;

Green – Good signal. This connection quality has to

be checked from within the Emotiv Control Panel.

The headset transmits the EEG signals wirelessly

to a Windows-based computer in the frequency of

2.4 GHz. The signals are filtered on the device

NCTA 2011 - International Conference on Neural Computation Theory and Applications

416

Table 2: Mental tasks.

Number Name Description

1 Rest Countdown from 20 to 0

2 Arm

Imagination of a repetitive

low circular movement

of the right arm

3 Song

Mentally singing the

“Happy Birthday” song

4 Math

Mentally performing

the Fibonacci series

5 Object Mentally rotating an object

before amplifying the data. A high-pass filter at 0.16

Hz cut-off frequency and a low-pass filter at 83 Hz

are applied. The internal sampling rate is 2.048 Hz.

Then, the data are filtered with a 5

order Sinc filter

to notch out the 50 Hz and 60 Hz frequencies, and

downsampled to 128 Hz.

2.2 Mental Tasks

Five cognitive processes or mental tasks have been

considered in the experiment. These mental tasks

have been taken into consideration due to the

placement of the electrodes in the Emotiv Epoc.

Table 2 shows the mental tasks considered.

2.3 Acquisition

A Matlab interface has been developed to register

the EEG signals. The interface gives us options to

connect and disconnect with the device, select the

configuration of the test, and start/stop the test. The

process used to register the data is as follows.

Each test is comprised of 25 trials, lasts 250

seconds and is repeated 4 times. There is a short

pause between tests. Once the test is started, the

mental tasks that the user has to perform are

displayed. Each of the five tasks is showed randomly

5 times, and lasts 10 seconds (Figure 2). In the first 2

seconds, a cross appears to indicate the user that a

new task is started; in the following 2 seconds, the

image of the mental task to perform is displayed; in

the last 6 seconds, the user performs the mental task.

Once the test is finished, 120 seconds of each task

have been recorded. A similar timing paradigm as

the one described in (Guger, et al., 2001) is used to

Using the Emotiv Control Panel, it has been

checked that the signal quality of each electrode was

always good (green colour).

EEG signals were acquired at a sample

frequency of 128 Hz. The EEG signals registered are

processed in sequences of 1 second of length,

including an overlap of 1/8 a second with the

previous sequence. Once the data are recorded, they

are preprocessed and a feature extraction algorithm

is applied.

2.4 Feature Extraction

Before extracting the main characteristics of the

registered data, a preprocessing has been applied to

the signals. The DC offset in all channels is

removed. Afterwards, the baseline of each electrode

is removed, eliminating the mean value of the signal

registered by each electrode. In this preprocessing,

all 14 electrodes have been used.

Following this, a feature extraction algorithm is

applied to the data in order to extract the main

characteristics of the EEG signals, to facilitate the

posterior classification. The algorithm used is based

on the frequency domain. The Fast Fourier

Transform (FFT) has been used to extract the

relevant characteristics of the EEG signals. The

frequency spectrum between 8 Hz and 30 Hz, with

2 Hz resolution (12 features), has been calculated in

order to analyze the rhythmic activity variations. All

the mental tasks are expected to be found in the EEG

signals in the alpha (8 Hz – 12 Hz) and beta waves

(12 Hz – 30 Hz). Each feature vector consists of 168

elements (12 features x 14 electrodes), obtaining 960

feature vectors by user.

Finally, a Surface Laplacian is applied in order to

improve the signal/noise ratio in each electrode

(Babiloni, et al., 2001).

2.5 Classifier

A Support Vector Machine (SVM) has been used in

this experiment as a classification algorithm. SVMs

are a very useful technique for data classification

(Hsu, Chang and Lin, 2003). It uses a hyperplane or

set of hyperplanes in a high or infinite dimensional

space to distinguish between object of different

classes.

A classification task usually consists in separating

data into trainings and testing sets. Each instance

Figure 2: Time distribution of each task.

DEVELOPMENT OF A LOW-COST SVM-BASED SPONTANEOUS BRAIN-COMPUTER INTERFACE

417

Figure 3: Image of a user performing the test.

in the training set contains a class label and several

features. The aim of SVM is to create a model,

based on the training data, which predicts the class

labels of the test data given its features.

The accuracy of the SVM depends on the

selection of the kernel type and the values of its

parameters. In a Brain-Computer Interface project,

the kernel generally used is the Gaussian or Radial

Basis Function (RBF) kernel (Lotte, et al., 2007).

There are two parameters for an RBF kernel, the

regularization parameter C and the parameter γ,

which determines the RBF width. It is not known

which C and γ are best for a specific problem. For

that reason, some kind of parameter search has to be

done. The aim of this search is to identify C and γ

values so that the classifier can accurately predict

the testing data. The best combination of C and γ is

often selected by a grid-search using cross-

validation with exponentially growing sequences of

C and γ. The one with the best cross-validation

accuracy will be selected.

3 EXPERIMENTAL RESULTS

In this section, the experimental results obtained are

showed. The offline analysis performed to the

registered data of five users is explained, using SVM

as a classification algorithm.

3.1 Participants

For the experiment, the participation of five healthy

right-handed male users with ages between 26 and

40 years old has been required. After informing the

users of the requirements and tests involved, the

volunteers agreed and gave their consent to take part

in the tests. All users had normal vision and hearing,

and no history of neurological or psychiatric

disorders. All tests were done in an isolated room

with the user sitting in front of a PC screen at a

distance of 1 meter (Figure 3).

3.2 SVM Classification

An offline analysis has been done to the registered

data. For this purpose, a Matlab application has been

developed.

The results of five users have been calculated.

Users 4 and 5 had not been previously involved in

any BCI experiment. For every user, all the possible

combinations of three and four tasks have been

calculated. For each combination of tasks, data may

be randomly extracted and separated into training

data and test data. 75% of the data has been used as

a training data, whereas the rest 25% is used as a test

data.

As a classification algorithm, the Matlab

interface of LIBSVM 2.9 library has been used

(Chang and Lin, 2001). LIBSVM, developed by the

National Taiwan University, is a free integrated

software for Support Vector classification,

regression, and distribution estimation. Features of

LIBSVM include, among many other, multi-class

classification, different SVM formulations and

various type of kernels.

In this BCI experiment, a C-SVM with a Radial

Basis Function (RBF) kernel has been used. To

identify C and γ parameters values so that the

classifier can more accurately predict the testing

data, a grid-search using cross-validation with

exponentially growing sequences of C (between 2

-3

and 2

) and γ (between 2

-13

and 2

) has been

perform. For the classification of three tasks, values

of C=512 and γ=0.0020 present the best results.

Classifying four tasks, values of C=1.024 and

γ=0.0020 have been selected.

Input values to the SVM were not normalized,

because normalization did not improve the

classification results.

3.3 Results

For each user and combination of three and four

tasks, the mean value of the accuracy achieved after

10 iterations of the application has been calculated,

in order to determine more accurately the success

percentage in the classification. Average percentage

of success for every set of tasks and user has been

calculated. Tables 3 and 4 show the results obtained

from the classification of three and four different

tasks.

NCTA 2011 - International Conference on Neural Computation Theory and Applications

418

Table 3: Percentage of correct classification using 3 mental tasks.

Tasks User 1 User 2 User 3 User 4 User 5 Average

Rest – Arm – Song 66.80 67.57 73.39 67.96 63.12 67.77

Rest – Arm – Math 74.08 72.80 72.78 67.55 67.18 70.88

Rest – Arm – Object 67.75 68.90 76.35 66.34 66.91 69.25

Rest – Song – Math 66.40 70.07 72.80 67.79 62.44 67.90

Rest – Song – Object 65.30 67.73 78.46 66.55 64.30 68.47

Rest – Math – Object 73.88 71.05 78.30 66.69 68.58 71.70

Arm – Song – Math 73.50 71.75 72.58 67.59 66.10 70.31

Arm –Song – Object 69.95 67.91 75.96 67.91 66.37 69.62

Arm – Math – Object 72.13 74.32 75.39 67.10 68.53 71.49

Song – Math – Object 71.83 71.21 76.35 68.05 66.86 70.86

Average 70.16 70.33 75.24 67.35 66.04

Table 4: Percentage of correct classification using 4 mental tasks.

Tasks User 1 User 2 User 3 User 4 User 5 Average

Rest – Arm – Song – Math 63.78 64.84 66.70 61.28 58.75 63.07

Rest – Arm – Song – Object 60.06 61.47 70.12 60.90 57.78 62.07

Rest – Arm – Math – Object 66.15 64.93 69.84 61.30 61.04 64.65

Rest – Song – Math – Object 62.38 63.45 70.57 61.38 58.60 62.28

Arm – Song – Math – Object 65.70 65.34 69.13 61.08 60.44 64.34

Average 63.62 64.01 69.27 61.19 59.32

On average, involving both three and four tasks,

user number 3 has achieved the best results. It is

noted that user number 3 was the one with more

experience in BCI experiments. In contrast, users 4

and 5, with no previous experience in BCI

experiments, have obtained a low success

percentage. This may probe that the more experience

you have in BCI experiments, the better results you

will obtain. Figure 4 shows the average accuracy of

each user.

As regards sets of three tasks, the combination

between the task “Rest – Math – Object” presents

the greatest accuracy, with a 71.70%. As for sets of

four tasks, the best results are classifying the tasks

“Rest – Arm – Math – Object”, with a 64.65% of

success. These combination of three and four tasks

can be used as cognitive processes to classify in

future works.

Furthermore, with the aim of obtaining more

information in the classification, tables 5 and 6 show

the average success percentage obtained by 5 users

on each task, for each combination of three and four

tasks. Results show that the task “Rest” has the

greatest success percentage (77.16% using 3 tasks

and 72.28% using 4 tasks) and the task “Math” has

the second best success percentage (69.72% using 3

tasks and 62.77% using 4 tasks). This is reflected in

the results showed in tables 3 and 4, where the

combination of tasks that achieves the best results

include these 2 tasks.

All results obtained indicate that an SVM

classification together with the Emotiv Epoc as a

Interface.

Figure 4: Average accuracy of each user.

4 CONCLUSIONS AND FUTURE

WORK

A spontaneous non-invasive Brain-Computer

Interface has been proposed. Using the Emotiv Epoc

device to register the EEG signals, this BCI allows

performing the classification of mental tasks. The

feature extraction process using Fast Fourier

Transform and the subsequent classification using a

Support Vector Machine has been explained.

Results obtained in the classification have been

showed. Success percentage for three and four tasks

DEVELOPMENT OF A LOW-COST SVM-BASED SPONTANEOUS BRAIN-COMPUTER INTERFACE

419

Table 5: Average success percentage by 5 users on each task using 3 mental tasks.

Tasks Rest Arm Song Math Object

Rest – Arm – Song 75.02 64.48 60.83 - -

Rest – Arm – Math 76.76 64.70 - 67.86 -

Rest – Arm – Object 80.44 62.28 - - 63.67

Rest – Song – Math 76.19 - 58.60 66.54 -

Rest – Song – Object 77.09 - 60.07 - 66.44

Rest – Math – Object 77.48 - - 68.53 67.35

Arm – Song – Math - 67.66 73.71 68.93 -

Arm –Song – Object - 65.69 77.35 - 65.09

Arm – Math – Object - 67.04 - 76.39 68.49

Song – Math – Object - - 73.86 70.05 66.86

Average 77.16 65.31 67.40 69.72 66.32

Table 6: Average success percentage by 5 users on each task using 4 mental tasks.

Tasks Rest Arm Song Math Object

Rest – Arm – Song – Math 71.83 60.10 53.46 62.02 -

Rest – Arm – Song – Object 73.03 58.24 54.28 - 57.70

Rest – Arm – Math – Object 73.24 58.27 - 63.01 58.31

Rest – Song – Math – Object 71.01 - 53.81 62.13 61.70

Arm – Song – Math – Object - 59.54 69.10 63.92 59.31

Average 72.28 59.04 57.66 62.77 59.26

indicates that the Emotiv Epoc is suitable to be used

in a Brain-Computer Interface. As a future work, the

implementation of an online application has been

proposed. Also, is expected to perform different tests

using volunteers with disabilities. In order to

compare results and verify if there is any

performance loss due to the Emotiv Epoc device, it

is expected to test the same experiments with a high-

quality research-oriented EEG system (gUSBamp

g.tec). Finally, the use of a new set of tasks (for

example, new motor tasks, tongue movement,

mental calculation or word formation) are suggested,

as well as the use of new classification algorithms.

ACKNOWLEDGEMENTS

This work has been supported by the Ministerio de

Ciencia e Innovación of the Spanish Government

through project DPI2008-06875-C03-03 and by the

Consellería d’Educació de la Generalitat Valenciana

through grant BEST/2010/047, grant

ACOMP/2011/066 and grant FPA/2011/058.

REFERENCES

American Electroencephalographic Society. (1991).

American Electroencephalographic Society Guidelines

for Standard Electrode Position Nomenclature.

Journal of Clinical Neurophysiology 8(2)

pp. 200-202.

AutoNOMOS project Freie Universität Berlin. (2011).

Brain Driver. [online] Available at: <http://

autonomos.inf.fu-berlin.de/news/brain-driver>

[Accessed 10 May 2011]

Babiloni, F., Cincotti, F., Bianchi, L., Pirri, G., Millán, J.

R., Mourio, J., Salinari, S. and Marciani, M.G. (2001).

Recognition of imagined hand movements with low

resolution surface laplacian and linear classifiers.

Medical Engineering Samp; Physics 23, pp. 323-328.

Bashashati, A., Fatourechi, M., Ward, R. K. and Birch, G.

E. (2007). A survey of signal processing algorithms in

brain–computer interfaces based on electrical brain

signals. Journal of Neural Engineering, 4(2), R32-57.

Bayliss, J. D. (2003). Use of the evoked potential P3

component for control in a virtual environment. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering 11, pp. 113-116.

Campbell, A. T., Choudhury, T., Hu, S., Lu, H., Mukerjee,

M. K., Rabbi, M. and Raizada, R. D. S. (2010).

NeuroPhone: Brain-Mobile Phone Interface using a

Wireless EEG Headset. [online] Available at

<http://sensorlab.cs.dartmouth.edu/publications.html>

[Accessed 15 May 2011]

Carmena, J. M., Lebedev, M. A., Crist, R. E., O’Doherty,

J. E., Santucci, D. M., Dimitrov, D. F., Patil, P. G,

Henriquez, C. S. and Nicolelis, M. A. L. (2003).

Learning to control a brain-machine interface for

reaching and grasping by primates. PloS Biology 1, pp.

193-208.

Chang, C.-C. and Lin, C.-J. (2001). LIBSVM: A Library

for Support Vector Machines. [online] (Updated April

NCTA 2011 - International Conference on Neural Computation Theory and Applications

420

2011) Available at <http://www.csie.ntu.edu.tw/

~cjlin/libsvm/index.html> [Accessed 12 May 2011]

Dornhege, G., Millán, J. R., Hinterberger, T., McFarland,

D. J. and Müller, K.-R. (2007). Towards Brain-

Computer Interfacing, MIT Press, Cambridge,

Massachussets.

Emotiv Systems. (2011). Brain Computer Interface

Technology. [online] Available at: <

http://www.emotiv.com/> [Accessed 12 May 2011]

Galán, F., Nuttin, M., Lew, E., Ferrez, P. W., Vanacker,

G., Philips J. and Millán, J. R. (2008). A brain-

actuated wheelchair: Asynchronous and non-invasive

Brain-computer interfaces for continuous control of

robots. Clinical Neurophysiology 119, pp. 2159-2169.

Garrett, D., Peterson, D. A., Anderson, C. W. and Thaut,

M. H. (2003). Comparison of Linear and Nonlinear

Methods for EEG Signal Classification. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, vol 11 (2), pp. 141-144.

Guger, C., Schlögl, A., Neuper, C., Walterspacher, D.,

Strein, T. and Pfurtscheller, G. (2001). Rapid

Prototyping of an EEG-Based Brain-Computer

Interface (BCI). IEEE Transactions on Rehabilitation

Engineering, vol. 9(1), pp. 49-58.

Guger, C., Holzner, C., Grönegress, C., Edlinger, G. and

Slater, M. (2008). Control of a smart home with a

Brain-Computer Interface. 4

International Brain-

Computer Interface Workshop and Training Course

2008, Graz, Austria.

Hsu, C.-W., Chang, C.-C. and Lin, C.-J. (2003). A

Practical Guide to Support Vector Classification.

[online] (Updated April 2010) Available at

<http://www.csie.ntu.edu.tw/~cjlin/libsvm/>

[Accessed 12 May 2011]

Iáñez, E., Azorín, J. M., Úbeda, A., Ferrández, J. M. and

Fernández, E. (2010). Mental Tasks-Based Brain-

Robot Interface. Robotics and Autonomous Systems,

vol. 58(12), pp. 1238-1245

Lotte, F., Congedo, M., Lécuyer, A., Lamarche, F. and

Arnaldi, B. (2007). A Review of Classification

Algorithms for EEG-Based Brain-Computer

Interfaces. Journal of Neural Engineering 4, R1-R13.

Serruya, M. D., Harsopoulos, N. G., Paninski, L., Fellows,

M. R. and Donoghue, K. (2002). Instant neural control

of a movement signal. Nature 416, pp. 141-142.

Sirvent, J. L., Azorín, J. M., Iáñez, E., Úbeda, A. and

Fernández, E. (2010). P300-based BCI for Internet

browsing. 8

International Conference on Practical

Applications of Agents and Multiagent Systems, IEEE

Spanish Chapter. Trends in PAAMS, AISC 71, pp.

615-622, ISS+: 1867;566

Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,

Pfurtscheller, G. and Vaughan, T. M. (2002). Brain-

Computer Interfaces for communication and control.

Clinical Neurophysiology 113, pp. 767-791.

DEVELOPMENT OF A LOW-COST SVM-BASED SPONTANEOUS BRAIN-COMPUTER INTERFACE

421