High Performance Multi-class Motor Imagery EEG Classification
Gul Hameed Khan, M. Asim Hashmi, Mian M. Awais, Nadeem A. Khan and Rushda Basir
School of Science and Engineering, Lahore University of Management Sciences (LUMS), Lahore, Pakistan
Keywords:
Motor Imagery, Brain Computer Interface, Electroencephalography, EEG Classification.
Abstract:
Use of Motor Imagery (MI) in Electroencephalography (EEG) for real-life Brain Computer Interface appli-
cations require high performance algorithms that are both accurate as well as less computationally intensive.
Common Spatial Pattern (CSP) and Filter Bank Common Spatial Pattern (FBSP) based methods of feature
extraction for MI-classification has been shown very promising. In this paper we have advanced this frontier
to present a new efficient approach whose variants out compete in accuracy (in terms of kappa values) with
the existing approaches with the same or smaller feature set. We have demonstrated that use of one mu band
and three beta sub-bands is very ideal both from the point-of-view of accuracy as well as computational com-
plexity. We have been able to achieve the best reported kappa value of 0.67 for Dataset 2a of BCI Competition
IV using our approach with a feature vector of length 64 directly composed out of FBCSP transformed data
samples without the need of further feature selection. The feature vector of size 32 directly composed from
FBCSP data is enough to outcompete existing approaches with regard to kappa value achievement. In this
paper we also have systematically reported experiments with different classifiers including kNN, SVM, LDA,
Ensemble, ANN and ANFIS and different lengths of feature vectors. SVM has been reported as the best
classifier followed by the LDA.
1 INTRODUCTION
Motor Imagery (MI) based Electroencephalography
(EEG) signals processing for the brain computer in-
terface (BCI) systems is one of the most progressing
technologies of current times. It’s numerous appli-
cations, extended to medical and non-medical fields
have concerned substantial attention over the recent
times and it is growing at an exponential rate (Meish-
eri, 2018).
In order to develop a BCI system for MI tasks, a
pre-processing scheme is mandatory because the EEG
signals captured from the brain are random and noisy
(Xie, 2018). Signal pre-processing incorporates arte-
facts removal such as noise and eye blinks in case of
MI data, selection of the appropriate channels band-
pass filtering of the signals. Numerous signal pre-
processing techniques have been developed for MI
tasks. (Cheng, 2004) addressed Common Average
Referencing (CAR) to de-noise the EEG signals by
averaging the signals across all the channels. Band-
pass filtering is also proposed by different authors
such as (Osalusi, 2018) and (Xie, 2018) for prepro-
cessing of EEG signals. It is usually done to obtain
the information only from the required band for fur-
ther processing.
For features extraction from the raw EEG signals,
there are many methodologies proposed in the litera-
ture which includes Common Spatial Patterns (CSP)
(Nguyen, 2018), Power Spectral Density (PSD) (Am-
jed, 2014), Discrete Wavelet Transforms (DWT) (Os-
alusi, 2018) etc. CSP is the most commonly used
technique, which computes Spatial Patterns in EEG
before extracting features for MI classification. CSP
calculates the filters, which maximizes the differ-
ence between two classes of the MI brain activities
(Nguyen, 2018), in terms of variance. It was first in-
troduced for two-class hand movement imagination
and then further refined for four class patterns. (Ab-
bas, 2018) Presented Filter Bank Common Spatial
Patterns (FBCSP) in which EEG data filtered into
multiple frequency bands is feed to CSP. Extracted
features in CSP domain are then used to compose the
feature vector. Fusion of spectral and temporal fea-
tures has also been utilized by (W.Abbas, 2018) in the
context of FBCSP approach. Dimensionality reduc-
tion technique, Principal Component Analysis (PCA),
has also been used in the literature (Lotte, 2018) for
MI-EEG classification.
Numerous approaches for classification have been
proposed which include k-nearest neighbor (KNN),
support vector machine (SVM), linear discriminant
Khan, G., Hashmi, M., Awais, M., Khan, N. and Basir, R.
High Performance Multi-class Motor Imagery EEG Classification.
DOI: 10.5220/0008864501490155
In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 149-155
ISBN: 978-989-758-398-8; ISSN: 2184-4305
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
149

analysis (LDA), Nave Bayes classifier (NBC), Ad-
aBoost ensemble learning, fuzzy logic and adap-
tive neuro-fuzzy systems (Meisheri, 2018), (Nguyen,
2018), (Baghli, 2014). Classification techniques like
artificial neural networks (ANN) and Deep learning
approaches have also been tested for EEG data classi-
fication in recent times (W. Abbas, 2018).
A high performance FBCSP-based technique for
feature extraction is proposed for 4-class MI-EEG
which has been tested with multiple classifiers. We
use only four filters for preprocessing as compared
to other FBCSP approaches such as (Xie, 2018) and
(Miao, 2017), which incorporates 9 bands in prepro-
cessing. Employing less number of bands (if the ac-
curacy is not compromised) provides opportunity to
yield a smaller features vector for classification. We
incorporate sensorimotor rhythms from mu and beta
ranges in our approach but use one band for mu and
three sub-bands for beta. These approaches also com-
putes other statistical parameters for feature selec-
tion such as Band power, Time domain parameters
and Mutual information. This adds on computational
complexity of the algorithm. We have composed our
feature vector directly from the FBCSP transformed
data as we are keeping our number of sub-bands just
four. This produces a feature vector of size 64 which
compares well with other FBCSP approaches also
performing statistical feature extraction like (Abbas,
2017) with 88 features, (Abbas, 2018) with 60 fea-
tures and (Miao, 2017) with 162 features.
Our approach based on FBCSP method is more
optimal with regard to accuracy, size of feature vec-
tor and computational complexity. CSP-based 4-class
MI-EEG classification approach like that of (Raza,
2016) has achieved a good kappa value of 0.58 but
has used a bank of 10 filters. Other methodologies
using CSP approach such as (Meisheri, 2018) and
(Nguyen, 2017) does have low feature vector size but
faces the disadvantage of low classification accuracy.
In our FBCSP approach, we managed to just out-
perform other existing approaches in terms of Kappa
value with feature vector size of 32 (Variant 2). This
offers the advantage of low computational complexity
as we are using only 4 bands. With a feature vector
of size 64 ( Variant 1), our approach leads other ap-
proach by high margin in accuracy and is still compu-
tationally light compared to many other approaches.
With regard to classification we will present results
with different classifiers and show that SVM performs
the best.
Rest of the paper is organized as follows: com-
plete workflow architecture of the proposed method-
ology is addressed in Section II. Section III encapsu-
lates the experimental paradigm with detailed discus-
sion on the obtained results. Finally, the conclusion
of this paper is incorporated in Section IV.
2 PROPOSED METHODOLOGY
Detailed work flow of the proposed MI-EEG data
classification methodology is presented in this sec-
tion. Fig. 1 summarizes the complete work flow ar-
chitecture with three different variants. The first vari-
ant (variant-1) all the FBCSP coefficients for classifi-
cation. Second variant (Variant-2) corresponds to us-
ing half of the features discussed in the next section
and third approach (Variant-3) also performs feature
reduction. We will highlight these differences in the
description of each variant. Further computation of
secondary features after applying FBCSP could also
be done as in (Abbas, 2018) but our goal is to just ex-
plore how many bands are optimal to utilize directly
for classification.
Figure 1: Proposed Methodology work flow.
2.1 Pre-processing
The brain oscillations for MI based BCI systems, the
EEG signals carry the sensorimotor rhythms in the
frequency band of 7-13 Hz (mu) and 13-30 Hz (beta)
(Nguyen, 2018). The raw EEG data is pre-processed
by developing four separate band pass filters as per
BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing
150

Nyquist Criteria, depicted in Fig. 1. The band pass
frequency for band 1 is 7 − 13 Hz, band 2 acquires
13 − 19 Hz, band 3 contains 19 − 25 Hz and 25 − 31
Hz is band 4. We have divided Beta band in three sub-
bands as the frequency range in beta band is large.
These frequency bands also provide the advantage of
artifacts rejection such as noise and eye blinks. Using
these bands in beta and mu rhythm produces adequate
feature for MI classification (Mahmood,, 2017).
2.2 Feature Extraction
Common Spatial Patterns (CSP) technique is utilized
for feature extraction from the pre-processed EEG
data. EEG classification performed by different au-
thors in the literature such as (Mahmood, 2017), (Ab-
bas, 2018) implies that CSP is an effective approach
for extracting features in EEG data classification.
CSP discriminates the EEG signals by decomposing
them into spatial patterns that increase the differences
between two separate classes. The main goal of ap-
plying CSP is to maximize the variance for one class
of EEG signals and decreases for the other ones. We
are using one vs rest approach of CSP by breaking the
multi-class problem to k binary classes.
The four classes of pre-processed EEG data
X1;X2;X3; and X4 indicates an N x G matrix where
N represents the number of EEG channels and G cor-
responds to number of data samples. The process of
the algorithm for multi-class CSP described in (Mah-
mood, 2017) is as follows. The normalized spatial
covariance C is presented as :
C
k
=
X
i
k
X
k
iT
trace(X
i
k
X
k
iT )
(1)
where X
iT
k
is the matrix transpose of X
i
k
for k
=1,2,3,4. Accumulating the all four spatial covari-
ance matrices as:
C =
4
∑
k=1
C
k
(2)
and covariance matrices for disjoint trials C
0
are
given as:
C
0
k
=
∑
j6=k
C
j
(3)
where j = 1,2,3,4. Matrix factorization o C is
done by eigenvalue decomposition
C = U
0
ΛU
T
0
(4)
Here, Λ is the square matrix of order N containing the
diagonal elements as eigenvalues and U
0
corresponds
to the matrix of eigenvectors. The data whitening pro-
cess is performed as:
P = Λ
−1/2
U
T
0
(5)
where P is the whitening matrix. The Covariance
matrices C
k
and C
0
k
are whitened to compute interme-
diate matrices S
k
and S
0
k
.
S
k
= PC
k
P
T
(6)
S
0
k
= PC
0
k
P
T
(7)
Both of these matrices share common eigenvec-
tors and sum of the eigenvalue matrices of both re-
sults to Identity. Consequently, the maximum eigen-
value for S
k
will acquire minimum eigenvalue for S
0
k
and vice versa. Therefore, maximizing the variance
between S
k
and S
0
k
by the transformation of matrix
X
k
onto eigenvector space. The matrix projection for
each of the class is computed as
W
k
= U
T
k
P (8)
where U
T
k
is the matirix of eigenvectors. The first and
last m rows of W
k
are obtained to develop a 2m × N
spatial filter W
kS
which is utilized to spatially filter X
i
k
.
Z
k
= W
kS
X
i
k
(9)
The W
kS
for four classes (k = 1 − 4) are con-
catenated resulting in one spatial filter W
S
of or-
der N × 8m. The band-passed filtered signals
x
band
(x
1
,x
2
,x
3
,x
4
) of order N × T for each trial are
spatially filtered and the log variance of these signals
is computed as feature vectors having the order of 1
× 8m.
F
band
=
log

diag(W
T
S
x
band
x
T
band
W
S
)]
trace(W
T
S
x
band
x
T
band
W
S
)
(10)
(Nguyen, 2018) then selects first two and last two
rows of these feature vectors. We applied multi-class
CSP algorithm discussed in (Nguyen, 2018) for four
classes. Instead of feeding the whole EEG data to
CSP as in (Nguyen, 2018), we have divided the signal
into 4 bands and then performed CSP transformation
on each band. Total of 64 features are obtained in the
same fashion by selecting the first and last two rows
respectively in each band by equation (10). These 64
data samples are directly utilized for classification in
the first variant of our approach (Variant-1). In order
to reduce the computational complexity of the pro-
posed algorithm, and considering the fact that max-
imum correlation and covariance lies in the first and
last row of the CSP transformed matrix, we choose
just the first and last row from the final spatial filter
(10), which corresponds to 8 features for each band.
This yields our second variant (Variant-2) having a
feature vector containing total of 32 features.
High Performance Multi-class Motor Imagery EEG Classification
151

Table 1: Accuracy(%) of different classifiers for 32 features proposed in our approach.
Subject kNN SVM LDA Ensemble ANN ANFIS
1 60.4 60.8 58.6 65.3 73.1 46.6
2 54.9 52.8 51.0 68.1 58.4 37.5
3 79.2 80.7 79.5 86.8 38.0 55.6
4 48.3 78.5 72.5 52.4 72.8 36.8
5 46.2 62.5 60.8 65.6 96.7 35.7
6 42.0 70.5 68.7 47.9 70.9 40.6
7 78.5 79.2 80.9 90.2 70.4 40.2
8 76.4 83.4 73.3 86.8 69.3 58.0
9 63.2 84.0 85.4 69.1 72.8 59.8
Mean 61.1 72.5 70.6 69.6 66.2 45.7
Table 2: Accuracy(%) of different classifiers for 16 features proposed in our approach with PCA.
Subject kNN SVM LDA Ensemble ANFIS
1 64.2 66.0 62.5 65.2 25.7
2 50.7 56.3 57.3 56.6 19.4
3 79.2 86.8 85.1 86.1 29.5
4 45.8 49.7 52.1 51.4 33.0
5 48.3 54.2 55.6 55.6 24.0
6 43.1 44.8 41.0 42.0 29.2
7 78.8 86.1 86.5 86.0 45.1
8 75.0 78.8 78.1 78.5 33.0
9 60.4 64.2 62.5 61.8 52.4
Mean 51.9 65.5 64.5 64.8 33.0
2.3 Feature Reduction
Principal Component Analysis (PCA) is a known
method for feature reduction. Although, we have se-
lected a very small number of EEG feature vectors
as compared to different algorithms discussed in the
literature such as (Abbas and Khan, 2018), (Abbas,
2018) and (Nguyen, 2018), PCA described in (Lotte,
2018) is also applied in one variant of our approach
to see it’s effectivity. 32 features obtained in the
previous section are further reduced to 16 features
(Variant-3).
2.4 Classification
After feature extracting process, the next step is to
utilize these features for signal classification. In or-
der to classify EEG signals, numerous classifiers have
been addressed in the literature (Meisheri, 2018). In
this paper, we are using built-in tools of MATLAB for
kNN, SVM, LDA, Neuro Fuzzy and Ensemble learn-
ing classifiers. SVM is used as weak classifier to de-
velop the ensemble learning scheme and parameter
setting kept the default in all the classifier tools. In
case of binary classification algorithms, we incorpo-
rate One Vs Rest approach by encapsulating K binary
classifiers. Results obtained by these classifiers are
addressed in next section.
3 RESULTS AND DISCUSSION
3.1 Dataset
The EEG Dataset we are using for classification in
this project is dataset 2a of BCI Competition IV
(Tangermann, 2012). The dataset comprises of MI
based EEG data from 9 different subjects. It contains
4 different motor imagery tasks, where the class 1 cor-
responds to imagination of left hand movement, class
2 relates to the movement of right hand, both feet are
associated with class 3 and tongue to class 4. The
recordings were made for total of 288 trails using 25
electrodes (22 EEG and 3 EoG). Signal sampling fre-
quency is 250 Hz and all the data is filtered in the band
of 0.5-100 Hz. An notch filter of the frequency of 50
Hz was applied to suppress the noise.
BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing
152

Table 3: Mean Kappa coefficient comparison with different techniques.
Method Year Feature Extraction Feature
Selec-
tion
Features
Used
Mean
Kappa
Value
Accuracy
(%)
(Guan, 2019) et al. 2019 Riemannian Geometry SJGDA - 0.60 70.0
(Xie, 2018) et al. 2018 Filter Bank ELM-TS-RE MIBIF - 0.62 71.5
(Meisheri,2018) et al. 2018 CSP with Fano’s inequality Nil - 0.38 53.5
(W. Abbas) et al. 2018 FBCSP FFTEM - 0.61 70.7
(Nguyen, 2018) et al. 2018 CSP Nil 16 0.53 64.7
(Abbas, 2018) et al. 2018 FBCSP BP
&TDP
60 0.59 69.2
(Abbas, 2017) et al. 2017 HALS-NNMF BP
&TDP
88 0.62 71.5
(Miao, 2017) et al. 2017 FBCSP STFSCSP 162 0.53 64.7
(Raza, 2016) et al. 2016 CSP Nil - 0.58 68.5
(Ang, 2012) et al. 2012 FBCSP MIBIF - 0.59 69.2
Proposed (Variant-3) - FBCSP PCA 16 0.54 65.5
Proposed (Variant-2) - FBCSP Nil 32 0.63 72.5
Proposed (Variant-1) - FBCSP Nil 64 0.67 75.2
3.2 Experimental Paradigm
Experiment is performed on dataset 2a of BCI com-
petition IV. All 22 EEG channels are used for classi-
fication. Each trial is 7.5 seconds long with 250 Hz
sampling rate. Besides all the challenges discussed
previously, choosing an optimized time window for
classification is also the requirement. We are using a
time window of 1s which ranges from 4s to 5s in each
trial for training data and 0.5 s window ranging from
2.5s to 3s in case of testing data for all the subjects.
3.3 Performance Evaluation
In order to evaluate the effectiveness of the proposed
algorithm, metrics to address the classification perfor-
mance is the Accuracy defined by (Nguyen, 2018).
Table1 addresses the percent classification Accu-
racy of evaluation data for all 9 subjects with different
classifiers. 32 CSP features are utilized for these clas-
sification results. The last row of the table represents
the average Accuracy of all the subjects. Classifica-
tion results presented in Table1 shows that SVM is the
most suitable classifier for proposed method. In order
to further reduce the computational complexity of the
classifiers, one variant of our approach uses Princi-
pal Component Analysis (PCA) algorithm. Total of
32 features obtained by CSP are further reduced to 16
features. Table 2 envelops the percent classification
Accuracy of all the subjects for evaluation data after
applying PCA algorithm with mean Accuracy as well.
Another metric used in the literature for perfor-
mance evaluation of different BCI classification al-
gorithms is the Cohen’s Kappa coefficient (Nguyen,
2018). Average Kappa value of all nine subjects ob-
tained by the proposed classification methodology is
compared with different MI classification techniques
in TABLE 3. We have included the size of feature set
of existing approaches where available and computed
the Kappa coefficient from stated classification accu-
racies where Kappa value was not provided. ”-” sign
in the table indicates that no exact feature size was
mentioned in the paper. The highest Kappa value ob-
tained in proposed method is achieved by SVM clas-
sifier, so we have mentioned it in comparison with
different number of features. If we choose first two
and last two rows from CSP (Nguyen, 2018), we ob-
tain 16 features for each class and there will be total
of 64 CSP features. This tends to improve the clas-
sification Accuracy to 75% by SVM. Kappa value in
this table demonstrates that the proposed methodol-
ogy provides the best classification results among all
the listed techniques.
3.4 Analysis and Discussion
Classification results of the methodologies in the lit-
erature address that dataset we are using in this study
is one of the most strenuous EEG data to classify.
Results comparison in terms of classification Accu-
racy score as well as for Kappa value with different
techniques has witnessed the effectiveness of the pro-
posed classification algorithm. Accuracy in case of
cross validation for all the subjects of the dataset and
for separate testing of evaluation data has shown that
the proposed method attained best results for SVM
High Performance Multi-class Motor Imagery EEG Classification
153

classification as we achieved 75.2 % mean accuracy
corresponding to the Kappa value of 0.67 on testing
dataset for 64 features. 72.5 % Accuracy with Kappa
value of 0.63 is achieved using 32 features only. Our
results are the best compared to all methods listed in
Table 3. Performance of ANN and Deep learning ap-
proaches can be further improved but these methods
require large amount of data and high computational
cost. Secondly, the results comparison with different
techniques which have been proposed in recent times,
has shown that the proposed method comes up with
best results for Kappa value as well as for classifica-
tion accuracy. Additionally, proposed technique gain
the dominance over other listed techniques is compu-
tational cost. We have achieved better results for less
number of features as compared to other methods.
4 CONCLUSION
A method for classifying MI signals of EEG for BCI
systems has been proposed in this paper and results
have been compared with different state of the art al-
gorithms. Proposed methodology presented in this
paper encapsulates 4 band-pass filters for preprocess-
ing. FBCSP algorithm is developed as feature ex-
traction technique to classify EEG signals for 4-class
motor imagery. Many classification algorithms have
been tested on our methodology with feature set of
different lengths where SVM has produced best clas-
sification results among all the listed approaches. Re-
sults are compared with some latest techniques of the
literature witnessing that the proposed methodology
addressed best results among all the listed techniques.
We have came with the conclusion that using a very
less number of CSP features extracted from only mu
and beta rhythms can provide a reliable feature ex-
traction source for multi class EEG data classifica-
tion. Since we have utilized the CSP features of a
limited number for classification task by a computa-
tionally very inexpensive classification technique, the
proposed algorithm can be applied for on-line classi-
fication of MI tasks in real-time BCI applications.
REFERENCES
H. Meisheri, N. Ramrao, S. K. Mitra, “Multiclass Common
Spatial Pattern for EEG based Brain Computer Inter-
face with Adaptive Learning Classifier”, ELSEVIER
Pattern Recognition Letters 2018
X. Xie, Z. L. Yu, Z. Gu, J. Zhang, L. Cen and Y. Li, ”Bi-
linear Regularized Locality Preserving Learning on
Riemannian Graph for Motor Imagery BCI,” in IEEE
Transactions on Neural Systems and Rehabilitation
Engineering, vol. 26, no. 3, pp. 698-708, March 2018.
M. Cheng, W. Jia, X. Gao, S. Gao and F. Yang, ”Mu
rhythm-based cursor control: an offine analysis,” Clin-
ical Neurophysiology, 2004.
B. Osalusi, A. Abraham, D. Aborisade, “EEG Classification
in Brain Computer Interface (BCI): A Pragmatic Ap-
praisal”, American Journal of Biomedical Engineer-
ing 2018, 8(1): 1-11
F. Lotte, L. Bougrain, A. Cichocki, M. Clerc, M. Congedo,
A. Rakotomamonjy, and F. Yger, “A review of clas-
sification algorithms for EEG-based brain–computer
interfaces: a 10 year update”, Journal of Neural Engi-
neering, 15 (2018) 031005 (28pp)
T. Nguyen, I. Hettiarachchi, A. Khatami, L. Gordon-Brown,
C. P. Lim and S. Nahavandi, ”Classification of Multi-
Class BCI Data by Common Spatial Pattern and Fuzzy
System,” in IEEE Access, vol. 6, pp. 27873-27884,
2018.
Amjed S. Al-Fahoum and Ausilah A. Al-Fraihat, ”Meth-
ods of EEG Signal Features Extraction Using Lin-
ear Analysis in Frequency and Time-Frequency Do-
mains,” ISRN Neuroscience, vol. 2014, Article ID
730218, 7 pages, 2014.
S. Kumar and F. Sahin, “A Framework for a Real Time In-
telligent and Interactive Brain Computer Interface”,
Computers and Electrical Engineering Volume 43,
April 2015.
H. Dose, J. S. Møller, S. Puthusserypady, H. K. Iversen
“A Deep Learning MI-EEG Classification Model for
BCIs”, 26th European Signal Processing Conference
(EUSIPCO) 2018
W. J. Bosl, H. Tager-Flusberg and C. A. Nelson, “EEG An-
alytics for Early Detection of Autism Spectrum Disor-
der: A data-driven approach”, Scientific Reports vol-
ume 8, Article number: 6828 (2018)
T. Nguyen, A. Khosravi, D. Creighton, S. Nahavandi
“EEG signal classification for BCI applications by
wavelets and interval type-2 fuzzy logic systems”, EL-
SEVIER, Expert Systems with Applications 42 (2015)
4370–4380
Qi Xu, H. Zhou, Y. Wang, J. Huang, ” Fuzzy support vec-
tor machine for classification of EEG signals using
wavelet-based features”, ELSEVIER Medical Engi-
neering and Physics 31 (2009) 858–865
F. Z. Baghli, Y. Lakhal, L. El Bakkali and O. Hamdoun,
”Design and simulation of Adaptive Neuro Fuzzy In-
ference System (ANFIS) controller for a robot manip-
ulator,” 2014 Second World Conference on Complex
Systems (WCCS), Agadir, 2014, pp. 298-303.
A. Mahmood, R. Zainab, R. B. Ahmad, M. Saeed and A. M.
Kamboh, ”Classification of multi-class motor imagery
EEG using four band common spatial pattern,” 2017
39th Annual International Conference of the IEEE En-
gineering in Medicine and Biology Society (EMBC),
Seogwipo, 2017, pp. 1034-1037.
M. Tangermann, K. M
¨
uller, Ad Aertsen, N. Birbaumer,
C.Braun, C. Brunner, R. Leeb, C. Mehring, K. J.
Miller, G. R. M
¨
uller-Putz, G. Nolte, G. Pfurtscheller,
H. Preissl, G. Schalk, A. Schl
¨
ogl, C. Vidaurre, S.
BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing
154

Waldert and B. Blankertz, “Review of the BCI com-
petition IV”, Frontiers in Neuroscience, 13 July 2012,
6, 55.
W. Abbas and N. A. Khan, ”A High Performance Approach
for Classification of Motor Imagery EEG,” 2018 IEEE
Biomedical Circuits and Systems Conference (Bio-
CAS), Cleveland, OH, 2018, pp. 1-4.
W. Abbas and N. A. Khan, ”FBCSP-based Multi-class Mo-
tor Imagery Classification using BP and TDP fea-
tures,” 2018 40th Annual International Conference of
the IEEE Engineering in Medicine and Biology Soci-
ety (EMBC), Honolulu, HI, 2018, pp. 215-218.
W. Abbas and N. A. Khan, ”DeepMI: Deep Learning for
Multiclass Motor Imagery Classification,” 2018 40th
Annual International Conference of the IEEE Engi-
neering in Medicine and Biology Society (EMBC),
Honolulu, HI, 2018, pp. 219-222.
K.K. Ang, Z.Y. Chin, C. Wang, C. Guan, H. Zhang, ”Filter
bank common spatial pattern algorithm on BCI com-
petition IV datasets 2a and 2b”, Frontiers in neuro-
science, vol. 6, pp. 39, 2012.
H. Raza, H. Cecotti, G. Prasad, ”A combination of trans-
ductive and inductive learning for handling non-
stationarities in motor imagery classification”, 2016
International Joint Conference on Neural Networks
(IJCNN), pp. 763-770, 2016.
M. Miao, H. Zeng, A. Wang, C. Zhao, F. Liu, ”Discrim-
inative spatial-frequency-temporal feature extraction
and classification of motor imagery EEG: An sparse
regression and Weighted Nave Bayesian Classifier-
based approach”, Journal of neuroscience methods,
vol. 278, pp. 13-24, 2017.
S. Guan, K. Zhao and S. Yang, ”Motor Imagery EEG
Classification Based on Decision Tree Framework
and Riemannian Geometry”,Computational Intelli-
gence and Neuroscience Volume 2019, Article ID
5627156
W. Abbas and N. A. Khan, ”A discriminative spectral-
temporal feature set for motor imagery classification,”
2017 IEEE International Workshop on Signal Process-
ing Systems (SiPS), Lorient, 2017, pp. 1-6.
High Performance Multi-class Motor Imagery EEG Classification
155