Artiﬁcial Neural Networks as a Tool for Recognition of Movements by

Electroencephalograms

Semen A. Kurkin, Elena N. Pitsik, Vyacheslav Yu. Musatov,

Anastasia E. Runnova and Alexander E. Hramov

Research and Education Center “Artiﬁcial Intelligence Systems and Neurotechnology”, Yuri Gagarin State Technical

University of Saratov, 77 Politechnicheskaya str., Saratov, Russia

Keywords:

Artiﬁcial Neural Networks, Brain-Computer Interface, EEG Analysis, Brain Activity, Multilayer Perceptron,

Radial Basis Function, Support Vector Machine.

Abstract:

Recognition of human brain activity associated with imaginary or real movements is a complex task that

requires an accurate and conscious choice of analysis approach. Recent researches revealed the great potential

of machine learning algorithms for electroencephalography data analysis due to the ability of these methods

to establish nonlinear and nonstationary correlations, and the most attention is focused on artiﬁcial neural

networks (ANNs). Here, we introduce the ANN-based method for recognition and classiﬁcation of patterns

in electroencephalograms (EEGs) associated with imaginary and real movements of untrained volunteers.

In order to get the fastest and the most accurate classiﬁcation performance of multichannel motor imagery

EEG-patterns, we propose our approach to selection of appropriate type, topology, learning algorithm and

other parameters of neural network. We considered linear neural network, multilayer perceptron, radial basis

function network and support vector machine. We revealed that appropriate quality of recognition can be

obtained by using particular groups of electrodes according to extended international 10−10 system. Besides,

pre-processing of EEGs by low-pass ﬁlter can signiﬁcantly increase the classiﬁcation performance. Obtained

results provide better insight on neural networks potential for integration in brain-computer interfaces that are

based on EEG patterns recognition.

1 INTRODUCTION

Development of the recognition methods of hu-

man brain activity associated with imaginary or real

movements is essential for brain-computer interfaces

(BCIs), which are highly demanded in many ﬁelds

of science and technology including medicine, high-

tech and industry (Kawase et al., 2017; Spuler, 2017;

Stacey and Litt, 2008). The most striking exam-

ples of possible application of BCIs are rehabilita-

tion of patients with cognitive and motor disabilities,

mental control of exoskeletons, manipulators, robots

and other complex technical devices (Peternel et al.,

2016), improving the education quality using BCI

with biological feedback, etc.

Modern BCI systems require effective processing

tools for EEG-patterns as a part of feedback mecha-

nism. Here, we introduce the approach based on ar-

tiﬁcial neural networks. Indeed, application of neural

networks for BCIs is actively studied issue in context

of EEG-data recognition (Hamedi et al., 2014; Manor

and Geva, 2015) due to a high efﬁciency and good

recognition performance provided by these methods.

In particular, we consider classiﬁcation of real and

imaginary movements of limbs by EEGs using differ-

ent types of neural networks and various methods of

training data representation (various channels selec-

tions and using of low-pass ﬁlter). Obtained results

can be successfully used for development of BCI-

based control systems for exoskeletons or anthropo-

morphic robots for therapy of patients with various

motor disorders after trauma or stroke (Nam et al.,

2018; Peternel et al., 2016).

2 METHODS

2.1 Experimental Setup

31-channel EEG was extracted during several ses-

sions of carefully planned experiments with 12 vol-

unteered participants, both male and female. All

166

Kurkin, S., Pitsik, E., Musatov, V., Runnova, A. and Hramov, A.

Artiﬁcial Neural Networks as a Tool for Recognition of Movements by Electroencephalograms.

DOI: 10.5220/0006860201660171

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 1, pages 166-171

ISBN: 978-989-758-321-6

subjects were healthy and were not participating in

any experimental work before, i.e. subjects were not

trained for execution of imagery movements. Two

types of experiments were carried out: the ﬁrst cor-

responded to real movements, and the second — to

the imagery. Each experiment was 30 minutes long

and included two types of tasks: movement of left or

right leg.

Two EEG-datasets corresponding, respectively, to

right and left leg imagery movements were formed,

each one containing 6000 samples. For ANN training,

we used fragments of 3 or 2.5 seconds length. Each

fragment was corresponding to one type of event.

Whole dataset was divided into training and test parts

in the ratio of 50% on 50% (Haykin, 2008).

Note, we collected also EEGs of real movements

of the limbs to perform four-class recognition. Our

goal was to obtain clear and stable classiﬁcation that

is appropriate for BCIs.

It is important to note that there are certain dif-

ﬁculties associated with EEG-data analysis extracted

during experiments with untrained subjects. Recent

studies revealed various advantages of BCI-training

for analysis of activity associated with motor im-

agery: for example, evoked motor responses were

larger for BCI-trained subjects (Mokienko et al.,

2013; Dijkerman et al., 2004). Despite that, devel-

opment of new methods of EEG-patterns recognition

and classiﬁcation of untrained subjects is essential for

BCI-based neurologic rehabilitation therapy because

of inability of patients to perform training (Jackson

et al., 2001). So, current study is dedicated to devel-

opment of ANN-based method that is able to perform

accurate classiﬁcation of imaginary and real move-

ments by EEG of untrained subjects.

2.2 Machine Learning Methods

In present study, we used four types of neural network

architectures, namely linear neural network, multi-

layer perceptron (MP), radial basis function (RBF)

network and support vector machine. This subsection

contains a short review of mathematical basis of these

algorithms.

Neural networks are biologically inspired models

that process input signal in the way the real neural

network of the brain does. Classical neural network

model consists of one input and one output layers and

one or more hidden layers (see Fig. 1)

This model, in particular, corresponds to mul-

tilayer perceptron (Bishop, 1995; Carling, 1992;

Fausett, 1994; Haykin, 2008). Elements with sig-

moidal activation functions are organized in layered

topology with forward signal transmission. This type

Input

Hidden layers

Output

n+1

Figure 1: General model of artiﬁcial neural network.

of network can be interpreted as input–output model,

where the weights and biases are free parameters.

Such network can model the function of connection

between inputs and outputs of almost any degree of

complexity, and the number of layers and elements in

each layer deﬁnes the complexity of the function.

The second network model that we considered

was radial basis function network (RBFN). RBFN has

a number of advantages over multilayer perceptron.

First, RBFN models a random nonlinear function of

connection between inputs and outputs, using only

one hidden layer, which makes unnecessary the se-

lection of number of layers. Second, parameters of

linear combination in input layer can be completely

optimized, using methods of linear modelling, which

work fast and have no difﬁculties with local minima

that interfere during multilayer perceptron training.

Thus, RBFN trains much faster than multilayer per-

ceptron.

On the other side, practical use of neural networks

shows that RBFN requires more number of elements

for correct modelling of functions, which means that

RBF-based model will work slower and require more

memory than corresponding MP. RBFN cannot ex-

trapolate the conclusions beyond the area of known

data. When the data is remote from the training set,

the value of response function drops to zero quickly.

On the contrary, MP provides more speciﬁc solutions

for processing highly deviant data (Patterson, 1996;

Ripley, 1996).

Along with the multilayer perceptron and RBFN,

support vector machines (SVMs) are universal ap-

proximators used to solve classiﬁcation problems.

The idea of SVM is constructing a hyperplane that

acts as the surface of solutions that maximally sep-

arates the different classes. SVM can provide good

quality of generalization in classiﬁcation task without

a priori knowledge about subject area of particular is-

sue. This feature is unique for SVM. According to

this method, the point in space is considered as a vec-

tor of dimension p, and classiﬁer is a hyperplane of

Artiﬁcial Neural Networks as a Tool for Recognition of Movements by Electroencephalograms

167

dimension p− 1, that divides points. The data can be

classiﬁed using differenthyperplanes,but the best one

provides the best division between two classes.

Initially, SVM is a linear classiﬁer, i.e. it can solve

only linearly separable tasks. Applying nonlinear ker-

nel, one can map initial data into the space of greater

dimension, where an optimal separating hyperplane

can exist. The following functions are often used as

the kernel ones:

• Linear function: K(x

, x) = x

• Sigmoid: K(x

, x) = tanh(k(x

, x)+c), k > 0, c < 0

• Radial basis function: K(x

, x) = e

−γkx

−xk

, γ > 0

During the analysis, we determined that the best

results in accuracy and speed of recognition of EEG-

patterns associated with motor or real imagery were

achieved with next conﬁgurations:

• RBFN with 251 neurons in hidden layer with

Gaussian activation function, 31 input and 1 out-

put linear neurons;

• Multilayer perceptron with one hidden layer con-

sisted of 15 neurons with hyperbolic tangent as an

activation function, 31 input linear neurons and

one output neuron with logistic activation func-

tion;

• SVM with nonlinear kernel based on radial basis

function with value 0.01 < γ < 0.1.

All results describing below were obtained us-

ing presented conﬁgurations of neural networks. For

greater representativeness we also used linear model,

that consists only of the input and output layer and

does not have any hidden layer. Such model is effec-

tive for establishing simple linear dependencies, but

we studied it additionally in order to increase under-

standing of how neural networks work with such non-

linear and nonstationary data as EEG.

The described ANNs were implemented using the

Matlab package. The method of error backpropaga-

tion was used to train the ANNs.

2.3 Dataset Optimization

Before training ANN, we performeddataset optimiza-

tion in order to improve classiﬁcation quality. The

idea was to reduce number of EEG channels and use

different channel sets for classiﬁcation until the com-

bination of both parameters, i.e. channels number and

classiﬁcation accuracy, is optimal. Channel sets asso-

ciate with brain areas where corresponding electrodes

were placed, namely, with frontal, central, parietal,

temporal and occipital lobes. During classiﬁcation,

we used 13 different channel sets including full place-

ment consisted of 31 channels. We also calculated the

results of classiﬁcation averaged on one electrode.

Besides the channel selection, we also performed

low-pass ﬁltering with cutoffs f

= 4 Hz or f

15 Hz. Pre-ﬁltering of EEG data is necessary for re-

ducing intrinsic noise and artifacts, such as eye move-

ments and blinks. It is known that appropriate ﬁlter

provides better classiﬁcation performance due to re-

ducing signal redundancy. However, the selection of

ﬁlter type, as well as development new ones, often be-

comes the study objective (Kumar et al., 2017; Gaur

et al., 2015). Here, we shortly describe the effect of

pre-ﬁltering on neural network classiﬁcation perfor-

mance.

3 RESULTS

The session of numerical experiments was conducted.

The full dataset that contained data from whole exper-

iment was splitting into the sets of duration 2.5 sec-

onds and 3 seconds, each one contained one real or

imaginary movement event. The qualities of classiﬁ-

cation of different ANN architectures and types were

compared.

3.1 Imaginary Movements

The Fig. 2 presents averaged over all subjects values

of recognition accuracy of imaginary movements of

legs using different groups of electrodes. One can

see, that the best results of classiﬁcation correspond

to RBFN: in the case without pre-ﬁltering (Fig. 2a

and 2b) accuracy reaches 80% when using all elec-

trodes and 70% — in average. Then goes multilayer

perceptron with 70% recognition accuracy maximum

and 65% in average. The linear network shows un-

stable recognition on the level of 58%. Comparison

of Fig. 2a and 2b corresponding to different dataset

lengths shows that this value does not affect signif-

icantly on recognition accuracy. Thus, we used 3-

second fragments in the following analysis.

Then we investigated the inﬂuence of pre-ﬁltering

of initial EEGs with low-pass ﬁlter with f

= 4Hz or

= 15 Hz. Fig. 2c and Fig. 2d show that pre-ﬁltering

of input data with low-pass ﬁlter allows to signiﬁ-

cantly increase the recognition accuracy (10 − 20%

on average), and the low-pass ﬁlter with f

= 4 Hz

demonstrates the best results and allows to achieve

the classiﬁcation accuracy up to 95%. From phys-

ical point of view, the last result means that signif-

icant increase of recognition accuracy due to low-

pass ﬁlter appears on account of cleaning the use-

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

168

50%

60%

70%

80%

90%

100%

F+Fp T T+C P P+C O+P P+O+C Fp+F+T Left Right Middle All

50%

60%

70%

80%

90%

100%

F+Fp T T+C P P+C O+P P+O+C Fp+F+T

50%

60%

70%

80%

90%

100%

F+Fp T T+C P P+C O+P P+O+C Fp+F+T

50%

60%

70%

80%

90%

100%

F+Fp T T+C P P+C O+P P+O+C Fp+F+T

Linear network Multilayer perceptron Radial Basis Function Support vector machine

Left Right Middle All

Figure 2: Recognition accuracy of legs motor imagery averaged over all subjects using different groups of electrodes (X-axis)

corresponding to different EEG channel sets: (a) without pre-ﬁltering, 3 seconds fragment length; (b) without pre-ﬁltering,

2.5 seconds fragment length; (c) with pre-ﬁltering with f

= 15Hz; (d) with pre-ﬁltering with f

= 4Hz. Four types of ANNs

were used (see the caption in the ﬁgure). We used next channel sets: full placement (All, 31 electrodes), right hemisphere (9

electrodes), left hemisphere (12 electrodes), parietal, occipital and central lobes (P+O+C, 9 electrodes), frontal and temporal

(F p+ F + T, 12 electrodes), parietal and occipital (P + O, 6 electrodes), parietal and central (P +C, 6 electrodes), central

and temporal lobes (C + T, 7 electrodes), frontal (F p+ F, 8 electrodes), middle (12 electrodes), temporal (T, 4 electrodes),

parietal (P, 3 electrodes).

ful low-frequency signal from high-frequency noise,

which appears during EEG recording. Note, that

SVM shows 2− 7% lower recognition accuracy than

RBFN.

It is obvious that the full placement (31 EEG

electrode) provides the best classiﬁcation result, de-

spite the high dimensionality of dataset and redun-

dant number of channels. However, it is possi-

ble to use less electrodes without signiﬁcant loss in

classiﬁcation accuracy: one can see that electrodes

placed on frontal and temporal lobes (12 electrodes,

Fp + F + T) and several other sets (6–9 electrodes,

P+ O+C, P+C, T +C, F + F p) provide ∼ 90% ac-

curacy. Thus, the selection of channel set depends

on particular goal. If the channel set used in BCI is

more important than classiﬁcation quality, then one

can choose one of proposed sets. However, it should

be noted that using channel sets above does not affect

accuracy signiﬁcantly.

We also calculated the “quality per channel” char-

acteristic, which is deﬁned as the ratio of the classiﬁ-

cation accuracy for the given channel set to the num-

ber of channels in the set (see Fig. 3).

One can see, that frontal channel set selected

above as optimal conﬁguration shows one of the

best quality of recognition per electrode. Despite

the fact that middle lobe, that includes 7 electrodes

from frontal, central, parietal and occipital lobes

(Fpz, Fz, FCz,Cz,Cpz, Pz, Oz), shows the best result

in Fig. 3, its integral classiﬁcation performance is

worse than for frontal channel set. This can be asso-

ciated with complex nature of EEGs corresponding to

imaginary movements, i.e. imagination of leg move-

ment can ﬁnd a response in electrical activity of re-

mote brain areas that are not not localized near the

middle.

10%

12%

All Right Fp+F+T Left P+O+C Frontal

(F+Fp)

Middle

Linear network Multilayer Perceptron

Radial Basis Function Support vector machine

Figure 3: The results of calculating “quality per channel”

characteristic.

3.2 Real Movements

In the previous section, we examined the optimal ap-

proaches to the choice of input data and channel set,

Artiﬁcial Neural Networks as a Tool for Recognition of Movements by Electroencephalograms

169

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

F+Fp P+O+C Fp+F+T Left Right Middle All

Linear Network

Multilayer Perceptron

Radial basis function

Figure 4: Recognition accuracy of legs real movements averaged over all subjects using different groups of electrodes (X-

axis) corresponding to different EEG channel sets. Four types of ANNs were used (see the caption in the ﬁgure). All data is

ﬁltered with low-pass ﬁlter with f

= 4 Hz.

which provide the best quality of recognition and

classiﬁcation of imaginary movements when using

ANNs. Here, we consider the effectiveness of the

proposed approaches in the recognition of real mo-

tions by EEGs. We represent the results of recogni-

tion accuracy of real movements of legs using dif-

ferent ANNs (see Fig. 4). Analogously, numerical

experiments revealed that pre-ﬁltering with low-pass

ﬁlter with f

= 4 Hz increases signiﬁcantly classiﬁca-

tion performance over all channel sets.

In this case, one of the most important results

is high classiﬁcation performance when using differ-

ent channel sets. In particular, electrodes placed on

frontal lobe (F p + F + T) provide 87% classiﬁcation

accuracy using RBFN, which can be considered as

good result.

4 CONCLUSIONS

We developed mathematical model based on ANN for

classiﬁcation of EEG-patterns corresponding to imag-

inary or real movements, which demonstrated high ef-

ﬁciency for untrained subjects. Achieved recognition

accuracy of movements was up to 90−95% for group

of subjects. RBFN demonstrated more accurate clas-

siﬁcation performance in both cases. Pre-ﬁltering of

input data using low-pass ﬁlter signiﬁcantly increases

recognition accuracy on 10− 20% in average, and the

low-pass ﬁlter with cutoff frequency 4 Hz shows the

best results. It was revealed that using different sets

of electrodes placed on different brain areas and con-

sisted of 6-12 channels, one can achieve close to max-

imal classiﬁcation accuracy. In particular, it is conve-

nient to use electrodes on frontal and temporal lobes

(Fp + F + T) for real movements, and several sets

containing 6-9 electrodes — in case with imaginary

movements (P+ O+ C, P + C, T + C, F + F p). This

result is important from practical point of view since

it allows to use more compact systems of registration

of EEGs keeping required recognition accuracy.

ACKNOWLEDGEMENTS

This work has been supported by Russian Science

Foundation (Grant 17-72-30003).

REFERENCES

Bishop, C. M. (1995). Neural Networks for Pattern Recog-

nition. Oxford University Press.

Carling, A. (1992). Introducing Neural Networks. Wilm-

slow, UK: Sigma Press.

Dijkerman, C. H., Ietswaart, M., Johnston, M., and

MacWalter, R. S. (2004). Does motor imagery train-

ing improve hand function in chronic stroke patients?

a pilot study. Clinical rehabilitation, 18(5):538–549.

Fausett, L. (1994). Fundamentals of Neural Networks: Ar-

chitectures, Algorithms and Applications. New York:

Prentice Hall.

Gaur, P., Pachori, R. B., Wang, H., and Prasad, G. (2015).

An empirical mode decomposition based ﬁltering

method for classiﬁcation of motor-imagery eeg sig-

nals for enhancing brain-computer interface. In Neu-

ral Networks (IJCNN), 2015 International Joint Con-

ference on, pages 1–7. IEEE.

Hamedi, M., Salleh, S., Noor, A., and Mohammad-

Rezazadeh, I. (2014). Neural network-based three-

class motor imagery classiﬁcation using time-domain

features for bci applications. In Region 10 Symposium,

2014 IEEE, pages 204–207. IEEE.

Haykin, S. (2008). Neural Networks: A Comprehensive

Foundation. Pearson; 3d edition.

Jackson, P. L., Laﬂeur, M. F., Malouin, F., Richards, C.,

and Doyon, J. (2001). Potential role of mental prac-

tice using motor imagery in neurologic rehabilita-

tion. Archives of physical medicine and rehabilitation,

82(8):1133–1141.

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

170

Kawase, T., Sakurada, T., Koike, Y., and Kansaku, K.

(2017). A hybrid bmi-based exoskeleton for paresis:

Emg control for assisting arm movements. Journal of

Neural Engineering, 14(1):016015.

Kumar, S., Sharma, A., and Tsunoda, T. (2017). An im-

proved discriminative ﬁlter bank selection approach

for motor imagery eeg signal classiﬁcation using mu-

tual information. BMC Bioinformatics, 18(16):545.

Manor, R. and Geva, A. B. (2015). Convolutional neural

network for multi-category rapid serial visual presen-

tation bci. Frontiers in computational neuroscience,

9:146.

Mokienko, O., Chervyakov, A., Kulikova, S., Bobrov, P.,

Chernikova, L., Frolov, A., and Piradov, M. (2013).

Increased motor cortex excitability during motor im-

agery in brain-computer interface trained subjects.

Frontiers in computational neuroscience, 7:168.

Nam, C. S., Nijholt, A., and Lotte, F. (2018). Brain-

Computer Interfaces Handbook: Technological and

Theoretical Advances. CRC Press.

Patterson, D. (1996). Artiﬁcial Neural Networks. Singa-

pore: Prentice Hall.

Peternel, L., Noda, T., Petric, T., Ude, A., Morimoto, J.,

and Babic, J. (2016). Adaptive control of exoskeleton

robots for periodic assistive behaviours based on emg

feedback minimisation. PLoS ONE, 11(2):1–26.

Ripley, B. D. (1996). Pattern Recognition and Neural Net-

works. Cambridge University Press.

Spuler, M. (2017). A high-speed brain-computer interface

(bci) using dry eeg electrodes. PLoS ONE, 12(2):1–

12.

Stacey, W. C. and Litt, B. (2008). Technology insight:

neuroengineering and epilepsydesigning devices for

seizure control. Nature Reviews Neurology, 4(4):190.

Artiﬁcial Neural Networks as a Tool for Recognition of Movements by Electroencephalograms

171