Research Article Feature Selection in Classification of ...downloads.hindawi.com/journals/cmmm/2014/713818.pdf · analysis of EOG signals using clearness based features, minimum redundancy

Research ArticleFeature Selection in Classification of Eye Movements UsingElectrooculography for Activity Recognition

S Mala and K Latha

Department of Computer Science and Engineering Anna University BIT Campus TiruchirappalliTamil Nadu 620 024 India

Correspondence should be addressed to S Mala ursmalagmailcom

Received 26 August 2014 Revised 10 November 2014 Accepted 13 November 2014 Published 9 December 2014

Academic Editor Chuangyin Dang

Copyright copy 2014 S Mala and K LathaThis is an open access article distributed under theCreative CommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Activity recognition is needed in different requisition for example reconnaissance system patient monitoring and human-computer interfaces Feature selection plays an important role in activity recognition data mining and machine learning Inselecting subset of features an efficient evolutionary algorithm Differential Evolution (DE) a very efficient optimizer is used forfinding informative features from eye movements using electrooculography (EOG) Many researchers use EOG signals in human-computer interactions with various computational intelligence methods to analyze eye movements The proposed system involvesanalysis of EOG signals using clearness based features minimum redundancy maximum relevance features and DifferentialEvolution based features This work concentrates more on the feature selection algorithm based on DE in order to improve theclassification for faultless activity recognition

1 Introduction

Many researches in activity recognition and computer visionadopt gesture recognition as forefront [1] Gesture recogni-tion is an ideal example of multidisciplinary research ges-tures convey meaningful information to interact with envi-ronment through body motions involving physical move-ments of fingers arms head face or body There are manytools for gesture recognition like statistical modelling com-puter vision pattern recognition image processing and soforth Feature extraction and feature selection are successfullyused for many gesture recognition systems Importanceof gesture recognition laid in human-computer interactionapplications from medical rehabilitation to virtual realityMitra and Acharya provided a survey on gesture recognitionparticularly on hand gesture and facial expressions [2]

The goal of activity recognition is to provide informationthat allows a system to best assist the user with his or hertask Activity recognition has become an important area for abroad range of applications such as patient monitoring vigi-lance system and human-computer interaction Bulling is thefirst to describe and apply eye base activity recognition to theproblem of recognition of everyday activities [3] Eye move-ments have plentiful information for activity recognition

Bulling also described and evaluated algorithms fordetecting 3 eye movement characteristics of EOG signalssaccades fixations blinks [4] One of the eye movementcharacters blink plays major role in activity recognitionBlink serves as the first line of ocular protection The wipingaction of the lid can remove dust pollens [5] According toOusler et al blink rate can indeed serve as an indicator ofthe cognitive or emotional state of the blinker [6] Cumminsstates blinking might also contribute to social interactionby strengthening mutual synchronization in movements andgestures [7] Leal amp Vrij represented blinks in increasedcognitive demand during lying [8] Ponder and Kennedyreported about blinks during emotional excitement [9] andHirokawa depicted more frequent blinks in highest nervous-ness [10] Cognitive processes have a substantial impact onblink rates mentally taxing activities like memorization ormathematical computation being associated with an increasein blink rate and inattentiveness (daydreaming) and stimulustracking being associatedwith lowblink rates Blinks togetherwith other eyemovementswould be possible sources of usefulinformation inmaking inference about peoplersquosmental states

One of the possibilities to detect eye movements is EOGwhich is a technique for measuring the resting potential of

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2014 Article ID 713818 9 pageshttpdxdoiorg1011552014713818

2 Computational and Mathematical Methods in Medicine

the retina [11 12] The following lists some of the wide rangeof applications of EOG

(i) Wijesoma et al used EOG for guiding and controllinga wheelchair for disabled people [13]

(ii) Usakli and Gurkan used EOG for using virtual key-board [14]

(iii) Deng et al used EOG for operating a TV remotecontrol and for a game [15]

(iv) Talaba used EOG for visual navigationmetaphor [16](v) Gandhi et al used EOG for controlling multitask

gadget [17](vi) Bulling et al used EOG for activity recognition [18]

Feature selection is important and necessary when a realworld application has to deal with training data of highdimensionalityThere aremany features selection approachesavailable in the literature Some of the hybrid approaches arelisted as follows

(i) Garcıa-Nieto et al presented a Differential Evolutionbased approach for efficient automated gene subsetselection using DLBCL Lymphoma and Colon tumorgene expression datasets The selected subsets areevaluated by means of the SVM classifier [19]

(ii) Li employed a DE-SVMmodel that hybridized DE ampSVM to improve the classification accuracy of roadicing forecasting using feature selection [20]

(iii) Xu and Suzuki proposed a feature selection methodbased on sequential forward floating selection toimprove performance of a classifier in the computer-ized detection of polyps in CT colonography (CTC)In this work feature selection is coupled with SVMclassifier and maximized the area under the receiveroperating characteristic curve [21]

(iv) Kuo et al proposed kernel based feature selectionmethod to improve the classification performance ofSVM using hyperspectral image datasets [22]

(v) Guven and Kara employed artificial neural networkanalysis of EOG signals for the purpose of distin-guishing between subnormal and normal eye [23]This enables the physician to make a quick judgmentabout the existence of eye disease with more confi-dence Only binary classification of EOG signal hasbeen performed in this work

From the literature review it can be observed that thereare a number of EOG applications being developed Stillnecessary feature selection algorithms need to be developedevaluated and used to produce substantial improvements incommunications with disabled people by using eye move-ments and to make inference about personrsquos cognitive state

This paper presents a hybrid feature selection techniquebased on DE for activity recognition using eye movements byEOG signals which can identify a subset of most informativeeye movement characteristics amongst all eye movementcharacteristics This method is used as an optimizer before

the classifier to EOG signal features for recognizing activitieslike read browse write video and copy The benefits of thefeature selection approach include improving the efficiencyof activity recognition since only a subset of eye movement isused and assisting SVM to attain satisfied accuracy

2 Methodology

Here we explain the preprocessing feature selection (withCBFS mRMR and DEFS) SVM classifier and the modelevaluation strategy used in this work

21 Dataset Description Recognition-of-Office-ActivitiesThe EOG data used in this study are collected from theAndreas Bullingrsquos ldquorecognition-of-office-activitiesrdquo data-set (httpswwwandreas-bullingdedatasetsrecognition-of-office-activities) Eight participants took part during thisstudy For about 30 minutes the participants were involved intwo continuous activity sequences The total dataset is abouteight hours The data aggregation of this work is groundedon office based activities performed in random order paperreading taking notes watching video and net browsingAdditionally the dataset includes a period of rest (the NULLclass) These activities are all generally performed during ausual working day

These experiments were carried out in a well-lit work-place during normal working hours Participants were seatedahead of 2 seventeen inch flat screens with a resolution of1280 times 1024 pixels on which a video player a browser andtext for copying in a word processor were on-screen Sheetsof paper and a pen were presented on the table close to theparticipants for reading and writing tasks No constraintswere forced with type of website and manner of interactionof browsing task

EOG signals were picked up using an array of five24mmAgAgCl wet electrodes from Tyco Healthcare placedaround the right eye The horizontal signal and verticalsignal were collected using two electrodes for each and thefifth electrode was placed on the forehead for the signalreference EOG data is captured employing a commercialEOG device known as TMSI (Twente Medical SystemsInternational) Mobi8 that integrates instrument amplifierswith 24 bit ADCs Mobi8 tends to have better signal qualityMobi8 was worn on a belt around each participantrsquos waistand recorded four channels EOG at a sampling rate of128Hz The behaviours of participants in specific phases(read browse write video copy and rest) were observed byan annotated activity changes with wireless remote controland their nature in daily life during regular working hoursContext Recognition Network Toolbox (CRNT) was used forhandling data recording and synchronization

22 Baseline Drifts Removal and Filtering EOG Signals Forpreprocessing this work adapts median filter and baselinedrift removal 1D wavelet decomposition at level nine usingDaubechies wavelets on the signal component [18]

Figure 1 shows EOG signals of horizontal before and afterbaseline drifts removal and filtering

Computational and Mathematical Methods in Medicine 3

Table 1 MSE and PSNR values

Activity Null Read Browse Write Video CopySamples number 320315 290963 302385 317092 304170 303441MSE

Before filtering 41736119890 + 06 31071119890 + 06 7550119890 + 05 21893119890 + 06 10267119890 + 06 31546119890 + 06After filtering 15957119890 + 05 53059119890 + 04 89856119890 + 04 18710119890 + 05 84765119890 + 04 11215119890 + 05

PSNRBefore filtering 304026 295266 303008 310553 305343 310118After filtering 449982 469136 396886 427312 422990 455411

0 1 2 3 4 5 6 7 8

232221

219

Time (ms)

Am

plitu

de

EOG horizontal signal (rawsignal)times105

Time (ms)0 1 2 3 4 5 6 7 8

012

Am

plitu

de

EOG horizontal signal (preprocessed)times104

minus2minus1

Figure 1 Before and after preprocessing (EOG) signals

Table 1 lists corresponding mean square error (MSE) andpeak signal to noise ratio (PSNR) for two participants withthousands of samples

23 Basic Eye Movement Type Detection Various activitiesusing eye movements by EOG can be portrayed as a regularpattern by a specific sequence of saccades and short fixationsof similar duration The amplitude change in signals variesfor various activities which can be used in identifying theregular office activities The reading activity using EOG ispatterned by small saccades and fixations No large changein amplitude is included in reading This is due to smalleye movement between the words and fast eye movementbetween ends of previous line and beginning of next lineWriting was similar to reading yet it required greater fixationduration and greater variance It was best described usingaverage fixation duration Copying activity includes normalback and forth eye movements which involves saccadesbetween screens This was reflected in the selection of smalland large horizontal saccade features as well as variance inhorizontal EOG fixations In contrast watching a video andbrowsing are highly unstructured These activities dependon the video or website being viewed These results proposethat for tasks that involve a known set of specific activityclasses recognition can be streamlined by only choosing eyemovement features known to best describe these classes

The steps in basic eye movement type detection forEOG based action recognition are (1) noise and baselinedrift removal (2) basic eye movement detection and (3)

feature extraction The basic eye movementsrsquo interpretationand detections are carried out using wavelet coefficients

Eye movement characteristics like saccades fixationsand blink patterns are separated from EOG signals usingcontinuous wavelet transform (CWT) 1D wavelet coefficientsusing aHaarmotherwavelet at scale 20 [18]ThepreprocessedEOG signal components (EOGh EOGv-denoised baselinedrift removed) are given as input to CWT

119862119887

119886= int

R

119904 (119905)1

radic119886120595(

119905 minus 119887

119886)119889119905 (1)

The wavelet coefficient (119862119887

119886) of signal (119904) at scale 119886 and

position 119887 is defined by (1)

119873119894 =

1 forall119894119862119894 (119904) lt minusthsdminus1 forall119894119862119894 (119904) gt thsd0 forall119894 minus thsd le 119862119894 (119904) le thsd

(2)

By applying a threshold thsd = 2000 on the coefficients119862119894(119904) = 119862

20

119894(119904) CWT creates a vector 119873 with elements

119873119894 based on (2) This step divides EOGh (horizontal EOG)and EOGv (vertical EOG) in saccadic (119873 = 1 minus1) andnonsaccadic (fixation) segment (119873 = 0)

We used the same wavelet coefficients to detect blinksin EOGv Features related to the eye movements using EOGsignal were calculated separately for two EOG (horizontaland vertical) signals for each participant Statistical featuressuch asmean variance andmaximumvalueminimumvaluebased on saccades fixations and blinks are extracted fromthis work Different characteristics of EOG signals result indifferent changes in the coefficients Totally 210 statisticalfeatures are extracted from this work

Eye movement characteristics such as saccades and blinkpatterns of EOG (horizontal and vertical) signals are shownin Figure 2

24 Feature Selection

241 Minimum Redundancy Maximum Relevance (mRMR)The minimum redundancy maximum relevance (mRMR)[24 25] algorithm is a sequential forward selection algo-rithm It uses mutual information to analyze relevance andredundancy The mRMR scheme selects the features thatcorrelate the strongest with a classification characteristic andcombined with selection features that are mutually different


0 50 100 150 200 250 300

0

2

4

Timestamp (s)

Am

plitu

de

Blink detection

EOGBlink

times104

minus2

minus4

minus6

minus8

(a)

0 1000 2000 3000 4000 5000

0

1

2 Marker vector small saccade

Timestamp (s)

Mar

ker v

ecto

r val

ue

minus1

minus2

Mar

ker v

ecto

r val

ue

0 500 1000 1500 2000 2500

0

115

05

Marker vector large saccade

Timestamp (s)

minus05

minus15

minus1

(b)

Figure 2 (a) Blink detection from EOGv (b) Saccades from EOGh

from each other having high correlation and it is denoted bythe following equation

119869 (119883119899) = 119868 (119883119899 119884) minus1

|119878|sum

119883119894isin119878

119868 (119883119899 119883119894) (3)

where 119868(119883119899 119884) is themeasure of dependence between feature119883119899 and objective 119884 119869 = 119868(1198831119899 119884) minus 119868(119883

1119899minus1 119884) is the

difference in informationwith andwithout119883119899 S is the featureset and |S| is the number of features

242 Clearness Based Feature Selection (CBFS) CBFS [26]calculates the distance between the objective sample and thecenter of every class and then compares the class of thenearest center with the class of the objective sample The

similarity ratio of all samples in a feature becomes a clearnessvalue for the feature [26]

Clearness based feature selection (CBFS) algorithm isa type of filter method Clearness means the detachmentbetween classes in a feature If (clearness of feature 1198912)gt (clearness of feature 1198911) then 1198912 is more useful forclassification than 1198911

Step 1 The centrist for read and write is calculated by averageoperation This is the median point of a class Med(119891119894 119895)represents the median point of class 119895 in the feature 119891119894 andit is calculated by the following equation

Med (119891119894 119895) =1

119896

119896

sum

119903=1

(119883119903119894 isin class 119895) (4)

where 119896 is a number of samples of class 119895

Step 2 For each 119909119894119895 the sample predicted class is calculatedAfter calculating the distance between 119909119894119895 and Med(119891119895 119888119894) forall classes this work takes the nearest centrist Med(119891119895 119904) and119904 is a predicted class label for 119909119894119895The distance between 119909119894119895 andMed(119891119895 119905) is calculated by the following equation

119863(119909119894119895Med (119891119895 119905)) =10038161003816100381610038161003816119909119894119895 minusMed (119891119895 119905)

10038161003816100381610038161003816 (5)

Step 3 Calculate 119899 times 119898 matrix 1198722 This matrix contains amatching result of predicted class label and correct class labelin 119862Score1198722(119894 119895) is calculated by

1198722 (119894 119895) = 1 if 1198721 (119894 119895) = 119862119894

0 if 1198721 (119894 119895) = 119862119894(6)

Step 4 Calculate 119862Score (119891119894) Initially we calculated 119862Score (119891119894)by

119862Score (119891119894) =1

119899

119899

sum

119903=1

1198722 (119903 119894) (7)

The range of 119862Score(119891119894) is [0 1] If 119862Score (119891119894) is close to 1 thenit indicates that classes in feature 119891119894 are grouped well andelements in 119891119894 can be clearly classified

25 Proposed Method The proposed method first identifiesessential features by applying a threshold (th119888) in correlatedvalues among features This stage reduces the feature spacedimensionality If correlation (119888) between the features satisfiesthe threshold (th119888 lt 08) those features are selected forthe next stage After removing tautological features approx-imately 184 features are selected for this work

The reduced feature space is given to the DEFS algorithmfor detecting the significant feature subset through machinelearning algorithm such as maximum a posterior approachHere for the projection of features into feature subspace we gofor kernelized Bayesian structure This method is used as anoptimizer before the classifier to EOG signal features for rec-ognizing activities like read browse write video and copy

Figure 3 illustrates the steps involved in the DifferentialEvolution process [27] The first step in the DE optimization


method is to generate a population (NPtimesD) of NPmemberseach of D-dimensional real-valued parameters In order toimprove the DE subset selection efficiency we are going tochange the fitness function in terms of maximum a posterioriprobability The kernel distribution is appropriate for EOGfeatures since it has a continuous distribution The EOGfeatures may be skewed and have multiple peaks we canuse a kernel which does not require a strong assumptionThe kernel needs additional time andmemory for computingthan the normal distribution For every feature we modelwith a kernel the Naive Bayes classifier or every categorybased on the training data for that class The default kernelis normal and also the classifier selects a width automaticallyfor every class and featureThe stopping criterionwas definedas reaching the maximum number of iterations The chosenfitness function was the classification error rates achieved bythe Naive Bayes classifier with kernel distribution

In thiswork parameter119865 is assigneddynamically CRwiththe value of 08 the number of dimensions of the problem(features) D is 15 the population used is 50 and the numberof generations is 10

V119895119894119892 = 1199091198951199030119892 + 119865 times (1199091198951199031119892 minus 1199091198951199032119892)

where

119894 = 0 to NP minus 1

119895 = 1 to D minus 1

1199091199031 1199091199032 = two population members

119865 isin (0 1) scale factor

119892 is generation

(8)

For each target vector mutant vector is created by using(8)

119906119895119894119892 = V119895119894119892 if rand (0 1) le 119862119903

119909119895119894119892 otherwise

where

119906119894+119895 is 119895th dimension from 119894th trial vector

119892 = population

119862119903 = Crossover probability

(9)

The parent vector is mixed with the mutated vector toproduce a trial vector as in (9)

DE employs uniform crossover Newly generated vectorresults in a lower objective function value (fitness) The ran-domly chosen initial population matrix of size (NP times DNF)containing NP randomly chosen vectors 119894 = (0 1 NPminus1)is created DNF is desired number of features to be selectedWemade search limited by the total number of features (NF =

15) Individual lower boundary of the search space is 119897 = 1

and upper boundary is H = NF = 15 Each new vector frominitial population is indexed as 0 toNPminus1The steps in theDEprocess are as follows the difference between two population

numbers (P1 P2) is added to a third population member(P3)The result (R4) is subject to crossover with candidate forreplacement (C5) to obtain a proposed (PR6) The proposedsystem is evaluated and replaces the candidate if it is foundto be the best In our scheme the probability of each featureis calculated and used as weighting to replace the duplicatedfeatures All the time the features are not in a linear mannerSo we only go for nonlinear kernel structure to enclose thefeatures in feature space The objective of this work is reduc-ing the complexities of evolutionary algorithmand improvingthe fitness validation bymachine learning probabilistic kernelclassifier by considering the nonlinearity of inputs

Table 2 shows the number of samples taken for eachactivity features identified from each activity and the per-formance summary for each activity with and without DEFSfeatures

26 SVM Activity classification involves a set of 119898 sam-ples labelled with a class A sample is a vector 119909119894 =

[1199091198941 1199091198942 119909119894119899] of 119899 features where 119909119894119895 [119891119895min 119891119895max] isthe activity recognition value for the jth feature of the 119894thsample (1 le 119894 le 119898 1 le 119895 le 119899) The label associatesa class value class i 119862 with the sample 119909119894 where 119862 =

class1 class2 class119896 is a set of class values for thesamples In this work we consider SVM to classify sampleswith six possible class values (119896 = 6)

SVM handles nonlinear data by using a kernel function[28] The kernel maps the data into a feature space Thenonlinear function is learned by a linear learning machinein a high dimensional feature space This is known as kerneltrick which implies that the kernel function transforms thedata into a higher dimensional feature space to form feasiblylinear separation [29]

Linear119870(119909119894 119909119895) = 119909119879

119894119909119895 (10)

The kernel function used in this work is linear kernelmeaning dot product which is shown in (10)

120596 (120572) =

119899

sum

119894=1

120572119894 minus1

2

119899

sum

119894=1119895=1

120572119894120572119895119870(119909119894 119909119895) (11)

120596 =

119899

sum

119894=1

120572119910119894119909119894 (12)

There is a class of functions 119870(119909 119910) with a linear space119878 and a function 120593 mapping 119909 to 119878 such that 119870(119909 119910) =

⟨120593(119909) 120593(119910)⟩The dot product takes place in the space 119878 Solve(11) and get the 120572119894 which maximize 120596 using (12)

119891 (119909new) = sign(119899

sum

119894=1

120572119910119894119870(119909119894119909new) + 119887) (13)

The classifier is defined by (13) We used 75 of eachactivity for training the classifierThe classifier performance isimproved by our proposed hybrid feature selection method


Base vector

Target vector Mutant vector

New population

Cross over target with mutant

Check for redundancy in features if redundancy exists use roulette wheel

Select trial or target

Populationvectors

Compute weighted difference

7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8

Original population Pxg Mutants population PgX0g

X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1

Figure 3 Differential Evolution operations

Table 2 Performance summary

Performance Null Read Browse Write Video Copy AllSamples number 320315 290963 302385 317092 304170 303441 1838366Number of features 232 211 218 225 218 218 210Accuracy (all features) 82 67 62 73 83 68 725Accuracy (with DEFS-15 features) 87 77 79 84 88 85 8333


Table 3 Detailed performance of the proposed approach with CBFS and mRMR features

Activity Features TP () FP () Precision () Recall () 119865-Measure ()

NullDEFS 7531 1420 8414 9224 8799mRMR 6550 1541 8095 9332 8670CBFS 6760 1430 8254 8989 8606

ReadDEFS 7160 1712 8070 9374 8673mRMR 5830 1829 7611 9179 8322CBFS 5729 1658 7756 8227 7984

BrowseDEFS 7517 1415 8416 9375 8869mRMR 6898 1438 8275 9517 8853CBFS 5400 1650 7660 8182 7912

WriteDEFS 8123 0992 8912 9705 9291mRMR 7910 0885 8894 9456 9219CBFS 6924 1210 8512 9039 8768

VideoDEFS 8551 1020 8934 9938 9409mRMR 6741 1309 8374 9023 8686CBFS 6237 1544 8016 8616 8305

CopyDEFS 8021 1318 8589 9761 9137mRMR 7034 1028 8525 8766 8745CBFS 5644 1117 8348 8188 8267

27 Performance Evaluation Criteria We adopted the follow-ing criteria to evaluate the performance of the classifier Weused Mat lab to calculate the accuracy of the classifier

Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)

119865-Measure = 2TP2TP + FP + FN

(14c)

Accuracy = TP + TNTP + TN + FP + FN

(14d)

where TP TN FP FN represents true positive true negativefalse positive and false negative

The sensitivity specificity true positive rate (TPR) falsenegative rate (FPR) and accuracy (ACC) were calculated byusing (14a)ndash(14d)

The detailed performance of the proposed approach withDEFS features is listed in Table 3 The precision for videoactivity is higher than other activities The recall for writingactivity is slightly higher than other two feature selectionmethods The detailed performance of the clearness basedfeatures with SVM and minimum redundancy maximumrelevance features with SVM approach is listed and is alsocompared with DEFS features The result shows that the pro-posed approach best performs with the other two methods

Figure 4 shows that precision for each activity is high forthe proposed DEFS based feature selection compared to theother two feature selection methods mRMR and CBFS Forwrite activity the proposed DEFS based feature selection andmRMR nearly provide the same results For copy activity allthe three feature selection methods provide almost the same

Null Read Browse Write Video Copy0

102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS

Figure 4 Precision for each activity by proposed DEFS basedfeatures with mRMR CBFS features

precision values For video activity the proposed methodoutperforms well Null and read activity precision shows thatthe CBFS features provide the best result when compared tomRMR features

3 Results and Discussion

Feature extracted addresses the problem of finding a moreinformative set of features The resultant feature subsetshows that the most informative features in EOG signalsare saccades and blinks The statistical measures for EOGsignal analysis proved to be very useful in searching thefeature space by using hybrid feature selection based ondifferential evolution The SVM with linear kernel is usedfor classification in EOG datasets before and after featureselection The results of 10-fold cross-validation are listed in


Table 3 In the table we report performance by accuracyThe dataset is divided into training set with three-fourth oforiginal and rest one-fourth of testing instances We focuson Differential Evolution based feature selection algorithmin order to improve the classification The detailed accuracyof class results is shown in Table 2 with an emphasis on thedifference before and after feature selection The results inTable 2 shows class accuracy is increased with minimumnumber of featuresThe selected features improve the perfor-mance in terms of lower error rates and the evaluation of afeature subset becomes simpler than that of a full set Whenthe results are compared with CBFS and mRMR featureselection the classifier (SVM) performance (ACC = 8333)is significantly improved with proposed features

4 Conclusions

A hybrid feature selection method was employed in EOGsignals based on the Differential Evolution and the proposedmethod is comparedwithCBFS andmRMR feature selectionThe differential evolutionary algorithm is utilized to givepowerful results in searching for subsets of features thatbest interact together through supervised learning approachEOG dataset with high dimensionality and number of targetclasses (119873 = 6) were considered to test the performanceof the proposed feature selection What better fits ourcase is lower classification error rate which is attained as016 The classification performance is significantly improvedwhen using the proposed feature selection algorithm thatuses differential evolution based on a posteriori probabilityThis method followed by wavelet feature extraction presentspowerful results in terms of accuracy (ACC = 8333) and anoptimal subset of features with size (119878) 15

Conflict of Interests

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

References

[1] A Bulling J A Ward H Gellersen and G Troster ldquoRobustrecognition of reading activity in transit using wearable elec-trooculographyrdquo in Proceedings of the 6th International Confer-ence Pervasive Computing pp 19ndash37 2008

[2] S Mitra and T Acharya ldquoGesture recognition a surveyrdquoIEEE Transactions on Systems Man and Cybernetics Part CApplications and Reviews vol 37 no 3 pp 311ndash324 2007

[3] M Vidal A Bulling and H Gellersen ldquoAnalysing EOG signalfeatures for the discrimination of eye movements with wearabledevicesrdquo in Proceedings of the 1st InternationalWorkshop on Per-vasive Eye Tracking and Mobile Eye-Based Interaction (PETMEIrsquo11) pp 15ndash20 ACM Beijing China September 2011

[4] D D Salvucci and J H Goldberg ldquoIdentifying fixations andsaccades in eye-tracking protocolsrdquo in Proceedings of the Sym-posium on Eye Tracking ResearchampApplications (ETRA rsquo00) pp71ndash78 November 2000

[5] M B Abelson ldquoItrsquos time to think about the blinkrdquo Review ofOphthalmology vol 18 no 6 pp 58ndash61 2011

[6] G W Ousler III K W Hagberg M Schindelar D Welch andM B Abelson ldquoThe ocular protection indexrdquo Cornea vol 27no 5 pp 509ndash513 2008

[7] F Cummins ldquoGaze and blinking in dyadic conversation a studyin coordinated behaviour among individualsrdquo Language andCognitive Processes vol 27 no 10 pp 1525ndash1549 2012

[8] A Vrij E Ennis S Farman and S Mann ldquoPeoples perceptionsof their truthful and deceptive interactions in daily liferdquo Journalof Forensic Psychology vol 2 pp 6ndash42 2010

[9] E Ponder andW P Kennedy ldquoOn the act of blinkingrdquoQuarterlyJournal of Experimental Physiology vol 18 pp 89ndash110 1927

[10] K Hirokawa A Yagi and Y Miyata ldquoAn examination of theeffects of linguistic abilities on communication stress measuredby blinking and heart rate during a telephone situationrdquo SocialBehavior and Personality vol 28 no 4 pp 343ndash354 2000

[11] A Ubeda E Ianez and J M Azorin ldquoWireless and portableEOG-based interface for assisting disabled peoplerdquo IEEEASMETransactions on Mechatronics vol 16 no 5 pp 870ndash873 2011

[12] A T Duchowski Eye Tracking Methodology Theory and Prac-tice vol 25 no 9 Springer New York NY USA 2007

[13] W S Wijesoma S W Kang C W Ong A P BalasuriyaT S Koh and K S Low ldquoEOG based control of mobileassistive platforms for the severely disabledrdquo in Proceedings ofthe IEEE International Conference on Robotics and Biomimetics(ROBIO rsquo05) pp 490ndash493 July 2005

[14] A B Usakli and S Gurkan ldquoDesign of a novel efficienthuman-computer interface an electrooculagram based virtualkeyboardrdquo IEEE Transactions on Instrumentation and Measure-ment vol 59 no 8 pp 2099ndash2108 2010

[15] L Y Deng C L Hsu T C Lin J S Tuan and S M ChangldquoEOG-based Human-computer Interface system developmentrdquoExpert Systems with Applications vol 37 no 4 pp 3337ndash33432010

[16] C-C Postelnicu F Girbacia and D Talaba ldquoEOG-basedvisual navigation interface developmentrdquo Expert Systems withApplications vol 39 no 12 pp 10857ndash10866 2012

[17] T Gandhi M Trikha J Santhosh and S Anand ldquoDevelopmentof an expert multitask gadget controlled by voluntary eyemovementsrdquo Expert Systems with Applications vol 37 no 6 pp4204ndash4211 2010

[18] A Bulling J A Ward H Gellersen and G Troster ldquoEyemovement analysis for activity recognition using electroocu-lographyrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 33 no 4 pp 741ndash753 2011

[19] J Garcıa-Nieto E Alba and J Apolloni ldquoHybrid DE-SVMapproach for feature selection application to gene expressiondatasetsrdquo in Proceedings of the 2nd International Symposium onLogistics and Industrial Informatics (LINDI rsquo09) pp 1ndash6 LinzAustria September 2009

[20] J Li ldquoA combination of DE and SVM with feature selection forroad icing forecastrdquo in Proceedings of the 2nd International AsiaConference on Informatics in Control Automation and Robotics(CAR rsquo10) vol 2 pp 509ndash512 Wuhan China March 2010

[21] J-W Xu and K Suzuki ldquoMax-AUC feature selection incomputer-aided detection of polyps in CT colonographyrdquo IEEEJournal of Biomedical and Health Informatics vol 18 no 2 pp585ndash593 2014

[22] B C Kuo H H Ho C H Li C C Hung and J S TaurldquoA kernel-based feature selection method for SVM with RBFkernel for hyperspectral image classificationrdquo IEEE Journalof Selected Topics in Applied Earth Observations and RemoteSensing vol 7 no 1 pp 317ndash326 2014


[23] A Guven and S Kara ldquoClassification of electro-oculogramsignals using artificial neural networkrdquo Expert Systems withApplications vol 31 no 1 pp 199ndash205 2006

[24] H Peng F Long and C Ding ldquoFeature selection based onmutual information criteria of Max-Dependency Max-Rele-vance and Min-Redundancyrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 27 no 8 pp 1226ndash12382005

[25] H Peng ldquomRMR Feature Selection Toolbox for MATLABrdquohttpresearchjaneliaorgpengprojmRMR

[26] M Seo and S Oh ldquoCBFS high performance feature selectionalgorithm based on feature clearnessrdquo PLoS ONE vol 7 no 7Article ID e40419 2012

[27] R N Khushaba A Al-Ani and A Al-Jumaily ldquoFeature subsetselection using differential evolution and a statistical repairmechanismrdquo Expert Systems with Applications vol 38 no 9 pp11515ndash11526 2011

[28] V N Vapnik Statistical Learning Theory Wiley-Interscience1998

[29] K Crammer and Y Singer ldquoUltraconservative online algo-rithms for multiclass problemsrdquo Journal of Machine LearningResearch vol 3 pp 951ndash991 2003

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


the retina [11 12] The following lists some of the wide rangeof applications of EOG

(i) Wijesoma et al used EOG for guiding and controllinga wheelchair for disabled people [13]

(ii) Usakli and Gurkan used EOG for using virtual key-board [14]

(iii) Deng et al used EOG for operating a TV remotecontrol and for a game [15]

(iv) Talaba used EOG for visual navigationmetaphor [16](v) Gandhi et al used EOG for controlling multitask

gadget [17](vi) Bulling et al used EOG for activity recognition [18]

Feature selection is important and necessary when a realworld application has to deal with training data of highdimensionalityThere aremany features selection approachesavailable in the literature Some of the hybrid approaches arelisted as follows

(i) Garcıa-Nieto et al presented a Differential Evolutionbased approach for efficient automated gene subsetselection using DLBCL Lymphoma and Colon tumorgene expression datasets The selected subsets areevaluated by means of the SVM classifier [19]

(ii) Li employed a DE-SVMmodel that hybridized DE ampSVM to improve the classification accuracy of roadicing forecasting using feature selection [20]

(iii) Xu and Suzuki proposed a feature selection methodbased on sequential forward floating selection toimprove performance of a classifier in the computer-ized detection of polyps in CT colonography (CTC)In this work feature selection is coupled with SVMclassifier and maximized the area under the receiveroperating characteristic curve [21]

(iv) Kuo et al proposed kernel based feature selectionmethod to improve the classification performance ofSVM using hyperspectral image datasets [22]

(v) Guven and Kara employed artificial neural networkanalysis of EOG signals for the purpose of distin-guishing between subnormal and normal eye [23]This enables the physician to make a quick judgmentabout the existence of eye disease with more confi-dence Only binary classification of EOG signal hasbeen performed in this work

From the literature review it can be observed that thereare a number of EOG applications being developed Stillnecessary feature selection algorithms need to be developedevaluated and used to produce substantial improvements incommunications with disabled people by using eye move-ments and to make inference about personrsquos cognitive state

This paper presents a hybrid feature selection techniquebased on DE for activity recognition using eye movements byEOG signals which can identify a subset of most informativeeye movement characteristics amongst all eye movementcharacteristics This method is used as an optimizer before

the classifier to EOG signal features for recognizing activitieslike read browse write video and copy The benefits of thefeature selection approach include improving the efficiencyof activity recognition since only a subset of eye movement isused and assisting SVM to attain satisfied accuracy

2 Methodology

Here we explain the preprocessing feature selection (withCBFS mRMR and DEFS) SVM classifier and the modelevaluation strategy used in this work

21 Dataset Description Recognition-of-Office-ActivitiesThe EOG data used in this study are collected from theAndreas Bullingrsquos ldquorecognition-of-office-activitiesrdquo data-set (httpswwwandreas-bullingdedatasetsrecognition-of-office-activities) Eight participants took part during thisstudy For about 30 minutes the participants were involved intwo continuous activity sequences The total dataset is abouteight hours The data aggregation of this work is groundedon office based activities performed in random order paperreading taking notes watching video and net browsingAdditionally the dataset includes a period of rest (the NULLclass) These activities are all generally performed during ausual working day

These experiments were carried out in a well-lit work-place during normal working hours Participants were seatedahead of 2 seventeen inch flat screens with a resolution of1280 times 1024 pixels on which a video player a browser andtext for copying in a word processor were on-screen Sheetsof paper and a pen were presented on the table close to theparticipants for reading and writing tasks No constraintswere forced with type of website and manner of interactionof browsing task

EOG signals were picked up using an array of five24mmAgAgCl wet electrodes from Tyco Healthcare placedaround the right eye The horizontal signal and verticalsignal were collected using two electrodes for each and thefifth electrode was placed on the forehead for the signalreference EOG data is captured employing a commercialEOG device known as TMSI (Twente Medical SystemsInternational) Mobi8 that integrates instrument amplifierswith 24 bit ADCs Mobi8 tends to have better signal qualityMobi8 was worn on a belt around each participantrsquos waistand recorded four channels EOG at a sampling rate of128Hz The behaviours of participants in specific phases(read browse write video copy and rest) were observed byan annotated activity changes with wireless remote controland their nature in daily life during regular working hoursContext Recognition Network Toolbox (CRNT) was used forhandling data recording and synchronization

22 Baseline Drifts Removal and Filtering EOG Signals Forpreprocessing this work adapts median filter and baselinedrift removal 1D wavelet decomposition at level nine usingDaubechies wavelets on the signal component [18]

Figure 1 shows EOG signals of horizontal before and afterbaseline drifts removal and filtering






0 1 2 3 4 5 6 7 8

232221

219

Time (ms)

Am

plitu

de


Time (ms)0 1 2 3 4 5 6 7 8

012

Am

plitu

de


minus2minus1







119862119887

119886= int

R

119904 (119905)1

radic119886120595(

119905 minus 119887

119886)119889119905 (1)




119873119894 =


(2)


20








0 50 100 150 200 250 300

0

2

4

Timestamp (s)

Am

plitu

de

Blink detection

EOGBlink

times104

minus2

minus4

minus6

minus8

(a)

0 1000 2000 3000 4000 5000

0

1


Timestamp (s)

Mar

ker v

ecto

r val

ue

minus1

minus2

Mar

ker v

ecto

r val

ue

0 500 1000 1500 2000 2500

0

115

05


Timestamp (s)

minus05

minus15

minus1

(b)



119869 (119883119899) = 119868 (119883119899 119884) minus1

|119878|sum

119883119894isin119878

119868 (119883119899 119883119894) (3)


1119899minus1 119884) is the






Med (119891119894 119895) =1

119896

119896

sum

119903=1

(119883119903119894 isin class 119895) (4)



119863(119909119894119895Med (119891119895 119905)) =10038161003816100381610038161003816119909119894119895 minusMed (119891119895 119905)

10038161003816100381610038161003816 (5)


1198722 (119894 119895) = 1 if 1198721 (119894 119895) = 119862119894

0 if 1198721 (119894 119895) = 119862119894(6)


119862Score (119891119894) =1

119899

119899

sum

119903=1

1198722 (119903 119894) (7)








V119895119894119892 = 1199091198951199030119892 + 119865 times (1199091198951199031119892 minus 1199091198951199032119892)

where

119894 = 0 to NP minus 1

119895 = 1 to D minus 1




(8)


119906119895119894119892 = V119895119894119892 if rand (0 1) le 119862119903

119909119895119894119892 otherwise

where


119892 = population


(9)











Linear119870(119909119894 119909119895) = 119909119879

119894119909119895 (10)


120596 (120572) =

119899

sum

119894=1

120572119894 minus1

2

119899

sum

119894=1119895=1

120572119894120572119895119870(119909119894 119909119895) (11)

120596 =

119899

sum

119894=1

120572119910119894119909119894 (12)



119891 (119909new) = sign(119899

sum

119894=1

120572119910119894119870(119909119894119909new) + 119887) (13)



Base vector


New population




Populationvectors


7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8


X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1














Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















0 1 2 3 4 5 6 7 8

232221

219

Time (ms)

Am

plitu

de


Time (ms)0 1 2 3 4 5 6 7 8

012

Am

plitu

de


minus2minus1







119862119887

119886= int

R

119904 (119905)1

radic119886120595(

119905 minus 119887

119886)119889119905 (1)




119873119894 =


(2)


20








0 50 100 150 200 250 300

0

2

4

Timestamp (s)

Am

plitu

de

Blink detection

EOGBlink

times104

minus2

minus4

minus6

minus8

(a)

0 1000 2000 3000 4000 5000

0

1


Timestamp (s)

Mar

ker v

ecto

r val

ue

minus1

minus2

Mar

ker v

ecto

r val

ue

0 500 1000 1500 2000 2500

0

115

05


Timestamp (s)

minus05

minus15

minus1

(b)



119869 (119883119899) = 119868 (119883119899 119884) minus1

|119878|sum

119883119894isin119878

119868 (119883119899 119883119894) (3)


1119899minus1 119884) is the






Med (119891119894 119895) =1

119896

119896

sum

119903=1

(119883119903119894 isin class 119895) (4)



119863(119909119894119895Med (119891119895 119905)) =10038161003816100381610038161003816119909119894119895 minusMed (119891119895 119905)

10038161003816100381610038161003816 (5)


1198722 (119894 119895) = 1 if 1198721 (119894 119895) = 119862119894

0 if 1198721 (119894 119895) = 119862119894(6)


119862Score (119891119894) =1

119899

119899

sum

119903=1

1198722 (119903 119894) (7)








V119895119894119892 = 1199091198951199030119892 + 119865 times (1199091198951199031119892 minus 1199091198951199032119892)

where

119894 = 0 to NP minus 1

119895 = 1 to D minus 1




(8)


119906119895119894119892 = V119895119894119892 if rand (0 1) le 119862119903

119909119895119894119892 otherwise

where


119892 = population


(9)











Linear119870(119909119894 119909119895) = 119909119879

119894119909119895 (10)


120596 (120572) =

119899

sum

119894=1

120572119894 minus1

2

119899

sum

119894=1119895=1

120572119894120572119895119870(119909119894 119909119895) (11)

120596 =

119899

sum

119894=1

120572119910119894119909119894 (12)



119891 (119909new) = sign(119899

sum

119894=1

120572119910119894119870(119909119894119909new) + 119887) (13)



Base vector


New population




Populationvectors


7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8


X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1














Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















0 50 100 150 200 250 300

0

2

4

Timestamp (s)

Am

plitu

de

Blink detection

EOGBlink

times104

minus2

minus4

minus6

minus8

(a)

0 1000 2000 3000 4000 5000

0

1


Timestamp (s)

Mar

ker v

ecto

r val

ue

minus1

minus2

Mar

ker v

ecto

r val

ue

0 500 1000 1500 2000 2500

0

115

05


Timestamp (s)

minus05

minus15

minus1

(b)



119869 (119883119899) = 119868 (119883119899 119884) minus1

|119878|sum

119883119894isin119878

119868 (119883119899 119883119894) (3)


1119899minus1 119884) is the






Med (119891119894 119895) =1

119896

119896

sum

119903=1

(119883119903119894 isin class 119895) (4)



119863(119909119894119895Med (119891119895 119905)) =10038161003816100381610038161003816119909119894119895 minusMed (119891119895 119905)

10038161003816100381610038161003816 (5)


1198722 (119894 119895) = 1 if 1198721 (119894 119895) = 119862119894

0 if 1198721 (119894 119895) = 119862119894(6)


119862Score (119891119894) =1

119899

119899

sum

119903=1

1198722 (119903 119894) (7)








V119895119894119892 = 1199091198951199030119892 + 119865 times (1199091198951199031119892 minus 1199091198951199032119892)

where

119894 = 0 to NP minus 1

119895 = 1 to D minus 1




(8)


119906119895119894119892 = V119895119894119892 if rand (0 1) le 119862119903

119909119895119894119892 otherwise

where


119892 = population


(9)











Linear119870(119909119894 119909119895) = 119909119879

119894119909119895 (10)


120596 (120572) =

119899

sum

119894=1

120572119894 minus1

2

119899

sum

119894=1119895=1

120572119894120572119895119870(119909119894 119909119895) (11)

120596 =

119899

sum

119894=1

120572119910119894119909119894 (12)



119891 (119909new) = sign(119899

sum

119894=1

120572119910119894119870(119909119894119909new) + 119887) (13)



Base vector


New population




Populationvectors


7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8


X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1














Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















V119895119894119892 = 1199091198951199030119892 + 119865 times (1199091198951199031119892 minus 1199091198951199032119892)

where

119894 = 0 to NP minus 1

119895 = 1 to D minus 1




(8)


119906119895119894119892 = V119895119894119892 if rand (0 1) le 119862119903

119909119895119894119892 otherwise

where


119892 = population


(9)











Linear119870(119909119894 119909119895) = 119909119879

119894119909119895 (10)


120596 (120572) =

119899

sum

119894=1

120572119894 minus1

2

119899

sum

119894=1119895=1

120572119894120572119895119870(119909119894 119909119895) (11)

120596 =

119899

sum

119894=1

120572119910119894119909119894 (12)



119891 (119909new) = sign(119899

sum

119894=1

120572119910119894119870(119909119894119909new) + 119887) (13)



Base vector


New population




Populationvectors


7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8


X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1














Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Base vector


New population




Populationvectors


7

26 0 32 15 19421

2251734

62713

3611308231052416

33120

14319

2218

29

728

1235 3

1

4

3

29

8


X1g

X2g

V0g

V1g

V2g

++

Trial vector U0g

Pxg+1

X0g+1

X1g+1

XNPminus2g

XNPminus1g

VNPminus2g

VNPminus1g

X

XNPminus2g+1

XNPminus1g+1














Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


























Precision =TP

TP + FP (14a)

Recall = TPTP + FN

(14b)


(14c)


(14d)






102030405060708090

Activity

Prec

ision

()

DEFSmRMRCBFS







4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















4 Conclusions




References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Documents

Research Article Feature Selection in Classification of ...downloads.hindawi.com/journals/cmmm/2014/713818.pdf · analysis of EOG signals using clearness based features, minimum redundancy