EEG Motor Imagery Classiﬁcation of Upper Limb Movements

Maria Claudia F. de Castro, Jo

ao Pedro de O. P. Galhianne and Esther Luna Colombini

Electrical Engineering Department, Centro Universit

ario da FEI, S

ao Bernardo do Campo, Brazil

Keywords:

EEG,

Band-power Extraction, Pattern Recognition, Linear Discriminant Analysis (LDA).

Abstract:

C EEG channel data are usually used when building systems that aim at distinguishing among right and left

hand movements. Few alternatives use multichannel systems when bigger sets of motor imagery are subject

to classiﬁcation and more inputs are required. In this context, this work proposes the use of 8 EEG channels

(F,C,P, and O), disposed in a non-conventional set up, to classify up to 4 motor imagery of the upper limbs

through a Linear Discriminant Analysis classiﬁer. A spatial feature selection, prior to classiﬁcation, is applied

in order to improve the classiﬁcation accuracy. For the many channel combinations tested, results suggest

that, in addition to the motor areas, other brain areas should be considered. For the proposed system, the best

classiﬁcation accuracy was achieved when distinguishing between left arm and left hand (89.74%) and using

only the electrodes in F areas. For the right versus left hand a 71.80% rate was obtained, with electrodes either

in P and O areas or in F and P areas. To discriminate between arms and hands, independently of the body side,

the best score was 83.33%, for F and P channels, whereas for right and left limbs the best score was 66.02%,

with only P channels. The best classiﬁcation accuracy for the 4 movement problem achieved 50.00%, using

all electrodes.

1 INTRODUCTION

Brain Computer Interfaces (BCI) are communication

systems that use the electrical brain activity as input

of a system that will translate them into a control sig-

nal, for an external device, that represents the sub-

ject’s wish. Originally, this technology was developed

for people with severe motor disabilities and common

applications include spelling devices, wheelchair, and

neuromotor prostheses control. However, nowadays,

entertainment applications for healthy users are gain-

ing space, as applications for games and virtual reality

interfaces (Hoffmann et al., 2007; Veen, 2009; Millan

et al., 2010). BCI systems can be classiﬁed into two

major categories depending on the signal that they

use. Some are based on Endogenous Potentials, such

as those used in imagined movement that are volun-

tarily generated by the user, and others are based on

Exogenous Potentials that are externally induced by

an stimuli (Veen, 2009).

The brain activity occurs in many regions of the

brain, either on the cortex, basal ganglia, cerebellum,

and thalamus, changing its oscillatory frequency ac-

cording to the mental and physical states of the sub-

ject. The main bands or rhythms typically observable

are: delta (1 - 4 Hz), theta (4 - 8 Hz), alpha (8 - 12 Hz),

Beta (12 -28 Hz), and Gama above 28 Hz (Hoffmann

et al., 2007; Hema et al., 2010).

The ﬁeld of motor imagery has shown a predom-

inant interest in the right and left hand movements

using EEG from C3, Cz, and C4 channels. (Xu

and Song, 2008) achieved 90%− 92% accuracy using

Discrete Wavelets Transforms (DWT) and Autore-

gressive (AR) Model features with a Linear Discrimi-

nant Analysis (LDA) classiﬁer, whereas (Hema et al.,

2009) applied an Elman Neural Network over EEG 4

band power features, achieving an average classiﬁca-

tion of 72%. (Huang and Wu, 2010) used the average

energy of the C3 and C4 channels, a Wavelet Package

Transform, and the Quadratic Discriminant Analysis

classiﬁer achieving a maximum rate of 88.71%. (Ku-

mar and Fumitoshi, 2010) proposed the use of a Rel-

ative Spectral Power as feature, applying it over the 5

EEG frequency bands. The classiﬁcation rate using a

LDA classiﬁer was 76.43%, whereas (Dolezal et al.,

2011), using the Power Spectral Density (PSD) with

a Support Vector Machine (SVM), achieved 75% of

classiﬁcation.

With a 14-electrode set up, located around the mo-

tor area, the best results achieved by (Higashi et al.,

2009), using correlation coefﬁcients based on Rhyth-

mic Component Extraction with a LDA classiﬁer,

ranged from 74.9% to 83%. Performing a spatial

feature selection from a 64-electrode set up, (Xiao



Claudia F. Castro M., Pedro de O. P. Galhianne J. and Luna Colombini E..

EEG Motor Imagery Classiﬁcation of Upper Limb Movements.

DOI: 10.5220/0004235003140317

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 314-317

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

et al., 2009) discriminated between two types of mo-

tor imagery among a set of four (right and left hand,

foot, and tongue) using Energy Entropy of the Short-

term Fourier Transform as feature. Comparing the

performance of a Linear Discriminant classiﬁer, a

Back-Propagation Neural Network, and a SVM, the

best accuracy was achieved by the Linear Discrim-

inant classiﬁer, with average classiﬁcations between

82.4% − 88%. With a set of 29-electrodes and us-

ing Independent Component Analysis prior to DWT

to obtain the features, and Bhattacharyya distance ma-

trices and scalp plots to classify, (Morash et al., 2008)

reached an average classiﬁcation around 35% to dis-

tinguish among the same four motor imagery classes.

It can be noticed that there is a prevalence of using

only C3 and C4 channels to distinguish between two

motor imagery, mostly restricted to the right and left

hand movements, with a few alternatives using mul-

tichannel systems aiming at classifying a bigger set

of motor imagery. In this context, this work proposes

the use of 8 EEG channels (F,C,P, and O), disposed in

a non-conventional set up, to classify up to 4 motor

imagery of the upper limbs with a LDA classiﬁer. A

spatial feature selection is applied, prior to classiﬁca-

tion, in order to verify the best electrode combination

to improve classiﬁcation performance.

2 MATERIALS AND METHODS

During the experiment, an able body subject was

seated on a comfortable arm chair, with the body re-

laxed. Six series of 5 repetitions of 4 random com-

mand sequences were given: close right hand, close

left hand, ﬂex right arm, ﬂex left arm, those which

should be imagined. EEG was recorded at 1000 Hz

using a Bioampliﬁer plus PowerLab 16/30 conﬁgura-

tion from AdInstruments, according to the approved

protocol (COEP - USJT - No.088/2011).

The signal, from each of the 8-channel set up, was

acquired transversally from electrodes Fz, Cz, Pz, Oz

to electrodes F3, F4, C3, C4, P3, P4, O1, and O2

(Figure 1). A segment of 2.5 s of each execution was

selected and the Spectral Power Magnitude Averages

in different frequency bands were computed as fea-

tures, as follows: alpha (8 - 12 Hz), Beta1 (12 - 16

Hz), Beta2 (16 - 20 Hz), Beta3 (20 - 28 Hz). For

the Gamma band, the effect of different conﬁgura-

tions were investigated: (1) 28 - 32 Hz; (2) 28 - 64

Hz; (3) 28 - 100 Hz; (4) Gamma1 (28 - 32 Hz), and

Gamma2 (32 - 64 Hz); (5) Gamma1 (28 - 32 Hz),

Gamma2 (32 - 64 Hz), and Gamma3 (64 - 100 Hz). A

spatial feature selection was also performed aiming to

ﬁnd efﬁcient electrode combinations that carry useful

and discriminative information.

Figure 1: Electrode channel set up.

A classiﬁer based on Linear Discriminant Analy-

sis (LDA) was used in order to distinguish different

groups of data. Experiments were performed to dis-

criminate each of the motor imagery: right hand ver-

sus left hand, right arm versus left arm, right arm ver-

sus right hand, left arm versus left hand, hands versus

arms, and right limb versus left limb. LDA is a well

known technique based on linear combinations of the

ratio between two covariances: between-class scatter

matrix and the within-class scatter matrix. It usually

results in a reliable classiﬁcation accuracy, requiring a

few number of samples and a low computational cost

if compared to other classiﬁcation methods such as

Neural Networks.

3 RESULTS AND DISCUSSION

The classiﬁcation accuracies achieved for each fre-

quency band when varying the boundaries of the

Gamma frequency band are shown in Figure 2. In

general, the impact of this variation was low, below

5%, specially when aiming at discriminating among

the four movements, right arm and right hand, left arm

and left hand, and between right and left limbs. Nev-

ertheless, the increase of the upper band boundary (28

- 64 Hz, 28 - 100 Hz) or the use of the total frequency

band (28 - 32 Hz, 32 - 64 Hz, 64 - 100 Hz) resulted

in a diminished classiﬁcation accuracy to distinguish

between right and left hand or right arm and right

hand. The increase of the upper boundary had also

a negative effect into the classiﬁcation rate to differ-

entiate between right and left arm. However, this sit-

uation generated the best discrimination rate between

left arm and left hand or arms versus hands. Although

small, the effect of the gamma band conﬁguration is

different for each experiment. This suggests an inves-

tigation a priori and an appropriate use depending on

the ultimate goal.

The classiﬁcation rates achieved by the applica-

tion of the LDA can be shown in Figures 3 and 4.

Figure 3 shows the classiﬁcation accuracies reached

EEGMotorImageryClassificationofUpperLimbMovements

315

Figure 2: Classiﬁcation accuracies for each frequency band

varying the Gamma frequency boundaries ( 1) 28 - 32 Hz;

2) 28 - 64 Hz; 3) 28 - 100 Hz; 4) Gamma1 (28 - 32 Hz), and

Gamma2 (32 - 64 Hz); 5) Gamma1 (28 - 32 Hz), Gamma2

(32 - 64 Hz), and Gamma3 (64 - 100 Hz).

considering all channels. The best average classiﬁ-

cation was 78.85%, obtained when distinguishing be-

tween motor imagery related to hands and arms, inde-

pendently of the body side, followed by a 75.64% of

classiﬁcation for left arm versus left hand. The classi-

ﬁcation accuracy when distinguishing between right

and left hands, a common problem presented in the

literature, was 56.43%, the same value achieved to

distinguish between right and left arms, whereas the

4 movements problem reached 50.00% accuracy.

Figure 3: Classiﬁcation accuracies using a LDA classiﬁer

over the F, C, P, and O channels.

The results for the spatial feature selection,

showed in Figure 4, improved all classiﬁcation accu-

racies except for the 4 Movement problem, which is

the most complex among those presented. The classi-

ﬁcation accuracy to distinguish between left arm and

left hand movement imagination achieved 89.74%,

using only the electrodes in the F areas, whereas for

the right versus left hands a 71.80% rate was obtained

with electrodes either in P and O areas or in F and

P areas. To distinguish between arms and hands,

independently of the body side, the best score was

83.33%, considering only F and P channels. For the

discrimination between right and left limbs, the best

score was 66.02% with only P channels. It is impor-

tant to mention that these two last experiments were

performed with groups assembled with mixed data:

there was no equivalent movement imagination for

both limbs.

Figure 4: Best classiﬁcation accuracies after spatial feature

selection.

In the predominant protocol for motor imagery,

the BCI systems use only C3, Cz, and C4 signals to

distinguish right versus left hands (Hema et al., 2009;

Xu and Song, 2008; Xiao et al., 2009; Huang and Wu,

2010; Kumar and Fumitoshi, 2010; Dolezal et al.,

2011). This is justiﬁed by the fact that when the sub-

ject performs an upper limb motor imagery the con-

sidered waves in the contra-lateral electrode concen-

trates more energy, while in the same side the energy

is suppressed (Xiao et al., 2009).

Despite the fact that this work has used a non-

conventional electrode set up, the spatial feature

selection process suggests that other brain regions

should be considered. Most of the experiments with

better classiﬁcation rates have not used the electrodes

in C areas after applying the spatial feature selection.

Speciﬁcally, the best average classiﬁcation accuracy

achieved to distinguish between right and left hands

used electrodes in F and P areas or in P and O areas.

To distinguish between right arm and right hand or

left arm and left hand, the best accuracy was achieved

using only F channels. When dealing with arms ver-

sus hands classiﬁcation, electrodes in F and P areas

demonstrated the best results, and for right versus left

limbs, only P channels were used to achieve a similar

performance. The frontal cortex (F area) is related to

activities of planning movements, the parietal cortex

(P area) is an area of association for proprioceptive in-

formation also related to movement control, whereas

the visual cortex (O area) could be used in order to

visualized the execution of the movement.

Another point that must be considered is that, ac-

cording to (Veen, 2009), the comparison among BCI

systems suggests that there is a trade-off between

speed and accuracy: slower systems that consider

a long period of data demonstrate higher accuracies

than faster ones. This can be related to the number

of features samples that are available to the classiﬁer,

supplying it with more useful information about the

motor imagery. Another point refers to the period of

time that the subject has to learn how to use motor

imagery. The use of BCI in real applications, during

device control, provides continuous feedback to the

subject regarding the action that is implemented. This

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

316

process enables the subject to learn how to better use

the motor imagery control. Some researchers consid-

ering a feedback stage and this training can provide a

performance improvement (Veen, 2009).

In this work no feedback was provided to the sub-

ject and a 2.5 s data period was considered, that com-

paratively to others (Xu and Song, 2008; Kumar and

Fumitoshi, 2010; Huang and Wu, 2010; Dolezal et al.,

2011) is a short period. Nevertheless, the classiﬁca-

tion rates were consistent or even higher than those

achieved by other systems.

On the other hand, the lower accuracies achieved

for the classiﬁcation between right and left arm, right

and left limbs, and also to distinguish the four motor

imagery set up, situations that were not found in the

literature for comparison, need further investigation.

Maybe a higher period of time and other electrodes

could be considered to provide more information to

the classiﬁer. Other features and classiﬁers might be

tested in order to evaluate their performances under

this environment set up.

4 CONCLUSIONS

This work presented a motor imagery classiﬁcation

system that uses a non-conventional electrode set up

and a spatial feature selection aiming at distinguish-

ing up to four upper limb motor imagery. The results

suggest that in addition to the motor areas (C3 and

C4) other brain areas should be considered. New sets

of experiments were proposed to classify between left

arm and left hand movement imagination and to dis-

criminate between arms and hands, resulting in high

classiﬁcation accuracy. Furthermore, the classiﬁca-

tion of 4 upper limb motor imagery was evaluated

and, for that, the results have shown that further im-

provements, such as the use of more electrodes, the

increase of the data period, and the use of other fea-

tures and classiﬁers are required. Finally, in order to

generalize the results, experiments with more subjects

are necessary.

ACKNOWLEDGEMENTS

The authors would like to thank FEI, CNPq and

FAPESP for supporting.

REFERENCES

Dolezal, J., Cerny, V., and Stastny, J. (2011). Constructing

a Brain-Computer Interface. In International Confer-

ence on Applied Electronics, pages 1–4.

Hema, C. R., Paulraj, M. P., Yaacob, S., Adom, A. H., and

Nagarajan, R. (2009). EEG Motor Imagery Classiﬁ-

cation of Hand Movements for a Brain Machine In-

terface. Biomedical Soft Computing and Human Sci-

ences, 14(2):49–56.

Hema, C. R., Paulraj, M. P., Yaacob, S., Adom, A. H., and

Nagarajan, R. (2010). An Analysis of the Effect of

EEG Frequency Bands on the Classiﬁcation of Mo-

tor Imagery Signals. Biomedical Soft Computing and

Human Sciences, 16(1):121–126.

Higashi, H., Tanaka, T., and Funase, A. (2009). Clas-

siﬁcation of single trial EEG during imagined hand

movement by rhythmic component extraction. In 31st

Annual International Conference of the IEEE EMBS,

pages 2482–5.

Hoffmann, U., Vesin, J.-M., and Ebrahimi, T. (2007). Re-

cent advances in brain-computer interfaces. In IEEE

International Workshop on Multimedia Signal Pro-

cessing, pages 1–8, Chania, Greece. IEEE.

Huang, S. and Wu, X. (2010). Feature extraction and clas-

siﬁcation of EEG for imagery movement based on

mu/beta rhythms. In 3rd International Conference on

Biomedical Engineering and Informatics, volume 2,

pages 891–894. IEEE.

Kumar, M. and Fumitoshi, M. (2010). Relative Spectral

Power ( RSP ) and Temporal RSP as Features for

Movement Imagery EEG Classiﬁcation with Linear

Discriminant Analysis. In SICE Annual Conference,

pages 439–448, Taipei, Taiwan.

Millan, J. D. R., Rupp, R., Muller-Putz, G. R., Murray-

Smith, R., Giugliemma, C., Tangermann, M., Vidau-

rre, C., Cincotti, F., Kubler, A., Leeb, R., Neuper, C.,

Muller, K. R., and Mattia, D. (2010). Combining

Brain-Computer Interfaces and Assistive Technolo-

gies: State-of-the-Art and Challenges. Frontiers in

Neuroscience, 4:161–193.

Morash, V., Bai, O., Furlani, S., Lin, P., and Hallett, M.

(2008). Classifying EEG signals preceding right hand,

left hand, tongue, and right foot movements and motor

imageries. Clinical Neurophysiology, 119(11):2570–

Veen, E. R. G. V. D. (2009). Survey of state-of-the-art

eeg-based bci systems as reliable computer interface

mechanisms. In 11th Twente Student Conference on

IT, pages 1–7. University of Twente.

Xiao, D., Mu, Z., and Hu, J. (2009). A Linear Discrimi-

nation Method Used in Motor Imagery EEG Classiﬁ-

cation. In Fifth International Conference on Natural

Computation, pages 94–98. IEEE.

Xu, B.-G. and Song, A.-G. (2008). Pattern recognition of

motor imagery EEG using wavelet transform. Journal

of Biomedical Science and Engineering, 1(May):64–

67.

EEGMotorImageryClassificationofUpperLimbMovements

317