Improving EEG-Based Motor Imagery Classification for Real ...downloads.hindawi.com/journals/cin/2017/9817305.pdf · Improving EEG-Based Motor Imagery Classification for Real-Time

Research ArticleImproving EEG-Based Motor Imagery Classification forReal-Time Applications Using the QSA Method

Patricia Batres-Mendoza1 Mario A Ibarra-Manzano23

Erick I Guerra-Hernandez1 Dora L Almanza-Ojeda23 Carlos R Montoro-Sanjose34

Rene J Romero-Troncoso35 and Horacio Rostro-Gonzalez136

1Laboratorio de Sistemas Bioinspirados Departamento de Ingenierıa Electronica DICISUniversidad de Guanajuato Carr Salamanca-Valle de Santiago Km 35 + 18 Km 36885 Salamanca GTO Mexico2Laboratorio de Procesamiento Digital de Senales Departamento de Ingenierıa Electronica DICISUniversidad de Guanajuato Carr Salamanca-Valle de Santiago Km 35 + 18 Km 36885 Salamanca GTO Mexico3Cuerpo Academico de Telematica DICIS Universidad de Guanajuato Carr Salamanca-Valle de Santiago Km 35 + 18 Km36885 Salamanca GTO Mexico4Departamento de Arte y Empresa DICIS Universidad de Guanajuato Carr Salamanca-Valle de Santiago Km 35 + 18 Km36885 Salamanca GTO Mexico5Departamento de Ingenierıa Electronica DICIS Universidad de Guanajuato Carr Salamanca-Valle de Santiago Km 35 + 18 Km36885 Salamanca GTO Mexico6Neuroscientific SystemTheory Department of Electrical and Computer EngineeringTechnical University of Munich Munich Germany

Correspondence should be addressed to Horacio Rostro-Gonzalez hrostrogugtomx

Received 14 July 2017 Revised 16 October 2017 Accepted 9 November 2017 Published 3 December 2017

Academic Editor Pedro Antonio Gutierrez

Copyright copy 2017 Patricia Batres-Mendoza et alThis is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in anymedium provided the originalwork is properly cited

We present an improvement to the quaternion-based signal analysis (QSA) technique to extract electroencephalography (EEG)signal features with a view to developing real-time applications particularly in motor imagery (IM) cognitive processes Theproposed methodology (iQSA improved QSA) extracts features such as the average variance homogeneity and contrast of EEGsignals related to motor imagery in a more efficient manner (ie by reducing the number of samples needed to classify the signaland improving the classification percentage) compared to the original QSA technique Specifically we can sample the signal invariable time periods (from 05 s to 3 s in half-a-second intervals) to determine the relationship between the number of samplesand their effectiveness in classifying signals In addition to strengthen the classification process a number of boosting-technique-based decision trees were implemented The results show an 8230 accuracy rate for 05 s samples and 7316 for 3 s samples Thisis a significant improvement compared to the original QSA technique that offered results from 3331 to 4082 without samplingwindow and from 3344 to 4107with sampling window respectivelyWe can thus conclude that iQSA is better suited to developreal-time applications

1 Introduction

In the last few years interest in inferring information fromthe human brain stemming from cognitive thoughts bymeans of electroencephalography (EEG) has expanded tovarious disciplines such as neuroscience robotics com-putational science physics and mathematics Research inthese areas tends to revolve around the development of new

communication and control technologies based on brain-computer interface (BCI) devices to support people withsevere neuromuscular conditions in ways that can enablethem to express their wishes or use devices as neuropros-thetics [1] wheelchairs [2 3] control a cursor on a computerscreen [4] or even a robot [5 6] Wolpaw et al [7] arguethat BCIs ldquogive their users communication and controlchannels that do not depend on the brainrsquos normal output

HindawiComputational Intelligence and NeuroscienceVolume 2017 Article ID 9817305 16 pageshttpsdoiorg10115520179817305

2 Computational Intelligence and Neuroscience

channels of peripheral nerves and musclesrdquo In other wordsBCI devices establish a communication channel betweenthe individual and a component (electromechanical devicesrobots software applications etc) to control it by meansof brain activity generated by the user to carry out someintended action [8] To that effect the user must call uponthose actions by means of a brain strategy known as motorimagery

Motor imagery (MI) is a conscious process defined as amental simulation of a particular movement [9] The motorimagery is endowed with the same functional relationship tothe imagined or represented movement and the same causalrole in the generation of the movement in question [10] Inother words MI is related to the intention and preparationof movements whereby the subject imagines carrying out aparticular action without making any real movements Thishas led to studies using motor imagery to decipher processesthat precede the execution of an action For instance Bai et al[11] claim that mental practice using motor imagery of limbmovementmay facilitatemotor recovery in personswho haveexperienced cerebrovascular injuries In additionMcFarlandet al [12] conducted a comparative analysis of EEG topogra-phies associated with actual hand movements and imaginedhand movements concluding that motor imagery plays animportant role in EEG-based communication and suggestingthatmu and beta rhythmsmight provide independent controlsignals

Similarly there have been studies focusing onMI supportand BCI systems proposing algorithms for feature extrac-tion and classification For instance Pfurtscheller et al [13]studied the reactivity of mu rhythms associated with theimagination of hand foot and tongue movements with 60EEG electrodes in nine able-bodied subjects (with a 6616performance rate) In turn Aghaei et al [14] argue for theuse of the separable common spatiospectral patterns (SCSSP)method to extract discriminant spatiospectral EEG featuresand a Laplacian filter of data set V of BCI competition III withthe following mental imagery tasks left-hand movementright-hand movement and generation of words beginningwith a random letter involving 3 subjects As for Gao etal [15] they conducted EEG signal analysis during left-hand movements right-hand movements and resting with10 subjects using the Kolmogorov complexity to extract thefeatures and an AdaBoost multiclass classifier achieving a795 accuracy rate In a separate study Schlogl et al [16]conducted a comparative study involving four classifiers todetermine the global separability of data in relation to fourdifferent MI tasks with 5 subjects modelling the EEG signalby means of an adaptive autoregressive (AAR) process whoseparameters were extracted through Kalman filtering Tradet al [17] used empirical mode decomposition (EMD) andband power (BP) to extract EEG signals and classify MIin experiments involving 10 subjects aged 22ndash35 as theyimagined left-hand and right-hand movements Choi andCichocki [18] implemented a MI-based algorithm to controla wheelchair using spatial filters to extract the features bymeans of a common spatial pattern (CSP) method andthe linear SVMs to classify feature vectors Three healthymen participated in the experiments where they had to

imagine clenching the right hand squeezing the left handand walking In [19] results of tests conducted with a 74accuracy rate to control a robot indoors on the basis ofthree mental states are presented In particular there areseveral studies focusing on the development of processingtechniques feature extraction and classification to improvethe BCI systems In Tables 1 and 2 we present the most com-monly used algorithms for these tasks from Lottersquo et al works[20] However we only extracted information related to ourwork that is motor imagery activities (imagined movementof the left hand imagined movement of the right handimagined movement of the foot imagined movement ofthe tongue relaxation and mental calculation) These worksuse different algorithms in the classification stages such asSVM KNN LDAMLP HMMGaussian classifier and Bayesquadratic including combinations of these which result inmost studies in acceptable performance rates However mostexperiments are performed on a small number of subjectswhich returns a low number of trials per session In thefeature extraction stagemost techniques used to analyze EEGsignals extract information within the frequency or time-frequency domain which may lead to information loss wheninformation is transformed In addition they require noise-elimination filters and frequency band localization to identifythe patterns of motor imagery

On the other hand in [21] the design of a motorimagery experiment was reported based on three mentalprocesses arrow moving to the left arrow moving to theright and waiting time lasting 5 seconds per image Thisstudy analyzed EEG signals with the implementation of thequaternion-based signal analysis (QSA) The quaternion-based signal analysis (QSA) method is a technique that usesEEG signals within the time domain because it is based onquaternion algebra The use of quaternion algebra with theQSA technique makes it possible to describe signals withinthe time domain by means of rotations and orientationsof 3D objects and represent multichannel EEG signals asa single entity preventing data ambiguity and producing amore accurate representation doing fewer calculations thanare needed with other techniques The offline analysis wasconducted in the feature extraction phase considering thetotal number of samples in each class which produced an8492 accuracy rate However while the analysis of offlinesignals is convenient and efficient offline analysis results maynot generalize their performance to online applications Inthe case of the QSAmethod the online analysis obtained wasjust 3331 considering window sizes of 05 seconds

Thus this paper presents an improvement to the QSAmethod that we shall call improved quaternion-based signalanalysis (iQSA) for use in the feature extraction and clas-sification phases whose contribution consists in providinga technique for use in real-time applications focusing onanalyzing EEG signals online reducing the sample sizesneeded to a tenth of the ones required by QSA resultingin a faster response and fewer delays to improve executiontimes in real-time actions The experiment involved usingan Emotiv-Epoc device to acquire the brain signals from themotor and visual regions of the cerebral cortex Similarlyduring the training and validation phases the EEG signal

Computational Intelligence and Neuroscience 3

Tabl

e1

Accu

racy

ofcla

ssifi

erin

motor

imag

eryba

sed

BCI

multic

lassTh

ecla

sses

are(T

1)left

imag

ined

hand

mov

emen

ts(T

2)rig

htim

agin

edha

ndm

ovem

ents

(T3)

imag

ined

foot

mov

emen

ts(T

4)im

agin

edtong

uem

ovem

ents

(T5)

relax(b

aselin

e)(

T6)w

ordge

neratio

nsta

rtin

gwith

aletters

pecific

and

(T7)

men

talc

alcu

lation

Datas

etAc

tivity

Subjec

tsTr

ials

Filte

rFe

atur

eCl

assifi

erAc

cura

cyRe

ferenc

es

BCIc

ompe

tition

III

T1T

2T3

T4

524

0to

360

Yes

AAR

Line

arSV

M63

[16]

KNN

4174

LD

A54

46

Mah

alan

obis

distan

ce53

50

T1T

2T6

3Ye

sNeu

ro-fu

zzyalgo

rithm

S-dF

asArt

8904

[23]

T1T

2T3

T4

324

0Ye

sCS

P

CSP+

SVM

79

[24]

CSP+

SVM

+KN

N+

LDA

69

PCA

+IC

A+

SVM

63

CBN

+SV

M91

T1T

2T6

312

Yes

SCSS

PLD

A

FBCS

P-NBP

W524

2plusmn694

[14]

FBCS

P-Li

n597

7plusmn997

SCSS

P-NBP

W532

1plusmn694

SCSS

P-Li

n593

3plusmn871

T1T

2T3

528

0Ye

sCS

PER

D-E

RSBS

SFOS

VM

7546

[25]

BCIc

ompe

tition

IV

T1T

2T3

T4

928

8Ye

sCS

P

CSP+

SVM

31

[24]

LDA

+SV

M30

CS

P+

SVM

29

CBN

+SV

M66

T1

T2

T3T

49

288

Yes

SCSS

PLD

ALD

A85

[14

]T1

T2

T3T

49

288

Yes

CSP

ERD-E

RSBS

SFOS

VM

8026

[25]

Oth

erda

tase

ts

T1T

2T3

T4

924

0Ye

sAAR

Kalm

anFilte

rM

DA

mdash[13]

T1T

2T5

912

0Ye

sSM

Rtw

o-dim

ensio

nallin

ear

classifi

er85

[2

6]

T1T

2T3

T4

828

8Ye

sIC

AC

SPFD

A33

ndash8

4[2

7]T1

T2

T33

480

Yes

(Sam

pEn)

SVM

(RBF

Kern

el)

6993

[28]

T1T

2T5

33Ye

sAAR

PCA

[12]

Epoc

T1T

2T3

310

0Ye

sSV

M66

16

[29]

T1T

25

Yes

ICAC

WT

Sim

plel

ogist

icM

eta

MLP

8040

[30]

T1T

2T4

175

Yes

Neu

raln

etwor

ks(P

SO)

91

[31]

T1T

2T3

T4

T73

40Ye

sBP

HM

M77

50

[32]

T1T

2T3

330

Yes

BPLD

A95

[33]


Table 2 Accuracy of classifier in motor imagery based BCI two classes The classes are (T1) left imagined hand movements and (T2) rightimagined hand movements

Dataset Activity Subjects Trials Filter Feature Classifier Accuracy References

BCIcompetition III

T1 T2 5 280 Yes LS-SVM (RBFkernel) 9572 [34]

T1 T2 4 140 Yes HMM 7750 + [35]

T1 T2 4 200 Yes CSP SVM (Gaussiankernel) 7750 [36]

T1 T2 7 100 Yes CSPEMDPCA KNN 858 [37]

T1 T2 4 100 Yes SVM 81 [29]

Other data sets


T1 T2 1 140 Yes CSPERD-ERS BSSFO-SVM 9757 [25]

T1 T2 80 Yes PLSRegression

Based on thedecoding principle 64 [1]

T1 T2 109 90 Yes CSP SUTCCSP 90 [38]

T1 T2 4 480 YesCSPERD-

ERSFFT

LDA SVM BPNN 84+ [39]

T1 T2 3 Yes BSS CSP SVM 92 [40]

Epoc

T1 T2 5 120 Yes CSPEMDMIDKRA

PSD Hjort CWTDWT 9779 [41]

T1 T2 8 100 Yes ERS ERD LDA 7037 [42]T1 T2 2 140 Yes CSP Naive Bayes 79 [43]

T1 T2 15 40 Yes Wavelet PSDEMD KNN 9180 [44]

database was strengthened by adding a greater number ofsubjects and by combining decision trees in the classificationphase based on use of the boosting technique

2 Materials and Methods

21 Quaternions Quaternions were proposed in 1843 byHamilton [22] as a set of four constituents (a real and threeimaginary components) as follows 119902 = 119908 + 119894119909 + 119895119910 + 119896119911where119908 119909 119910 119911 isin R and 119894 119895 119896 are symbols of three imaginaryquantities known as imaginary unitsThese units follow theserules

1198942 = 1198952 = 1198962 = 119894119895119896 = minus1119894119895 = 119896119895119896 = 119894119896119894 = 119895119895119894 = minus119896119896119895 = minus119894119894119896 = minus119895

(1)

A quaternion can be described as

119902 = (119904 + a) a = (119909 119910 119911) (2)

where 119904 and a are known as the quaternionrsquos scalar and vectorrespectively When 119904 = 0 119902 is known as pure quaternion

Based on the expanded Eulerrsquos formula the rotation forquaternion around the axis 119899 = [119899119909 119899119910 119899119911] by angle theta isdefined as follows (see [21 45] for further details)

A rotation of angle 120579 around a unit vector a = [119886119909 119886119910 119886119909]is defined as follows

119902 = cos(1205792) + (ax sdot i + ay sdot j + az sdot k) sin(1205792) (3)

Furthermore the operation to be performed on a vectorr to produce a rotated vector r1015840 is

r1015840 = 119902r119902minus1 = (cos(1205792) + a sin(1205792))sdot r(cos(1205792) minus a sin(1205792))

(4)

Equation (4) is a useful representation that makes therotation of a vector easier We can see that r is the originalvector r1015840 is the rotated quaternion and 119902 is the quaternionthat defines the rotation

22 iQSAAlgorithm The iQSA is amethod that improves theperformance and precision of the QSA method which is atechnique to analyze EEG signals for extracting features based


Table 3 Statistical features extracted using quaternions

Statistical features Equation

Mean (120583) = sum( 119902mod)119873119904Variance (1205902) = (sum(119902mod)2 minus 120583)2 + sum(119902mod)22119873119904Contrast (con) = sum(119902mod)2119873119904Homogeneity (119867) = sum 1

1 + (119902mod)2

on rotations and orientations bymeans of quaternion algebraWith iQSA we can conduct real-time signal analysis as thesignals are being acquired to difference of QSA method whoperforms an offline analysis

The iQSA approach consists of three modules quater-nion classification and learning which are described asfollows

(1) Quaternion module in this module a features matrix(119872) is defined from the description of the quaternion119902 and a vector 119877 where each of them correspondsto an array of 4 and 3 EEG channels respectivelyThis module is divided into three steps which aredescribed as follows

(a) Sampling window here we define the samplesize (119899119904) to be analyzed and a displacement ofthe window in the signal (119905 disp) That is theiQSA method performs the sampling by meansof superposing producing a greater number ofsamples to reinforce the learning stage of thealgorithm

(b) Calculate rotation and module then a rotatedvector 119902rot is calculated using the quaternion119902 and vector 119877 where 119902 is a 119898-by-4 matrixcontaining 119898 quaternions and 119877 is an 119898-by-3matrix containing 119898 quaternions displaced onthe basis of a 119889119905 value Later the modulus isapplied to the quaternion 119902rot resulting in thevector 119902mod

(c) Building an array of features finally the array119902mod is used to form a matrix with119872119894119895 featureswhere 119894 corresponds to the analyzed segmentand 119895 is one of the 4 features to be analyzedusing the equations included in Table 3 thatis mean (120583) variance (1205902) contrast (con) andhomogeneity (119867)

Equation (5) shows matrix 119872 with its features vec-tor In this matrix rows correspond to samples andcolumns to features

119872 =[[[[[[[[

1205831 12059021 con1 11986711205832 12059022 con2 1198672 120583119894 1205902119894 con119894 119867119894

]]]]]]]] (5)

(2) Classification module the aim of this module is tocreate a combination of models to predict the valueof a class according to its characteristics to do this weuse the boosting method adapted to the QSA modelTo be more specific we combine ten decision treesand a weight is obtained for each of them which willbe used during the learning to obtain the predictionby a majority rule Here we take 70 of samples from119872 of which the 80 are used for training and the restfor validation as follows

(a) Training the subset of training data (80 sam-ples) is assessed using the models of decisiontrees to obtain amatrixwith accurately classifiedsamples (119866) and a matrix with inaccuratelyclassified samples (119861) The learning of each treeis done by manipulating the training data setand partitioning the initial set in several subsetsaccording to the classification results obtainedthat is a new subset is formed generated andcreated with the accurately classified samplestwice that number of inaccurately classifiedsamples and the worst-classified samples fromthe original training data set Later when thetrees of classification are created these are usedwith the test data set to determine a weight infunction of the accuracy of their predictions

(b) Validation in this block we obtain a matrixwith the reliability percentages given by thedecision trees The subset of validation data(20 samples) is assessed using the decisiontrees to obtain an array with reliability valuesgiven by the processing of decision trees and theclasses identified in the training process

(3) Learning module in this module the subset oftest data of 119872 (30 samples) is assessed to obtainreliability values matrix (119877) and the prediction bymajority rule (119862) That is the learning process willbe conducted assessing features matrix of test usingthe trees that have been generated in classificationprocess The final predictor comes from a weightedmajority rule of the predictors from various decisiontrees Equation (6) shows the recognition and errorrates obtained in each of the prediction models Inthis matrix RT corresponds to accuracy rate ETcorresponds to error rate and 120572 corresponds toreliability value of each decision tree

119877 =[[[[[[[

RT1 ET1 1205721RT2 ET2 1205722

RT119894 ET119894 120572119894

]]]]]]] (6)

In Algorithm 1 we present the pseudocode for themain elements of the iQSA algorithm towards real-timeapplications


Algorithm 1 iQSA methodRequire quat = 1199021 1199022 1199023 1199024 119889119905 119899119904 119905 disp task(1) 119910(119905) larr segments of signals(2) for each 119899119904 in 119910119894(119905) do119902(119899119904) larr quat(119899119904)119903(119899119904) larr quat(119899119904 minus 119889119905)119902rot(119899119904) larr 119899rot (119902(119899119904) 119903(119899119904))119902mod(119899119904) larr mod(119902rot(119899119904))119872119894119895 larr 119891119895 (119902mod(119899119904)) 119895 = 1 119898119888119894 larr 119888 = (1 2 3 119899) | 119910119894(119905) isin 119888119899119904 = 119899119904 + 119905disp

end for(3) 119872119896119895 larr 119872119894119895 | 119896119894 = 119905(4) 119872119897119895 larr 119872119894119895 | 119897 notin 119896 119897119894 = 1 minus119905(5) [120573BTree] = Classify module(119872119896119895 119888119896)(6) [119877 119862 119881119898] = Learning module(119872119897119895 119888119897 120573119894)(7) 119903119905 = 119862119896 | 119862119896 = 119862119896119862119896(8) 119903V = 119862119897 | 119862119897 = 119862119897119862119897Function Classify module (M c)(1) 119872119886119895 larr 119872119894119895 | 119886119894 = 119904(2) 119872119887119895 larr 119872119894119895 | 119887 notin 119886 119887119894 = 1 minus119905lowast119879119903119886119894119899119894119899119892 119901119903119900119888119890119904119904lowast(3) for 119896 = 1 to 10 do[119866119898119895 119861119899119895BTree] larr training(119872119886119895 119888119886)119872119886119895 larr 119872119886119895 + 119861119899119895 minus 119866119898119895

end forlowast119881119886119897119894119889119886119905119894119900119899 119901119903119900119888119890119904119904lowast(4) for each BTree do[119862119894] larr validation(BTree119872119887119895 119888119887)120573119894 larr 119862119894 | 119862119894 = 119862119894119862119894end for(5) return 120573 119862BTree

Function Learning module (M c 120573 B Tree)(1) for each BTree do[119877119894 119881119898] larr classify(BTree119872119894119895 120573119894 119888119894)end for(2) return 119877 119862119894 119881119898

Algorithm 1 iQSA algorithm

221 iQSA versus QSA Methods In Figure 1 we showthrough block diagrams the main differences between theiQSA and QSA methods It can be observed that QSAmethod considers quaternion and classification modulesOn the other hand iQSA method considers an improvedclassification module and one more for learning

In Table 4 we present some of the main improvementsmade in the iQSA algorithm

The QSA method by its characteristics is an excellenttechnique for the offline processing of EEG signals consid-ering large samples of data and being inefficient when suchdata is smallThis last becomes essential for online processingbecause the operations depend on the interaction in real timebetween the signals produced by the subject and the EEGsignals translated with the aid of the algorithm In this regardthe iQSAmethod considers small samples of data and createsa window with the superposition technique for a better data

Table 4 Main modules in QSA and iQSA

Modules QSA iQSASampling window X ∙Quaternion module ∙ ∙Classification module ∙ ∙Boosting technique X ∙Learning module X ∙Superposition technique X ∙Weight normalization X ∙Majority voting X ∙Processing type Batch Real-time

classification during feature extraction strengthening theclassification phase by means of the boosting technique

In this way the signals produced by the iQSA algorithmaffect the posterior brain signals which in turn will affectthe subsequent outputs of the BCI Figure 2 presents thetime system diagram of iQSA which indicates the timelineof events in the algorithm It follows the process of threeEEG data blocks (1198731 1198732 1198733) obtained from the Emotiv-Epoc device The processing block covers the analysis andprocessing of the data with the iQSA method until obtainingan output (OP) in this case the class to which each119873119894 blockbelongs

It is also possible to observe the start and duration of thenext two sets of data where the start of block1198732 is displacedfrom the progress of block 1198731 by a time represented as 119879dispbetween 119905minus1 and 1199050 therefore while the block 1198731 reachesthe output at 1199052 the process of 1198732 continues executing untilreaching its respective output

23 BCI System A brain-computer interface is a system ofcommunication based on neural activity generated by thebrain A BCI measures the activity of an EEG signal byprocessing it and extracting the relevant features to interactin the environment as required by the user An exampleof this device is the Emotiv-Epoc headset (Figure 3(a)) anoninvasive mobile BCI device with a gyroscopic sensor and14 EEG channels (electrodes) and two reference channels(CMSDRL)with a 128Hz sample frequencyThedistributionof sensors in the headset is based on the international 10ndash20-electrode placement system with two sensors as reference forproper placement on the head with channels labeled as AF3F7 F3 FC5 T7 P7 O1 O2 P8 T8 FC6 F4 F8 and AF4(Figure 3(b))

An advantage of the Emotiv-Epoc device is its ability tohandle missing values very common and problematic whendealing with biomedical data

24 Decision Trees and Boosting Decision trees (DT) [46ndash48] are a widely used and easy-to-implement technique thatoffers high speed and accuracy rates DT are used to analyzedata for prediction purposes In short they work by settingconditions or rules organized in a hierarchical structurewhere the final decision can be determined following condi-tions established from the root to its leaves


iQSA method QSA method

Quaternion module Quaternion module

Samplingwindow

Calculaterotation andmodule


Building an arrayof features(70ndash30)


Training matrix Test matrix

Classication module

Classication module

Learning module

Quaternion Quaternion(q1 q2 q3 q4) (q1 q2 q3 q4)

Training Validation

Boos

ting

Building Treewith trainingmatrix

Evaluate thetest matrix

Building newsubset andcalculate weight

Prediction by amajority rule

Normalizationof recognitionrate

Building a matrixwith the reliabilitypercentages

Building an array offeatures to training andvalidation(80ndash20)

Evaluate the testmatrix with themodels

nalrecognition

nalrecognition

Versus

Training(KNN SVM DT)

Validation(KNN SVM DT)

Evaluate the

matrixvalidation

Figure 1 iQSA and QSA comparison

Recently several alternative techniques have been pre-sented to construct sets of classifiers whose decisions arecombined in order to solve a task and to improve the resultsobtained by the base classifier There exist two populartechniques to build sets bagging [49] and boosting [50]Both methods operate under a base-learning algorithm thatis invoked several times using various training sets Inour case we implemented a new boosting method adaptedto QSA method whereby we trained a number of weakclassifiers iteratively so that each new classifier (weak learner)focuses on finding weak hypothesis (inaccurately classified)In other words boosting calls the weak learner 119908 times thusdetermining at each iteration a random subset of trainingsamples by adding the accurately classified samples twicethat number of inaccurately classified samples and the worst-classified samples from the original training data set to forma new weak learner 119908

As a result inaccurately classified samples of the previousiteration are given an (120572) weight in the next iteration forcingthe classifying algorithm to focus on data that are harder to

classify in order to correct classification errors of the previousiteration Finally the reliability percentage of all the classifiersis added and a hypothesis is obtained by a majority votewhose prediction tends to be the most accurate

25 Channel Selection for iQSA From the 14 electrodes thatthe Emotiv-Epoc device provides we decided to perform ananalysis on the electrodes located in three different regionsof the cerebral cortex looking for those with better perfor-mance to form the quaternion (1) electrodes located on themotor cortex which generates neural impulses that controlmovements (2) those on the posterior parietal cortex wherevisual information is transformed into motor instructions[51] and (3) the ones on the prefrontal cortex which appearas a marker of the anticipations that the body must make toadapt to what is going to happen immediately after [52]

In this regard in Table 5 we present the performance foreach of the sets of channels that were selected and furtheranalyzed by the iQSA algorithm


Block processing duration

Sample block duration



Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1

Figure 2 Timeline of events in the iQSA algorithm Here the iQSA performs the three aforementioned modules and OP represents the classobtained

(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)

Figure 3 BCI system (a) Emotiv-Epoc headset and (b) Emotiv-Epoc electrode arrangement

Comparing the data sets shown in Figure 4 it is observedthat all have a similar behavior in each sample size For a64-sample analysis data sets 2 and 4 obtained 8230 and8233 respectively The channels for data set 2 are relatedto the frontal and motor areas of the brain and the channelsfor data set 4 are related to the parietal andmotor areas Fromthese results anddue to the fact thatwemainly focus onmotorcontrol tasks set 2 (F3 F4 FC5 and FC6) has been chosen

3 Experiment

As seen in earlier sections one of the objectives of thispaper is to analyze EEG signals towards real-time appli-cations Therefore it is necessary to reduce the numberof samples of EEG signals that have been acquired bysubjecting them to a process based on the iQSA method todecode motor imagery activities while keeping or improving


Table 5 Accuracy rate to several data sets of channel blocks

Data sets Samples384 320 256 192 128 64

FC5 FC6 P7 P8 7416 7413 8007 8192 8160 8223F3 F4 FC5 FC6 7316 7488 8023 8139 8165 8230F3 F4 F7 F8 7348 7387 7969 8174 8158 8214F3 F4 P7 P8 7347 7392 8006 8168 8187 8233AF3 AF4 FC5 FC6 7417 7410 8031 8182 8199 8226

384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)

Comparison between data sets

07

075

08

085

09

095

1

Data set 1Data set 2Data set 3

Data set 4Data set 5

Figure 4 Comparison of data sets with different sample sizes Dataset 1 FC5 FC6 P7 P8 data set 2 F3 F4 FC5 FC6 data set 3 F3F4 F7 F8 data set 4 F3 F4 P7 P8 data set 5 AF3 AF4 FC5 FC6

accuracy results achieved with this technique In this waythe experiment was designed to register EEG signals fromseveral individuals and to identify three motor-imagery-related mental states (think left motion think right motionand waiting time)

31 Description of the Experiment The experiment was con-ducted in three sessions Session 1 involved motor imagerywith tagged visual support (119881+119871) that is to say an arrowwiththe word LEFT or RIGHT written on it Session 2 involvedmotor imagery with visual (119881) support only In Session 3 thesubject only received a tactile stimulus (119879) to evoke motorimagery while keeping his or her eyes closed In each sessionthe aim was to identify three mental states (think left motionthink right motion and waiting time) The visual 119881 and 119881119871stimuli were provided using a GUI interface developed inPython 27 to show themovement of a red arrow for 5 secondsfor each of the motor brain actions and a fixed cross at thecenter of the GUI to indicate a rest period for 3 seconds(the third brain action) The tactile stimulus was provided bytouching the leftright shoulder of the individual to indicatethe brain action to be performed Each session lasted for 5minutes and was done on separate days

To start with the participant was asked to sit on acomfortable chair in front of a computer screen As the

FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)

(ink le motion) Time inseconds

121110987654321

Figure 5 Methodology for activities in motor-cognitive experi-ments

participant was given instructions regarding the test anEmotiv-Epoc headset was placed on his or her head makingsure that each of the Emotiv device electrodes was makingproper contact with the scalp (Figure 5) Once the participantwas ready he or she was asked not to make sudden bodymovements that could interfere with the signal acquisitionresults during the experiment

The training paradigm consisted of a sequential repetitionof cue-based trials (Figure 5) Each trial startedwith an emptyblank screen during the time 119905 = 0 to 119905 = 3 s a cross wasdisplayed to indicate to the user that the experiment hadstarted and that it was time to relaxThen at second 3 (119905 = 3 s)an arrow appearing for 5 s pointed either to the left or to theright Each position indicated by the arrow instructed thesubject to imagine left or right movement respectively Thenext trial started at 119905 = 8 s with a cross This process wasrepeated for 5minutes displaying the arrows 32 times and thecross 33 times within each runThus the data set recorded forthe three runs consisted of 96 trials

32 iQSA Method Implementation After data acquisitionthe next stage consists in extracting EEG signal features inorder to find the required classes that is think left motionthink right motion and waiting time To start with the iQSAmethod was implemented to represent the four EEG signalswithin a quaternion and to carry out the feature extractionrelated to the stimuli presented To that effect data set 2 wasprepared to assess their performance when time (and thenumber of samples) was reduced Under these conditions 3seconds (384 samples) 25 seconds (320 samples) 2 seconds(256 samples) 15 seconds (192 samples) 1 second (128


ink le motion (class 1)ink right motion (class 2)Waiting time (class 0)

iQSA method

FeaturesM CT HG VR Class 1

Class 2

Class 0

Sample blockduration

qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)

Figure 6 Graphic description of the iQSA method 119902rot represent the rotation of 119902 with respect to 119903 119902mod is the module of 119902rot class 1 class2 and class 0 represent the motor activities evaluated in this case ldquothink left motionrdquo ldquothink right motionrdquo and ldquowaiting timerdquo respectively

samples) and 05 seconds (64 samples) were consideredValues 0 1 and 2 were used to refer to the three mentalactivities waiting time (0) think left motion (1) and thinkright motion (2)

So a segments matrix was obtained to detect suddenchanges between classes for each data set The segmentsmatrix consisted of samples from 32 trials from left andright classes and samples from 33 trials for the waiting timeclass In addition the quaternion was created with the blocksignals proposed (F3 F4 FC5 and FC6) considering F3 asthe scalar component and F4 FC5 and FC6 as the imaginarycomponents (Figure 6) From the samples to be analyzedseveral segments were created to generate the quaternion 119902and vector 119877 with a displacement 119889119905 = 4 and thus obtain therotation (119902rot) and modulus (119902mod)

Once the module was obtained the mean (120583) contrast(con) homogeneity (119867) and variance (1205902) features werecalculated to generate the matrix119872 and the vector 119888 with therequired classes Later we returned to the current segmentand a displacement of 64 samples for the signal was effectedto obtain the next segment

Later M was used in the processing stage using 70of data from training and 30 from test here each classhas the same sample size and was randomly chosen In theclassification module we generate the matrix 1198721 with the80 from training data of 119872 and 1198722 with the remaining

20 of the training data which will be used for trainingand validation where 10 training trees were created to forcethe algorithm to focus on the inaccurately classified dataHere for each iteration we generate new subsets of sampleswith the double of inaccurately classified samples and feweraccurately classified samples along with a reliability rate foreach tree Later with matrix 1198722 the data were validated toobtain the percentage of reliability of each tree Finally withthe remaining data of 119872 a prediction by majority rule isvoted and a decision is reached based on the reliability rateof each classification tree

4 Results

As said earlier one of the aims of this study was to reduce theassessment time (sample size) without loss of the accuracyrate Therefore a comparison of the iQSA and QSA tech-niques was performed to show its behavior when the numberof samples decreases

41 Comparison between iQSA and QSA Methods To evalu-ate and compare the performance of the iQSA online versusQSA offline algorithms the same sets of data from the 39participants were used considering the different sample sizes(384 320 256 192 128 and 64 samples) as input for thealgorithm


Sample size Error rate

384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)

Error rate for sample size with QSA method

QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)

Figure 7 QSA method behavior (a) Table of behavior of QSA method (b) graphic of error rate for sample size

Table 6 Table of accuracy rate for iQSA and QSA method withdifferent number samples

Sample size QSA (1) QSA (2) iQSA (3)384 4082 4107 7316320 4074 4029 7487256 4356 4396 8023192 4241 4271 8139128 3745 3828 816564 3331 3344 8230

In spite of the good performance rates reported withthe QSA algorithm for offline data analysis the classificationresults were not as good when the number of sampleswas reduced up to 64 samples as shown in Figure 7(a)Graphically as can be seen the error rate increases graduallywhen the sample size is reduced for analysis reaching up to6656 error for a reading of 64 samples In this way it wasnecessary to make several adjustments to the algorithm insuch away as to support an online analysis with small samplessizes without losing information and sacrificing the goodperformance rates provided by the offline QSA algorithm

Thus the data of the 39 subjects were evaluated consider-ing three cases (1)QSAmethodwithoutwindow samples (2)QSA method with window samples and (3) iQSA proposedmethod Comparing the results of Table 6 the precisionpercentages for the iQSA method range from 7316 for 384samples to 8230 for 64 samples unlike the original QSAmethod whose percentage decreases to 3331 for 64 samplesin case 1 and 3344 in case 2 although the sampling windowhas been implemented

Figure 8 shows the behavior of the iQSA technique inblue and QSA technique in red where the mean maximumand minimum of each technique are shown Comparingboth techniques it is observed that the data analyzed withthe original QSA method drastically loses precision as thenumber of samples selected decreases This was not the

Behavior of iQSA and QSA

iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples

Figure 8 Behavior of the iQSA and QSA methods with differentsample sizes

case for data sets using the iQSA technique whose precisiondid not vary too much when the number of samples forclassification was reduced improving even its performance

42 iQSA Results Figure 9 shows the behavior of 39 partici-pants under both techniques (iQSA and QSA) consideringall the sample sizes The values obtained using the iQSAmethod are better than values obtained with QSA methodwhose values are below 50 Numerical results show that theiQSA method provides a higher accuracy over the originalQSA method for tests with a 64-sample size

Given the above results a data analysis was conductedusing 64 samples to identify the motor imagery actionsThe performances obtained with the set of classifiers werecompared using various assessment metrics such as recog-nition rate (RT) and error rate (ET) and sensitivity (119878)

12 Computational Intelligence and NeuroscienceAc

cura

cy (n

orm

aliz

ed

)

iQSA versus QSA accuracy rates per subjects

iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects

384 samples320 samples256 samples


Figure 9 iQSA andQSAmethod performances per sample size andsubjects

and specificity (Sp) The sensitivity metric shows that theclassifier can recognize samples from the relevant class andthe specificity is also known as the real negative rate becauseit measures whether the classifier can recognize samples thatdo not belong to the relevant class

RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)

As Table 7 shows the highest accuracy percentage was8450 and the lowest 6875 In addition the sensitivityaverage for class 0 (1198780) associated with the waiting timemental state was 8253 and the sensitivity average for class1 (1198781) and class 2 (1198782) associated with think left motionand think right motion was 8107 and 8165 respectivelywhich indicates that the classifier had no problem classifyingclasses In turn the specificity rate for class 0 (Sp0) showsthat 8136 of the samples classified as negative were actuallynegative while class 1 (Sp1) performed at 8209 and class 2(Sp2) at 8180

To evaluate the performance of our approach we havecarried out a comparison with other methods such as FDCSP[53] MEMD-SI-BCI [54] and SR-FBCSP [55] using data set2a from BCI IV [56] and the results are shown in Table 8Our method shows a slight improvement (109) comparedto the SR-FBCSP method which presents the best results ofthe three

To analyze our results we performed a significant sta-tistical test making use of the STAC (Statistical Tests forAlgorithms Comparison) web platform [57] Here we chosethe Friedman test with a significance level of 010 to get

Visual (tagged) TactileVisualExperiment

Behavior of subjects with three types of experiment

07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

Figure 10 Graph of behavior of subjects with three types ofexperiments tagged visual visual and tactile stimulation

a ranking of the algorithms and check if the differencesbetween them are statistically significant

Table 9 shows the Friedman test ranking results obtainedwith the 119901 value approach From such table we can observethat our proposal gets the lowest ranking that is iQSA hasthe best results in accuracy among all the algorithms

In order to comparewhether the difference between iQSAand the other methods is significant a Li post hoc procedurewas performed (Table 10) The differences are statisticallysignificant because the 119901 values are below 010

In addition the classification shown in Table 8 is achievedby the iQSA in real time and compared to the other methodsa prefiltering process is not required

In Table 11 we show the time required for each ofthe tasks performed by the iQSA algorithm (1) featuresextraction (2) classification and (3) learningThe EEG signalanalysis in processes 1 and 2 was responsible for obtainingthe quaternion from the EEG signals learning trees andtraining Process 3 evaluates and classifies the signal based onthe generated decision trees as should be done in real-timeanalysis

According to the experiment for offline classification(processes 1 and 2) the processing time required was 03089seconds However the time required to recognize the motorimagery activity generated by the subject was of 00095seconds considering that the learning trees were alreadygenerated and even when the processing phase was done inreal time it would take 03184 seconds to recognize a patternafter the first reading With this our proposed method canbe potentially used in several applications such as controllinga robot manipulating a wheelchair or controlling homeappliances to name a few

As said earlier the experiment consisted of 3 sessionswith each participant (visual tagged visual and tactile)The average accuracy obtained from the 3 experiments isbetween 76 and 78 Figure 10 shows results from the


Table 7 Comparison of performance measures for the decision tree classifier using 64 samples

Subject ER RT 1198780 1198781 1198782 Sp0 Sp1 Sp2(1) 017667 082333 083000 081500 082500 082000 082750 082250(2) 018625 081375 081375 079500 083250 081375 082312 080437(3) 017542 082458 085125 079000 083250 081125 084188 082063(4) 019145 080855 081923 081410 079231 080321 080577 081667(5) 020875 079125 077750 079125 080500 079812 079125 078437(6) 015792 084208 086500 081875 084250 083063 085375 084187(7) 018417 081583 083500 079875 081375 080625 082438 081687(8) 018333 081667 082125 080625 082250 081437 082188 081375(9) 017125 082875 082875 083625 082125 082875 082500 083250(10) 016458 083542 086375 084375 079875 082125 083125 085375(11) 018803 081197 080641 081154 081795 081474 081218 080897(12) 015500 084500 083125 084000 086375 085187 084750 083562(13) 019542 080458 083125 078250 080000 079125 081563 080688(14) 015792 084208 085000 082625 085000 083813 085000 083813(15) 019167 080833 080625 079875 082000 080937 081312 080250(16) 019458 080542 082500 076875 082250 079563 082375 079688(17) 016292 083708 085000 083500 082625 083063 083813 084250(18) 017137 082863 084615 081154 082821 081987 083718 082885(19) 015875 084125 084000 083250 085125 084187 084562 083625(20) 018250 081750 081500 080375 083375 081875 082438 080937(21) 031250 068750 065625 068625 072000 070312 068812 067125(22) 016708 083292 086250 082125 081500 081812 083875 084188(23) 018917 081083 082375 081625 079250 080437 080813 082000(24) 016958 083042 084250 081500 083375 082438 083813 082875(25) 017500 082500 084079 083947 079474 081711 081776 084013(26) 019583 080417 080750 080250 080250 080250 080500 080500(27) 016500 083500 083000 082750 084750 083750 083875 082875(28) 020083 079917 078500 079625 081625 080625 080063 079062(29) 018947 081053 083421 079474 080263 079868 081842 081447(30) 016083 083917 085875 083250 082625 082937 084250 084563(31) 017875 082125 083375 080625 082375 081500 082875 082000(32) 019083 080917 082125 081125 079500 080312 080812 081625(33) 019333 080667 082125 079750 080125 079937 081125 080937(34) 018917 081083 080250 082000 081000 081500 080625 081125(35) 016333 083667 086500 082750 081750 082250 084125 084625(36) 016833 083167 083250 082250 084000 083125 083625 082750(37) 018833 081167 083125 081000 079375 080188 081250 082063(38) 017000 083000 081125 086750 081125 083937 081125 083938(39) 019083 080917 082125 080375 080250 080313 081187 081250Mean 018247 081753 082533 081071 081656 081363 082095 081802STD 002544 002544 003451 002796 002404 002315 002667 00291

visual stimulation session which produced the lowest rateat 7663 followed by tagged visual session which reached7698 and the tactile session 7728

In Figure 10 it is shown that themental activity generatedwith the aid of a tactile stimulus is slightlymore accurate thanvisual and visual tagged stimuli where recognition is moreimprecise and slow with visual stimulus

To summarize the results of tests conducted with 39participants using this newmethod to classify motor imagery

brain signals 20 times with each participant considering70ndash30 of the data have been presented and creatingseveral subsets of 80ndash20 for the classification processThe average performance accuracy rate was 8175 whenusing 10 decision trees in combination with the boostingtechnique at 05-second sampling rates The results showthat this methodology for monitoring representing andclassifying EEG signals can be used for the purposes of havingindividuals control external devices in real time


Table 8 Comparison results

Subject iQSA FDCSP MEMD-SI-BCI SR-FBCSP(1) 08543 09166 09236 08924(2) 08296 06805 05833 05936(3) 08273 09722 09167 09581(4) 08509 07222 06389 07073(5) 07879 07222 05903 07810(6) 08284 07153 06736 06998(7) 08172 08125 06042 08824(8) 08466 09861 09653 09532(9) 08425 09375 06667 09192Average 083164 08294 07292 08207Std 002054 01238 01581 01302

Table 9 Friedman test ranking results

Algorithm RankingiQSA 153333FDCSP 155556SR-FBCSP 165556MEMD-SI-BCI 265556

Table 10 Li post hoc adjusted 119901 values for the test error ranking inTable 9

Comparison Adjusted 119901 valueiQSA versus datasets 000118iQSA versus SR-FBCSP 096717iQSA versus FDCSP 097137iQSA versus MEMD-SI-BCI 070942

Table 11 Execution time of iQSA method

Process Time in secondsiQSA quaternion (1) 02054iQSA classification (2) 01035iQSA learning (3) 00095

5 Conclusions

Feature extraction is one of the most important phasesin systems involving BCI devices In particular featureextraction applied to EEG signals for motor imagery activitydiscrimination has been the focus of several studies in recenttimes This paper presents an improvement to the QSAmethod known as iQSA for EEG signal feature extractionwith a view to using it in real timewithmental tasks involvingmotor imagery With our new iQSA method the raw signalis subsampled and analyzed on the basis of a QSA algorithmto extract features of brain activity within the time domainby means of quaternion algebra The feature vector madeup of mean variance homogeneity and contrast is used inthe classification phase to implement a set of decision-treeclassifiers using the boosting technique The performanceachieved using the iQSA technique ranged from 7316 to8230 accuracy rates with readings taken between 3 seconds

and half a second This new method was compared to theoriginal QSA technique whose accuracy rates ranged from4082 to 3331 without sampling window and from 4107to 3334 with sampling window We can thus conclude thatiQSA is a promising technique with potential to be used inmotor imagery recognition tasks in real-time applications

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This research has been partially supported by the CONACYTproject ldquoNeurociencia Computacional de la teorıa al desar-rollo de sistemas neuromorficosrdquo (no 1961) Horacio Rostro-Gonzalez acknowledges the University of Guanajuato for thesupport provided through a sabbatical year

References

[1] P Ofner and G R Muller-Putz ldquoUsing a noninvasive decodingmethod to classify rhythmicmovement imaginations of the armin two planesrdquo IEEE Transactions on Biomedical Engineeringvol 62 no 3 pp 972ndash981 2015

[2] R Chai S H Ling G P Hunter and H T Nguyen ldquoTowardfewer EEG channels and better feature extractor of non-motorimagery mental tasks classification for a wheelchair thoughtcontrollerrdquo in Proceedings of the 34th Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety (EMBC) pp 5266ndash5269 San Diego CA USA August2012

[3] H S KimMH ChangH J Lee andK S Park ldquoA comparisonof classification performance among the various combinationsof motor imagery tasks for brain-computer interfacerdquo in Pro-ceedings of the 2013 6th International IEEE EMBS Conferenceon Neural Engineering NER 2013 pp 435ndash438 San Diego CAUSA November 2013

[4] J R Wolpaw D J McFarland G W Neat and C A FornerisldquoAn EEG-based brain-computer interface for cursor controlrdquoElectroencephalography and Clinical Neurophysiology vol 78no 3 pp 252ndash259 1991

[5] A Vourvopoulos and F Liarokapis ldquoRobot navigation usingbrain-computer interfacesrdquo in Proceedings of the 11th IEEEInternational Conference on Trust Security and Privacy inComputing and Communications TrustCom-2012 pp 1785ndash1792 Liverpool UK June 2012

[6] R Upadhyay P K Kankar P K Padhy and V K Gupta ldquoRobotmotion control using Brain Computer Interfacerdquo in Proceedingsof the 2013 IEEE International Conference on Control Automa-tion Robotics and Embedded Systems CARE 2013 5 1 pagesJabalpur India December 2013

[7] J R Wolpaw N Birbaumer W J Heetderks et al ldquoBrain-computer interface technology a review of the first inter-national meetingrdquo IEEE Transactions on Neural Systems andRehabilitation Engineering vol 8 no 2 pp 164ndash173 2000

[8] J R Wolpaw N Birbaumer D J McFarland G Pfurtschellerand T M Vaughan ldquoBrain-computer interfaces for communi-cation and controlrdquo Clinical Neurophysiology vol 113 no 6 pp767ndash791 2002


[9] BGraimannAGrendan andG PfurtschellerBrain-ComputerInterfaces Revolutionizing Human-Computer InteractionSpringer-Verlag Berlin Germany 2010

[10] M Jeannerod ldquoMental imagery in the motor contextrdquo Neu-ropsychologia vol 33 no 11 pp 1419ndash1432 1995

[11] O Bai D Huang D Fei and R Kunz ldquoEffect of real-timecortical feedback in motor imagery-based mental practicetrainingrdquo Neurorehabilitation vol 34 pp 355ndash363 2014

[12] D J McFarland L A Miner T M Vaughan and J R WolpawldquoMu and beta rhythm topographies during motor imagery andactual movementsrdquo Brain Topography vol 12 no 3 pp 177ndash1862000

[13] G Pfurtscheller C Brunner A Schlogl and F H Lopes daSilva ldquoMu rhythm (de)synchronization and EEG single-trialclassification of different motor imagery tasksrdquo NeuroImagevol 31 no 1 pp 153ndash159 2006

[14] A S Aghaei M S Mahanta and K N Plataniotis ldquoSeparablecommon spatio-spectral patterns for motor imagery BCI sys-temsrdquo IEEE Transactions on Biomedical Engineering vol 63 no1 pp 15ndash29 2016

[15] L Gao W Cheng J Zhang and J Wang ldquoEEG classificationfor motor imagery and resting state in BCI applications usingmulti-class Adaboost extreme learning machinerdquo Review ofScientific Instruments vol 87 no 8 Article ID 085110 2016

[16] A Schlogl F Lee H Bischof and G Pfurtscheller ldquoCharac-terization of four-class motor imagery EEG data for the BCI-competition 2005rdquo Journal of Neural Engineering vol 2 no 4pp 14ndash22 2005

[17] D Trad T Al-Ani and M Jemni ldquoMotor imagery signal clas-sification for BCI system using empirical mode decompositionand bandpower feature extractionrdquo Broad Research in ArtificialIntelligence and Neuroscience (BRAIN) vol 7 2 pp 5ndash16 2016

[18] K Choi and A Cichocki Control of a Wheelchair by MotorImagery in Real Time vol 5326 Springer-Verlag 2008

[19] J D R Millan F Renkens J Mourino and W GerstnerldquoNoninvasive brain-actuated control of a mobile robot byhuman EEGrdquo IEEE Transactions on Biomedical Engineering vol51 no 6 pp 1026ndash1033 2004

[20] F Lotte M Congedo A Lecuyer F Lamarche and BArnaldi ldquoA review of classification algorithms for EEG-basedbrainndashcomputer interfacesrdquo Journal of Neural Engineering vol4 no 2 pp R1ndashR13 2007

[21] P Batres-Mendoza C R Montoro-Sanjose E I Guerra-Hernandez et al ldquoQuaternion-based signal analysis for motorimagery classification from electroencephalographic signalsrdquoSensors vol 16 no 3 article no 336 2016

[22] W R Hamilton ldquoOn quaternionsrdquo Proceedings of the Royal IrishAcademy vol 3 pp 1ndash16 1847

[23] M Almonacid J Ibarrola and J-M Cano-Izquierdo ldquoVotingStrategy to Enhance Multimodel EEG-Based Classifier SystemsforMotor Imagery BCIrdquo IEEE Systems Journal vol 10 no 3 pp1082ndash1088 2016

[24] L He D HuMWan YWen KM VonDeneen andM ZhouldquoCommon Bayesian Network for Classification of EEG-BasedMulticlass Motor Imagery BCIrdquo IEEE Transactions on SystemsMan and Cybernetics Systems vol 46 no 6 pp 843ndash854 2016

[25] H-I Suk and S-W Lee ldquoA novel bayesian framework fordiscriminative feature extraction in brain-computer interfacesrdquoIEEE Transactions on Pattern Analysis andMachine Intelligencevol 35 no 2 pp 286ndash299 2013

[26] S Bermudez I Badia A Garcıa Morgade H Samaha and PF M J Verschure ldquoUsing a hybrid brain computer interfaceand virtual reality system to monitor and promote corticalreorganization through motor activity and motor imagerytrainingrdquo IEEE Transactions on Neural Systems and Rehabilita-tion Engineering vol 21 no 2 pp 174ndash181 2013

[27] M Naeem C Brunner R Leeb B Graimann and GPfurtscheller ldquoSeperability of four-class motor imagery datausing independent components analysisrdquo Journal of NeuralEngineering vol 3 no 3 pp 208ndash216 2006

[28] L Wang G Xu J Wang S Yang M Guo and W YanldquoMotor imagery BCI research based on sample entropy andSVMrdquo in Proceedings of the 2012 6th International Conferenceon Electromagnetic Field Problems and Applications ICEFrsquo2012pp 1ndash4 June 2012

[29] L Schiatti L Faes J Tessadori G Barresi and L MattosldquoMutual information-based feature selection for low-cost BCIsbased on motor imageryrdquo in Proceedings of the 38th AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society EMBC 2016 pp 2772ndash2775 Orlando FlUSA August 2016

[30] E Abdalsalam M M Z Yusoff N Kamel A Malik andM Meselhy ldquoMental task motor imagery classifications fornoninvasive brain computer interfacerdquo in Proceedings of the2014 5th International Conference on Intelligent and AdvancedSystems ICIAS 2014 pp 1ndash5 Kuala Lumpur Malaysia June2014

[31] N Prakaksita C-Y Kuo and C-H Kuo ldquoDevelopment of amotor imagery based brain-computer interface for humanoidrobot control applicationsrdquo in Proceedings of the IEEE Interna-tional Conference on Industrial Technology ICIT 2016 pp 1607ndash1613 Taipei Taiwan March 2016

[32] B Obermaier C Neuper C Guger and G PfurtschellerldquoInformation transfer rate in a five-classes brain-computerinterfacerdquo IEEE Transactions on Neural Systems and Rehabili-tation Engineering vol 9 no 3 pp 283ndash288 2001

[33] R Scherer G R Muller C Neuper B Graimann and GPfurtscheller ldquoAn asynchronously controlledEEG-based virtualkeyboard Improvement of the spelling raterdquo IEEE Transactionson Biomedical Engineering vol 51 no 6 pp 979ndash984 2004

[34] S Siuly and Y Li ldquoImproving the separability of motor imageryEEG signals using a cross correlation-based least square supportvector machine for brain-computer interfacerdquo IEEE Transac-tions on Neural Systems and Rehabilitation Engineering vol 20no 4 pp 526ndash538 2012

[35] S Solhjoo A M Nasrabadi and M R H GolpayeganildquoClassification of chaotic signals using HMM classifiers EEG-based mental task classificationrdquo in Proceedings of the 13thEuropean Signal Processing Conference EUSIPCO2005 pp 257ndash260 September 2005

[36] C Park D Looney N Ur Rehman A Ahrabian and D PMandic ldquoClassification of motor imagery BCI using multi-variate empirical mode decompositionrdquo IEEE Transactions onNeural Systems and Rehabilitation Engineering vol 21 no 1 pp10ndash22 2013

[37] J Hurtado-Rincon S Rojas-Jaramillo Y Ricardo-CespedesA M Alvarez-Meza and G Castellanos-Dominguez ldquoMotorimagery classification using feature relevance analysis AnEmotiv-based BCI systemrdquo in Proceedings of the 2014 19thSymposium on Image Signal Processing and Artificial VisionSTSIVA 2014 September 2014


[38] C Park C C Cheong-Took and D P Mandic ldquoAugmentedcomplex common spatial patterns for classification of noncir-cular EEG from motor imagery tasksrdquo IEEE Transactions onNeural Systems and Rehabilitation Engineering vol 22 no 1 pp1ndash10 2014

[39] Z Tang S Sun S Zhang Y Chen C Li and S Chen ldquoAbrain-machine interface based on ERDERS for an upper-limbexoskeleton controlrdquo Sensors vol 16 no 12 article no 20502016

[40] K Choi and A Cichocki ldquoControl of a Wheelchair by MotorImagery in Real Timerdquo in Proceedings of the roceedings of the9th International Conference on Intelligent Data Engineering andAutomated Learning vol 5326 pp 330ndash337 Springer-VerlagDaejeon South Korea 2008

[41] L Arias-Mora L Lopez-Rıos Y Cespedes-Villar L FVelasquez-Martinez A M Alvarez-Meza and G Castellanos-Dominguez ldquoKernel-based relevant feature extraction tosupport Motor Imagery classificationrdquo in Proceedings of the20th Symposium on Signal Processing Images and ComputerVision STSIVA 2015 Bogota Colombia September 2015

[42] S Dharmasena K Lalitharathne K Dissanayake A Sampathand A Pasqual ldquoOnline classification of imagined hand move-ment using a consumer grade EEG devicerdquo in Proceedings ofthe 2013 IEEE 8th International Conference on Industrial andInformation Systems ICIIS 2013 pp 537ndash541 Peradeniya SriLanka December 2013

[43] V N Stock and A Balbinot ldquoMovement imagery classificationin EMOTIV cap based system byNaıve Bayesrdquo in Proceedings ofthe 38th Annual International Conference of the IEEE Engineer-ing inMedicine and Biology Society EMBC 2016 pp 4435ndash4438Orlando Fl USA August 2016

[44] DMMann-Castrillon S Restrepo-AgudeloH J Areiza-LaverdeA E Castro-Ospina and L Duque-Munoz Exploratory analy-sis of motor imagery local database for BCI systems

[45] A Janota V Simak D Nemec and J Hrbcek ldquoImproving theprecision and speed of Euler angles computation from low-costrotation sensor datardquo Sensors vol 15 no 3 pp 7016ndash7039 2015

[46] S Kotsiantis ldquoA hybrid decision tree classifierrdquo Journal ofIntelligent amp Fuzzy Systems Applications in Engineering andTechnology vol 26 no 1 pp 327ndash336 2014

[47] R L White ldquoAstronomical applications of oblique decisiontreesrdquo AIP Conference Proceedings vol 1082 pp 37ndash43 2008

[48] A A Elnaggar and J S Noller ldquoApplication of remote-sensingdata and decision-tree analysis to mapping salt-affected soilsover large areasrdquo Remote Sensing vol 2 no 1 pp 151ndash165 2010

[49] L Breiman ldquoBagging predictorsrdquoMachine Learning vol 24 no2 pp 123ndash140 1996

[50] Y Freund andR E Schapire ldquoExperiments with a new boostingalgorithmrdquo in Proceedings of the 13th International Conferenceon Machine learning (ICML) pp 148ndash156 1996

[51] T Jiralerspong C Liu and J Ishikawa ldquoIdentification of threemental states using a motor imagery based brain machineinterfacerdquo in Proceedings of the 2014 IEEE Symposium on Com-putational Intelligence in Brain Computer Interfaces (CIBCI) pp49ndash56 Orlando FL USA December 2014

[52] J M Fuster ldquoPrefrontal neurons in networks of executivememoryrdquo Brain Research Bulletin vol 52 no 5 pp 331ndash3362000

[53] J Wang Z Feng and N Lu ldquoFeature extraction by commonspatial pattern in frequency domain for motor imagery tasksclassificationrdquo in Proceedings of the 2017 29th Chinese Control

And Decision Conference (CCDC) pp 5883ndash5888 ChongqingChina May 2017

[54] P Gaur R B Pachori H Wang and G Prasad ldquoA multivari-ate empirical mode decomposition based filtering for subjectindependent BCIrdquo in Proceedings of the 27th Irish Signals andSystems Conference ISSC 2016 pp 1ndash7 Londonderry UK June2016

[55] H V Shenoy A P Vinod and C Guan ldquoShrinkage estimatorbased regularization for EEG motor imagery classificationrdquo inProceedings of the 10th International Conference on InformationCommunications and Signal Processing ICICS 2015 SingaporeDecember 2015

[56] B Blankertz G Dornhege M Krauledat K-R Muller andG Curio ldquoThe non-invasive Berlin brain-computer interfacefast acquisition of effective performance in untrained subjectsrdquoNeuroImage vol 37 no 2 pp 539ndash550 2007

[57] I Rodrıguez-Fdez A Canosa M Mucientes and A BugarınldquoSTAC a web platform for the comparison of algorithmsusing statistical testsrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems pp 1ndash8 Istanbul Turkey August2015

Submit your manuscripts athttpswwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



channels of peripheral nerves and musclesrdquo In other wordsBCI devices establish a communication channel betweenthe individual and a component (electromechanical devicesrobots software applications etc) to control it by meansof brain activity generated by the user to carry out someintended action [8] To that effect the user must call uponthose actions by means of a brain strategy known as motorimagery

Motor imagery (MI) is a conscious process defined as amental simulation of a particular movement [9] The motorimagery is endowed with the same functional relationship tothe imagined or represented movement and the same causalrole in the generation of the movement in question [10] Inother words MI is related to the intention and preparationof movements whereby the subject imagines carrying out aparticular action without making any real movements Thishas led to studies using motor imagery to decipher processesthat precede the execution of an action For instance Bai et al[11] claim that mental practice using motor imagery of limbmovementmay facilitatemotor recovery in personswho haveexperienced cerebrovascular injuries In additionMcFarlandet al [12] conducted a comparative analysis of EEG topogra-phies associated with actual hand movements and imaginedhand movements concluding that motor imagery plays animportant role in EEG-based communication and suggestingthatmu and beta rhythmsmight provide independent controlsignals

Similarly there have been studies focusing onMI supportand BCI systems proposing algorithms for feature extrac-tion and classification For instance Pfurtscheller et al [13]studied the reactivity of mu rhythms associated with theimagination of hand foot and tongue movements with 60EEG electrodes in nine able-bodied subjects (with a 6616performance rate) In turn Aghaei et al [14] argue for theuse of the separable common spatiospectral patterns (SCSSP)method to extract discriminant spatiospectral EEG featuresand a Laplacian filter of data set V of BCI competition III withthe following mental imagery tasks left-hand movementright-hand movement and generation of words beginningwith a random letter involving 3 subjects As for Gao etal [15] they conducted EEG signal analysis during left-hand movements right-hand movements and resting with10 subjects using the Kolmogorov complexity to extract thefeatures and an AdaBoost multiclass classifier achieving a795 accuracy rate In a separate study Schlogl et al [16]conducted a comparative study involving four classifiers todetermine the global separability of data in relation to fourdifferent MI tasks with 5 subjects modelling the EEG signalby means of an adaptive autoregressive (AAR) process whoseparameters were extracted through Kalman filtering Tradet al [17] used empirical mode decomposition (EMD) andband power (BP) to extract EEG signals and classify MIin experiments involving 10 subjects aged 22ndash35 as theyimagined left-hand and right-hand movements Choi andCichocki [18] implemented a MI-based algorithm to controla wheelchair using spatial filters to extract the features bymeans of a common spatial pattern (CSP) method andthe linear SVMs to classify feature vectors Three healthymen participated in the experiments where they had to

imagine clenching the right hand squeezing the left handand walking In [19] results of tests conducted with a 74accuracy rate to control a robot indoors on the basis ofthree mental states are presented In particular there areseveral studies focusing on the development of processingtechniques feature extraction and classification to improvethe BCI systems In Tables 1 and 2 we present the most com-monly used algorithms for these tasks from Lottersquo et al works[20] However we only extracted information related to ourwork that is motor imagery activities (imagined movementof the left hand imagined movement of the right handimagined movement of the foot imagined movement ofthe tongue relaxation and mental calculation) These worksuse different algorithms in the classification stages such asSVM KNN LDAMLP HMMGaussian classifier and Bayesquadratic including combinations of these which result inmost studies in acceptable performance rates However mostexperiments are performed on a small number of subjectswhich returns a low number of trials per session In thefeature extraction stagemost techniques used to analyze EEGsignals extract information within the frequency or time-frequency domain which may lead to information loss wheninformation is transformed In addition they require noise-elimination filters and frequency band localization to identifythe patterns of motor imagery

On the other hand in [21] the design of a motorimagery experiment was reported based on three mentalprocesses arrow moving to the left arrow moving to theright and waiting time lasting 5 seconds per image Thisstudy analyzed EEG signals with the implementation of thequaternion-based signal analysis (QSA) The quaternion-based signal analysis (QSA) method is a technique that usesEEG signals within the time domain because it is based onquaternion algebra The use of quaternion algebra with theQSA technique makes it possible to describe signals withinthe time domain by means of rotations and orientationsof 3D objects and represent multichannel EEG signals asa single entity preventing data ambiguity and producing amore accurate representation doing fewer calculations thanare needed with other techniques The offline analysis wasconducted in the feature extraction phase considering thetotal number of samples in each class which produced an8492 accuracy rate However while the analysis of offlinesignals is convenient and efficient offline analysis results maynot generalize their performance to online applications Inthe case of the QSAmethod the online analysis obtained wasjust 3331 considering window sizes of 05 seconds

Thus this paper presents an improvement to the QSAmethod that we shall call improved quaternion-based signalanalysis (iQSA) for use in the feature extraction and clas-sification phases whose contribution consists in providinga technique for use in real-time applications focusing onanalyzing EEG signals online reducing the sample sizesneeded to a tenth of the ones required by QSA resultingin a faster response and fewer delays to improve executiontimes in real-time actions The experiment involved usingan Emotiv-Epoc device to acquire the brain signals from themotor and visual regions of the cerebral cortex Similarlyduring the training and validation phases the EEG signal


Tabl

e1

Accu

racy

ofcla

ssifi

erin

motor

imag

eryba

sed

BCI

multic

lassTh

ecla

sses

are(T

1)left

imag

ined

hand

mov

emen

ts(T

2)rig

htim

agin

edha

ndm

ovem

ents

(T3)

imag

ined

foot

mov

emen

ts(T

4)im

agin

edtong

uem

ovem

ents

(T5)

relax(b

aselin

e)(

T6)w

ordge

neratio

nsta

rtin

gwith

aletters

pecific

and

(T7)

men

talc

alcu

lation

Datas

etAc

tivity

Subjec

tsTr

ials

Filte

rFe

atur

eCl

assifi

erAc

cura

cyRe

ferenc

es

BCIc

ompe

tition

III

T1T

2T3

T4

524

0to

360

Yes

AAR

Line

arSV

M63

[16]

KNN

4174

LD

A54

46

Mah

alan

obis

distan

ce53

50

T1T

2T6

3Ye

sNeu

ro-fu

zzyalgo

rithm

S-dF

asArt

8904

[23]

T1T

2T3

T4

324

0Ye

sCS

P

CSP+

SVM

79

[24]

CSP+

SVM

+KN

N+

LDA

69

PCA

+IC

A+

SVM

63

CBN

+SV

M91

T1T

2T6

312

Yes

SCSS

PLD

A

FBCS

P-NBP

W524

2plusmn694

[14]

FBCS

P-Li

n597

7plusmn997

SCSS

P-NBP

W532

1plusmn694

SCSS

P-Li

n593

3plusmn871

T1T

2T3

528

0Ye

sCS

PER

D-E

RSBS

SFOS

VM

7546

[25]

BCIc

ompe

tition

IV

T1T

2T3

T4

928

8Ye

sCS

P

CSP+

SVM

31

[24]

LDA

+SV

M30

CS

P+

SVM

29

CBN

+SV

M66

T1

T2

T3T

49

288

Yes

SCSS

PLD

ALD

A85

[14

]T1

T2

T3T

49

288

Yes

CSP

ERD-E

RSBS

SFOS

VM

8026

[25]

Oth

erda

tase

ts

T1T

2T3

T4

924

0Ye

sAAR

Kalm

anFilte

rM

DA

mdash[13]

T1T

2T5

912

0Ye

sSM

Rtw

o-dim

ensio

nallin

ear

classifi

er85

[2

6]

T1T

2T3

T4

828

8Ye

sIC

AC

SPFD

A33

ndash8

4[2

7]T1

T2

T33

480

Yes

(Sam

pEn)

SVM

(RBF

Kern

el)

6993

[28]

T1T

2T5

33Ye

sAAR

PCA

[12]

Epoc

T1T

2T3

310

0Ye

sSV

M66

16

[29]

T1T

25

Yes

ICAC

WT

Sim

plel

ogist

icM

eta

MLP

8040

[30]

T1T

2T4

175

Yes

Neu

raln

etwor

ks(P

SO)

91

[31]

T1T

2T3

T4

T73

40Ye

sBP

HM

M77

50

[32]

T1T

2T3

330

Yes

BPLD

A95

[33]




BCIcompetition III


T1 T2 4 140 Yes HMM 7750 + [35]



T1 T2 4 100 Yes SVM 81 [29]

Other data sets







ERSFFT



Epoc









(1)


119902 = (119904 + a) a = (119909 119910 119911) (2)







(4)







1 + (119902mod)2








119872 =[[[[[[[[

1205831 12059021 con1 11986711205832 12059022 con2 1198672 120583119894 1205902119894 con119894 119867119894

]]]]]]]] (5)





119877 =[[[[[[[

RT1 ET1 1205721RT2 ET2 1205722

RT119894 ET119894 120572119894

]]]]]]] (6)






















Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in




Tabl

e1

Accu

racy

ofcla

ssifi

erin

motor

imag

eryba

sed

BCI

multic

lassTh

ecla

sses

are(T

1)left

imag

ined

hand

mov

emen

ts(T

2)rig

htim

agin

edha

ndm

ovem

ents

(T3)

imag

ined

foot

mov

emen

ts(T

4)im

agin

edtong

uem

ovem

ents

(T5)

relax(b

aselin

e)(

T6)w

ordge

neratio

nsta

rtin

gwith

aletters

pecific

and

(T7)

men

talc

alcu

lation

Datas

etAc

tivity

Subjec

tsTr

ials

Filte

rFe

atur

eCl

assifi

erAc

cura

cyRe

ferenc

es

BCIc

ompe

tition

III

T1T

2T3

T4

524

0to

360

Yes

AAR

Line

arSV

M63

[16]

KNN

4174

LD

A54

46

Mah

alan

obis

distan

ce53

50

T1T

2T6

3Ye

sNeu

ro-fu

zzyalgo

rithm

S-dF

asArt

8904

[23]

T1T

2T3

T4

324

0Ye

sCS

P

CSP+

SVM

79

[24]

CSP+

SVM

+KN

N+

LDA

69

PCA

+IC

A+

SVM

63

CBN

+SV

M91

T1T

2T6

312

Yes

SCSS

PLD

A

FBCS

P-NBP

W524

2plusmn694

[14]

FBCS

P-Li

n597

7plusmn997

SCSS

P-NBP

W532

1plusmn694

SCSS

P-Li

n593

3plusmn871

T1T

2T3

528

0Ye

sCS

PER

D-E

RSBS

SFOS

VM

7546

[25]

BCIc

ompe

tition

IV

T1T

2T3

T4

928

8Ye

sCS

P

CSP+

SVM

31

[24]

LDA

+SV

M30

CS

P+

SVM

29

CBN

+SV

M66

T1

T2

T3T

49

288

Yes

SCSS

PLD

ALD

A85

[14

]T1

T2

T3T

49

288

Yes

CSP

ERD-E

RSBS

SFOS

VM

8026

[25]

Oth

erda

tase

ts

T1T

2T3

T4

924

0Ye

sAAR

Kalm

anFilte

rM

DA

mdash[13]

T1T

2T5

912

0Ye

sSM

Rtw

o-dim

ensio

nallin

ear

classifi

er85

[2

6]

T1T

2T3

T4

828

8Ye

sIC

AC

SPFD

A33

ndash8

4[2

7]T1

T2

T33

480

Yes

(Sam

pEn)

SVM

(RBF

Kern

el)

6993

[28]

T1T

2T5

33Ye

sAAR

PCA

[12]

Epoc

T1T

2T3

310

0Ye

sSV

M66

16

[29]

T1T

25

Yes

ICAC

WT

Sim

plel

ogist

icM

eta

MLP

8040

[30]

T1T

2T4

175

Yes

Neu

raln

etwor

ks(P

SO)

91

[31]

T1T

2T3

T4

T73

40Ye

sBP

HM

M77

50

[32]

T1T

2T3

330

Yes

BPLD

A95

[33]




BCIcompetition III


T1 T2 4 140 Yes HMM 7750 + [35]



T1 T2 4 100 Yes SVM 81 [29]

Other data sets







ERSFFT



Epoc









(1)


119902 = (119904 + a) a = (119909 119910 119911) (2)







(4)







1 + (119902mod)2








119872 =[[[[[[[[

1205831 12059021 con1 11986711205832 12059022 con2 1198672 120583119894 1205902119894 con119894 119867119894

]]]]]]]] (5)





119877 =[[[[[[[

RT1 ET1 1205721RT2 ET2 1205722

RT119894 ET119894 120572119894

]]]]]]] (6)






















Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in






BCIcompetition III


T1 T2 4 140 Yes HMM 7750 + [35]



T1 T2 4 100 Yes SVM 81 [29]

Other data sets







ERSFFT



Epoc









(1)


119902 = (119904 + a) a = (119909 119910 119911) (2)







(4)







1 + (119902mod)2








119872 =[[[[[[[[

1205831 12059021 con1 11986711205832 12059022 con2 1198672 120583119894 1205902119894 con119894 119867119894

]]]]]]]] (5)





119877 =[[[[[[[

RT1 ET1 1205721RT2 ET2 1205722

RT119894 ET119894 120572119894

]]]]]]] (6)






















Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in







1 + (119902mod)2








119872 =[[[[[[[[

1205831 12059021 con1 11986711205832 12059022 con2 1198672 120583119894 1205902119894 con119894 119867119894

]]]]]]]] (5)





119877 =[[[[[[[

RT1 ET1 1205721RT2 ET2 1205722

RT119894 ET119894 120572119894

]]]]]]] (6)






















Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in























Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in






Samplingwindow






Classication module

Classication module

Learning module


Training Validation

Boos

ting









nalrecognition

nalrecognition

Versus



Evaluate the

matrixvalidation












Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in








Tdisp

Tdisp

Tdisp

t1 t2

t0

EEG samples

EEG samples

EEG samples

iQSA

iQSA

iQSA

OP

OP

OP

N1

N2

N3

tminus1


(a)

AF3 AF4

F3 F4F7 F8

FC5 FC6

T7 T8

P7 P8

O1 O2

CMS DRL

(b)



3 Experiment






384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in







384 320 256 192 128 64Samples

Accu

racy

(nor

mal

ized

)


07

075

08

085

09

095

1







FC5 FC6F3 F4

Stimulation

Waitingtime Random

arrow

(ink right motion)


121110987654321







iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in





iQSA method


Class 2

Class 0


qr

qmod(q r)

qrot(q r)

q (channel FC56)

q (channel FC5)

q (channel F4)

q (channel F3)

q (channel FC56)

q (channel FC5)

q (channel F4)







4 Results





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in





384 5892

320 5971

256 5604

192 5728

128 6171

64 6656

(a)

Accu

racy

(nor

mal

ized

)


QSA method

05052054056058

06062064066068

07

320 256 192 128 64384Samples

(b)








iQSAQSA

Accu

racy

(nor

mal

ized

)

02

03

04

05

06

07

08

09

1

320 256 192 128 64 384Samples






cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in




cura

cy (n

orm

aliz

ed

)


iQSA method

QSA method

0010203040506070809

1

5 10 15 20 25 30 35 400Subjects





RT = 119888 | 119888 = 119888 119888 (7)

ET = 119888 | 119888 = 119888 119888 (8)

119878119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (9)

Sp119889 = 119888 | 119888 = 119889 119888 = 119888 119888 | 119888 = 119889 (10)






07

072

074

076

078

08

082

Accu

racy

(nor

mal

ized

)

























5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in



















5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in












5 Conclusions





Acknowledgments


References




































































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in






























































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in
































Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in










Advances in

FuzzySystems


Volume 2014












Journal of



Advances in

Multimedia


Biomedical Imaging


Advances in


RoboticsJournal of










Advances in



Documents

Improving EEG-Based Motor Imagery Classification for Real ...downloads.hindawi.com/journals/cin/2017/9817305.pdf · Improving EEG-Based Motor Imagery Classification for Real-Time