Channel Selection for Motor Imagery Task Classiﬁcation using

Non-linear Separability Measurement

Stuti Chug and Vandana Agarwal

Department of Computer Science and Information Systems, BITS Pilani, Pilani Campus, India

Keywords:

Brain Computer Interface, Channel Selection, Non Linearity Measurement, Support Vector Machine.

Abstract:

The EEG based motor imagery task classiﬁcation requires only those channels which contribute to the maxi-

mum separability of the training data of different classes. The irrelevant channels are therefore not considered

in the formation of feature vectors used in classiﬁcation. In this paper, we propose a novel algorithm for

efﬁcient channel selection (NLMCS). The algorithm computes the proposed metric λ for non-linearity mea-

surement (NLM) and uses this for channel selection. The algorithm is validated on the benchmarked BCI

competition IV datasets IIa and IIb. The selected channels are then used for extracting Haar wavelet features

and subjected for classiﬁcation using Support vector Machine. The minimum value of λ corresponds to the

optimal channel selection resulting in the best accuracy of motor imagery task classiﬁcation. The mean Kappa

coefﬁcient computed for BCI competition IV IIa dataset using the proposed algorithm is 0.65 and it outper-

forms some existing approaches.

1 INTRODUCTION

Brain computer interface (BCI) systems are used in

various application areas such as assistive technology

support based on classiﬁcation of motor imagery of

EEG patterns of thoughts, communication and gam-

ing to provide a level of independence to people suf-

fering from neuromuscular disorder. Brain computer

interface provides a platform where brain signals are

connected with computer. BCI system uses various

type of signals such as electroencephalogram (EEG),

Electrocorticography (ECog), magnetoencephalogra-

phy (MEG) and functional magnetic resonance imag-

ing (fMRI). However, EEG signals are widely used

in non-invasive BCI system. The basic approach

of BCI is used to convert user brain activity pattern

into corresponding command. BCI system comprises

of different modules such as signal acquisition, pre-

processing for artifact and noise removal, channel

selection, feature extraction and classiﬁcation. BCI

applications mainly focus on needs of people who

are not able to communicate directly to the world

e.g. medical applications such as deep brain stim-

ulation for Parkinson disease and cochlear implants

for the deaf. BCI introduced an efﬁcient way to con-

vert thoughts into machine commands. Thought pat-

https://orcid.org/0000-0002-8942-5114

terns have also been examined in applications such

as controlling a cursor on the screen by brain, choos-

ing a letter from a virtual keyboard, browsing internet,

emotion recognition, motor imagery and gaming (Ab-

dulkader et al., 2015).

The raw EEG signals are captured through a large

number of channels, mounted on the scalp of the per-

son, of which some channels are highly active during

a particular mental activity such as imagining about

the movement of an arm or of a leg. The selection of

such active channels providing useful and discrimina-

tory information is a challenging task. These channels

once selected, are used to extract the features to rep-

resent the thought samples. These samples of EEG

signals are represented as points in the highly multi-

dimensional space. These points belonging to differ-

ent pattern classes may be highly overlapping caus-

ing misclassiﬁcation or may be separable using higher

order non linear polynomial functions. Linear sep-

arability of these sample points plays important role

in classiﬁcation of thought patterns. If classiﬁcation

data is linearly separable then it is easier to classify as

compared to the data which is non-linearly separable.

In literature, highly non-linearly separable EEG data

is classiﬁed using an extremely complex deep neu-

ral network that uses a large number of hidden layers

(Zhang et al., 2019). This induces an extremely large

parametric space to work with, especially for com-

Chug, S. and Agarwal, V.

Channel Selection for Motor Imagery Task Classiﬁcation using Non-linear Separability Measurement.

DOI: 10.5220/0010787300003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 171-178

ISBN: 978-989-758-549-4; ISSN: 2184-4313

171

(a)

(b)

Figure 1: Sample input space (a) sample data is linearly

separable (b) sample data is not separable linearly.

putation of synaptic weights. The optimal parame-

ters thus are computationally intensive and sometimes

difﬁcult to be obtained due to the limitations of al-

gorithms such as Gradient Descent and Genetic al-

gorithms, which get trapped in the local optima. In

this paper, we attempt to explore the linear separabil-

ity of the data to be able to work with linear classiﬁers

such as Support Vector Machine to save the time from

complex hidden layers and classify thought patterns

more efﬁciently and accurately. Linear separability of

data is computed using computational geometry, lin-

ear programming, neural network and quadratic pro-

gramming (Elizondo, 2006; Elizondo et al., 2012).

Our Contribution in This Paper: In this paper, we

propose a novel algorithm for Non-Linearity Mea-

surement (NLM) metric to compute the best channels

in-order to improve the motor imagery task classiﬁca-

tion. This algorithm uses the raw EEG data from all

channels initially and computes the proposed NLM

metric for all classes pairwise for the given combina-

tion of channels. The minimum value of NLM metric

is then used to select the best channels for classiﬁ-

cation. The proposed algorithm is validated with the

bench marked BCI competition IV datasets IIa and

IIb. The algorithm performance has been measured in

terms of accuracy and Kappa coefﬁcient, and it out-

performs some of the existing work in the same do-

main.

This paper is organized as follow: Section 2 de-

scribes the brief literature survey on channel section,

Section 3 presents the basic framework and the pro-

posed algorithm, section 4 discusses the results and

section 5 presents the conclusion.

2 RELATED WORK

The aim of using a channel selection algorithm is to

enhance the classiﬁcation accuracy by reducing over-

ﬁtting issue and also reducing computational com-

plexity while using EEG data. Channel selection is

usually considered as part of the feature extraction

process.

Channel selection algorithms are based on prior

information regarding scalp region of interest of mo-

tor imagery task and iteratively optimize the most rel-

evant channels (Shenoy and Vinod, 2014). The au-

thors showed that their algorithm gives 90.77% ac-

curacy on BCI Competition III dataset IVa data us-

ing ten channels. A real-time feedback based Iter-

RelCen method using relief algorithm gives 85% ac-

curacy on self-data (Shan et al., 2015). This algo-

rithm is not suitable for the multi-class model and

is time-consuming. ReliefF-based channel selection

is used for reducing number of channels for conve-

nience in practical usage in DEAP dataset (Zhang

et al., 2016). CKSCSP method obtained a minimum

set of relevant channels along with the spatial region

of brain, with 84.55% accuracy on BCI Competition

III dataset (Kirar and Agrawal, 2017). Cohen’s d ef-

fect size CSP (E-CSP) channel selection algorithm

eliminates those channels which do not give any use-

ful information by using z-score (Baig et al., 2020;

Das and Suresh, 2015). It shows 85.85% classiﬁca-

tion accuracy on BCI Competition III dataset IVa by

using 9 channels. Sequential Floating Forward Se-

lection (SFFS) technique adds and removes channels

according to its Signal Characteristics Advantage Dis-

advantage P300 (Baig et al., 2020; Qiu et al., 2016).

In ICA-BCI system subject speciﬁc minimal channel

subset was selected to produce high classiﬁcation ac-

curacy but, in this algorithm, the minimal set of chan-

nels is greater than 5-8 (Zhou et al., 2019). A detailed

review on channel section techniques can be found in

(Baig et al., 2020).

The raw EEG data from selected channels is sub-

jected for feature extraction using various methods

such as Fourier transform, Discrete Wavelet transform

(DWT) & Haar wavelet (Nicolas-Alonso and Gomez-

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

172

Gil, 2012). Various classiﬁcation algorithms are used

for motor imagery classiﬁcation namely: linear clas-

siﬁer, nonlinear Bayesian classiﬁer, nearest neighbor

classiﬁer, neural network and combination of classi-

ﬁers. Linear classiﬁers use linear functions for dis-

tinguishing the classes such as Linear discriminant

analysis (LDA) and support vector machine (SVM)

(Gonzalez et al., 2013; Nicolas-Alonso and Gomez-

Gil, 2012).

Power spectral density (PSD) technique adopted

for feature extraction on various frequency transfor-

mations enhances the classiﬁcation performance us-

ing LDA classiﬁer (Alam et al., 2021). In ”Dis-

tance Preservation to Local Mean” [DPLM] approach

for dimensionality reduction the local architecture of

the features is preserved by preserving distances to

local means (Davoudi et al., 2017). (Gaur et al.,

2018) introduced a pre-processing ﬁlter approach

”Subject Speciﬁc Multivariate EMD Filter” [SS-

MEMDBF]. The ﬁlters based on MEMD reduce the

non-stationaries caused by inter and intrasubject dif-

ferences, thus obtaining enhanced EEG signals. For

classiﬁcation, the author used Riemannian mean com-

putation for all classes. A multi class time-frequency

CSP algorithm for motor imagery feature extraction

with three different classiﬁers named Linear Discrim-

inant Analysis (LDA), Na

ıve Bayes (NVB), and Sup-

port Vector Machine (SVM) was proposed by (Zhang

and Eskandarian, 2020). Sparse time-frequency seg-

ment common spatial pattern (STFSCSP) algorithm is

used for features selection in multiple time-frequency

segments and the Weighted Na

ıve Bayesian Clas-

siﬁer (WNBC) was used as classiﬁer (Miao et al.,

2017). Bilinear sub-manifold learning (BSML) clas-

siﬁcation algorithm was proposed to classify the data

on a learned intrinsic sub-manifold based on high-

dimensional Riemannian manifold (Xie et al., 2016).

3 PROPOSED ALGORITHM

In the present work, we propose to compute the non

linear separability of the training samples belonging

to different motor imagery classes. Each sample is

represented by a d-dimensional point, where d is the

total number of features. The j

training sample of

class is represented by the feature vector T

p j

given

as T

p j

= (x

p j

, x

p j

, x

p j

, . . . x

p j

, . . . x

p j

). If the number

of intersection points between two classes is less, then

the data is highly linear separable. A new metric for

NLM for ﬁnding the optimal number of channels is

proposed. All combinations of the channels are gen-

erated and the NLM value for each combination is

computed.

Algorithm 1: NLM pairwise (T

,p,k).

Input: T

: Feature matrix obtained using

equation (9), class p, class k

Output: λ

D=s*t // D is the total no of features

from ’s’ selected channels

=0 // Number of overlapping feature

points between class p & k

for class u = [p,k] do

for each dimension i = 1 to D do

Min(u, i) = Min

u j

);

Max(u, i) = Max

u j

);

for u =[p,k] do

for v = [p,k] do

for each training feature vector

au j

...T

au j

] for class u

count=0 // Initialize count as 0

if u 6= v

for each dimension i = 1 to D do

if x

u j

>= Min(v,i)&&x

u j

Max(v, i) then

count = count + 1;

if count == D then

= N

+ 1 // training sample

adds to overlapping

Compute λ

// N

:Number of

samples in class k

Consider a two class sample data represented us-

ing 2-dimensional feature vectors [Fig.1]. The class

1 data is represented by red dots and has 11 samples,

while class 2 is represented by blue dots which has 12

samples [Fig.1a]. The notion used for sample size of a

class p is N

as is explained after equation (1). In this

example, N

=11 and N

=12. The term x

p j

represents

the i

feature value extracted from the j

sample of

the class p where p represents class index (p = 1, 2),

j represents the sample number ( j = 1, 2, ..m) and i

represents the feature number (i = 1, 2, ...d).

In Fig.1, we have shown samples for two classes

p = 1 and p = 2, where m = 11 for p = 1, and m = 12

for p = 2. The axes of coordinate plane are f

and

. The range of feature values is computed for each

class in each dimension along the axes f

and f

re-

spectively. Each sample data T

p j

is represented by its

feature vector (x

p j

) and is shown as a point. The

range of class p samples is deﬁned in f

dimension by

Min(p,1) and Max(p,1), where Min and Max repre-

sent the minimum and maximum of all data samples

p j

( j = 1, 2) in f

dimension. Similarly, Min(p,2)

and Max(p,2) represent the minimum and maximum

Channel Selection for Motor Imagery Task Classiﬁcation using Non-linear Separability Measurement

173

Algorithm 2: NLMCS (T).

Input: Feature matrix T of training data set

Output: C={C

, ...C

}

All Channels = [c

, c

, ...c

] // q number

of channels initially

Min NLM=10.0; C=All channels;

// Initialize

for a = 1 to 2

− 1 do

ch = a

combination of channel set

, ...C

} using (7)

Sum = 0;

= feature matrix obtained using (9) for

channel set

for p= 1: (no. of classes-1) do

for k= p+1: no. of classes do

= NLM pairwise[T

, p, k]

Sum = Sum + λ

= Sum; // using (2)

if λ

<= Min NLM then

Min NLM = λ

; C = ch;

return C // C corresponds to Min NLM

(a)

(b)

Figure 2: The proposed NLM metric is not affected by the

class imbalance data. Example class size N

= 6, and N

21. It is affected by the overall spread of overlapping data

across other classes. In (a)λ

= 6/(6+21) = 0.222 (b) λ

20/(21+6) = 0.74.

values in f

direction.

A data sample T

p j

from class p is investigated if it

overlaps with data samples of any other class k (where

k 6= p) by checking if x

p j

>= Min (k, i) and x

p j

Max (k, i) for all i = 1, 2, ...d. If the overlap is in all di-

mensions, the data sample T

p j

from class p is consid-

ered to be overlapping with class k samples, causing

non linear separability. This contributes in the pro-

posed metric (λ

) of NLM between classes p and k

and is deﬁned as follows.

+ N

(1)

Where N

is the total number of overlapping

points between classes p and k, and N

and N

are

the total number of samples in classes p and k respec-

tively [Algorithm 1]. The proposed metric of non lin-

earity measurement is given below

λ =

∑

p,k=1

p6=k

(2)

Where n is the total number of classes in given

dataset.

The metric (λ

) is effectively used for the class

balanced data set used in this study. However, the

metric does not impose any penalty for the class im-

balanced data set [Fig.2]

The feature matrix T of the training data involv-

ing all channels prior to channel selection is shown in

equation (3).

T =

...

: : : : :

...

: : : : :

...

: : : : :

...

: : : : :

...













class 1: N

feature vectors

class 2: N

feature vectors

class n: N

feature vectors

(3)

The size of feature matrix T is r ∗d, where r is the

number of rows in the feature matrix and is equal to

the total number of samples in the training data given

as.

r = N

+ N

+ ... + N

(4)

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

174

The number of columns in T is d, which is equal

to the total number of channels multiplied by the

number of features extracted for a channel [assuming

equal number of features extracted from each chan-

nel]. Total number of classes is n and N

is the num-

ber of training samples in class p. The raw feature

matrix T given in equation (3) involves all channels

prior to channel selection. If there are a total of q

channels (say C

...C

) initially and if t features

are extracted from each channel, then the total num-

ber of features d is equal to q ∗ t. Then the matrix T

in equation (3) can further be illustrated as

T =

...

(q−1)t+1

...

(q−1)t+1

...

: : : : : : :

...

(q−1)t+1

...

(q−1)t+1

...

(q−1)t+1

...

: : : : : : :

...

(q−1)t+1

...

: : : : : : :

...

(q−1)t+1

...

(q−1)t+1

...

: : : : : : :

...

(q−1)t+1

...













Channel C

(5)

The class labels along with corresponding rows

are used in the NLM algorithm for the purpose of

computing λ

for classes p and k. The features ma-

trix in equation (5) above involves all channels which

is reﬁned by selecting the channels appropriately. An

inclusion of channel consists of retaining the appro-

priate columns. For example, if the channel C

is se-

lected, then all t columns are used for the j

sample

of class p and is given by

(s−1)t+1

p j

(s−1)t+2

p j

(s−1)t+3

p j

... x

p j

The combination of selected channels of all q chan-

nels is represented by the bit string of 0

and 1

where

0 represents the absence and 1 represents the pres-

ence of a channel in an experimental run. For ex-

ample a combination 1011001 represents the selec-

tion of C

and C

channels. This bit string is

also equivalent to a unique decimal number (64 + 0 +

16 + 8 + 0 + 0 + 1 = 89). The position of bit for each

channel is ﬁxed. The channel numbering can corre-

spond to either left to right or right to left in the bit

string and remains ﬁxed throughout the experiment.

In order to form the combinations, we systematically

iterate from 1 to 2

−1. This keeps generating equiva-

lent bit string in each iteration which is used to extract

the relevant columns from the feature matrix deﬁned

in equation (5). A feature matrix T

is obtained by

retaining all columns corresponding to the included

channels C represented as 1

in the binary equivalent

of the number a. Recalling the above example of a

taken as 1011001, the selected channels are written as

a set {C

}. If q = 10 (say) then,

} ⊆ {C

} (6)

To generalize, if a total of s channels

, ..C

} are selected using the binary

equivalent of a, then the selected channels can be

represented as a subset of all q channels.

, ..C

} ⊆ {C

...C

} (7)

where s ≤ q. Let the set of selected channels

, ...C

}, obtained using binary equivalent of

a, be represented as C. Then the feature matrix

formed by retaining the corresponding t*s number of

columns in T from equation (5) can be denoted by T

Let us deﬁne,

ap j

−1)t+1

p j

−1)t+2

p j

... x

−1)t+t

p j

1×t

(8)

Then,

a11

...

a11

a12

...

a12

: : : :

a1N

...

a1N

a21

...

a21

a22

...

a22

: : : :

a2N

...

a2N

: : : :

an1

...

an1

an2

...

an2

: : : :

anN

...

anN













Class 1

Class 2

Class n

(9)

The size of matrix T

is given as r ∗ st, where r is ob-

tained using equation (4), s is the number of channels

selected equation (7) and t is the number of features

extracted from each channel. Each of T

ap j

in equa-

tion (9), for all classes p=1,2,...n and their samples

Channel Selection for Motor Imagery Task Classiﬁcation using Non-linear Separability Measurement

175

indexed by j=1,2,...,N

, is then replaced by its corre-

sponding vector as deﬁned in equation (8). The al-

gorithm NLMCS() [Algorithm 2] returns the selected

channels to be used for obtaining T

using (9) to train

the classiﬁer for motor imagery task classiﬁcation.

4 EXPERIMENTAL SETUP

4.1 Data Set Description

In this paper, two bench marked datasets namely

BCI Competition IV dataset IIa and IIb were used.

BCI Competition IV dataset IIa has 4 classes of

motor imagery tasks. The EEG signals were col-

lected from nine volunteer participants including four

classes namely left hand, right hand, tongue and feet.

Two sessions of motor imagery tasks were recorded

from each subject, one for training and the other for

evaluation. Each session contains 288 trials of data

recorded with 25 channels (22 EEG Channels and

3 EOG channels). BCI Competition IV dataset IIb

has 2 classes of motor imagery tasks. The EEG

signals were collected from nine volunteer partici-

pants including two classes namely left hand and right

hand movements. Two sessions of motor imagery

tasks were recorded from each subject one for train-

ing and the other for evaluation. Each session con-

tains 120 trials of data recorded with 3 channels [

http://www.bbci.de /competition/iv/] .

4.2 Experimental Evaluation

The performance of the proposed algorithm is eval-

uated on BBCI Competition IV datasets IIa and IIb

as discussed above. The BCI Competition IV IIa

dataset has nine subjects as is mentioned above and

each subject data is divided into training and test-

ing. We train the model on training dataset and val-

idate the accuracy on testing dataset. Out of the 25

channels, we initially used 10 channels subset (i.e

11,13,15,19,20,21,22,23,24,25) for this experiment to

reduce the exponential time complexity. For these 10

channels, 1023 combinations (2

-1) are generated.

Thirty features were extracted using Haar wavelet

with parameters n=5 and m=4 (where n and m are lev-

els of decomposition) from each of the 10 channels on

training dataset. The feature vector size for each train-

ing sample class is taken as 600 (60 features* number

of channels). The best channels were selected using

the proposed NLMCS algorithm and used for classi-

ﬁcation using Support Vector Machine (SVM). Two

performance measures were used to evaluate the per-

formance of the proposed algorithm; cohen’s kappa

(a)

(b)

Figure 3: Channel selection using NLMCS() (a) NLM met-

ric λ versus accuracy scatter graph (b) Number of times

channels contributed in high accuracy.

cofﬁcient(k) and accuracy (Acc) computed using the

diagonal values of confusion matrix (Table 3). The

accuracy (Acc) was computed as the ratio.

Acc =

∑

t=1

∑

t=1

∗ 100 (10)

Cohen’s kappa coefﬁcient given below is used for

evaluating the performance of the proposed algo-

rithm.

κ =

− P

1 − P

(11)

∑

t=1

∗ n

(12)

where, P

and P

represent observed agreement and

chance agreement on test samples respectively. N is

the total number of test samples, n is the total number

of classes, n

is the total number of samples predicted

to be belonging to class t and n

is the total number of

samples from actual class t. The proposed algorithm

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

176

Table 1: NLM and accuracy of each subject with respective channels selected using proposed algorithm.

BCI Comp IV dataset IIa BCI Comp IV dataset IIb

Subject

Channels

NLM

Acc(%)

Channels

NLM

Acc(%)

11 13 15 19 20 21 22 23 24 25 1 2 3

Sub 1 0 0 1 1 1 1 1 1 1 1 0.438 61.458 0 1 0 0.870 55

Sub 2 1 1 1 1 0 0 1 1 0 1 0.544 57.916 1 1 1 0.808 66.32

Sub 3 0 1 1 0 0 0 1 1 0 1 0.640 56.944 1 0 1 0.708 60.32

Sub 4 1 1 0 1 0 0 1 1 0 0 0.435 55.55 0 1 1 0.879 54.07

Sub 5 1 1 0 1 0 0 1 1 1 1 0.448 84.680 1 0 1 0.735 67.20

Sub 6 1 1 1 0 0 1 0 1 1 1 0.324 54.166 0 0 1 0.996 65.77

Sub 7 1 1 0 0 1 0 1 1 1 1 0.397 78.819 1 1 0 0.801 61.43

Sub 8 0 0 1 0 1 0 1 1 1 1 0.398 75.694 1 1 1 0.829 52.61

Sub 9 0 0 0 0 0 0 0 1 1 1 0.317 62.152 1 1 1 0.892 56.73

Table 2: Kappa values of the proposed approach and existing approaches for BCI competition IV IIa dataset.

Study Approach #C 1 2 3 4 5 6 7 8 9 Mean

(Xie et al., 2016) TSSM+LDA 22 0.77 0.33 0.77 0.51 0.35 0.36 0.71 0.72 0.83 0.59

(Davoudi et al., 2017) DPLM 22 0.75 0.49 0.76 0.49 0.34 0.36 0.68 0.76 0.76 0.60

(Miao et al., 2017) DSFTP 22 0.63 0.43 0.74 0.54 0.19 0.26 0.63 0.62 0.69 0.53

(Gaur et al., 2018) SS-

MEMDBF

22 0.86 0.24 0.70 0.68 0.36 0.34 0.66 0.75 0.82 0.60

(Zhang and Eskandarian,

2020)

TFCSP 22 0.62 0.36 0.76 0.40 0.29 0.34 0.59 0.57 0.62 0.51

(Alam et al., 2021) FE-PSD 22 0.47 0.48 0.92 0.60 0.61 0.50 0.53 0.72 0.68 0.61

Proposed Algorithm NLMCS <=10 0.61 0.49 0.61 0.58 0.81 0.53 0.78 0.77 0.67 0.65

Table 3: Confusion matrix for subject 1.

Predicted class

1 2 3 4

1 35 16 13 8

2 14 44 9 5

3 8 8 50 6

Actual

class

4 7 6 7 52

was implemented using MATLAB R2020b and all ex-

periments were performed on Intel(R) Core(TM) i5-

4590 CPU Processor(3.30GHz).

4.3 Results and Discussion

This section presents the results of classiﬁcation accu-

racy using the proposed NLMCS() algorithm on both

BCI Competition IV datasets IIa and IIb. These re-

sults indicate that if data is more linearly separable

then classiﬁcation accuracy will enhance. Fig.2(a)

displays the effect of NLM on accuracy where each

point is a unique combination of channels. The ﬁgure

depicts the scatter plot of 2

(=1024) points repre-

senting unique combinations of 10 different channels.

To visually depict the channel selection, in this ﬁg-

ure the training data used is of subject 5, while sim-

ilar plots for other subject are not displayed here to

avoid redundant ﬁgures. It is observed from the ﬁgure

that the maximum accuracy (84.680%) is associated

with the minimum NLM value (0.448) on one chan-

nel combination and the minimum accuracy (29%) is

associated with with the maximum NLM value (5.9)

on another channel combination. All combinations

of channels that attribute to high accuracy (greater

than 77% in the ﬁgure) and low NLM metric λ (lesser

than 6.0) are displayed as blue circles in Fig. 2(a).

Of these 316 such different combinations out of to-

tal 1024 combinations, a few channels heavily con-

tributed, while other channels were not involved in

high accuracy. This observation is presented in Fig.

2(b). It is observed that channels 15 and 19 were least

active, while channels 23 and 24 contributed the most

for high accuracy. Table 1 summarises the NLM met-

ric (λ) values along with the selected channels to pro-

duce maximum accuracy for both the datasets. The

accuracy (Acc) and Kappa coefﬁcient (K) where com-

puted using equations (10) and (11).

The proposed algorithm achieved a mean kappa

values of 0.65 for all subjects on less than 10 chan-

nels. This value was cross-validated on 10 indepen-

dent runs. The comparative analysis with previous

studies of BCI competition IV dataset IIa shows that

the proposed algorithm achieves better results com-

pared to other studies in literature as in Table 2. The

proposed channel selection algorithm (NLMCS) com-

putation time is 13.107 sec, training computation time

through SVM is 0.599 sec and total testing computa-

tion time for 288 samples is 2.867 sec. Per sample,

the testing time comes out 0.0097 sec.

Channel Selection for Motor Imagery Task Classiﬁcation using Non-linear Separability Measurement

177

5 CONCLUSIONS

In this paper, a computational geometry-based algo-

rithm is proposed for selecting the optimal number

of channels which is based on non-linear separabil-

ity measurement. The proposed algorithm is evalu-

ated on both BCI competition IV datasets IIa and IIb.

In this work, an effect of low NLM metric (λ) value

on high accuracy has been investigated by using the

EEG channels that are used to compute such λ. Al-

though, the most effective channels out of the 10 ini-

tial channels have been selected, yet the potential of

all 25 channels has not been explored. Based on pro-

posed NLMCS() algorithm, we are motivated to apply

clustering to compute the λ value for clusters between

two classes. In future, the work will be extended for

all channels and a subject independent model will be

built on features taken from the selected channels de-

sign for use in self evolving neural network for EEG

classiﬁcation for improved accuracy.

REFERENCES

Abdulkader, S. N., Atia, A., and Mostafa, M.-S. M. (2015).

Brain computer interfacing: Applications and chal-

lenges. Egyptian Informatics Journal, 16(2):213–230.

Alam, M. N., Ibrahimy, M. I., and Motakabber, S. (2021).

Feature extraction of eeg signal by power spectral den-

sity for motor imagery based bci. In 2021 8th Interna-

tional Conference on Computer and Communication

Engineering (ICCCE), pages 234–237. IEEE.

Baig, M. Z., Aslam, N., and Shum, H. P. (2020). Filtering

techniques for channel selection in motor imagery eeg

applications: a survey. Artiﬁcial intelligence review,

53(2):1207–1232.

Das, A. and Suresh, S. (2015). An effect-size based chan-

nel selection algorithm for mental task classiﬁcation

in brain computer interface. In 2015 IEEE Interna-

tional Conference on Systems, Man, and Cybernetics,

pages 3140–3145. IEEE.

Davoudi, A., Ghidary, S. S., and Sadatnejad, K. (2017). Di-

mensionality reduction based on distance preservation

to local mean for symmetric positive deﬁnite matrices

and its application in brain–computer interfaces. Jour-

nal of neural engineering, 14(3):036019.

Elizondo, D. (2006). The linear separability problem: Some

testing methods. IEEE Transactions on neural net-

works, 17(2):330–344.

Elizondo, D. A., Birkenhead, R., Gamez, M., Garcia, N.,

and Alfaro, E. (2012). Linear separability and classiﬁ-

cation complexity. Expert Systems with Applications,

39(9):7796–7807.

Gaur, P., Pachori, R. B., Wang, H., and Prasad, G. (2018).

A multi-class eeg-based bci classiﬁcation using multi-

variate empirical mode decomposition based ﬁltering

and riemannian geometry. Expert Systems with Appli-

cations, 95:201–211.

Gonzalez, A., Nambu, I., Hokari, H., Iwahashi, M., and

Wada, Y. (2013). Towards the classiﬁcation of single-

trial event-related potentials using adapted wavelets

and particle swarm optimization. In 2013 IEEE In-

ternational Conference on Systems, Man, and Cyber-

netics, pages 3089–3094. IEEE.

Kirar, J. S. and Agrawal, R. (2017). Composite kernel sup-

port vector machine based performance enhancement

of brain computer interface in conjunction with spa-

tial ﬁlter. Biomedical Signal Processing and Control,

33:151–160.

Miao, M., Zeng, H., Wang, A., Zhao, C., and Liu, F.

(2017). Discriminative spatial-frequency-temporal

feature extraction and classiﬁcation of motor im-

agery eeg: An sparse regression and weighted na

ıve

bayesian classiﬁer-based approach. Journal of neuro-

science methods, 278:13–24.

Nicolas-Alonso, L. F. and Gomez-Gil, J. (2012). Brain

computer interfaces, a review. sensors, 12(2):1211–

1279.

Qiu, Z., Jin, J., Lam, H.-K., Zhang, Y., Wang, X., and Ci-

chocki, A. (2016). Improved sffs method for channel

selection in motor imagery based bci. Neurocomput-

ing, 207:519–527.

Shan, H., Xu, H., Zhu, S., and He, B. (2015). A novel

channel selection method for optimal classiﬁcation in

different motor imagery bci paradigms. Biomedical

engineering online, 14(1):1–18.

Shenoy, H. V. and Vinod, A. P. (2014). An iterative opti-

mization technique for robust channel selection in mo-

tor imagery based brain computer interface. In 2014

IEEE International Conference on Systems, Man, and

Cybernetics (SMC), pages 1858–1863. IEEE.

Xie, X., Yu, Z. L., Lu, H., Gu, Z., and Li, Y. (2016).

Motor imagery classiﬁcation based on bilinear sub-

manifold learning of symmetric positive-deﬁnite ma-

trices. IEEE Transactions on Neural Systems and Re-

habilitation Engineering, 25(6):504–516.

Zhang, C. and Eskandarian, A. (2020). A computation-

ally efﬁcient multiclass time-frequency common spa-

tial pattern analysis on eeg motor imagery. In 2020

42nd Annual International Conference of the IEEE

Engineering in Medicine & Biology Society (EMBC),

pages 514–518. IEEE.

Zhang, J., Chen, M., Hu, S., Cao, Y., and Kozma, R. (2016).

Pnn for eeg-based emotion recognition. In 2016 IEEE

international conference on Systems, Man, and Cy-

bernetics (SMC), pages 002319–002323. IEEE.

Zhang, R., Zong, Q., Dou, L., and Zhao, X. (2019). A novel

hybrid deep learning scheme for four-class motor im-

agery classiﬁcation. Journal of neural engineering,

16(6):066004.

Zhou, B., Wu, X., Ruan, J., Zhao, L., and Zhang, L. (2019).

How many channels are suitable for independent com-

ponent analysis in motor imagery brain-computer in-

terface. Biomedical Signal Processing and Control,

50:103–120.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

178