Motor Imagery signal Classification for BCI System Using Empirical Mode

Décomposition and Bandpower Feature Extraction

Dalila Trad

ESSTT, LaTICE 5, Avenue Taha Hussein, B. P. : 56, Bab Menara, 1008 Tunisia
trad.dalila@gmail.com

Tarik Al-Ani

Dep. Informatique, Cit Descartes-BP 99, 93162 Noisy-Le-Grand, UVSQ, LISV, 10/12 Avenue de

/Europe,78140 Velizy, France and ESIEE-Paris, France.
t.alani@esiee.fr

Mohamed Jemni

ESSTT, LaTICE 5, Avenue Taha Hussein, B. P. : 56, Bab Menara, 1008 Tunisia
mohamed.jemni@alecso.org.tn

Abstract

The idea that brain activity could be used as a communication channel has rapidly

developed. Indeed, Electroencephalography (EEG) is the most common technique to measure the

brain activity on the scalp and in real-time. In this study we examine the use of EEG signals in Brain

Computer Interface (BCI). This approach consists of combining the Empirical Mode Decomposition

(EMD) and band power (BP) for the extraction of EEG signals in order to classify motor imagery

(MI). This new feature extraction approach is intended for non-stationary and non-linear

characteristics MI EEG. The EMD method is proposed to decompose the EEG signal into a set of

stationary time series called Intrinsic Mode Functions (IMF). These IMFs are analyzed with the

bandpower (BP) to detect the characteristics of sensorimotor rhythms (mu and beta) when a subject

imagines a left or right hand movement. Finally, the data were just reconstructed with the specific

IMFs and the bandpower is applied on the new database. Once the new feature vector is rebuilt, the

classification of MI is performed using two types of classifiers: generative and discriminant. The

results obtained show that the EMD allows the most reliable features to be extracted from EEG and

that the classification rate obtained is higher and better than using the direct BP approach only. Such

a system is a promising communication channel for people suffering from severe paralysis, for

instance, people with myopathic diseases or muscular dystrophy (MD) in order to help them move a

joystick to a desired direction corresponding to the specific motor imagery.

Keywords: Brain Computer Interface, motor imagery, Bandpower, Empirical Mode

Decomposition, Hidden Markov Model, Support Vector Machines, Cohens kappa coefficient.

1. Introduction

A BCI system is a communication and control pathway between a brain and an external

device. It does not require any external devices or muscle intervention to issue commands and

complete the interaction (Abdulkader et al., 2015). This system primarily concerns the medical

domain and mainly the field of disability. So, the major goal of the BCI research is to provide

severely disabled people with a new communication channel, which is not based on the traditional

motor output channels (Van Erp et al., 2012).

A BCI system is represented as a system in a continuous closed, generally composed of six

steps, (see fig 1): 1. Brain activity measurement, 2. Preprocessing, 3. Feature Extraction, 4.

Classification, 5. Translation into a command and Feedback.

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

Figure 1. General architecture of an online (BCI)

One major challenge of our BCI system is to describe the signals EEG by a few relevant

values called features i.e. step 3 in Fig (1). The success of the mental imagery classification depends

on the choice of features used to characterize the raw EEG signals. These features can then be used

in step 4 in order to classify the user’s mental state. Several approaches for feature extraction have

been proposed in literature.

These approaches are based on time, frequency or time-frequency methods (Al-ani & Trad,

2010; Pfurtscheller, 2004). A common way to gain BCI control is to use motor imagery of left and

right hand movement, which is based on Event-related desynchronization/synchronisation

(ERD/ERS) in specific frequency bands. For instance, imagination of left or right hand movement

results in amplitude causes an attenuation (Event-related desynchronization (ERD)) of  (8-13Hz)

and central  (13-30Hz) rhythms at the contra-lateral sensorimotor representation area and, in an

amplitude increase (event-related synchronization (ERS)) within the  band (30-40Hz) at the ipsi-

lateral hemisphere (Pfurtscheller, 1999; Neuper & Pfurtscheller, 1999). Several common band

power techniques were employed in the BCI literature. Herman et al. (Herman et al., 2008)

demonstrated that the Yule and Welch PSD approaches, mainly dominate the other studied ones.

These approaches are based essentially on some linearity and stationary hypothesis such as

the use of fast Fourier transform (FFT) spectrum in a short-time segment of data. The accuracy of

the FFT calculation is closely related to the choice of the duration of the signal segment. However,

the nature of the EEG signal is non stationary and nonlinear. The main non-stationary and nonlinear

feature extraction technique is the Wavelet Transform (WT) (Samar et al., 1999). Various kinds of

wavelets have been used for BCI, such as Morlet wavelets (Lemm et al., 2004) and wavelet packet

decomposition (Hettiarachchi et al. 2014). Despite being more effective than the FFT, WT approach

shows a much bigger ambiguity in signal decomposition. However, it cannot provide higher

resolution both in time and frequency domain, besides, the decomposition of signal is not adaptive.

In this paper, we applied a recent technique proposed by Huang et al. (Huang et al. 1998),

called the empirical mode decomposition (EMD) for nonlinear and non-stationary time series data

for pattern extraction from motor imagery EEG of left and right hand imaginary movement. This

method EMD is a data driven approach (i.e. one does not need to define a mother wavelet

beforehand) that can be used to decompose adaptively a signal into a finite number of mono-

component signals, which are known as intrinsic mode functions (IMF s) or modes. It considers

signals at their local oscillations, but they are not necessarily considered in the sense of Fourier

harmonics. Their extraction is non-linear, whereas their recombination for exact reconstruction of

the signal is linear. The IMFs admit well-behaved Hilbert transforms (HT) (Long et al. 1995) and

they satisfy the following properties: they are symmetric, different IMFs yield different

instantaneous local frequencies as functions of time that give sharp identifications of embedded

structures. In this work, we propose a hybrid approach combining the EMD and BP for feature

extraction from the EEG signals. We first apply the EMD to select only the IMFs corresponding to

sensorimotor rhythms (mu and beta) using Welch-based PSD to extract the reliable information of

EEG during left and right hand movement imagination. Based on these new features, the

experimental using two classifiers hidden Markov models (HMMs) and Support Vector Machines

(SVM) results show that the proposed method improves recognition rate greatly.

D. Trad, T. Al-Ani, M. Jemni - Motor Imagery signal Classification for BCI System Using Empirical Mode Décomposition
and Bandpower Feature Extraction

2. Methods

2.1. Participants and experimental paradigm

The EEG signals used for this experiment were recorded by a real-time EEG acquisition

hardware g.GAMMAsys active electrode System along with a g.USBamp amplifier g.tec, Guger

Technologies. In this study, the EEG data of ten subjects (two young females and seven young

males with ages ranging from about 22 to 35 years), recorded during imagination of the left and

right hand movements. A subject training session in our work consisted of one experimental run of

40 trials with randomized direction of the cues (20 left-hand imagination and 20 right-hand

imagination). For each subject, we used only 4 out of the best sessions (i.e. 160 trials/subject).

Therefore, each subject is seated in a comfortable armchair 150 cm in front of a computer screen.

At the beginning of each trial (t = 0 s), a fixation cross-appeared on the black screen. After

two seconds a warning stimulus was given in the form of a beep. From 3 to 4.25s, an arrow (cue

stimulus), pointing to the left or right, was shown on the screen. The subject was instructed to

imagine a left or right hand movement until the end of the trial, depending on the direction of the

arrow. The EEG was sampled and classified on-line throughout the session. Between 4.25 and 8s,

the classification result was used to give a continuously updated feedback stimulus in the form of a

horizontal bar that appeared in the center of the screen. The paradigm is illustrated in fig (2).

Figure 2. Timing of one trial of the experiment with continuous feedback (Guger et al, 2001)

2.2. Preprocessing

Two bipolar EEG channels were measured over C3 and C4 according to the international 10-

20 System (Jasper 1958). The left ear served as reference and the FPz as ground. The signals were

sampled at 256 Hz and bandpass-filtered between 0.5 Hz and 30 Hz. An additional 50 Hz notch

filter was enabled to suppress line noise.

2.3. Feature extraction

Once the raw EEG data is recorded and pre-processed, the following step is to extract its

intrinsic features. This step is essential in the functioning of BCIs due to the important measured

amount of brain activities. The aim of this step is then to find a better representation of the EEG

signal while keeping the most relevant properties corresponding to the performed mental imagery.

This information is called features. In this work, we propose a robust method bases on BP and EMD

to extract the relevant EEG features corresponding to MI.

2.3.1. Bandpower

The features may be extracted from the EEG signals by estimating the power distribution of

the EEG in predefined frequency bands. In general, the band power is estimated by digitally band-

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

pass filtering a signal in a given frequency band, then squaring the filtered signal and finally

averaging the obtained values over a given time window. Pfurtscheller et al. (Pfurtscheller et

al.,1997) used the BP and demonstrated that for each subject, different frequency components were

found in the  and  bands which provided best discrimination between left and right hand

movement imagination. These frequency bands varied between 9 and 14 Hz and between 18 and 26

Hz. Such features have been notably used with success for motor imagery classification.

2.3.2. Empirical mode decomposition (EMD):

The traditional EMD was recently proposed as an adaptive time-frequency data analysis

method (Huang et al., 1998). An algorithm based on an empirical framework defines it. The basic

EMD is defined by a process called sifting to break down any multimodal signal to a sum of basis

components called intrinsic mode functions (IMFs). The IMFs are zero-mean AM-FM signals which

must satisfy two conditions: the first one is that the number of extrema and that of zero-crossing

must differ at most by one; the second one is that the mean value between the upper and lower

envelopes are equal to zero at any point. Conceptually, the establishment of this method is quite

simple: one needs to consider a signal at its local oscillation level, remove the fastest oscillation and

iterate the process on the residue considered as a new signal. At the end of the sifting processes, a

given signal x(t) can be written as a sum of a finite number of IMFs, Im(t), m = 1. 2, .... M, and a

final residue rM(t) (Equation1):

(1)

The decomposition is stopped at step M, if either the residue is a mono-component signal or

has only 2 extrema. The stopping criterion must be set to ensure that the obtained signal satisfies the

properties of an IMF while limiting the number of iterations. For more details about the different

steps of the sifting process for the calculation of the IMFi as well as the stopping criterion definition

see (Huang et al., 1998). Since the decomposition into IMFs is based on the local characteristic time

scale of the data, it applies to nonlinear and non-stationary processes. The IMFs admit well-behaved

Hilbert transforms (HT) and they satisfy the following properties: they are symmetric, different

IMFs yield different instantaneous local frequencies as functions of time that give sharp

identifications of embedded structures. The decomposition is done linearly or non-linearly

depending on the data. This complete and almost orthogonal decomposition is empirically realized

by identifying the physical local characteristic time scales intrinsic to these data, which is the lapse

between successive extrema.

2.3.3. EMD and BP for motor imagery

In this work, we propose a direct nonlinear approach to extract the more relevant IMFs

corresponding to the different frequency components in the  and  bands and then obtain the BP in
order to use them as features for mental task classification (see Fig. 3). The feature vector pi used for

the demonstration in this paper is composed, for each sample I, 1 < i < 2048, in a given trial (among

a total of 160 trials) of four bandpower, calculated of the rhythms  and  in positions C3 and C4
(Trad et al., 2011).

Fig. 4 shows the EMD decomposition result of one-trial (left hand movement imagination)

for subject 2 in the channels C3 and C4 respectively (the pre-filtered EEG signal used for this

illustration is not corrupted by blinking artifact.). Each channel is decomposed into ten IMFs and

one residue.

)()()(
1

trtItx
M

m
 



D. Trad, T. Al-Ani, M. Jemni - Motor Imagery signal Classification for BCI System Using Empirical Mode Décomposition
and Bandpower Feature Extraction

Welch based PSD method was applied to analyze the different characteristics of each IMF.

This method was applied to each IMF to find the active frequency bands such as the  and 
rhythms. Fig. 5 shows the PSD in each IMF shown in Fig.4. The major advantage of EMD is that

the input signal may be decomposed directly and adaptively to basic functions (IMFs), each with a

distinct time scale. The IMFs are ordered in increasing time scales, i.e., decreasing frequency. Based

on this property, we can notice that the characteristics of the active frequency bands corresponding

to  [8-12Hz] and  [13-30Hz] are located only in IMF 1, IMF 2 on C3 and C4.
Therefore, the new signal is reconstructed by keeping only the two first IMFs. EMD also

allows eliminating the artifacts in the EEG during the recording sessions like eye blinks and eyeball

movements. In Fig 5, we noted that ocular artifact frequency is generally low around 5Hz with high

amplitude. This artifact appears mainly in IMF3 and IMF4. Finally, band power was applied for the

new signal. As a last step, the logarithm of the BP is calculated in order to transform the distribution

of this feature to a more Gaussian like shape, because the classifiers we used, such as HMMs and

SVM assume normally distributed features.

Figure 3. Hybrid EMD-BP approach for one trail feature extraction

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

Figure 4. The EMD decomposition results for subject 2 when he imagines left hand movement. From top-left

to down-right: the raw signal, the ten IMFs and the residue in channel C4 (a) and in C3 (b)

D. Trad, T. Al-Ani, M. Jemni - Motor Imagery signal Classification for BCI System Using Empirical Mode Décomposition
and Bandpower Feature Extraction

Figure 5. PSD (dB/Hz) vs freaquency (Hz) of each IMF showen in fig 4 in channel C4 (a) and in C3 (b)

2.4. Classification

The goal of the classification step is to assign automatically a class to the feature vector

previously extracted. In this study, four feature sets for each feature extraction approach, BP and

EMD+BP, were used for classification and for test. Two sets of feature vectors; one for LMI and the

other for RMI were used to train the classifier (creating one model for each MI) while the two other

sets of feature vectors, one for left and the other for right MI were used for test. Each set contains 40

feature matrices each of dimension 4x2048. We evaluated our feature extraction approach by using

2 classifiers.

2.4.1. Hidden Markov Model (HMM)

The HMMs approach is very efficient nonlinear technique used for the classification of time

series (Rabiner, 1989). It necessitates two stages: a training stage where the stochastic process

models are estimated through extensive observation corpus and decoding or detection stage where

the model may be used off/on-line to obtain the likelihoods of the given test sequence evaluated by

each model. A HMM is defined by the following compact notation to indicate the complete

parameter set of the model  = (II, A, B), where II, A and B are the initial State distribution vector,

matrix of State transition probabilities and the set of the observation probability distribution in each

State, respectively: 

II = [II 1, II 2, ..., II Ns], II i = P(q1 = si), A = [aij ], aij = P(qt+1 = sj | qt = si). Where 1 ≤ i, j ≤ N,

si, sj  S, S = {s1, s2, ..., sN}, t  {1, 2, ..., T }. The observation at time (or index) t, Ot, is considered

in this paper as continuous or real-valued K-dimensional vector Ot  R
K
, 1 ≤ t ≤ T, 1 ≤ t ≤ T.

For a continuous observation, the State conditional probability of the observation bi(Ot) may be

defined by a finite mixture of any log-concave or elliptically symmetric probability density function

(pdf), e.g. Gaussian pdf, with State conditional observation mean vector µi and State conditional

observation covariance matrix Si. In this paper we consider only a single multivariate Gaussian pdf,

so B may be defined as B = {µi, Si} i = 1,2, ... ,Ns. At each instant of time t, the model is in one of the

States i, 1 ≤ i ≤ Ns. It outputs Ot according to a density function bi(Ot) and then jumps to State j, 1 ≤ j ≤

Ns with probability aij. The State transition matrix defines the structure of the HMM (Rabiner, 1989).

The model  may be obtained off-line by a given training procedure. In practice, given an

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

observation sequence O = {01, O2, ..., Ot}, and an initial model , the HMMs need three fundamental

problems to be solved:

1. How to calculate the likelihood P(O|)? The solution to this problem provides a score of

how O belongs to .

2. How to determine the most likely State sequence that corresponds to a given observation

sequence O? The solution to this problem provides the sequence of the hidden States

corresponding to the given observation sequence O.

3. How to adjust the model  in order to maximize P(O|)? This is the problem of estimating

the model parameters given a corpus of training observations sequences.

Problems 1 and 2 are solved in the decoding or classification stage using the forward or the

Viterbi algorithms (Rabiner, 1989), while problem 3 is solved during the training phase using either

a conventional algorithm such as the Baum-Welch algorithm (Rabiner, 1989) or other optimization-

based algorithm, e.g., (Al-ani & Hamam, 1978).

Our training scheme is based on Baum-Welch training algorithm and the Bayesian Inference

Criterion (BIC). This scheme makes the training procedure independent of the a priori knowledge of

the structure of each HMM needed in the Baum-Welch algorithm. The decoding or classification

may be realised on-line using the stochastic dynamic programming based Viterbi algorithm. This

algorithm needs all the constructed HMMs to classify on-line a given observation sequence.

In our work, two HMMs were built for RMI and LMI using 40 training feature sequences

from each MI. Each sequence is composed of T = 2048 feature vectors of dimension K = 4 each.

2.4.2. Support Vector Machines (SVM)

SVM was first introduced by Vapnik (Vapnik, 1998) (Vapnik, 2000). It is based on two

fundamental principles:

1. Maximum margin which is the distance between the border of separation and the closest

samples. In SVMs, the boundary of separation is chosen as the one that maximizes the

margin. This is justified by the theory of Vapnik-Chervonenkis (or statistical theory of

learning) (Vapnik, 1998). The problem is to find the optimal separating boundary from a

given training set. This is done by formulating the problem as a quadratic optimization

problem for which there are known algorithms.

2. In order to deal with cases where the data are not linearly separable, we increase of the size

of the input space. To be able to handle nonlinear separable data, we transform the space of

input data to another space with larger size in which the probability that a linear boundary

exists is higher. This transformation is realised by using a nonlinear kernel function. Several

kernel functions may be used to design the SVMs. In our work, the Gaussian Radial Basis

Function (RBF):



K(x,y) exp
||xy ||

 2













 (2)

With a scaling factor, , of 1 was selected as it gives the most accurate results.
In our work, two classes SVM was built for RMI and LMI using 40 training feature sequences

from each MI. Each sequence is composed of T = 2048 feature vectors of dimension K = 4 each.

2.5. Statistical performance evaluation method: Cohen’s kappa coefficient

Training the classifiers presented above was designed to minimize the classification error or

to increase the classification rate usually measured by the rate of correctly classified trials. To

interpret the classification performance, we take as reference the percentage rate reached by a

random classification. Indeed, evaluating the performance of a classifier, based on a given features,

is an important issue since its performance may be used for simplifying human training. The

evaluation the performance of our feature extraction is based on the performance of the

D. Trad, T. Al-Ani, M. Jemni - Motor Imagery signal Classification for BCI System Using Empirical Mode Décomposition
and Bandpower Feature Extraction

classification results on RMI and LMI. In order to compare our new feature extraction method

results with the traditional feature extraction method results, we used the Cohen’s kappa coefficient

(Cohen, 1960). It allows to measure the agreement between two classifiers in classification into two

categories. The value of the Cohens K ranges between 1 and -1, where 1 corresponds to perfectly

correct classification and values less than 1 imply less than perfect agreement.

Let P0 the observed agreement among two classifiers and Pe the expected agreement. The

Cohens kappa K is defined by:



K 
Po Pe

1Pe
(3)

3. Results and discussion

3.1. Confusion matrices

The recognition rate calculated by HMM and SVM for the 10 subjects is represented by the

confusion matrix (Table 1) for the two feature extraction methods direct BP and EMD + BP. It is

clearly seen that the combination of the two features methods, EMD and BP gives the best

classification rates for almost all subjects. Based on these matrices, the performance of our approach

was evaluated using K scores.

Table 1. Confusion matrix: recognition rates calculated by HMMs and SVM in the case of RMI and LMI.

These results are showen for both methods of feature extraction (BP and EMD+BP) for all subjects.
Subject FE true class HMM - predicted class SVM –predicted class

predicted class 1 RMI LMI RMI LMI

BP RMI 70 30 86.25 13.75

LMI 2.5 97.5 36.25 63.75

EMD+BP RMI 90 10 85 15

LMI 2.5 97.5 2.50 97.50

BP RMI 77.5 22.5 76.25 23.75

LMI 7.5 923 40 60
EMD+BP RMI 75 25 75 25

LMI 2.5 97.5 25 75

BP RMI 57.5 42.5 61.25 38.75

LMI 35 65 33.75 66.25

EMD+BP RMI 72.5 27.5 61.87 38.13

LMI 40 60 25 75

BP RMI 71.42 28.58 55.62 44.38

LMI 37.15 62.85 42.50 75.50

EMD+BP RMI 71.42 28.58 77.50 22.50

LMI 42.85 57.15 82.75 71.25

BP RMI 52.5 47.5 76.25 23.75

LMI 42.5 57.5 33.13 61.87

EMD+BP RMI 72.5 27.5 76.25 23.75

LMI 40 60 27.50 72.50

BP
RMI 80 20 67.50 32.50

LMI 50 50 32.50 76.50

EMD+BP RMI 85 15 68.12 31.88

LMI 47.5 523 17.50 8230

RMI 80 20 70.62 29.38
LMI 35 65 28.75 71.25

EMD+BP RMI 80 20 78.12 21.88

LMI 32.5 67.5 29.38 70.62

BP
RMI 87.5 12.5 57.50 42.50

LMI 32.5 67.5 45.63 54.37

EMD+BP RMI 90 10 38.75 16.25

LMI 37.5 62.5 33.75 66.25

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

BP RMI 60 40 78.12 21.88
LMI 45 55 28.13 71.87

EMD+BP RMI 82.5 17.5 77.50 22.50

LMI 25 75 7.50 72.50

BP RMI 60 40 66.87 33.13

LMI 42.5 57.5 36.88 63.12

EMD+BP RMI 75 25 75 25

LMI 35 65 25 75

3.2. Cohen’s Kappa coefficient

Table 2 gives the Kappa- values for each one of the ten subjects in order to evaluate our

method of feature extraction based on EMD+BP. The average score obtained by the HMM classifier

when the feature vectors are generated by EMD+BP is equal to 0.54 while that obtained when the

feature vectors are generated by BP is equal to 0.4, i.e., an increase of 35%. For the SVM classifier,

the average score increases from 0.34 to 0.52, i.e., an increase of 53% using the methods BP and

EMD+BP respectively. Therefore, the use of the EMD+BP approach with the two classifiers gives

to the Kappa scores a significant superiority compared to the direct BP approach.

Table 2. K-values obtained by the two types of classifiers (C) HMMs and SVM with the two types

of features extraction (FE) methods: BP and EMD+BP for 10 subjects during RMI and LMI as well

as their mean values

C FE 1 2 3 4 5 6 7 8 9 10 Mean k
HMM BP 0.6 0.8 0.4 0.2 0.2 0.4 0.4 0.6 0.2 0.2 0.40

EMD+BP 0.8 0.8 0.6 0.4 0.4 0.4 0.4 0.6 0.6 0.4 034
SVM BP 0.6 0.4 0.2 0.2 0.4 0.4 0.4 0.2 0.4 0.2 0.34

EMD+BP 0.8 0.6 0.4 0.2 0.4 0.6 0.4 0.6 0.6 0.6 032

3.3. Translation into a command

Once the motor imagery is identified, a command may be associated to this mental task in

order to control a machine (Prataksita et al., (2014)) (Guger et al., 1999). In this work, we

constructed a new Simuhnk/MathWork model to translate on-line the EEG signals into low-level

commands. Fig. 6 shows our experimental EEG-based BCI System (Trad et al., 2015).

1. Off-line phase: In the first step, the EEG signals are recorded while subjects imagine a

right and left hand movement. The second step is the preprocessing and feature extraction of

EEG data. In this step, we implemented our method to extract the relevant features of the

EEG. This method is based on the combination of EMD and BP. In the third step, we

implemented the HMM classifier to assign a model to each motor imagery task: l for LMI

and 2 for RMI.

2. On-line phase: Once the motor imagery is identified by one of the two models, a low-

level command can be then associated to this mental task. This mechanism was

implemented with Simulink/Mathworks. EMD + BP and Viterbi algorithm (Rabiner 1989)

are implemented as an embedded function in Simulink in order to identify the motor

imagery on-line. Viterbi algorithm has two inputs data, the first input is the EEG data and

the second is the two models already constructed in the first step (offline). The Viterbi-based

recognition result is translated into a command to reinforce the movement of the joystick

(right or left) in order to help persons with myopathic diseases or muscular dystrophy to

move this joystick to a desired direction.

D. Trad, T. Al-Ani, M. Jemni - Motor Imagery signal Classification for BCI System Using Empirical Mode Décomposition
and Bandpower Feature Extraction

4. Conclusion

The work presented in this paper concerns the steps of feature extraction and classification

for motor imagery in BCI framework. These steps are essential for the design of BCI Systems since

these Systems represent a direct communication channel between the brain of a subject and a

machine without any direct muscular intervention. Our approach is based on indirect, independent

and asynchronous BCI Systems. Ten healthy subjects participated in the realization of this process

by imagining two hand movements. During our work, we have studied the changes in frequency and

amplitude of the EEG from each subject participating in our experiment. Changes in frequency

distribution within the bands of sensorimotor  and  rhythms vary from one individual to another
and evolve strongly over time. For this reason it is essential to carry a preliminary study for each

subject. Once the values of sensorimotor rhythms are determined for each subject, it is easier to

control subject's brain activity when one realizes motor imagery. This is achieved through the

feedback returned to subject. EMD is particularly important in extracting directly from the EEGs the

time courses of  and  rhythms. This facilitates the detection of the reactive ERD/ERS bands for
individual subjects, i.e., no need to choose an individual subject specific frequency band.

Another advantage of EMD is not only that it is a data driven approach but also it has the

advantage of removing or reducing the artifacts and the noise affecting the EEG signal.

Therefore, we deduce that the classification rate in the two movements imagination are better using

EMD+BP approach than the direct BP approach. When subject motor imagery is recognized, we

translated the EEG signal to a low-level command. This System could allow subjects who suffer

from severe motor disabilities to better reinforce a joysticks movement to right or left.

References
Abdulkader, S. N., Atia, A., & Mostafa, M.-S. (2015). Brain computer interfacing: Applications

and challenges. Egyptian Informatics Journal 16, 213–230

Van Erp, J., Lotte, F., & Tangermann, M. (2012). Brain-computer interfaces: Beyond medical

applications. IEEE Computer 45, 26–34

Al-ani, T. & Trad, D. (2010). Signal processing and classification approaches for brain-computer

interface. In Intelligent and Biosensors, ed. V. S. Somerset (InTech), 25–66

Pfurtscheller, G. (2004). Brain-computer interface - state of the art and future prospects. In

Proceedings of European Signal Processing Conference (Vienna, Austria), 509–510

Figure 6. The general conception of our asynchronous system BCI (offline - online) for reinforcement

of a joystick movement

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 7, Issue 2, June 2016, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

Pfurtscheller, G. (1999). Eeg event-related desynchronization (erd) and eventrelated synchronization

(ers). Electroencephalography: Basic Principles, Clinical Applications and Related Fields, 4th

ed , 958–967

Neuper, C. & Pfurtscheller, G. (1999). Motor imagery and erd related desyncronization. In

Eventrelated Synchronisation, eds. G. Pfurtscheller and E. F. H. Lopes da Silva (Amsterdam

Handbook of Electroencephalography and Clinical Neurophysiology. Revised Edition:

Elsevier), 303–525

Herman, P., Prasad, G., McGinnity, T., & Coyle, D. (2008). Comparative analysis of spectral

approaches to feature extraction for eeg-based motor imagery classification. IEEE

Transactions on Neural Systems and Rehabilitation Engineering 16, 317–326

Samar, V., Bopardikar, A., Rao, R., & Swartz., K. (1999). Wavelet analysis of neuroelectric

waveforms: A conceptual tutorial. Brain and Language 66, 7–60

Lemm, S., Schafer, C., & Curio, G. (2004). BCI competition 2003data set iii: probabilistic modeling

of sensorimotor mu rhythms for classification of imaginary hand movements. IEEE

Transactions on Biomedical Engineering 51, 10771080

Hettiarachchi, I., Nguyen, T., & Nahavandi, S. (2014). Motor imagery data classification for bci

application using wavelet packet feature extraction. In Neural Information Processing

(Springer International Publishing), 519–526

Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., et al. (1998). The empirical

mode decomposition and the hilbert spectrum for nonlinear and non-stationary time series

analysis. In Proceedings of the Royal Society. vol. 454, 903–995

Long, S. R., Huang, N., Tung, C., Wu, M., Lin, R., & Mollo-Christensen, Y., E Yuan (1995). The

Hilbert techniques: an alternate approach for non-steady time series analysis. IEEE Geoscience

Remote Sensing Soc. Lett. 3, 6–11

Guger, C., Schlogl, A., Neuper, C., Walterspacher, D., Strein, T. & Pfurtscheller, G. (2001) Rapid

prototyping of an eeg-based brain-computer interface (bci), Electroencephalography and

Clinical Neurophysiology 9, 49-58.

Jasper, H. (1958). The ten-twenty electrode system of the international federation. In

Electroencephalography and Clinical Neurophysiology. vol. 10, 370–375

Pfurtscheller, G., Neuper, C., Flotzinger, D. & Pregenzer, M. (1997). Eegbased discrimination

between imagination of right and left hand movement. In Electroencephalography and Clinical

Neurophysiology. vol. 103, 642–651

Trad, D., Al-ani, T., Monacelli, E. & Jemni, M. (2011). Nonlinear and nonstationary framework for

feature extraction and classification of motor imagery. International Conference on

Rehabilitation Robotics, IEEE, 1-6

Rabiner, L. R. (1989). A tutorial on hidden markov models and selected application in speech

recognition. In Proceedings IEEE. vol. 77, 267–296

Al-ani, T. & Hamam, Y. (1978). A low complexity simulated annealing approach for training

hidden markov models. International Journal of Operational Research (IJOR) 8, 483–510

Vapnik, V. N. (1998). Computational Learning Theory (John Wiley & Sons)

Vapnik, V. N. (2000). The Nature of Statistical Learning Theory (2nd Ed., Springer)

Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological

Measurement 20, 37–46

Prataksita, N., Lin, Y-T., Chou, H-C. & Kuo, C-H. (2014). Brain-robot control interface:

Development and application. Bioelectronics and Bioinformatics (ISBB), IEEE, 1-4

Guger, C., Harkam, W., Hertnaes, C. & Pfurtscheller, G. (1999). Prosthetic control by an EEG-

based brain-computer interface (BCI). Proceedings AAATE, vol 45, 799-804.

Trad, D., Al-ani, T., & Jemni, M. (2015). A feature extraction technique of eeg based on emd-bp for

motor imagery classification in bci. In 2015 5th International Conference on Information

Communication Technology and Accessibility (ICTA) (IEEE), 1–6