Content uploaded by Md. Asadur Rahman
Author content
All content in this area was uploaded by Md. Asadur Rahman on Feb 27, 2019
Content may be subject to copyright.
2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), 7-9 February, 2019
978-1-5386-9111-3/19/$31.00 ©2019 IEEE
Towards the Effective Intrinsic Mode Functions for
Motor Imagery EEG Signal Classification
Tahnia Nazneen
Department of Biomedical Engineering
Khulna University of Engineering &
Technology (KUET)
Khulna-9203, Bangladesh
tahnian.k08@gmail.com
Md. Asadur Rahman
Department of Biomedical Engineering
Khulna University of Engineering &
Technology (KUET)
Khulna-9203, Bangladesh
bmeasadur@gmail.com
Md. Nurunnabi Mollah
Department of EEE
Khulna University of Engineering &
Technology (KUET)
Khulna-9203, Bangladesh
nurunnabim12@gmail.com
Abstract— To better utilize one of the most powerful signal
decomposition methods called Empirical Mode Decomposition
(EMD) in the field of a brain-computer interface, a better
understanding of its components is necessary. In analyzing
two-class motor imagery Electroencephalogram, one or more
specific intrinsic mode function (IMF) plays a major role in the
classification and therefore, utilization of the signal. This
research work investigates the most effective IMF for motor
imagery EEG signal by PCA based Hilber-Huang
transformation. To implement the research work, motor
imagery (left and right hand) EEG signals of eight participants
are taken into consideration. The signals of central lobes are
transformed into three dimensions to one dimension by PCA.
These transformed signals are decomposed into 8 IMFs.
Significant features are extracted from the signals and
classified by ANN and LDA method and tabulated. From the
results, we found that IMF4 provides the most significant
classification accuracies.
Keywords— Empirical mode decomposition, intrinsic mode
function, EEG, motor imagery, ANN, LDA.
I. INTRODUCTION
Electroencephalography (EEG) is the noninvasive
recording and measuring the electrical activity of the cerebral
cortex performed via electrodes placed on the scalp [1]. One
of the signals that EEG picks up from the brain is the motor
imagery signal which is acquired from the mental process by
which an individual simulates a given action in his/her mind
without practically performing the movement in reality [2].
This signal is similar in nature to the signal that is acquired
during the practical performance of the task; varying only in
the state of being blocked at a cortico-spinal level.
Originating from the neuronal changes in the primary
sensorimotor cortex, motor imagery signal can alter mu
rhythm (noted µ, with frequencies 8−12Hz) and beta rhythm
(noted β, with frequencies > 13Hz); thus, it can be used to
drive brain-computer interfaced devices which are used in
rehabilitation and gaming industry.
One of the most effective signal decomposition methods
in Brain-Computer Interface (BCI) research is the Empirical
Mode Decomposition (EMD). The Empirical Mode
Decomposition technique, first introduced by Huang et al.
[3], decomposes a signal adaptively into the simplest
intrinsic oscillatory modes known as Intrinsic Mode
Functions (IMFs). The advantage of EMD over other
methods such as Fourier Transform (FT) or Wavelet
Transform (WT) is that it can be used for the analysis of
nonstationary and nonlinear signals such as EEG, giving it a
much-anticipated time-frequency analysis ability [4-5]. One
of the six limitations of EMD included in [3] the
optimization problem that deals with the selection of the best
IMF and its uniqueness. Methods such as Multivariate EMD
[6], Ensemble EMD [7] and Variational Mode
Decomposition [8] assist in solving the other limitations such
as processing bi-variate or tri-variate signals, eliminating
mode mixing present in the original EMD and removing
EMD’s sensitivity to noise and sampling. In these cases, a
previously acquired knowledge of the most effective IMF
plays a major role in reducing the computational complexity
and time which may result in designing the more user-
friendly BCI system for rehabilitation systems. However,
there is no specific documentation of the IMF which can be
analyzed to yield the highest accuracy in case of
differentiating between a two-class BCI using motor imagery
EEG signals.
In this study, we analyzed the IMF most responsible for
differentiating between gross left hand and gross right hand
movement from the motor imagery signal. The basic
contribution of this work is to find the most effective IMFs
for motor imagery EEG signal classification. For this
purpose, we collected EEG signal from the scalp of eight
individuals by a 9-channel Wireless B-Alert X10 EEG
headset. The data acquisition was performed in such a way
that frontal, central, and parietal lobes were covered by this
system. Since, the most effective area for the motor imagery
is central lobe, only the data C3, Cz, and C4 positions
(international 10/20) are considered for further analysis.
With necessary preprocessing, principal component
analysis (PCA) was performed on the channels C3, Cz, and
C4 to make it single channel data with considering the 1st
principal component. Then, the empirical mode
decomposition method was applied to the signals to
decompose into eight IMFs. Each of the IMFs were then
taken into account for feature extraction. Using the time-
frequency domain features of the IMF signals, the two events
(left hand and right hand imagery movement) were classified
using the artificial neural network (ANN) and linear
discriminant analysis (LDA) algorithms. Here we have used
simple type classifier as LDA and the complex one is ANN
so that we can understand the ability of the features to be
classified in both environments. The classification accuracies
from the classifiers for every individual and every IMF were
tabulated to find the most effective IMF for the concerning
event classifications. For selecting the most effective IMF we
considered both the average classification accuracies with
their standard deviations regarding the participants and the
consistencies of the results between two classifiers (nonlinear
ANN and linear LDA).
Fig. 1. Electrode positions of C3, Cz and C4 on the scalp according to the
International 10/20 method. The green colored labels (9 channels) indicate
the channels used in data acquisition.
Fig. 2. EEG signal acquisition from the scalp of the subject by 9-channel
wireless B-Alert X-10 device where the subject performs the mental motor
imagery task with the guide of the graphical protocol aiding software.
This paper is organized as follows: Section II notes the data
acquisition procedure of this study. Section III describes the
proposed methodology with proper mathematical
interpretations. The results and discussions are presented in
Section IV. Finally, we draw our conclusion regarding the
total work with future research perspective in Section V.
II. DATA ACQUISITION
We captured EEG data from eight healthy and right-
handed participants at 256Hz sampling rate using B-Alert
X10 EEG System. The headset contains 9 channels of EEG
data. The subjects sat upright in a chair with the electrodes
placed on their heads in the specified positions as shown in
Fig. 1. Each subject had undergone 20 trials for the recording
of brain activity during imagery movement. The task was to
simulate the action of abducting the hand from resting
condition and adducting it back to the resting condition using
solely the arm joint mentally without truly moving the arms.
After 5 seconds of rest initially, the subjects imagined
moving the right hand in aforementioned manner within 5
seconds and it was followed by 5 seconds’ rest. The same
procedure was followed for the left hand. Visual trigger
signals using a graphical protocol aiding software [9] were
presented on a screen to indicate the start of a trial and a task.
The data acquisition procedure has been given in Fig. 2. The
signal was filtered by an IIR notch filter to remove 50 Hz
power line noises. For data acquisition, AcqKnowledge 4.4
software was used. Furthermore, data acquisition procedure
was performed in the Advanced Bio-signal Processing
Laboratory, Dept. of Biomedical Engineering of Khulna
University of Engineering & Technology (KUET).
III. METHODOLOGY
The complete methodology has been presented by the
block diagram given in Fig. 3. After starting with the process
of loading the 9-channel EEG signal, a number of steps have
been deployed. The steps are associated with some
processing algorithms. These individual algorithms are
discussed briefly to explain the procedure to find the most
effective IMF regarding the motor imagery EEG signal
classification. Although some methods inside the proposal
like ANN, LDA, and EMD are well known, according to the
significance regarding this research work, these methods are
re-explained briefly. It is done due to provide a
comprehensive concept with the mathematical interpretations
about the fundamental structures of the investigating
procedure and the corresponding findings of this work.
Start
Load the 9 channel EEG data
Preprocessing: dimensionality
reduction (PCA), filtering, and
signal decomposition (EMD)
IMF= 1:8
IMF=IMF+1
Separate the data as the intrinsic
mode function (IMF) and extract
the concerning features
Divide the features for
training and testing Store 20% data
for testing
Train the network for
modeling
Tabulate all the data
from the storage
Store the predicted
results
End
Yes
No
Test
Predictive
Model
IMF≤ 8
Fig. 3. Flowchart of the proposed methodology that describes all steps to
find the most effective IMF for motor imagery EEG signal classification.
A. Dimensionality Reduction
Dimensionality reduction is a crucial step in the field of
BCI signal processing to reduce the probable high
computational cost and prevent overfitting. PCA is a linear
transformation approach which seeks a projection that best
represents the data without losing much of its information.
This unsupervised method maximizes the variance criterion.
If the first two or three principle components explain a
percentage of variance, then, entire dataset can be visualized
in this two or three-dimensional space respectively [10]. In
this study, the left hand motor imagery and right hand
imagery EEG data has been transformed using PCA and the
first projection vector is extracted to be further subjected to
EMD. Since the C3, C4, and Cz EEG data are most
responsible for the motor imagery movement (see Fig. 4),
these 3 channels are consider for dimension reduction from 3
to 1 using PCA operation.
B. Signal Decomposition
The raw sets of signals need an initial decomposition into
its linear or non-linear simple intrinsic mode oscillations
rather than a set of coefficients. As mentioned before, this
adaptive decomposition method further allows separating the
oscillatory mode that is generated when a motor imagery
task is performed.
The EMD procedure for an input signal
)(tx
is
described below. This is known as the sifting process [4].
a) Assume
)()(
1txth
b) Find out all the local extrema (minima and
maxima).
c) Determine the upper and lower envelope
)(teU
and
)(teL
respectively, which connect all local
maxima and local minima with a cubic spline
interpolation.
d) Calculate the mean of
)(teU
and
)(teL
i.e.,
2)()(
)( tete
tLU
(1)
e) Generate the first IMF candidate,
)()()( 111 tthth
(2)
f) Check whether
)(
1th
is an IMF by checking with
the two basic criteria: the mean value criterion and
the S-number criterion.
g) Repeat the steps 2 to 6 until an IMF
)(
1th
is
determined.
h) If the first IMF is found out, consider
)()( 11 thtD
i) In order to find out the remaining components, find
out the residue,
)()()( 11 tDtxtr
(3)
j) The sifting process is continued until the final
residue becomes a function from which an IMF
extraction is not possible.
Fig. 4. The topoplot of the power distribution of the 9-channel EEG
signals of imagery movement of the left hand (left one) and right hand
(right one). These topoplot has been prepared by the MATLAB toolbox
given in [11].
It is the highest frequency component of
)(tx
. Eventually,
the signal
)(tx
is decomposed into M number of IMFs
where
)(tDp
is pth IMF and
)(tr
is the final residue [3]:
M
pptrtDtx 1)()()(
(4)
C. Feature Extraction
EMD method generates IMFs whose length (duration) is
the same as the original input signal. The ability to convert
this data into a reduced set of features represents a very
important step in any classification for developing BCI.
Since these features characterize the behavior of the EEG
signals for BCI, their selection is of crucial importance. In
this work, five significant features regarding EEG signal
classification [12] were extracted from each IMF signal
those are mean absolute deviation (
MAD
), total power
(
x
P
), L2 Norm (
2
L
), entropy (
x
H
), and Skewness (
).
Suppose, the data of a typical IMF is
i
D
, where indicates
the IMF number. Each IMF contains
n
number of sampling
points where the
D
represents the average value of that
significant IMF,
i
D
. The mathematical interpretation of the
feature extraction procedure from the
i
D
has been given by
(5)-(9).
n
DD
MAD
n
ii
1
(5)
n
iix D
n
P1
2
1
(6)
n
ii
DL 1
2
2
(7)
)(log).(10
1i
n
iix DPDPH
(8)
n
i
iDD
n13
3
)(1
(9)
Furthermore, the notation
represents the standard
deviation of the IMF,
i
D
.These features were used for two
imagery movement related EEG signal classification.
D. Classification
For classification purposes, two different types of
classifiers were used in this research work. A nonlinear type
classifier (ANN) and a linear type classifier (LDA) were
chosen for classification purpose.
a) Artificial Neural Network
Inspired by biological neural networks, artificial neural
networks (ANNs) are a family of statistical learning
algorithms used to estimate functions and predict classes
that depend on a large number of inputs [13]. ANNs are
represented as systems of interconnected neurons which are
capable of mapping input to its output activation by taking
values as its input [14]. A Feed-Forward network with 2
hidden layers was used in this research. It was trained using
the Levenberg-Marquardt algorithm.
b) Linear Discriminant Analysis
Linear discriminant analysis or LDA finds a linear
combination of features that separates two or more classes
of data [10]. Firstly, the d-dimensional mean vectors for the
different classes from the dataset. It is followed by the
calculation of the scatter matrix along with the eigenvectors
and corresponding eigenvalues. The eigenvectors are then
sorted by decreasing eigenvalues. N eigenvectors are chosen
with the largest eigenvalues to form a d×N dimensional
matrix. This matrix transforms the samples onto the new
subspace. This can be summarized by the matrix
multiplication: YY=XX×WW (where XX is an n×d
dimensional matrix representing the n samples, and YY are
the transformed n×N dimensional samples in the new
subspace) [15].
Therefore, a number of steps were performed to
complete the method of the proposed research work. The
total procedure can be presented by the following flowchart
given in Fig. 3. From the figure we already got that all steps
are presented with the chronological order. In the
classification phase, we applied 5-fold cross-validation
technique to present the classification accuracies. Therefore,
from 20 trials, 16 set features were used to train and 4 set
were used to test. According to the procedure of the 5-fold
cross validation, the training and testing features are
repeated 5 times and evaluated the performances 5 times.
The results are then averaged to present as the final
outcomes of the classification accuracies.
IV. RESULTS AND DISCUSSIONS
Judging from the complexity of the brain signals, it’s not
surprising that the long and arduous process may yield
unsatisfactory results. The raw signal, when plotted, doesn’t
show any difference between the signals corresponding to
left-hand or right-hand movement. Furthermore, high
dimensionality is always a threat to classification.
Therefore, high dimensional was reduced by applying PCA.
The three channels (C3, Cz, and C4) were reduced into one
channel information. A typical example of the PCA has
been given in Fig. 6 where three channels were reduced to
one channel considering the only first principal component
to reconstruct the signal (since the score of 1st principal
component was 0.85).
After performing PCA, we have decomposed the EEG
signal by the EMD method as described by the proposed
methodology. A typical EMD of an EEG signal has been
given in Fig. 7. From the figure, we find that the EEG signal
has been decomposed into 7 IMFs. From all 8 subjects’ data
gave 7 IMFs each. Consequently, all of the IMFs were
considered for feature extraction. The feature extraction
procedure was conducted as the proposed methodology. The
extracted features were separated based on the class label.
After that, the classifiers were used as 5-fold cross-
validation approach to classifying the signals. Based on the
features MAD, L2 norm, and total power, a scatter plot has
been given in Fig. 5 to show the differentiability. Though
here only three dimensions of the features were utilized, all
five features will definitely create more differentiability
between the two class features for classifications.
To classify the features of two classes, ANN and LDA
were utilized. For ANN, 5 input layer and two output layer
used to design the ANN with 3 hidden layers. Both the LDA
and ANN were set to classify the features based on the 5-
fold cross-validation technique. The classification results for
eight subjects were tabulated with sequential manner.
Fig. 5. The raw motor imagery signal of (a) left and (b) right hand
movement of Subject 4 with the duration of 5 seconds each.
Fig. 6. 3-dimensional scatter plot of the features MAD, L2 Norm, and
Total Power. The triangles indicate left hand whereas the circles indicate
right hand features.
Fig. 7. A typical hand imagery movement related EEG signal and its IMFs through EMD analysis.
TABLE I. CLASSIFICATION ACCURACIES REGARDING ALL THE
SUBJECTS OF EACH IMF BY LDA CLASSIFIER
TABLE II. CLASSIFICATION ACCURACIES REGARDING ALL THE
SUBJECTS OF EACH IMF BY ANN CLASSIFIER
Subject
Serial Number of IMFs and the classification accuracy (%)
1
2
3
4
5
6
7
8
1
72
78
76
82
79
71
61
58
2
74
70
72
78
72
69
65
55
3
63
65
68
74
65
66
55
50
4
75
78
89
92
79
75
63
68
5
77
76
76
82
80
74
70
72
6
82
88
88
97
81
82
68
65
7
62
65
68
76
70
65
55
58
8
70
72
78
85
72
78
65
55
Average
71.87
74.00
76.87
83.25
74.75
72.50
62.75
60.12
St. dev.
6.79
7.69
8.06
7.90
5.80
5.87
5.52
7.47
The classification accuracies regarding different IMFs
have been presented in Table I and Table II. The
classification accuracies of 8 subjects by LDA are given in
Table I and the corresponding results of ANN are given in
Table II. From the results, it is found that consistently IMF 4
is providing the highest classification accuracies for both the
ANN and LDA classifiers. To present such clear results the
average classification accuracies regarding all the IMFs
have been presented by the Bar Graph in Fig.8. In this
result, the standard deviations of the results regarding the 8
subjects have also been added. We found that the
classification accuracy of IMF 4 reaches up to 97% where
ANN outperforms the LDA.
In addition, it is also found that the classification
accuracies of IMF3 and IMF5 are also convincing.
Combination approaches of these IMFs may be used to
improve the classification accuracies of motor imagery EEG
signal classification.
Fig. 8. The average classification accuracies with the standard deviation
with respect to the IMFs. Results suggest that IMF 4 showing the highest
classification accuracy for both the LDA and ANN.
Subject
Serial Number of IMFs and the classification accuracy (%)
1
2
3
4
5
6
7
8
1
68
75
70
78
65
62
65
58
2
66
62
66
75
62
62
60
63
3
63
65
60
76
63
63
63
62
4
75
72
78
84
69
71
69
65
5
70
82
76
85
78
72
71
67
6
78
84
87
95
85
85
75
68
7
65
72
73
76
72
65
61
63
8
69
68
76
80
68
70
64
65
Average
69.25
72.50
73.25
81.12
70.25
68.75
66.0
63.87
*St. dev.
5.06
7.70
8.15
6.72
7.88
7.74
5.20
3.13
*St. dev. =standard deviation
V. CONCLUSION
The most effective IMF in the classification of right
hand versus left hand motor imagery signal was determined
using EMD. The results showed that, no matter what the
variance, the 4th IMF always showed the highest accuracy.
This finding may help to reduce the memory needed to run
such processing programs and the processor power and
processing time as well. However, more motor imagery
tasks, such as feet vs. hands or the classification among the
feet are yet to be documented. An algorithm for
optimization of the EMD method by finding the most
effective IMF maybe developed in future.
We also hope to conduct further research to combine
two or more IMFs to improve the classification accuracies
for multiple class problems regarding the EEG signal
classification. In addition, the time-frequency and frequency
domain properties will be considered to select the most
effective IMF’s for the EEG signal classification of other
types of stimuli.
References
[1] G. Light, L. Williams, F. Minow, J. Sprock, A. Rissling, R. Sharp, N.
Swerdlow, and D. Braff, “Electroencephalography (EEG) and event-
related potentials (ERPs) with human participants,” Current Protocols
in Neuroscience, vol. 52, pp. 21–24, July 2010.
[2] M. Lotze, and L. G. Cohen, “Volition and imagery in
neurorehabilitation,” Cognitive and Behavior Neurology, vol. 19,
no.3, pp. 135–140, September 2006.
[3] N. E. Huang, Z. Shen, S. R. Long, M. L. Wu, H. H. Shih, Q. Zheng,
N. C. Yen, C. C. Tung, H. H. Liu, “The empirical mode
decomposition and Hilbert spectrum for nonlinear and nonstationary
time series analysis,” Proc. R. Soc. London A, vol. 454, pp. 903-995,
1998.
[4] N. Zhuang, Y. Zeng, L. Tong, C. Zhang, H. Zhang, and B. Yan,
“Emotion Recognition from EEG Signals Using Multidimensional
Information in EMD Domain,” BioMed Research International, vol.
2017, Article ID 8317357, 9 pages, 2017.
[5] M. E. Alam, “Electroencephalographic signal processing and
classification techniques for noninvasive motor imagery based brain-
computer interface,” M.Sc. Theses, Georgia Southern University,
2017.
[6] P. Özel, A. Akan, and B. Yılmaz, “Noise-Assisted Multivariate
Empirical Mode Decomposition Based Emotion Recognition”,
Electrica, vol. 18, no: 2, pp. 263-274, 2018..
[7] H. Purba, J. T. Musu, S. A. Diria, W. Permono, O. Sadjati, I. Sopandi,
F. Ruzi, “Completed ensemble empirical mode decomposition: a
robust signal processing tool to identify sequence strata,” IOP Conf.
Series: Earth and Environmental Science 132, 2018.
[8] R. Ram and M. N. Mohanty, “Performance analysis of adaptive
variational mode decomposition approach for speech enhancement,”
International Journal of Speech Technology, vol. 21, no. 2, pp. 369–
381, 2018.
[9] M. A. Rahman, “Matlab Based Graphical Protocol,” MATLAB
Central File Exchange. Retrieved October 20, 2018. Available in:
https://www.mathworks.com/matlabcentral/fileexchange/69162-
matlab-based-graphical-protocol.
[10] M. A. Rahman, M. M. Haque, A. Anjum, M. N. Mollah, and M.
Ahmad, “Classification of motor imagery events from prefrontal
hemodynamics for BCI application,” International Joint Conference
on Computational Intelligence (IJCCI), 2018.
[11] M. A. Rahman, “Topoplot for B-Alert X-10 9-Channel EEG Signal”
MATLAB Central File Exchange. Retrieved January 14, 2019.
Available in :
https://www.mathworks.com/matlabcentral/fileexchange/69991
[12] M. A. Rahman, M. M. Or Rashid, F. Khanam, M. K. Alam, and M.
Ahmad, “EEG based brain alertness monitoring by Statistical and
Artificial Neural Network Approach,” vol. 9, no. 12, 2019.
[13] M. A. Rahman and M. Ahmad “Movement-related events
classification from functional near-infrared spectroscopic signal,” 19th
International Conference on Computer and Information Technology
(ICCIT), 2016, Dhaka, Bangladesh, pp. 207-212.
[14] R. Ahmmed, M. A. Rahman, and M. F. Hossain, “An advanced
algorithm combining SVM and ANN classifiers to categorize tumors
with position from brain MRI images,” Advances in Science,
Technology and Engineering Systems Journal, vol. 3, no. 2, pp. 40-
48, March 2018.
[15] M. F. Kabir, S. M. Rabiul Islam, and M. A. Rahman, “Accuracy
improvement of fNIRS based motor imagery movement classification
by standardized common spatial pattern,” 4th International Conference
on Electrical Engineering and Information & Communication
Technology (ICEEICT), 13-15 September 2018, Dhaka, Bangladesh,
pp. 1-6.
