A Novel EEG Classification for Baseline Motor Cortex using

Adaboost

Said Abenna

, Mohammed Nahid

and Abderrahim Bajit

Faculty of Sciences and Technology, Hassan II University, Casablanca, Morocco

National School of Applied Sciences, Ibn Tofail University, Kenitra, Morocco

Keywords: Brain-computer interface (BCI), Electroencephalogram (EEG), Data Analysis, Classification, Optimization

Abstract: This paper contains a new method for EEG classification and specifically for baseline hand motor cortex

recognition. This work is based on the determination of links between all electrodes when moving hands using

covariance and correlation matrices, also the importance degree for every electrode from the best to the bad

in the classification stage, and thus the application of differential evolution (DE) optimizer to increase the

prediction accuracy of AdaBoost algorithm to the best state. The results obtained show that the prediction

accuracy value for hand EEG motor cortex classification takes the value of 100% when using more than six

electrodes without using feature extraction algorithms, knowing that the related work has a maximum average

accuracy value of 99.1% using the PNN algorithm. Therefore, this work has a very important role in increasing

the EEG signals prediction quality, either on the side of improving the classification algorithms or minimizing

the number of acquisition channels needed.

1 INTRODUCTION

The Brain-Computer Interface (BCI) transmits brain

activity acquired from the human scalp to a computer

for controlling external devices and assisting the

handicapped in regaining organ abilities (Abenna et

al., 2021a). It's almost research for the use of

electroencephalogram (EEG) in the controls

intelligent in robotic arms and other external devices.

Compared to other signal types, EEG signals have

several different direct communications between a

human brain and a computer (Jin et al., 2015; Li et al.,

2016; Tang et al., 2020; Zhang et al., 2018). The

collected brain signals vary depending on the

structure of the human brain and the subject's mental

state, and these brain activities of each subject are

unique. EEG signals are not woody and non-stable,

which means that the EEG signal properties change

over time (Khosla et al., 2020; Tang et al., 2020; Yin

and Zhang, 2017). Furthermore, the recorded EEG

signals are frequently intermingled with noise,

making analysis difficult. As a result, efficient steps

to enhance the signal-to-noise ratio (SNR) of EEG

data should be taken (Michelmann et al., 2018; Tang

et al., 2020; Whitmore and Lin, 2016). EEG waves

https://orcid.org/0000-0002-4172-7439

convert brain waves, which means that it is a

continuous record of the brain’s electrical activity by

placing metal electrodes on the scalp (Jasper, 1958;

Khosla et al., 2020; Patel et al., 2018). Neurons

communicate spontaneously with each other by

generating electrical currents and remain active at all

times, even when a person is asleep or relaxed. Low

cost, high time resolution, high flexibility, usability,

non-invasive, portability, and safe nature make the

EEG a powerful tool compared to other functional

neuroimaging techniques such as magneto-

encephalogram (MEG), positron emission

tomography (PET), functional magnetic resonance

imagery (fMRI), and transcranial magnetic

stimulation (TMS). This work is interested in

developing a new method more efficient for the

predicting system of EEG signals, in this sense, this

work uses the covariance and correlation matrices to

determine the acquisition system quality used and

finds electrical leaks between all electrodes, thus

optimization algorithms like DE used to maximize the

quality of the results found, and ultimately the

AdaBoost classification algorithm used to generate

the prediction models (Abenna et al., 2021a), Fig. 1

presents a basic architecture of the prediction system

Abenna, S., Nahid, M. and Bajit, A.

A Novel EEG Classiﬁcation for Baseline Motor Cortex using AdaBoost.

DOI: 10.5220/0010729600003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 125-129

ISBN: 978-989-758-559-3

125

Figure 1: Illustrate of EEG acquisition and processing.

for an EEG signal. This method has been applied for

hand EEG motor cortex, in such a way the prediction

accuracy values are usually 100% using more than 6

acquisition electrodes, compared to related work that

has accuracy values less than 99.1% using PNN.

The rest of the article is organized as follows:

Section 2 presents all algorithms of classification and

optimization proposed for the system. Results and

discussion are presented in section 3, while section 4

provides conclusions and an overview of the future

work.

2 METHODS

2.1 Dataset

'Projectbci-1D' dataset: The motive is a 21-year-old

right arm with no known health condition. An EEG

consists of a random movement of the actual left and

right hands, such as the recorded is with closed eyes.

The electrodes used in this work (FP1, FP2, F3, F4,

C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, FZ,

CZ, and PZ) are distributed as following the

international system 10-20 as illustrated in Fig. 2.a.

The recording sampling at 500 Hz with NeuroFax

EEG device using 19 electrodes for acquisition. The

data were exported with a common reference

'eemagine-EEG', where the AC line operates at 50 Hz.

Fig. 3 illustrates some of the EEG signals acquired by

the NeuroFax device using 19 electrodes when eyes

are closed and the left-hands move. We notice that the

EEG signals are very noisy and misunderstood.

2.2 Adaptive Boost

Freund and Chapyle invented AdaBoost in 1996 as an

iterative boost procedure. The primary goal of this

recovery effort is to focus on situations that are

difficult to categorize. First, each instance is assigned

the

same weight. Iteration increases all weights of

(a) Location of all

electrodes

(b) Nerve system between the

lef

hand and the central lobe

Figure 2: Positioning of all electrodes used to acquire the

EEG signals for the motor cortex of hands.

Figure 3: A part of EEG signals acquired using NeuroFax

device.

lower-ranked instances and reduces correctly ranked

instances, more details are in (Chatterjee et al., 2019).

The AdaBoost is supported by the Algo. 1. Such as

the 𝑥



and 𝑦



represent the feature set and the

corresponding decision class label for the ith instance.

It represents the variance vector with n size because

there is a T iteration, and each instance starts with a

distribution of 1/n. 𝐴



is calculated in the n training

instance (Chatterjee et al., 2019). The weak learner is

applied at each step, and AdaBoost employs the

exponential loss function 𝑒𝑥𝑝𝑎



𝑦



𝜙



𝑥



 to

calculate a weighted error epsilon of 𝐴



. 𝑎



, 𝑦



, and

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

126

Algorithm 1: Adaptive boosting.

- Inputs: Given 𝑥



,𝑦



,...,𝑥



,𝑦



 ,

where n is the total number of

training data instances, 𝑥



is the

feature data and 𝑦



is the

associated decision class,

𝑥

𝑒𝑠𝑡

the testing data;

- Initialize: 𝐴



1/n, where 𝑖 

1,...,𝑛;

for each iteration t in T do

Use distribution to train weak

learners 𝐴



;

Choose: 𝑎



12∗𝑙𝑜𝑔1𝜖



/𝜖



;

Loss: 𝑒𝑥𝑝𝑎



𝑦



𝜙



𝑥



;

Update: 𝐴



1𝑖𝐴



𝑖 ∗ 𝑙𝑜𝑠𝑠/𝑔



where 𝑔



is a new normalization

factor;

𝑔





∑

𝐴



𝑖 ∗ 𝑒𝑥𝑝𝑎



𝑦



𝜙



𝑥









;

End.

eturn

𝑓

𝑥



←𝑠𝑔𝑖𝑛

∑

𝑎



𝜙



𝑥











;

𝜙



𝑥



 denote an n-dimensional weight vector, a

vector containing the actual decision class of n cases,

and a vector containing the anticipated outcome for

the ith instance, respectively. The weight

combination sign of the lower classifier is calculated

by the final classifier f (Chatterjee et al., 2019).

2.3 Optimization

The optimization advantage is to select a valid

classifier and to reduce the functionality used in

restricted classifications to increase the prediction

accuracy. This process is further optimized by the DE

algorithm to determine optimal synthesis conditions

(Estimators number (NE), learning rate (LR), and

random-state (RS)) to maximize reaction yield, more

details are in (Abenna et al., 2021a; Rodrigues et al.,

2018).

3 RESULTS AND DISCUSSION

The experiments were conducted on the 2.4 GHz

desktop and 6 GB of RAM with four Intel®Core

(TM) i5 CPUs and 64 bit/Windows 10 operating

system, and python.3.6 for programming.

3.1 Analysis Results

Fig. 4 illustrates the matrix of covariance between the

19 acquisition electrodes in the state of the left hand

(Fig. 4.a) and of the right-hand movement (Fig. 4.b).

In this figure, we can observe easily the existence of

more connection between all electrodes when the

right-hand moves compared to the left hand,

indicating a large change in the distribution of

electrical signals in the brain between the two hands

state, which also shows that has facilitated the

classification of these signals.

Table 1: Symbol and description.

Symbol Description

E n-estimators

R learning-rate

random-state

EEG The numbe

of electrodes

Accurac

OL Zero One Loss

Tc Classification time

Tp Prediction time

Optimization time

(a) Left hand (b) Right hand

Figure 4: The covariance matrix for each moved hand.

3.2 Feature Selection Results

Fig. 5 shows a classification of electrodes according

to their importance for the classification stage using

the AdaBoost algorithm (Abenna et al., 2021b),

knowing that each electrode has been represented by

its degree of importance from 0 to 1, in this figure we

notice that the best electrodes used are FP1, PZ, and

FP2, knowing that the number of electrodes can be

decreased up to 3 or 1 single electrode without a great

degradation of the system accuracy. Fig. 5.b

illustrates the correlation matrix between the

electrodes signal to detect all links between them and

avoid electrical leaks that degrade the quality of the

acquired signals, such that we notice that all degrees

are low except between some electrodes such as T5,

O1, O2, CZ, and FP2, which implies the best quality

of the acquisition system, and we are not needed to

use any spatial filter to decrease the correlation

between the channels (Whitmore and Lin, 2016).

A Novel EEG Classiﬁcation for Baseline Motor Cortex using AdaBoost

127

(a) Channels selection using AB (b) Correlation matrix

Figure 5: Channels selection and correlation matrix for EEG signals classification of hands moved.

3.3 Evaluation Metrics

Accuracy and Zero-One-Loss are typically metrics

used to measure the performance of biomedical and

complex data during classification.

𝐴𝐶 





(1)

𝑍𝑂𝐿  FP  FN (2)

Where True Positive (TP) refers to a circumstance

in which an alarm is generated although the left hand

has moved during testing. The term TN (True

Negative) describes a circumstance in which the

right-hand moves but no alarm is generated. When the

left hand is employed, the alert is not raised, which is

referred to as FP (False Positive). The term FN (False

Negative) describes a circumstance in which the

right-hand moves but no alarm is triggered (Abenna

et al., 2021b).

3.4 Classification Results

Table 2 shows a main parameter optimization of

AdaBoost to well improve the quality of prediction,

as it can find during testing large combinations of

AdaBoost parameters for that gives precision values

of 100%, knowing that we can choose only those

corresponding to a low value of NE, to guarantee a

high speed of classification and prediction, thus a

good quality of prediction. Table 3 shows a size

improvement of the acquisition system without

degradation of prediction performance, such that we

do the recognition of EEG signals during testing

using these AdaBoost parameters (NE = 257, LR =

0.9969, and RS = 911), so we choose only the best

NEEG-channel have been selected in Figure 5.a, and

we notice that the accuracy value remains at the level

of 100% when NEEG ≥ 6, we notice that the use of

just two electrodes FP2 and PZ gives an accuracy of

97%, indicating the possibility of developing new

devices for the baseline hand EEG motor cortex

prediction with a small size, as well as Tc and Tp,

decreases when we decrease the NEEG. In table 4, the

work of Hossain et al. (Hossain et al., 2015) who uses

the BP and PNN algorithms for the EEG

classification finds that the precision value cannot

exceed 99.1% but at the work of this paper the

accuracy value take 100% when using more than six

acquisition electrodes.

Table 2: AdaBoost parameters optimization using DE.

AdaBoost parameters Optimization results

NE LR RS AC (%) ZOL To (s)

257 0.9969

911 100.0 0 97.73

319 1.5301

559 100.0 0 124.94

Table 3: Improving the number of channels used for hands

EEG classification.

NEEG

Classification results

AC (%) ZOL Tc(s) Tp(s)

100.0 1 48.47 1.43

100.0 0 33.39 1.17

99.58 108 22.91 1.62

97.00 771 18.40 1.00

90.10 2547 15.03 1.00

Table 4: Comparative results with related work.

Work Accurac

Methods

Hossain et al.

(

2015

)

88.9% BP

99.1% PNN

This wor

100% AdaBoost

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

128

4 CONCLUSIONS

In conclusion, the method used in this work shows its

efficiency in predicting the hand moved from the

EEG signals without any mistake, such that this work

uses the covariance matrices to show all changes in

the distribution of the brain activities when moving

every single hand (left or right), so the correlation

matrix used to determine the electrical leaks between

all electrodes, also the use of AdaBoost algorithm to

classifying the EEG signals and the minimization of

the number of channels (NEEG). Also, the use of the

DE optimizer improves the classification

performances, knowing that the accuracy value in this

work takes the value of 100% when using more than

six electrodes. We hope that this work will help other

researchers to develop a good EEG signals prediction

system. In future work, our team focuses on

developing new and more efficient methods and

instigating this work for real-time applications of the

BCI systems.

REFERENCES

Abenna, S., Nahid, M., Bajit, A., 2021. “BCI: Classifiers

Optimization for EEG Signals Acquiring in Real

Time”. In 2020 6th IEEE Congress on Information

Science and Technology (CiSt), IEEE.

Abenna, S., Nahid, M., Bajit, A., 2021. “Chapter 18 Brain

Computer Interface: A Novel EEG Classification for

Baseline Eye States Using LGBM Algorithm”,

Springer Science and Business Media LLC.

Chatterjee, R., Datta, A., Sanyal, D.K., 2019. “Ensemble

learning approach to motor imagery EEG signal

classification”. Machine Learning in Bio-Signal

Analysis and Diagnostic Imaging.

Hossain, A.A., Rahman, M.W., Riheen, M.A., 2015. “Left

and right hand movements EEG signals classification

using wavelet transform and probabilistic neural

network”. International Journal of Electrical and

Computer Engineering.

Jasper, H., 1958. “Report of the committee on methods of

clinical examination in electroencephalography.

Electroencephalogr”. Clin. Neurophysiol.

Jin, J., Sellers, E., Zhou, S., Zhang, Y., Wang, X.,

Cichocki, A., 2015. “A p300 brain-computer interface

based on a modification of the mismatch negativity

paradigm”. International Journal of Neural Systems.

Khosla, A., Khandnor, P., Chand, T., 2020. “A comparative

analysis of signal processing and classification methods

for different applications based on EEG signals”.

Biocybernetics and Bio-medical Engineering, 1–2.

Li, Y., Pan, J., Long, J., Yu, T., Wang, F., Yu, Z., Wu, W.,

2016. “Multimodal BCIs: Target detection, multi-

dimensional control, and awareness evaluation in

patients with disorder of consciousness”. Proceedings

of IEEE.

Michelmann, S., Treder, M.S., Griffiths, B., Kerre´n, C.,

Roux, F., Wimber, M., Rollings, D., Sawlani, V.,

Chelvarajah, R., Gollwitzer, S., Kreiselmeyer, G.,

Hamer, H., Bowman, H., Staresina, B., Hanslmayr, S.,

2018. “Data-driven re-referencing of intracranial EEG

based on independent component analysis (ICA)”.

Journal of Neuroscience Methods journal.

Patel, R., Gireesan, K., Sengottuvel, S., 2018. “Decoding

non-linearity for effective extraction of the eye-blink

artifact pattern from EEG recordings”. Pattern

Recognition Letters. doi:10.1016/j.patrec. 2018.01.022.

Rodrigues, H.C., Herskovits, J., Soares, C.M.M., Arau´jo,

A.L., Guedes, J.M., Folgado, J.O., Moleiro, F.,

Madeira, J.F.A., 2018. “EngOpt 2018 proceedings of

the 6th international conference on engineering

optimization”. Springer Nature Switzerland.

Tang, X., Li, W., Li, X., Ma, W., Dang, X., 2020. Motor

imagery EEG recognition based on conditional

optimization empirical mode decomposition and multi-

scale convolutional neural network. Expert Systems

With Applications.

Whitmore, N., Lin, S., 2016. “Unmasking local activity

within local field potentials (LFPs) by removing distal

electrical signals using independent component

analysis”. NeuroImage.

Yin, Z., Zhang, J., 2017. “Cross-session classification of

mental workload levels using EEG and an adaptive

deep learning model”. Biomedical Signal Processing

and Control.

Zhang, Y., Wang, Y., Zhou, G., Jin, J., Wang, B., Wang,

X., Cichocki, A., 2018. “Multi-kernel extreme learning

machine for EEG classification in brain-computer

interfaces”. Expert Systems With Applications.

A Novel EEG Classiﬁcation for Baseline Motor Cortex using AdaBoost

129