Classiﬁcation of Hand Movement in EEG using ERD/ERS and

Multilayer Perceptron

Pavel Mochura and Pavel Mautner

Department of Computer Science and Enginnering, University of West Bohemia, 22 Univerzitn

ı, Pilsen, Czech Republic

Keywords:

Electroencephalography, ERD/ERS, Neural Network, EEG Signal Classiﬁcation, Feature Vectors.

Abstract:

Continuous EEG activity in the measured subjects includes different patterns depending on what activity the

subject performed. ERD and ERS are examples of such patterns related to movement, for example of a hand,

ﬁnger or foot. This article deals with the detection of motion based on the ERD/ERS patterns. By linking

ERD/ERS, feature vectors which are later classiﬁed by neural network are created. The resulting neural

network consists of one input and one output layer and two hidden layers. The ﬁrst hidden layer contains

3,000 neurons and the second one 1,500 neurons. A training set of feature vectors is used for the training of

this neural network and the back-propagation algorithm is used for the subsequent adjustment of the weights.

With this setting and training, the neural network is able to classify motion in an EEG record with an average

accuracy of 79.92%.

1 INTRODUCTION

Since the beginning of time humanity has been

plagued by many diseases and medical conditions.

One of those medical conditions is a stroke, which

can be caused, for example, by high blood pressure.

People who have survived a stroke may be perma-

nently partially paralyzed and therefore have limited

limb mobility. This paper seeks to help these patients,

speciﬁcally ones with limited hand mobility.

Electroencephalography is commonly used as

a recording technique for non-invasive BCI (Brain-

computer interface) systems. BCI provides a connec-

tion between a human and an external device or ap-

plication using neurophysiolohical signals. BCI sys-

tems can be divided into two types, invasive and non-

invasive. Invasive BCI is directly implanted in brain

tissue, while non-invasive BCI uses electrophysiolog-

ical records.(Birbaumer and Cohen, 2007)

BCI technology can increase the effectiveness of

rehabilitation and thus improve muscle control for

stroke patients and other patients with limited mobil-

ity. This can be done by detecting ERD/ERS in the

brain activity and supply the patient’s impaired mus-

cle control, e. g. send trigger ﬁg. 1.

Artiﬁcal neural network (ANN) was used for

classiﬁcation of the movement. Accuracy of the

ERD/ERS pattern classiﬁcation by this neural net-

work was aproximately 80% in our case. After a mi-

nor modiﬁcation (converting the classiﬁer to online

mode) it is possible to link the classiﬁer to virtual re-

Figure 1: BCI system that is capable of detecting ERD/ERS

from spontaneous imaginary movements. The intended

movement will be induced by electrical stimulation.

ality (VR) kit. The affected patient will just put on

a VR kit, with a book loaded, in which the patient

will be able to browse at will only by thinking of hand

movement.

For data processing and neural network classiﬁ-

cation, the Python programming language with MNE

and Keras tools was used.

The structure of this article is as following. Chap-

Mochura, P. and Mautner, P.

Classiﬁcation of Hand Movement in EEG using ERD/ERS and Multilayer Perceptron.

DOI: 10.5220/0009167007130717

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 5: HEALTHINF, pages 713-717

ISBN: 978-989-758-398-8; ISSN: 2184-4305

713

Figure 2: This chart shows ﬁltered, squared, averaged and

normalized epochs (54 epochs found). The X-axis indicates

the time that is deﬁned from -2.5 to 0.5. The Y-axis indi-

cates the power of signal.

ter 3 describes what ERD/ERS is and how it can be

calculated. This chapter also describes the creation of

feature vectors from the obtained ERD/ERS. Chapter

4 describes in detail the scenario used in the measure-

ment, the course of the measurement and the hard-

ware used for the measurement. The chapter 5 de-

scribes the structure of the artiﬁcial neural network

and it’s conﬁguration in detail. Obtained results are

also discused here. Conclusions and future work are

mentioned in the last chapter 6

2 STATE OF THE ART

The idea of BCI was originally proposed by Jaques

Vidal in (Vidal, 1973) where he proved that signals

recorded from brain activity could be used to effec-

tively represent a user’s intent.

The author of (Sepulveda, 7 05) used features

produced by Motor Imagery to control a robot arm.

Features such as the band power in speciﬁc fre-

quency bands (alpha: 8-12Hz and beta: 13-30Hz)

were mapped into right and left limb movements. In

addition, they used similar features with Motor Im-

agery, which are the ERD/ERS comparing the sig-

nal’s energy in speciﬁc frequency bands with respect

to the mentally relaxed state. It was shown in (Mo-

hamed, 2011) that the combination of ERD/ERS and

Movement-Related Cortical Potentials improves EEG

classiﬁcation as this offers an independent and com-

plimentary information.

A single trial right/left hand movement classiﬁca-

tion is reported in (Kim et al., 2003). The authors an-

alyzed both executed and imagined hand movement

EEG signals and created a feature vector consisting

of the ERD/ERS patterns of the mu and beta rhythms

and the coefﬁcients of the autoregressive model. Arti-

ﬁcial Neural Networks is applied to two kinds of test-

ing datasets and an average recognition rate of 93% is

achieved.(H. et al., 2013)

Linear Discriminant Analysis was used to clas-

sify ERD/ERS patterns associated with Motor Im-

agery. (Pfurtscheller et al., 2000) used brain oscilla-

tions (ERS) to control an electrical driven hand or-

thosis (open or close) for restoring the hand grasp

function. The subjects imagined left versus right

hand movement, left and right hand versus no spe-

ciﬁc imagination, and both feet versus right hand, and

chieved an average classiﬁcation accuracy of approx-

imately 65%, 75% and 95%, respectively.

3 EVENT-RELATED

DESYNCHRONIZATION AND

EVENT-RELATED

SYNCHRONIZATION

Certain events can block or desynchronize the ongo-

ing alpha activity (Pfurtscheller and da Silva, 1999).

These types of changes are time-locked to the event

but not phase-locked, and thus cannot be extracted by

a simple linear method, but may be detected by a fre-

quency analysis or a Fourier Transform (Pfurtscheller,

1977). This means that these events may be either

decreases or increases of power in given frequency

bands.

The ﬁrst case is called Event-related desynchro-

nization (or ERD) and the second one is called Event-

related synchronization (ERS). Of course both ERD

and ERS phenomena are not only found on EEG

recordings but also on MEG recordings (Pfurtscheller,

2001). ERD/ERS phenomena can be viewed as gener-

ated by changes in one ore more parameters that con-

trol oscillations in neuronal networks.

One of the basic features of ERD/ERS measure-

ments is that the EEG/MEG power within identiﬁed

frequency bands is displayed relative to the power of

the same EEG/MEG derivations recorded during the

stimulation or resting phases a few second before the

event occurs (Krause et al., 2008) (in our case move-

ment with left or right hand). Because event-related

changes in ongoing EEG/MEG need time to develop

and to recover, especially when alpha band rhythms

are involved, the interval between two consecutive

events should last at least 10 seconds.

3.1 Computing ERD/ERS

There are multiple ways to calculate ERD/ERS from

EEG data. I chose one of the simpler methods de-

scribed below.

To calculate ERD/ERS it is necessary to ﬁlter the

input EEG data. Because ERD is located at frequen-

cies from 8 Hz to 12 Hz and ERS at frequencies from

HEALTHINF 2020 - 13th International Conference on Health Informatics

714

Figure 3: Microcontroller board STM324F429I-DISCO

and EKG/EMG shield from company Olimex.

14 Hz to 22 Hz, the coresponding bandpass ﬁlters are

applied to input data. Afterwards, the ﬁltered data

are squared and coresponding epochs are found in this

data. The epochs are locked to synchronization marks

saved in EED. Each mark determines the beginning of

moving activity of the measured subject. The epochs

starts 2.5s before synchronization mark and the ends

0.5s after synchronization mark. It is important to

choose the correct EEG channel in which the data

for ERD/ERS calculation will be used. For right/left

hand movement the EEG channels C3/C4 was used.

In the next step the epoched data for channel are av-

eraged and ERD is evaluated by following expression

1. (Formaggio et al., 2013)

ERD

(%) =

Act

− R

100 (1)

Act represents the averaged epochs and R can be

calculated using eq. 2.

R =

k + 1

k+r

∑

i=r

Act

(2)

According to the equation we can say that R is

averaging of values in the interval [r

, r

+ k], which

is approximately two seconds before the event and

+ k is approximately 0.5 seconds after the event.

The result of this procedure for ERD can be seen

in the ﬁg. 2 (MACHIDA and TANAKA, 2018)

3.2 Creating Feature Vectors

ERD / ERS can be imagined as two vectors contain-

ing just as many items as the given epoch contains

milliseconds. The easiest way to prepare EEG data

for classiﬁcation is to create a vector with a size of the

sum of the ERD (2501 samples) and ERS (2501 sam-

ples) vector sizes. After creating the vector the ERD

vector is saved into it and the ERS vector is placed

behind ERD vector. This is a relatively simple way of

create feature vectors from the ERD and ERS, which

can then be used in most classiﬁers. This feature vec-

tor contains exactly 5002 items.

When ERD / ERS was calculated from the stim-

ulation phase, the number 1 was placed at the end of

the feature vectors. When ERD / ERS was calculated

from the resting phase, the number 0 was placed at

the end of the feature vectors.

4 MEASURING EEG

The whole EEG scenarion for measurement consists

of 4 cycles, each cycle containing a resting and stim-

ulating phase. At the start of each cycle, the sub-

ject is in the resting phase for one minute, where he

must sit completely at rest withnout any movement.

This includes reducing blinking to a minimum if his

eyes are open. After the resting phase, the subject en-

ters a stimulation phase that lasts 2 minutes, where

the subject moves wrist movement with left or right

hand. In the stimulation phase, the subject performs

the given task after a ﬁve second interval. The sub-

ject is notiﬁed of the phase change by a green LED

placed in front of the subject. When the LED is on,

the subject is in the stimulation phase and performs

the task. When the LED is off, the subject is in the

resting phase. The phases are then alternated this way

and each of them is repeated three times. This means

that each cycle lasts exactly 9 minutes.

As mentioned above, the whole measurement con-

sists of 4 cycles. The cycles differ from each other by

the task performed by the subject in the stimulation

phase, optionally combined with alternating open or

closed eyes. List of the cycles:

1. Movement with left hand with open eyes.

2. Movement with left hand with close eyes.

3. Movement with right hand with open eyes.

4. Movement with right hand with close eyes.

Classiﬁcation of Hand Movement in EEG using ERD/ERS and Multilayer Perceptron

715

Figure 4: Structure of a multilayer binary perceptron that

contains one input and output layer and two hidden layers.

4.1 Measurement Process

To this day twelve healthy people were measured

(men of age 21-26 and women of age 18-23).

Before starting the measurement, it was explained

to each subject how the whole measurement will be

done and before each cycle it was speciﬁed how the

cycle will be done. Meanwhile the nurse attached an

EEG cap on the subject’s head, ﬁtted with Ag / AgCl

electrodes according to a 10-20 system. Afterwards,

she attached 2 electrodes to the subject’s hand and one

ground electrode below the elbow, because the dis-

tance to the bone is smallest there. Lastly a reference

electrode of EEG cap was attached to the earlobe.

After ﬁnishing the preparation, the subject was

placed in a dark sound-proof chamber to avoid distur-

bance by the surroudings during measurement. Be-

fore closing the chamber all electrodes were checked

if their resistance is less than 5KΩ. This check was

followed by a quick reminder of what the subject

sould do in the stimulation phase. After that, the mea-

surement of the ﬁrst cycle began. After each cycle,

the chamber door was opened and it was explained to

the subject what to do in the next cycle.

The EEG data were recorded by the BrainAmp

DC ampliﬁer with BrainVision recorder software. For

EMG recording, synchronization pulses generation

and driving stimulation scenario the microcontroller

STM324F429I-DISCO board with EKG/EMG shield

of Olimex company were used. Fig. 3.

5 CLASSIFICATION RESULTS

To classify the created feature vectors, I chose an arti-

ﬁcial neural network programmed in Python using the

Keras module.

Training and classiﬁcation data are loaded from

ﬁles into arrays that represent the input layer.

The entire artiﬁcial neural network is made up of

the Sequential model, which is essentially a linear

stack of layers. This can be imagined as a layer list,

where each additional item in the list represents an

additional layer. After extensive testing a multilayer

perceptron was used, with one input layer, one out-

put layer and two hidden layers. The threshold used

was Keras default (0.5). This is a binary neural net-

work, which means that the output layer has only one

neuron.

A sigmoidal activation function is used in all the

hidden layers and in the output layer, deﬁned by the

Eq. 3. Binary Cross-Entropy is used as a loss func-

tion.

(u) =

1 + e

−u

(3)

Next a back-propagation training algorithm is

used, with the number of iterations set to 100.

The biggest problem was to determine the opti-

mal number of neurons in hidden layers. The artiﬁcial

neural network was started a hundred times with each

setting and the results averaged. Testing has shown

that the best possible number of neurons for the ﬁrst

hidden layer is 3 000 and 1 500 for the second hidden

layer. The accuracy of the classiﬁer with this setting

ranged between 75.00 % - 85.42 %, where the aver-

age value of all measured results is 79.92 %. This

accuracy interval is determined by random setting of

weights when training artiﬁcial neural network. The

numbers of neurons and the resulting accuracy can be

seen in tab. 1

The resulting simpliﬁed network architecture can

be seen in Fig. 4.

Table 1: Adjusted parameters of artiﬁcial neural network

and their minimum, maximum and average accuracy. If the

number of neurons in the second layer is 0, it means that the

network had only one hidden layer.

Number of neurons in

hidden layers

Min.

acc.

Max.

acc.

Average

acc.

1. layer 2. layer

250 0 64.58% 83.33% 76.23%

500 0 70.83% 83.33% 77.35%

50 25 68.75% 85.42% 76.42%

100 50 68.75% 85.42% 77.42%

200 100 68.75% 87.50% 77,77%

500 250 70.83% 83.33% 77.35%

1 000 500 68.75% 85.42% 79.23%

2 500 1 250 75.00% 85.42% 79.90%

3 000 1 500 75.00% 85.42% 79.92%

HEALTHINF 2020 - 13th International Conference on Health Informatics

716

6 CONCLUSIONS AND FUTURE

WORK

The suggested ANN has proven to be suitable for

classiﬁcation of movement in EEG data. As men-

tioned in the introduction, after a minor modiﬁcation

it is possible to link the classiﬁer to VR kit. The af-

fected patient will just put on a VR kit, with a book

loaded, in which the patient will be able to browse at

will only by thinking of hand movement.

Taking into account of the results of the other

works published in Chapter 2, we plan to do in the

future:

• Modiﬁcation of ANN architecture and use of

Deep learning to improve the results of classiﬁ-

cation.

• Complementing the scenario with the possibility

of measuring and detecting imagiantion of move-

ment.

• Connection with VR kit.

ACKNOWLEDGEMENTS

This work was supported by the University speciﬁc

research project SGS-2019-018 Processing of hetero-

geneous data and its specialized applications.

REFERENCES

Birbaumer, N. and Cohen, L. G. (2007). Brain-computer

interfaces. The Journal of Physiology, 579(3):621–

636.

Formaggio, E., Storti, S. F., Galazzo, I. B., Gandolﬁ, M.,

Geroin, C., Smania, N., Spezia, L., Waldner, A., Fi-

aschi, A., and Manganotti, P. (2013). Modulation

of event-related desynchronization in robot-assisted

hand performance. Journal of NeuroEngineering and

Rehabilitation, 10(1).

H., M., Samaha, A., and AlKamha, K. (2013). Automated

classiﬁcation of l/r hand movement eeg signals us-

ing advanced feature extraction and machine learning.

International Journal of Advanced Computer Science

and Applications, 4(6).

Kim, J., Hwang, D., Cho, S., and Han, S. (2003). Sin-

gle trial discrimination between right and left hand

movement with eeg signal. Proceedings of the 25th

Annual International Conference of the IEEE Engi-

neering in Medicine and Biology Society (IEEE Cat.

No.03CH37439), pages 3321–3324.

Krause, C. M., Boman, P.-A., Sillanm

aki, L., Varho, T., and

Holopainen, I. E. (2008). Brain oscillatory eeg event-

related desynchronization (erd) and -sychronization

(ers) responses during an auditory memory task are

altered in children with epilepsy. Seizure, 17(1):1–10.

MACHIDA, R. and TANAKA, H. (2018). Visualiza-

tion of erd/ers on leg motor imagery. International

Symposium on Affective Science and Engineering,

ISASE2018(0):1–6.

Mohamed, A.-K. (2011). Towards improved eeg interpre-

tation in a sensorimotor bci for the control of a pros-

thetic or orthotic hand. (p. 144).

Pfurtscheller, G. (1977). Graphical display and statistical

evaluation of event-related desynchronization (erd).

Electroencephalography and Clinical Neurophysiol-

ogy, 43(5):757–760.

Pfurtscheller, G. (2001). Functional brain imaging based on

erd/ers. Vision Research, 41(10-11):1257–1260.

Pfurtscheller, G. and da Silva, F. L. (1999). Event-

related eeg/meg synchronization and desynchroniza-

tion. Clinical Neurophysiology, 110(11):1842–1857.

Pfurtscheller, G., Guger, C., M

uller, G., Krausz, G., and Ne-

uper, C. (2000). Brain oscillations control hand ortho-

sis in a tetraplegic. Neuroscience Letters, 292(3):211–

214.

Sepulveda, F. (2011-07-05). Brain-actuated control of robot

navigation. Advances in Robot Navigation.

Vidal, J. J. (1973). Toward direct brain-computer commu-

nication. Annual Review of Biophysics and Bioengi-

neering, 2(1):157–180.

Classiﬁcation of Hand Movement in EEG using ERD/ERS and Multilayer Perceptron

717