Review Article Methods of EEG Signal Features...

Review ArticleMethods of EEG Signal Features Extraction Using LinearAnalysis in Frequency and Time-Frequency Domains

Amjed S Al-Fahoum1 and Ausilah A Al-Fraihat2

1 Biomedical Systems and Informatics Engineering Department Hijjawi Faculty for Engineering TechnologyYarmouk University Irbid 21163 Jordan

2 Biomedical Engineering Department Faculty of Engineering Hashemite University Zarqa 13115 Jordan

Correspondence should be addressed to Amjed S Al-Fahoum amjedalfahoumrcapgmailcom

Received 20 October 2013 Accepted 9 January 2014 Published 13 February 2014

Academic Editors A Grant J A Hinojosa and M S Oliveira

Copyright copy 2014 A S Al-Fahoum and A A Al-Fraihat This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Technically a feature represents a distinguishing property a recognizable measurement and a functional component obtainedfrom a section of a pattern Extracted features are meant to minimize the loss of important information embedded in the signal Inaddition they also simplify the amount of resources needed to describe a huge set of data accurately This is necessary to minimizethe complexity of implementation to reduce the cost of information processing and to cancel the potential need to compressthe information More recently a variety of methods have been widely used to extract the features from EEG signals amongthese methods are time frequency distributions (TFD) fast fourier transform (FFT) eigenvector methods (EM) wavelet transform(WT) and auto regressive method (ARM) and so on In general the analysis of EEG signal has been the subject of several studiesbecause of its ability to yield an objective mode of recording brain stimulation which is widely used in brain-computer interfaceresearches with application in medical diagnosis and rehabilitation engineering The purposes of this paper therefore shall bediscussing some conventional methods of EEG feature extraction methods comparing their performances for specific task andfinally recommending the most suitable method for feature extraction based on performance

1 Introduction

In recent years brain computer interface and intelligentsignal segmentation have attracted a great interest rangingfrommedicine tomilitary objectives [1ndash6] To facilitate brain-computer interface assembly a professionalmethod of featureextraction from EEG signal is desired

The brain electrical activity is represented by the elec-troencephalogram (EEG) signalsMany neurological diseases(ie epilepsy) can be diagnosed by studying the EEG signals[7ndash9]The recoding of the EEG signals is performed by fixingan electrode on the subject scalp using the standardizedelectrode placement scheme (Figure 1) [10ndash12] Howeverthere are many sources of artifacts The signal noise whichcan set in when signal is being captured will adversely affectthe useful feature in the original signal The major sourcesof the artifact are muscular activities blinking of eyes duringsignal acquisition procedure and power line electrical noise

[13] Many methods have been introduced to eliminate theseunwanted signals Each of them has its advantages anddisadvantages Nevertheless there is a common path for EEGsignal processing (Figure 2) The first part is preprocessingwhich includes acquisition of signal removal of artifactssignal averaging thresholding of the output enhancement ofthe resulting signal and finally edge detection The secondstep in the operation is the feature extraction scheme whichis meant to determine a feature vector from a regular vectorA feature is a distinctive or characteristic measurementtransform structural component extracted from a segment ofa pattern [14] Statistical characteristics and syntactic descrip-tions are the two major subdivisions of the conventionalfeature extraction modalities Feature extraction scheme ismeant to choose the features or informationwhich is themostimportant for classification exercise [15ndash17] The final stage issignal classification which can be solved by linear analysisnonlinear analysis adaptive algorithms clustering and fuzzy

Hindawi Publishing CorporationISRN NeuroscienceVolume 2014 Article ID 730218 7 pageshttpdxdoiorg1011552014730218

2 ISRN Neuroscience

Nasion

Anion

L R

T3

F7

C2 C4 T4

F3 Fz F4

F8

C3

P3 P4PzT5 T6

O1 O2

A2 A3

Pp2Pp1

Figure 1 Standardized electrode placement scheme [11]

Raw EEG

Segment detection

Feature extraction

Classification

Output

Figure 2 Stages of EEG signal processing

techniques and neural networks This is done by exploitingthe algorithmic characteristics of the feature vector of the datainput and thus gives rise to a hypothesis [10 15]

This paper presents a short review of mathematicalmethods for extracting features fromEEG signalsThe reviewconsiders five different methods for EEG signal extractingThe adopted approach is such that a full literature reviewis introduced for the five different techniques summarizingtheir strengths and weaknesses

2 Methods

Different articles were used to extract advantages and disad-vantages of selectedmethods by thoroughly reviewing chosenarticles including the main methods for linear analysis ofone-dimensional signals in the frequency or time-frequency

domain Different common methods of interest were com-pared and the general advantages and disadvantages of thesemodalities were discussed

21 Fast Fourier Transform (FFT)Method Thismethod emp-loys mathematical means or tools to EEG data analysisCharacteristics of the acquired EEG signal to be analyzedare computed by power spectral density (PSD) estimation inorder to selectively represent the EEG samples signal How-ever four frequency bands contain the major characteristicwaveforms of EEG spectrum [18]

The PSD is calculated by Fourier transforming the esti-mated autocorrelation sequence which is found by nonpara-metric methods One of these methods is Welchrsquos methodThe data sequence is applied to data windowing producingmodified periodograms [19]The information sequence 119909

119894(119899)

is expressed as

119909119894 (119899) = 119909 (119899 + 119894119863) 119899 = 0 1 2 119872 minus 1

while 119894 = 0 1 2 119871 minus 1

(1)

take 119894119863 to be the point of start of the 119894th sequence Then 119871

of length 2119872 represents data segments that are formed Theresulting output periodograms give

asymp

119875

(119894)

119909119909(119891) =

1

119872119880

1003816100381610038161003816100381610038161003816100381610038161003816

119872minus1

sum

119899=0

119909119894(119899)119908(119899)119890

minus1198952120587119891119899

1003816100381610038161003816100381610038161003816100381610038161003816

2

(2)

Here in the window function 119880 gives normalization factorof the power and is chosen such that

119880 =

1

119872

119872minus1

sum

119899=0

1199082(119899) (3)

where 119908(119899) is the window function The average of thesemodified periodograms gives Welchrsquos power spectrum asfollows

119875119882

119909119909=

1

119871

119871minus1

sum

119894=0

asymp

119875

(119894)

119909119909(119891) (4)

22 Wavelet Transform (WT) Method WT plays an impor-tant role in the recognition and diagnostic field it compressesthe time-varying biomedical signal which comprises manydata points into a small few parameters that represents thesignal [14]

As the EEG signal is nonstationary [7] the most suitableway for feature extraction from the raw data is the use ofthe time-frequency domain methods like wavelet transform(WT) which is a spectral estimation technique in whichany general function can be expressed as an infinite seriesof wavelets [20ndash22] Since WT allows the use of variablesized windows it gives a more flexible way of time-frequencyrepresentation of a signal In order to get a finer low-frequency resolution WT long time windows are used incontrast in order to get high-frequency information shorttime windows are used [13]

ISRN Neuroscience 3

2

g[n]

g[n]

g[n]

h[n]

h[n]

h[n]

X[n]

D2

D1

D3A1

A2

A3middot middot middot

darr 2

darr 2

darr 2

darr 2

darr 2

darr 2

Figure 3 Implementation of decomposition of DWT [14]

Furthermore WT only involves multiscale structure andnot single scale This method is just the continuation ofthe orthodox Fourier transform method [23] Moreover itis meant to resolve issues of nonstationary signals such asEEG [14] In the WT method the original EEG signal isrepresented by secured and simple building blocks known aswavelets The mother wavelet gives rise to these wavelets aspart of derived functions through translation and dilationthat is (shifting) and (compression and stretching) opera-tions along the time axis respectively [24] There are twocategories for the WT the first one is continuous while theother one is discrete [14]

221 Continuous Wavelet Transform (CWT) Method Thiscan be expressed as

CWT (119886 119887) = int

infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (5)

119909(119905) stands for the unprocessed EEG where 119886 stands fordilation and 119887 represents translation factor The 120595

119886119887(119905) den-

otes the complex conjugate and can be calculated by

120595119886119887 (

119905) =

1

radic|119886|

120595(

119905 minus 119887

119886

) (6)

where 120595(119905) means wavelet However its major weakness isthat scaling parameter 119886 and translation parameter 119887 of CWTchange continuously Thus the coefficients of the wavelet forall available scales after calculationwill consume a lot of effortand yield a lot of unused information [14]

222 DiscreteWavelet Transform (DWT) In order to addressthe weakness of the CWT discrete wavelet transform (DWT)has been defined on the base of multiscale feature rep-resentation Every scale under consideration represents aunique thickness of the EEG signal [23] The multiresolutiondecomposition of the raw EEGdata 119909(119899) is shown in Figure 3Each step contains two digital filters 119892(119899) and ℎ(119899) and twodownsamplers by 2The discretemotherwavelet119892(119899) is a highpass in nature while its mirror image is ℎ(119899) is a low-pass innature

As shown in Figure 3 each stage output provides a detailof the signal 119863 and an approximation of the signal 119860 wherethe latest becomes an input for the next step The number oflevels to which the wavelet decomposes is chosen dependingon the component of the EEG data with dominant frequency[14]

The relationship between WTs and filter ℎ that is lowpass can be represented as follows

119867(119911)119867 (119911minus1) + 119867 (minus119911)119867 (minus119911

minus1) = 1 (7)

Here 119867(119911) represents filterrsquos ℎ 119911-transform The high-passfilterrsquos complementary 119911-transform is expressed as

119866 (119911) = 119911119867 (minus119911minus1) (8)

By precisely describing the features of the signal segmentwithin a specified frequency domain and localized time dom-ain properties there are a lot of advantages that overshadowthe high computational and memory requirement of theconventional convolution based implementation of the DWT[14 23]

23 Eigenvectors These methods are employed to calculatesignalsrsquo frequency and power from artifact dominated mea-surements The essence of these methods is the potential ofthe Eigen decomposition to correlate even artifact corruptedsignal There are a few available eigenvector methods amongthem are Pisarenkorsquos method MUSIC method and mini-mum-norm method [25 26]

231 Pisarenkorsquos Method Pisarenkorsquos method is among theavailable eigenvector approaches used to evaluate powerspectral density (PSD) To calculate the PSD the mathemati-cal expression 119860(119891) is employed and given as [27 28]

119860 (119891) =

119898

sum

119896=0

119886119896119890minus1198952120587119891119896

(9)

In the equation above 119886119896stands for coefficients of the defined

equation and 119898 defines eigenfilterrsquos 119860(119891) order [25 26]Pisarenko method uses signal desired equation to estimatethe signalrsquos PSD from eigenvector equivalent to theminimumeigenvalue as follows

119875PISARENKO =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (10)

232 MUSIC Method This method eradicates issues relatedto false zeros by the help of the spectrarsquos average equivalentto artifact subspace of the whole eigenvectors [28] Resultingpower spectral density is therefore obtained as

119875MUSIC (119891) =

1

(1119870)sum119870minus1

119894=0

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (11)

233 Minimum Norm Method This method makes falsezeros in the unit circle to separate them from real zeros to beable to calculate a demanded noise subspace vector 119886 fromeither the noise or signal subspace eigenvectors Howeverwhile the Pisarenko technique form application of only thenoise subspace eigenvector corresponding to the minimumeigenvalue the minimum norm technique picks a linear

4 ISRN Neuroscience

combination of the whole of noise subspace eigenvectors[25 26] This technique is depicted by

119875MIN (119891119870) =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (12)

All the aforementioned eigenvector methods can bestaddress the signal that is composed of many distinctivesinusoids embedded in noise Consequently they are proneto yield false zeros and thus resulting in a relatively poorstatistical accuracy [26]

24 Time-Frequency Distributions These methods requirenoiseless signals to provide good performance Thereforevery restricted preprocessing stage is necessary to get rid ofall sorts of artifacts Being time-frequency methods they dealwith the stationary principle windowing process is thereforerequired in the preprocessing module [29] The definition ofTFD for a signal 119909(119899) was generalized by Cohen as [30]

119875 (119905 119908) =

1

2120587

int

infin

minusinfin

int

infin

minusinfin

119860 (120579 120591)Φ (120579 120591) 119890minus119895120579119905minus119895120596120591

119889120579 119889120591

(13)

where

119860 (120579 120591) =

1

2120587

int

infin

minusinfin

119909(119906 +

120591

2

) 119909lowast(119906 minus

120591

2

) 119890119895120579119906

119889119906 (14)

119860(120579 120591) is popularly known as ambiguity Function andΦ(120579 120591) refers to kernel of the distribution while 119903 and 119908 aretime and frequency dummy variables respectively

Smooth pseudo-Wigner-Ville (SPWV) distribution is avariant method which incorporates smoothing by indepen-dent windows in time and frequency namely 119882

119908(120591) and

119882119905(119905) [29]

SPWV (119905 119908)

= int

infin

minusinfin

119882119908 (

120591) [int

infin

minusinfin

119882119905 (119906 minus 119905) 119909 (119906 +

120591

2

)

times 119909lowast(119906 minus

120591

2

) 119889119906] 119890minus119895120596120591

119889120591

(15)

The feature extraction using this method is based on theenergy frequency and the length of the principal track Eachsegment gives the values 119864

119896 119865119896 and 119871

119896 The EEG signal is

firstly divided into 119896 segments then the construction of athree-dimensional feature vector for each segment will takeplace Energy of each segment 119896 can be calculated as follows

119864119896= int

infin

minusinfin

int

infin

minusinfin

120599119896(119905 119891) 119889119905 119889119891 (16)

where 120599119896(119905 119891) stands for the time-frequency representation

of the segment However to calculate the frequency of eachsegment 119896 we make use of the marginal frequency as follows

119865119896= int

infin

minusinfin

120599119896(119905 119891) 119889119905 (17)

Finally for achieving good results noiseless EEG signalsor a well-denoised signal should be used for TFD [30]

Table 1 Comparison between FFT and AR [8]

Method Frequency resolution Spectral leakageFFT Low HighAR High LowWT High Low

25 Autoregressive Method Autoregressive (AR) methodsestimate the power spectrum density (PSD) of the EEG usinga parametric approach Therefore AR methods do not haveproblem of spectral leakage and thus yield better frequencyresolution unlike nonparametric approach Estimation ofPSD is achieved by calculating the coefficients that is theparameters of the linear system under consideration Twomethods used to estimate AR models are briefly describedbelow [18 19]

251 Yule-WalkerMethod In this method AR parameters orcoefficients are estimated by exploiting the resulting biasedapproximate of the autocorrelation data functionThis is doneby subsequently finding theminimization of the least squaresof the forward prediction error as given below [31]

[

[

[

[

[

119903 (0)119909119909119903 (minus1)119909119909

sdot sdot sdot 119903(minus119901 + 1)119909119909

119903 (1)119909119909119903 (0)119909119909

119903(minus119901 + 2)119909119909

d

119903 (119901 minus 1)

119909119909119903 (119901 minus 2)

119909119909sdot sdot sdot 119903 (0)119909119909

]

]

]

]

]

times

[

[

[

[

[

119886 (1)

119886 (1)

119886 (119901)

]

]

]

]

]

(18)

where 119903119909119909

can be defined by

119903119909119909 (

119898) =

1

119873

119873minus119898minus1

sum

119873=0

119909lowast(119899) 119909 (119899 + 119898) 119898 ge 0 (19)

Calculating the above set of (119901 + 1) linear equations the ARcoefficients can be obtained

119875119861119880

119909119909(119891) =

1205902

119908119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (20)

while 119908119901

gives the approximated lowest mean square errorof the 119901th-order predictor given as follows

1205902

119908119901= 119864119891

119901= 119903119909119909 (

0)

119901

prod

119896=1

[1 minus1003816100381610038161003816119886119896 (

119896)1003816100381610038161003816

2] (21)

252 Burgrsquos Method It is an AR spectral estimation basedon reducing the forward and backward prediction errorsto satisfy Levinson-Durbin recursion [8] Burgrsquos methodestimates the reflection coefficient directly without the needto calculate the autocorrelation functionThismethod has thefollowing strength Burgrsquos method can estimate PSDrsquos datarecords to look exactly like the original data value It canyield intimately packed sinusoids in signals once it containsminimal level of noise

ISRN Neuroscience 5

Table 2 Comparison between performances of EEG methods

Method name Advantages Disadvantages Analysismethod Suitability

Fast fouriertransform

(i) Good tool for stationary signalprocessing(ii) It is more appropriate for narrowbandsignal such as sine wave(iii) It has an enhanced speed overvirtually all other available methods inreal-time applications

(i) Weakness in analyzing nonstationarysignals such as EEG(ii) It does not have good spectralestimation and cannot be employed foranalysis of short EEG signals(iii) FFT cannot reveal the localizedspikes and complexes that are typicalamong epileptic seizures in EEG signals(iv) FFT suffers from large noisesensitivity and it does not have shorterduration data record

Frequencydomain

Narrowbandstationarysignals

Wavelettransform

(i) It has a varying window size beingbroad at low frequencies and narrow athigh frequencies(ii) It is better suited for analysis ofsudden and transient signal changes(iii) Better poised to analyze irregulardata patterns that is impulses existing atdifferent time instances

Needs selecting a proper mother waveletBoth time andfreq domainand linear

Transient andstationarysignal

Eigenvector Provides suitable resolution to evaluatethe sinusoid from the data

Lowest eigenvalue may generate falsezeros when Pisarenkorsquos method isemployed

Frequencydomain

Signal buriedwith noise

Timefrequencydistribution

(i) It gives the feasibility of examininggreat continuous segments of EEG signal(ii) TFD only analyses clean signal forgood results

(i) The time-frequency methods areoriented to deal with the concept ofstationary as a result windowing processis needed in the preprocessing module(ii) It is quite slow (because of thegradient ascent computation)(iii) Extracted features can be dependenton each other

Both time andfrequencydomains

Stationarysignal

Autoregressive

(i) AR limits the loss of spectral problemsand yields improved frequency resolution(ii) Gives good frequency resolution(iii) Spectral analysis based on AR modelis particularly advantageous when shortdata segments are analyzed since thefrequency resolution of an analyticallyderived AR spectrum is infinite and doesnot depend on the length of analyzed data

(i) The model order in AR spectralestimation is difficult to select(ii) AR method will give poor spectralestimation once the estimated model isnot appropriate and modelrsquos orders areincorrectly selected(iii) It is readily susceptible to heavybiases and even large variability

Frequencydomain

Signal withsharp spectral

features

Thedifference betweenmethod of Yule-Walker andBurgrsquosmethod is in the way of calculating the PSD For Burgrsquosmethod the PSD is estimated as follows

119875119861119880

119909119909(119891) =

_119864119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (22)

Parametric methods like autoregressive one reduce thespectral leakage issues and yield better frequency resolutionHowever selecting the proper model order is a very seriousproblem Once the order is too high the output will inducefalse peaks in the spectra If the order is too low the result willproduce smooth spectra [32]

3 Performance of Methods

The general aim of this review is to shed light on EEGsignal feature extraction and to show how fast the methodused for the signal extraction and how reliable it will be theextracted EEG signal features Moreover how these extractedfeatures would express the states of the brain for differentmental tasks and to be able to yield an exact classificationand translation of mental tasks The speed and accuracy ofthe feature extraction stage of EEG signal processing aretherefore very crucial in order not to lose vital informationat a reasonable time So far in the discussed literature waveletmethod is introduced as a solution for unstable signals itincludes the representation by wavelets which are a group offunctions derived from the mother wavelet by dilation andtranslation processes The window with varying size is the

6 ISRN Neuroscience

most significant specification of this method since it ensuresthe suitable time frequency resolution in all frequency range[26] Autoregression analysis suffers from speed and is notalways applicable in real-time analysis while FFT appears tobe the least efficient of the discussed methods because of itsinability to examine nonstationary signals The strength ofAR method can be emphasized by further comparing itsperformance with that of classical FFT as shown in Table 1

It is highly recommend to use ARmethod in conjunctionwith more conservative methods such as periodograms tohelp to choose the correct model order and to avoid gettingfooled by spurious spectral features [32]

Themost important application for eigenvectors is to eva-luate frequencies and powers of signals from noise corruptedsignal the principle of this method is the decomposition ofthe correlation matrix of the noise corruptedThree methodsfor eigenvectors module were discussed Pisarenko multiplesignal classification (MUSIC) and minimum norm [27] Thegood thing about the eigenvector method is that it producesfrequency spectra of high resolution evenwhen the signal-to-noise ratio (SNR) is low However this method may producespurious zeros leading to poor statistical accuracy [26]

The TFD method offers the possibility to analyze rela-tively long continuous segments of EEG data even when thedynamics of the signal are rapidly changing At the same timea good resolution both in time and frequency is necessarymaking this method not preferable to use in many cases [30]

Table 2 shows the summary of advantages and disad-vantages of the above-mentioned methods their accuraciesspeeds and suitability to make it easier to compare theirperformances

4 Conclusion

Five of the well-known methods for frequency domain andtime-frequency domain methods were discussed Acclaimabout the definite priority of methods according to theircapability is very hard The findings indicate that eachmethod has specific advantages and disadvantages whichmake it appropriate for special type of signals Frequencydomain methods may not provide high-quality performancefor some EEG signals In contrast time-frequency methodsfor instance may not provide detailed information on EEGanalysis asmuch as frequency domainmethods It is crucial tomake clear the of the signal to be analyzed in the applicationof the method whenever the performance of analyzingmethod is discussed Considering this the optimummethodfor any application might be different

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] L Mayaud M Congedo A van Laghenhove et al ldquoA compar-ison of recording modalities of P300 Event Related Potentials(ERP) for Brain-Computer Interface (BCI) paradigmrdquo ClinicalNeurophysiology vol 43 no 4 pp 217ndash227 2013

[2] X-Y Wang J Jin Y Zhang et al ldquoBrain control human-computer integration control based on brain-computer inter-face approachrdquo Acta Automatica Sinica vol 39 no 3 pp 208ndash221 2013

[3] G Al-Hudhud ldquoAffective command-based control system inte-grating brain signals in commands control systemsrdquo Computersin Human Behavior vol 30 pp 535ndash541 2014

[4] J L S Blasco E Ianez A Ubeda and J M Azorın ldquoVisualevoked potential-based brain-machine interface applications toassist disabled peoplerdquo Expert Systems with Applications vol 39no 9 pp 7908ndash7918 2012

[5] H Cecotti ldquoSpelling with non-invasive brain-computer inter-facesmdashcurrent and future trendsrdquo Journal of Physiology vol 105no 1ndash3 pp 106ndash114 2011

[6] I S Kotchetkov B Y Hwang G Appelboom C P Kellnerand E S Connolly Jr ldquoBrain-computer interfaces militaryneurosurgical and ethical perspectiverdquo Neurosurgical Focusvol 28 no 5 article E25 2010

[7] E D Ubeyli ldquoStatistics over features EEG signals analysisrdquoComputers in Biology and Medicine vol 39 no 8 pp 733ndash7412009

[8] R Agarwal J Gotman D Flanagan and B Rosenblatt ldquoAuto-matic EEG analysis during long-term monitoring in the ICUrdquoElectroencephalography and Clinical Neurophysiology vol 107no 1 pp 44ndash58 1998

[9] A Meyer-Lindenberg ldquoThe evolution of complexity in humanbrain development an EEG studyrdquo Electroencephalography andClinical Neurophysiology vol 99 no 5 pp 405ndash411 1996

[10] A Subasi and E Ercelebi ldquoClassification of EEG signals usingneural network and logistic regressionrdquo Computer Methods andPrograms in Biomedicine vol 78 no 2 pp 87ndash99 2005

[11] A Subasi ldquoEEG signal classification usingwavelet feature extra-ction and amixture of expertmodelrdquo Expert Systemswith Appli-cations vol 32 no 4 pp 1084ndash1093 2007

[12] H Jasper and L D Proctor Reticular Formation of the Brain1958

[13] M R N Kousarrizi A A Ghanbari M Teshnehlab M Ali-yari and A Gharaviri ldquoFeature extraction and classificationof EEG signals using wavelet transform SVM and artificialneural networks for brain computer interfacesrdquo in Proceedingsof the International Joint Conference on Bioinformatics SystemsBiology and Intelligent Computing (IJCBS rsquo09) pp 352ndash355August 2009

[14] D Cvetkovic E D Ubeyli and I Cosic ldquoWavelet transform fea-ture extraction from human PPG ECG and EEG signal respo-nses to ELF PEMF exposures a pilot studyrdquoDigital Signal Proce-ssing vol 18 no 5 pp 861ndash874 2008

[15] H Kordylewski D Graupe and K Liu ldquoA novel large-memoryneural network as an aid in medical diagnosis applicationsrdquoIEEE Transactions on Information Technology in Biomedicinevol 5 no 3 pp 202ndash209 2001

[16] I Guler and E D Ubeyli ldquoAdaptive neuro-fuzzy inference sys-tem for classification of EEG signals using wavelet coefficientsrdquoJournal of Neuroscience Methods vol 148 no 2 pp 113ndash1212005

[17] E D Ubeyli ldquoWaveletmixture of experts network structurefor EEG signals classificationrdquo Expert Systems with Applicationsvol 34 no 3 pp 1954ndash1962 2008

[18] A Subasi M K Kiymik A Alkan and E Koklukaya ldquoNeuralnetwork classification of EEG signals by using AR with MLEpreprocessing for epileptic seizure detectionrdquoMathematical andComputational Applications vol 10 no 1 pp 57ndash70 2005

ISRN Neuroscience 7

[19] O Faust R U Acharya A R Allen and C M Lin ldquoAnalysisof EEG signals during epileptic and alcoholic states using ARmodeling techniquesrdquo IRBM vol 29 no 1 pp 44ndash52 2008

[20] A Prochazka J Kukal and O Vysata ldquoWavelet transform usefor feature extraction and EEG signal segments classificationrdquoin Proceedings of the 3rd International Symposium on Commu-nications Control and Signal Processing (ISCCSP rsquo08) pp 719ndash722 March 2008

[21] I Daubechies ldquoWavelet transform time-frequency localizationand signal analysisrdquo IEEE Transactions on Information Theoryvol 36 no 5 pp 961ndash1005 1990

[22] S Soltani ldquoOn the use of the wavelet decomposition for timeseries predictionrdquo Neurocomputing vol 48 no 1 pp 267ndash2772002

[23] N Hazarika J Z Chen A C Tsoi and A Sergejew ldquoClassifi-cation of EEG signals using thewavelet transformrdquo Signal Proce-ssing vol 59 no 1 pp 61ndash72 1997

[24] M Unser and A Aldroubi ldquoA review of wavelets in biomedicalapplicationsrdquo Proceedings of the IEEE vol 84 no 4 pp 626ndash638 1996

[25] E D Ubeyli ldquoAnalysis of EEG signals by implementingeigenvector methodsrecurrent neural networksrdquoDigital SignalProcessing vol 19 no 1 pp 134ndash143 2009

[26] E D Ubeyli ldquoAnalysis of EEG signals by combining eigenvectormethods andmulticlass support vectormachinesrdquoComputers inBiology and Medicine vol 38 no 1 pp 14ndash22 2008

[27] E D Ubeyli and I Guler ldquoFeatures extracted by eigenvectormethods for detecting variability of EEG signalsrdquo Pattern Reco-gnition Letters vol 28 no 5 pp 592ndash603 2007

[28] S A AwangM Paulraj and S Yaacob ldquoAnalysis of EEG signalsby eigenvector methodsrdquo in Proceedings of the IEEE EMBSConference on Biomedical Engineering and Sciences (IECBESrsquo12) pp 778ndash783 December 2012

[29] C Guerrero-Mosquera and A N Vazquez ldquoNew approach infeatures extraction for EEG signal detectionrdquo in Proceedings ofthe Annual International Conference of the IEEE Engineering inMedicine and Biology Society (EMBC rsquo09) pp 13ndash16 September2009

[30] L CohenTime-FrequencyAnalysis PrenticeHall Upper SaddleRiver NJ USA 1995

[31] P Stoica and R L Moses Introduction to Spectral AnalysisPrentice hall Upper Saddle River NJ USA 1997

[32] M Polak and A Kostov ldquoFeature extraction in developmentof brain-computer interface a case studyrdquo in Proceedings of the20th Annual International Conference of the IEEE Engineeringin Medicine and Biology Society (EMBC rsquo98) pp 2058ndash2061November 1998

Submit your manuscripts athttpwwwhindawicom

Neurology Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Alzheimerrsquos DiseaseHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


BioMed Research International


Research and TreatmentSchizophrenia

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Neural Plasticity


Parkinsonrsquos Disease


Research and TreatmentAutism

Sleep DisordersHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Neuroscience Journal

Epilepsy Research and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Psychiatry Journal


Computational and Mathematical Methods in Medicine

Depression Research and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Brain ScienceInternational Journal of

StrokeResearch and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Neurodegenerative Diseases


Journal of

Cardiovascular Psychiatry and NeurologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

2 ISRN Neuroscience

Nasion

Anion

L R

T3

F7

C2 C4 T4

F3 Fz F4

F8

C3

P3 P4PzT5 T6

O1 O2

A2 A3

Pp2Pp1

Figure 1 Standardized electrode placement scheme [11]

Raw EEG

Segment detection

Feature extraction

Classification

Output

Figure 2 Stages of EEG signal processing

techniques and neural networks This is done by exploitingthe algorithmic characteristics of the feature vector of the datainput and thus gives rise to a hypothesis [10 15]

This paper presents a short review of mathematicalmethods for extracting features fromEEG signalsThe reviewconsiders five different methods for EEG signal extractingThe adopted approach is such that a full literature reviewis introduced for the five different techniques summarizingtheir strengths and weaknesses

2 Methods

Different articles were used to extract advantages and disad-vantages of selectedmethods by thoroughly reviewing chosenarticles including the main methods for linear analysis ofone-dimensional signals in the frequency or time-frequency

domain Different common methods of interest were com-pared and the general advantages and disadvantages of thesemodalities were discussed

21 Fast Fourier Transform (FFT)Method Thismethod emp-loys mathematical means or tools to EEG data analysisCharacteristics of the acquired EEG signal to be analyzedare computed by power spectral density (PSD) estimation inorder to selectively represent the EEG samples signal How-ever four frequency bands contain the major characteristicwaveforms of EEG spectrum [18]

The PSD is calculated by Fourier transforming the esti-mated autocorrelation sequence which is found by nonpara-metric methods One of these methods is Welchrsquos methodThe data sequence is applied to data windowing producingmodified periodograms [19]The information sequence 119909

119894(119899)

is expressed as

119909119894 (119899) = 119909 (119899 + 119894119863) 119899 = 0 1 2 119872 minus 1

while 119894 = 0 1 2 119871 minus 1

(1)

take 119894119863 to be the point of start of the 119894th sequence Then 119871

of length 2119872 represents data segments that are formed Theresulting output periodograms give

asymp

119875

(119894)

119909119909(119891) =

1

119872119880

1003816100381610038161003816100381610038161003816100381610038161003816

119872minus1

sum

119899=0

119909119894(119899)119908(119899)119890

minus1198952120587119891119899

1003816100381610038161003816100381610038161003816100381610038161003816

2

(2)

Here in the window function 119880 gives normalization factorof the power and is chosen such that

119880 =

1

119872

119872minus1

sum

119899=0

1199082(119899) (3)

where 119908(119899) is the window function The average of thesemodified periodograms gives Welchrsquos power spectrum asfollows

119875119882

119909119909=

1

119871

119871minus1

sum

119894=0

asymp

119875

(119894)

119909119909(119891) (4)

22 Wavelet Transform (WT) Method WT plays an impor-tant role in the recognition and diagnostic field it compressesthe time-varying biomedical signal which comprises manydata points into a small few parameters that represents thesignal [14]

As the EEG signal is nonstationary [7] the most suitableway for feature extraction from the raw data is the use ofthe time-frequency domain methods like wavelet transform(WT) which is a spectral estimation technique in whichany general function can be expressed as an infinite seriesof wavelets [20ndash22] Since WT allows the use of variablesized windows it gives a more flexible way of time-frequencyrepresentation of a signal In order to get a finer low-frequency resolution WT long time windows are used incontrast in order to get high-frequency information shorttime windows are used [13]

ISRN Neuroscience 3

2

g[n]

g[n]

g[n]

h[n]

h[n]

h[n]

X[n]

D2

D1

D3A1

A2


darr 2

darr 2

darr 2

darr 2

darr 2

darr 2




CWT (119886 119887) = int

infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (5)


119886119887(119905) den-


120595119886119887 (

119905) =

1

radic|119886|

120595(

119905 minus 119887

119886

) (6)






minus1) = 1 (7)


119866 (119911) = 119911119867 (minus119911minus1) (8)




119860 (119891) =

119898

sum

119896=0

119886119896119890minus1198952120587119891119896

(9)



119875PISARENKO =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (10)


119875MUSIC (119891) =

1

(1119870)sum119870minus1

119894=0

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (11)


4 ISRN Neuroscience


119875MIN (119891119870) =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (12)



119875 (119905 119908) =

1

2120587

int

infin

minusinfin

int

infin

minusinfin

119860 (120579 120591)Φ (120579 120591) 119890minus119895120579119905minus119895120596120591

119889120579 119889120591

(13)

where

119860 (120579 120591) =

1

2120587

int

infin

minusinfin

119909(119906 +

120591

2


120591

2

) 119890119895120579119906

119889119906 (14)



119908(120591) and

119882119905(119905) [29]

SPWV (119905 119908)

= int

infin

minusinfin

119882119908 (

120591) [int

infin

minusinfin

119882119905 (119906 minus 119905) 119909 (119906 +

120591

2

)


120591

2

) 119889119906] 119890minus119895120596120591

119889120591

(15)


119896 119865119896 and 119871



119864119896= int

infin

minusinfin

int

infin

minusinfin

120599119896(119905 119891) 119889119905 119889119891 (16)



119865119896= int

infin

minusinfin

120599119896(119905 119891) 119889119905 (17)






[

[

[

[

[

119903 (0)119909119909119903 (minus1)119909119909


119903 (1)119909119909119903 (0)119909119909

119903(minus119901 + 2)119909119909

d

119903 (119901 minus 1)

119909119909119903 (119901 minus 2)

119909119909sdot sdot sdot 119903 (0)119909119909

]

]

]

]

]

times

[

[

[

[

[

119886 (1)

119886 (1)

119886 (119901)

]

]

]

]

]

(18)

where 119903119909119909

can be defined by

119903119909119909 (

119898) =

1

119873


sum

119873=0

119909lowast(119899) 119909 (119899 + 119898) 119898 ge 0 (19)


119875119861119880

119909119909(119891) =

1205902

119908119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (20)

while 119908119901


1205902

119908119901= 119864119891

119901= 119903119909119909 (

0)

119901

prod

119896=1

[1 minus1003816100381610038161003816119886119896 (

119896)1003816100381610038161003816

2] (21)


ISRN Neuroscience 5






Frequencydomain


Wavelettransform






Frequencydomain






Stationarysignal

Autoregressive



Frequencydomain


features


119875119861119880

119909119909(119891) =

_119864119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (22)




6 ISRN Neuroscience






4 Conclusion




References



















ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of


ISRN Neuroscience 3

2

g[n]

g[n]

g[n]

h[n]

h[n]

h[n]

X[n]

D2

D1

D3A1

A2


darr 2

darr 2

darr 2

darr 2

darr 2

darr 2




CWT (119886 119887) = int

infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (5)


119886119887(119905) den-


120595119886119887 (

119905) =

1

radic|119886|

120595(

119905 minus 119887

119886

) (6)






minus1) = 1 (7)


119866 (119911) = 119911119867 (minus119911minus1) (8)




119860 (119891) =

119898

sum

119896=0

119886119896119890minus1198952120587119891119896

(9)



119875PISARENKO =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (10)


119875MUSIC (119891) =

1

(1119870)sum119870minus1

119894=0

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (11)


4 ISRN Neuroscience


119875MIN (119891119870) =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (12)



119875 (119905 119908) =

1

2120587

int

infin

minusinfin

int

infin

minusinfin

119860 (120579 120591)Φ (120579 120591) 119890minus119895120579119905minus119895120596120591

119889120579 119889120591

(13)

where

119860 (120579 120591) =

1

2120587

int

infin

minusinfin

119909(119906 +

120591

2


120591

2

) 119890119895120579119906

119889119906 (14)



119908(120591) and

119882119905(119905) [29]

SPWV (119905 119908)

= int

infin

minusinfin

119882119908 (

120591) [int

infin

minusinfin

119882119905 (119906 minus 119905) 119909 (119906 +

120591

2

)


120591

2

) 119889119906] 119890minus119895120596120591

119889120591

(15)


119896 119865119896 and 119871



119864119896= int

infin

minusinfin

int

infin

minusinfin

120599119896(119905 119891) 119889119905 119889119891 (16)



119865119896= int

infin

minusinfin

120599119896(119905 119891) 119889119905 (17)






[

[

[

[

[

119903 (0)119909119909119903 (minus1)119909119909


119903 (1)119909119909119903 (0)119909119909

119903(minus119901 + 2)119909119909

d

119903 (119901 minus 1)

119909119909119903 (119901 minus 2)

119909119909sdot sdot sdot 119903 (0)119909119909

]

]

]

]

]

times

[

[

[

[

[

119886 (1)

119886 (1)

119886 (119901)

]

]

]

]

]

(18)

where 119903119909119909

can be defined by

119903119909119909 (

119898) =

1

119873


sum

119873=0

119909lowast(119899) 119909 (119899 + 119898) 119898 ge 0 (19)


119875119861119880

119909119909(119891) =

1205902

119908119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (20)

while 119908119901


1205902

119908119901= 119864119891

119901= 119903119909119909 (

0)

119901

prod

119896=1

[1 minus1003816100381610038161003816119886119896 (

119896)1003816100381610038161003816

2] (21)


ISRN Neuroscience 5






Frequencydomain


Wavelettransform






Frequencydomain






Stationarysignal

Autoregressive



Frequencydomain


features


119875119861119880

119909119909(119891) =

_119864119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (22)




6 ISRN Neuroscience






4 Conclusion




References



















ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of


4 ISRN Neuroscience


119875MIN (119891119870) =

1

1003816100381610038161003816119860 (119891)

1003816100381610038161003816

2 (12)



119875 (119905 119908) =

1

2120587

int

infin

minusinfin

int

infin

minusinfin

119860 (120579 120591)Φ (120579 120591) 119890minus119895120579119905minus119895120596120591

119889120579 119889120591

(13)

where

119860 (120579 120591) =

1

2120587

int

infin

minusinfin

119909(119906 +

120591

2


120591

2

) 119890119895120579119906

119889119906 (14)



119908(120591) and

119882119905(119905) [29]

SPWV (119905 119908)

= int

infin

minusinfin

119882119908 (

120591) [int

infin

minusinfin

119882119905 (119906 minus 119905) 119909 (119906 +

120591

2

)


120591

2

) 119889119906] 119890minus119895120596120591

119889120591

(15)


119896 119865119896 and 119871



119864119896= int

infin

minusinfin

int

infin

minusinfin

120599119896(119905 119891) 119889119905 119889119891 (16)



119865119896= int

infin

minusinfin

120599119896(119905 119891) 119889119905 (17)






[

[

[

[

[

119903 (0)119909119909119903 (minus1)119909119909


119903 (1)119909119909119903 (0)119909119909

119903(minus119901 + 2)119909119909

d

119903 (119901 minus 1)

119909119909119903 (119901 minus 2)

119909119909sdot sdot sdot 119903 (0)119909119909

]

]

]

]

]

times

[

[

[

[

[

119886 (1)

119886 (1)

119886 (119901)

]

]

]

]

]

(18)

where 119903119909119909

can be defined by

119903119909119909 (

119898) =

1

119873


sum

119873=0

119909lowast(119899) 119909 (119899 + 119898) 119898 ge 0 (19)


119875119861119880

119909119909(119891) =

1205902

119908119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (20)

while 119908119901


1205902

119908119901= 119864119891

119901= 119903119909119909 (

0)

119901

prod

119896=1

[1 minus1003816100381610038161003816119886119896 (

119896)1003816100381610038161003816

2] (21)


ISRN Neuroscience 5






Frequencydomain


Wavelettransform






Frequencydomain






Stationarysignal

Autoregressive



Frequencydomain


features


119875119861119880

119909119909(119891) =

_119864119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (22)




6 ISRN Neuroscience






4 Conclusion




References



















ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of


ISRN Neuroscience 5






Frequencydomain


Wavelettransform






Frequencydomain






Stationarysignal

Autoregressive



Frequencydomain


features


119875119861119880

119909119909(119891) =

_119864119901

100381610038161003816100381610038161003816

1 + sum119875

119896=1

_119886119901(119896)119890minus1198952120587119891119896

100381610038161003816100381610038161003816

2 (22)




6 ISRN Neuroscience






4 Conclusion




References



















ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of


6 ISRN Neuroscience






4 Conclusion




References



















ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of


ISRN Neuroscience 7



























Neural Plasticity










Psychiatry Journal









Journal of














Neural Plasticity










Psychiatry Journal









Journal of


Review Article Methods of EEG Signal Features...

Documents

Transcript of Review Article Methods of EEG Signal Features...