feature selection and classifier parameters estimation for

Research ArticleFeature Selection and Classifier Parameters Estimation for EEGSignals Peak Detection Using Particle Swarm Optimization

Asrul Adam1 Mohd Ibrahim Shapiai2 Mohd Zaidi Mohd Tumari3

Mohd Saberi Mohamad4 and Marizan Mubin1

1 Applied Control and Robotics (ACR) Laboratory Department of Electrical Engineering Faculty of EngineeringUniversity of Malaya 50603 Kuala Lumpur Malaysia

2 Faculty of Electrical Engineering Universiti Teknologi Malaysia 81310 Johor Bahru Malaysia3 Faculty of Electrical and Electronic Engineering Universiti Malaysia Pahang 26600 Pekan Pahang Malaysia4 Faculty of Computing Universiti Teknologi Malaysia 81310 Johor Bahru Malaysia

Correspondence should be addressed to Asrul Adam asruladamsiswaumedumy

Received 18 June 2014 Accepted 30 July 2014 Published 19 August 2014

Academic Editor Shifei Ding

Copyright copy 2014 Asrul Adam et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Electroencephalogram (EEG) signal peak detection is widely used in clinical applications The peak point can be detected usingseveral approaches including time frequency time-frequency and nonlinear domains depending on various peak features fromseveral models However there is no study that provides the importance of every peak feature in contributing to a good andgeneralizedmodel In this study feature selection and classifier parameters estimation based on particle swarm optimization (PSO)are proposed as a framework for peak detection on EEG signals in time domain analysis Two versions of PSO are used in the study(1) standard PSO and (2) random asynchronous particle swarm optimization (RA-PSO) The proposed framework tries to find thebest combination of all the available features that offers good peak detection and a high classification rate from the results in theconducted experiments The evaluation results indicate that the accuracy of the peak detection can be improved up to 9990 and9859 for training and testing respectively as compared to the framework without feature selection adaptation Additionally theproposed framework based on RA-PSO offers a better and reliable classification rate as compared to standard PSO as it produceslow variance model

1 Introduction

The peak detection algorithms have significantly been usedon different types of biological signals such as electroocu-logram (EOG) electrocardiogram (ECG) and electroen-cephalogram (EEG) EOG signal is generated by human eyeECG signal is generated by heart EEG signal is generated bybrain The peak detection in the EOG signal has been usedfor detecting the eye blink [1 2] In the EOG based signal anumber of electrodes are placed around the eyes If the eyesmove in vertical direction positive or negative peak pointswill arise For the ECG signal peak detection is typically usedto detect the combination ofQ R and Swaves or the so-calledQRS complex [3]The QRS complex is a peak model for ECGsignal including Q-valley point R-peak point and S-valley

point Other important peak points in ECG signal are P-peakpoint and T-peak point The detection of the QRS complex iscritical part in numerous ECG signal processing system Thedifferent pattern of QRS complex will determine the patientheart syndrome Additionally the peak detection for the EEGsignal has been widely used to detect P300 response [4 5]and epilepsy response [6] P300 is a brain response measuredby electrodes covering the parietal lobe in the presence ofvisual and auditory stimuli A brain with chronic disorderwill respond with epilepsy Therefore the utilization of peakdetection algorithm for the biological signals is compatible inthis study

To date variety approaches of peak detection algorithmshave been proposed These algorithms can be categorizedinto four main approaches based on time domain [7ndash15]

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 973063 13 pageshttpdxdoiorg1011552014973063

2 The Scientific World Journal

frequency domain [16] time-frequency domain [10 17] andnonlinear [18] In time domain approach the peaks are ana-lyzed in time In frequency domain approach the peaks areanalyzed in frequency In time-frequency domain approachthe peaks are analyzed in both time and frequency domainIn nonlinear approach some statistical parameters of thepeaks are analyzed The general framework of peak detectionalgorithm usually involves several processes which are signalpreprocessing peak candidate detection feature extractionand classification Various signal preprocessing methodshave been employed such as data compression [19] wavelettransform [6] Kalman filter [20] and Hilbert transform [15]Two methods for peak candidate detection have been usedwhich are three point sliding windowmethod [8] and k-pointnonlinear energy operator (k-NEO) method [21] Variousfeature extraction techniques have been proposed which aremodel-based [21] wavelet analysis [22] template matching[23] and power spectra analysis [24] Several classifiers havebeen used which are rule-based [8 24] artificial neuralnetwork (ANN) [10 11 25 26] support vector machine(SVM) [7 27] and expert system [10] The highlightedpurposes in designing the framework are to achieve thehighest performance and to reduce the computational timeAlmost all studies in the EEG peak detection literature focuson the problem of detecting peaks in epileptic EEG signals Areview of peak detection algorithms that is employed to theepileptic EEG signal is presented in [28]Thepeak detection isjust a first step in epileptic event detectionThemain goal is todetermine the epileptic spikes not thewhole peaksThereforefor an epileptic event detection system the epileptic spikedetection performance not the peak detection performanceis the performance of interest

In time domain approach fourteen different peak featuresare recognized from different peak models [7ndash10] The peakmodel is a set of peak features that represents a peak by itsamplitude width and slope Most algorithms [7ndash13 21] intime domain approach consider different peak models andthe different styles of framework The peak model is chosenbased on the experiences of EEG expert To date there isno any peak detection framework that automatically findsthe finest existing peak model The use of the finest peakmodel will give a chance for the algorithm to achieve a goodperformance On the other hand the chosen peak model isnot necessarily suitable for different types of biological signalMoreover the finest peak model represents somemeaningfulinformation on the signal to be evaluated Therefore theadaptation of feature selection technique is important inthis study to automatically find the finest peak model Theutilization of feature selection on peak detection algorithmwill also reduce the computational time

In this study feature selection and classifier parame-ters estimation method based on standard particle swarmoptimization (PSO) and random asynchronous PSO (RA-PSO) algorithm are employed The process to find the finestpeak model and classifier parameter estimation is executedsimultaneously The peak features will be evaluated by a rule-based classifier The role of the classifier is to distinguishbetween peak point and non-peak point Rule-based classifieris employed due to the ability to provide an outstanding

interpretation for the obtained decisions [24] In addition theparameter values are tricky to be estimated manually A PSOalgorithm is considered to be appropriate for addressing theproblem based on the reason in which the feature selectionis a binary search problem and determination of classifierparameter is a continuous search problem [29]

11 Peak Model in Time Domain Analysis Peak model is a setof peak features that represents a peak by its amplitude widthand slope In time domain analysis fourteen different peakfeatures are recognized from different peak models [8ndash10]The earliest peak model was introduced by Dumpala et al in1982 [8] The peak model comprises four features which are(1) the amplitude between the magnitude of peak point andthe magnitude of valley point at the first half wave (2) thewidth between valley point of first half point and valley pointat second half wave (3) and (4) two slopes between a peakpoint and valley point in the first half wave and second halfwave A similar definition of the peak amplitude and slopesare also been used in [7 11 13]

An additional feature of peak amplitude and two featuresof peak width have been introduced by Acir et al [7 11]The additional peak amplitude is the amplitude between themagnitude of peak point and themagnitude of valley point ofthe second half wave The peak widths are the width betweenpeak point and valley point of first half wave and secondhalf wave The total features that are introduced by Acir etal are six features Acir et al did not use the width featurethat was introduced by Dumpala et al A similar definition ofthe peak amplitudes widths and slopes has also been used in[21] In [21] an additional peak feature is added with a set offeatures that is introduced in [7 11] which is the area of peakHowever the definition of area integration is not presented inthe paper

In addition Liu et al [10] have introduced eleven peakfeatures The proposed peak model consists of four ampli-tudes (1) the amplitude between themagnitude of peak pointand the magnitude of valley point at the first half wave (2)the amplitude between the magnitude of peak point and themagnitude of valley point of the second half wave (3) theamplitude between themagnitude of peak and themagnitudeof turning point at the first half wave and (4) the amplitudebetween themagnitude of peak and themagnitude of turningpoint at the second half wave The turning point is definedas the point where the slope decreases more than 50 ascompared to the slope of the preceding pointThemodel alsoconsists of three widths (1) the width between valley point atfirst half point and valley point at second half wave (2) thewidth between turning point at first half wave and turningpoint at second half wave and (3) the width between halfpoint at first half wave and half point at second half waveThere are four slopes that are also measured (1) and (2) twoslopes between a peak point and valley point in the first halfwave and second half wave (3) and (4) two slopes betweenpeak point and turning point at first half wave and secondhalf wave

Another peak model consists of four features which hasbeen proposed by Dingle et al [9] The peak amplitude is

The Scientific World Journal 3

Table 1 Summary of different peak models on different style of framework

Peak model Type of signal Description of framework

Dumpala et al(1982) [8]

Electrical controlactivity (ECA)

The theory of maxima and minima using three-point sliding window approach has beenapplied to detect a candidate peak Two flowcharts of peak detection have been proposed Apredicted peak can be identified if the feature values satisfied the decision threshold values Thestrength and weakness of the proposed approach are described as follows (1) strength theauthors claimed that the proposed peak detection algorithm can be used for other biologicalsignals (2) weakness the utilization of peak-to-peak amplitude on the peak model is hard todistinguish between noise and actual peak In addition large variation of peak width in thesignal may drop the classification performance

Dingle et al(1993) [9] Epileptic EEG

Based on the defined peak model the features are grouped into two (1) epileptiform transientparameters and (2) background activity parameters Two-threshold systems have beenemployed to detect a candidate peak or candidate epileptiform transient Expert system whichconsidered both spatial and temporal contextual information has been used to reject theartifacts and classify the transient events The strength and weakness of the proposed approachare described as follows (1) strength moving average amplitude is good in rejecting false peakpoints The employed features are claimed to offer good performance in the proposed expertsystem (2) weakness inconsistency of feature slope information as the proposed work claimedthat the proposed framework fails to provide slope information

Liu et al (2002)[10] Epileptic EEG

Wavelet transform has been used to decompose the EEG signal Based on the decomposedsignals and the defined peak model seven features are calculated These features are used as theinput of ANN classifier Expert system which considered both spatial and temporal contextualinformation has been used to reject the artifact Several heuristic rules have been employed todistinguish the type of artifact After all artifacts are recognized and rejected the decision willbe made to classify the epileptic events The strength and weakness of the proposed approachare described as follows (1) strength the employed features is claimed to offer goodperformance in the proposed expert system (2) weakness it considers that almost all thefeatures may deteriorate the classification performance

Acir et al (2005)[11] Epileptic EEG

A three-stage procedure based on ANN is proposed for the detection of epileptic spikes TheEEG signal is transformed into time-derivative signal Several rules have been used to detect apeak candidate The features of peak candidate are calculated based on the defined peak modelThese features are fed into two discrete perceptron classifiers to classify into three groupsdefinite peak definite non-peak and possiblepossible non-peak The peak that belongs in thethird group is going to be further processed by nonlinear classifier The strength and weaknessof the proposed approach are described as follows (1) strength the employed features areclaimed to offer good performance in the proposed system (2) weakness inconsistency offeature slope information as the proposed work claimed that the proposed framework fails toprovide slope information

Acir (2005) [26] Epileptic EEG

A two-stage procedure based on a modified radial basis function network (RBFN) is proposedfor the detection of epileptic spikes The EEG signal is transform into time-derivative signalSeveral rules have been used to detect a peak candidate The features of peak candidate arecalculated based on the defined peak model These features are fed into discrete perceptronclassifiers to classify into two groups definite non-peak and peak-like non-peak The peak thatbelongs to the second group requires further process by modified RBFN classifier The strengthand weakness of the proposed approach are described as follows (1) strength the employedfeatures are claimed to offer good performance in the proposed system (2) weaknessinconsistency of feature slope information as the proposed work claimed that the proposedframework fails to provide slope information


A two-stage procedure is proposed for the detection of epileptic spike k-NEO has been used todetect a candidate peak The peak features are calculated based on the defined peak modelThese features are then used as the input of the AdaBoost classifier The strength and weaknessof the proposed approach are described as follows (1) strength the peak model considersfeature based on peak area (2) weakness the definition of area integration is not presented inthe paper

the difference between the peak point and the floating meanThe floating mean is the average EEG which is centeredat the peak point that is also called moving average curve(MAC) [12] The width is calculated based on the differencebetween the valley point at the first half wave and the

valley point at the second half wave The two slopes are theslopes between a peak point and valley point in the firsthalf wave and second half wave Summary of different peakmodels on different style of framework is briefly describedin Table 1 The strength and weakness are also highlighted


Rule-based

classifier

Rule-based

classifier

Peakcandidatedetection

Featureextraction

Predictedpeak pointlocations

Peak candidatedetection

Featureextraction

FilteredEEG

signalPeak features

FilteredEEG

signal

Training

Testing

PSOalgorithm

Thresholds

Feature selection andparameters estimation

Finest peakmodel

Optimal decisionthresholds

Figure 1 Feature selection and parameters estimation framework for peak detection algorithm

in Table 1 Generally the authors claimed that the selectedpeak feature offers good classification performance on theproposed framework However the previous works did notprovide the justification on the selected features

2 Methodology

Figure 1 shows the framework of the proposed techniquesfor EEG signal peak detection There are two phases of theprocess which are training and testing phases The trainingphase is firstly run to find the finest peak model and theoptimal decision threshold values Next the testing phase isutilized for unseen EEG signal

The framework can be divided into four stages peakcandidate detection features extraction of peak candidatefeature selection and parameters estimation and classifica-tion In the first stage the detection of peak candidates isperformed to differentiate between a peak candidate and anon-peak candidate The second stage is the extraction ofpeak candidate features In the third stage PSO algorithm isadapted during the training phase for feature selection andclassifier parametersrsquo estimation Finally the peak candidatesare classified between predicted peak and predicted non-peakat particular locations by rule-based classifier

21 Peak Candidate Detection The first step to detect peaksis to find candidate peaks Consider a discrete-time signal119909(119868) of 119871 points The 119894th candidate peak point PP

119894 as shown

in Figure 2 is identified using three-points sliding windowmethod [8] Those three-points are denoted as 119909(119868 minus 1) 119909(119868)and 119909(119868 + 1) for 119868 = 1 2 119871 A candidate peak point isidentified when 119909(PP

119894minus 1) lt 119909(PP

119894) gt 119909(PP

119894+ 1) and two

associated valley points VP1119894and VP2

119894 are in between as

shown in Figure 2 Both valley points exist when 119909(VP1119894minus

PPi

HP1i

TP1i

VP1i

VP2i

HP2i

TP2i

MAC (PPi)

Figure 2 Model-based parameters

1) gt 119909(VP1119894) lt 119909(VP1

119894+ 1) and 119909(VP2

119894minus 1) gt 119909(VP2

119894) lt

119909(VP2119894+ 1)

22 Feature Extraction Based on the existing peak modelsthe total peak features are fourteen The peak features of apeak candidate are calculated based on the eightmodel-basedparameters as shown in Figure 2 The parameters consistof the 119894th candidate peak point PP

119894 the two associated

valley points VP1119894and VP2

119894 the half point at first half

wave (HP1119894) the half point at second half wave (HP2

119894) the

turning point at first half wave (TP1119894) the turning point

at second half wave (TP2119894) and the moving average curve

(MAC(PP119894)) The peak features can be categorized into three

groups amplitude width and slope There are five differentamplitudes five different widths and four different slopesthat can be calculated based on the model-based parametersAll equations and description of peak features are tabulatedin Table 2 Referring to Table 3 the peak model which is


Table 2 Equations and descriptions of peak features

Peak feature Equation Description

Amplitudes

1198911=1003816100381610038161003816119909(PP119894) minus 119909(VP1

119894)1003816100381610038161003816

Amplitude between the magnitude of peak and the magnitude of valley at the firsthalf wave

1198912=1003816100381610038161003816119909(PP119894) minus 119909(VP2

119894)1003816100381610038161003816

Amplitude between the magnitude of peak and the magnitude of valley of thesecond half wave

1198913=1003816100381610038161003816119909(PP119894) minus 119909(TP1

119894)1003816100381610038161003816

Amplitude between the magnitude of peak and the magnitude of turning point atthe first half wave

1198914=1003816100381610038161003816119909(PP119894) minus 119909(TP2

119894)1003816100381610038161003816

Amplitude between the magnitude of peak and the magnitude of turning point atthe second half wave

1198915=1003816100381610038161003816119909(PP119894) minusMAC(PP

119894)1003816100381610038161003816 Amplitude between the magnitude of peak and the magnitude of moving average

Widths

1198916=1003816100381610038161003816VP1119894minus VP2

119894

1003816100381610038161003816 Width between valley point of first half point and valley point at second half wave

1198917=1003816100381610038161003816PP119894minus VP1

119894

1003816100381610038161003816 Width between peak point and valley point at first half wave

1198918=1003816100381610038161003816PP119894minus VP2

119894

1003816100381610038161003816 Width between peak point and valley point of second half wave

1198919=1003816100381610038161003816TP1119894minus TP2

119894

1003816100381610038161003816

Width between turning point at first half wave and turning point at the second halfwave

11989110=1003816100381610038161003816HP1119894minusHP2

119894

1003816100381610038161003816 Width between half point of first half wave and half point of second half wave

Slopes

11989111=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP1

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816

Slope between a peak point and valley point at the first half wave

11989112=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP2

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816

Slope between a peak point and valley point at the second half wave

11989113=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP1

119894)

PP119894minus TP1

119894

10038161003816100381610038161003816100381610038161003816

The slope between peak point and turning point at the first half wave

11989114=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP2

119894)

PP119894minus TP2

119894

10038161003816100381610038161003816100381610038161003816

The slope between peak point and turning point at the second half wave

Table 3 List of different peak models and sets of features

Peak model Set of features Number of featuresDumpala et al (1982) [8] 119891

1 1198916 11989111 11989112

4Acir et al (2005) [7 11 26] 119891

1 1198912 1198917 1198918 11989113 11989114

6Liu et al (2002) [10] 119891

1 1198912 1198913 1198914 1198916 1198919 11989110 11989111 11989112 11989113 11989114

11Dingle et al (1993) [9] 119891

5 1198916 11989111 11989112

4

introduced byDumpala et al [8] andDingle et al [9] consistsof four featuresThe peakmodel which is specified by Acir etal [7 11] consists of six features The peak model which isspecified by Liu et al [10] consists of eleven features

23 Feature Selection and Parameters Estimation Using Par-ticle Swarm Optimization In this stage the peak featuresand classifier parameters are simultaneously found using twodifferent PSO algorithms which are standard PSO and RA-PSO algorithms At the end of this stage the finest peakmodel and the optimal classifier parameters are obtainedTheoptimal classifier parameters represent the optimal decisionthreshold values

The PSO algorithm was firstly introduced by Kennedyand Eberhart in 1995 [30] The PSO algorithm has beennumerously enhanced fundamentally [31 32] and appliedin many fields [33ndash35] Fundamentally the PSO algorithmfollows several steps as described in Algorithm 1 (1) initial-ization (2) calculation of the fitness function (3) updatingthe personal best (pbest) for each particle and global best(gbest) (4) updating the particlersquos velocity and the particlersquos

(1) Initialization(2) while not stopping criteria do(3) for each 119894th particle in a population do(4) calculate fitness function(5) update pbest and gbest(6) end for(7) for each particle in a population do(8) update the 119894th particlersquos velocity and(9) update the 119894th particlersquos position(10) end for(11) end while

Algorithm 1 Standard PSO Algorithm

position and (5) performing termination based on a stoppingcriterion

In PSO particles search for the best solution and updatethe position information from iteration to iteration Eachparticle in the population consists of a vector position andvector velocity in 119889 dimension The position of particle 119894 at


(1) Initialization(2) while not stopping criteria do(3) while not meet 119873 times do(4) Randomly choose 119894th particle in a population(5) for 119894th particle in a population do(6) calculate fitness function(7) update pbest and gbest(8) update the 119894th particlersquos velocity and(9) update the 119894th particlersquos position(10) end for(11) end while(12) end while

Algorithm 2 Random Asynchronous PSO (RA-PSO)

Table 4 Representation of particle position

Particle Peak features (binary type) Thresholds (continuous type)1 2 sdot sdot sdot 119899119891 119899119891 + 1 119899119891 + 2 sdot sdot sdot 119899119891 times 2

119904119896

119894119909119896

1198941119909119896

1198942sdot sdot sdot 119909

119896

119894119889119909119896

1198941119909119896


119896

119894119863

iteration 119896 is denoted as 119904119896119894= 119909119896

1198941 119909119896

1198942 119909119896

1198943 119909

119896

119894119889 while

the velocity of particle 119894 at iteration 119896 is denoted as V119896119894=

V1198961198941 V1198961198942 V1198961198943 V119896

119894119889 The pbest of particle 119894 is represented as

119901119896

119894= 119901119896

1198941 119901119896

1198942 119901119896

1198943 119901

119896

119894119889 and the gbest is denoted as 119901119896

119892=

119901119896

1198921 119901119896

1198922 119901119896

1198923 119901

119896

119892119889 To obtain the updated position of a

particle 119904119896+1119894

each particle changes its velocity as the follows

V119896+1119894= 120596V119896119894+ 11988811199031(119901119896

119894minus 119909119896

119894) + 11988821199032(119901119896

119892minus 119909119896

119894) (1)

where 1198881is a cognitive coefficient 119888

2is a social coefficient 119903

1

and 1199032are random values [0 1] and 120596 is a decrease inertial

weight [36 37] calculated as follows

120596 = 120596max minus (120596max minus 120596min119896max

) times 119896 (2)

where 120596max and 120596min denote the maximum and minimumvalues of inertia weight respectively and 119896max is the maxi-mum iteration Then the particlersquos position is updated basedon (3) Note that this equation is only valid for continuousversion of PSO algorithm

119904119896+1

119894= 119904119896

119894+ V119896+1119894 (3)

For a binary version of PSO [38] the particle position isupdated based on the following equation

119879 (V119896+1119894) =

10038161003816100381610038161003816tanh (V119896+1

119894)

10038161003816100381610038161003816 (4)

119904119896+1

119894=

(119904119896

119894)

minus1

if rand lt 119879 (V119896+1119894)

119904119896

119894rand ge 119879 (V119896+1

119894)

(5)

Equation (4) is a transfer function which is the main partof the binary version Several studies have proven that thistransfer function significantly improves the performance of

the standard binary PSO Equation (5) is used to update theparticle position according to the given rules where 119904119896

119894and

V119896119894represent the vector position and velocity of 119894th particle at

iteration 119896 and (119904119896119894)

minus1

is the complement of 119904119896119894 The particle

position maintains the current position when the velocity islower than random value and its complement the positionwhen the velocity is greater than random value This methodhas been introduced by Mirjalili and Lewis (2013) that is alsonamed as v-shaped transfer function [39]

Synchronous update in standard PSO algorithm indicatesthat all particles move to their new position after all particlesare evaluated as described in Algorithm 1 However in RA-PSO [40] a particle immediately updates its position afterit is evaluated without the need to wait until the evaluationof all particles is completed Moreover an 119894th particle in apopulation is randomly chosen with a total 119873 times before119894th particle is evaluated 119873 is the total number of particlesSome particles might be chosen more than once while someparticles might not be chosen at all The RA-PSO algorithmis described in Algorithm 2

To perform the feature selection and parameters esti-mation simultaneously both versions of PSO algorithm areemployed to the standard PSO and RA-PSO algorithmsTable 4 illustrates the representation of particle position The119894th particle at iteration 119896 119904119896

119894 in PSO represents two types

of dimensions which are binary and continuous type ofdimension [29] 119904119896

119894= 119909

119896

1198941 119909119896

1198942 119909

119896

119894119889 119909119896

1198941 119909119896

1198942 119909

119896

119894119863

The 119889 = 1 2 3 119899119891 is a 119889th dimension of binary type andthe119863 = 119899119891 + 1 119899119891 + 2 119899119891 + 3 119899119891 times 2 is a119863th dimensionof continuous type 119899119891 is the total number of peak featuresThe particle dimension is a two times number of featuresThenumber of thresholds is equal of the number of features

In the initialization stage of PSO algorithm some of theparameters are initialized (1) the initial PSO parameters and(2) the initial particle position The initial PSO parameters


consist of the maximum inertia weight 120596max the minimuminertia weight 120596min the velocity clamping Vmax the velocityvector for each particle the pbest score for each particle gbestscore the cognitive component 119888

1 and the social component

1198882 The random values 119903

1and 1199032 are randomly distributed

values from 0 to 1 All particles are randomly placed withinthe search space

For the calculation of fitness function geometric mean(Gmean) is employed The Gmean is calculated as follows

TPR = TPTP + FN

TNR = TNTN + FP

Gmean = radicTPR times TNR

(6)

where true peak (TP) is a correctly detected peak point truenon-peak (TN) is a correctly detected non-peak point falsepeak (FP) is a wrongly detected the non-peak point falsenon-peak (FN) is a wrongly detected peak point TPR is a truepeak rate and TNR is a true non-peak rate

24 Rule-Based Classifier A rule-based classifier is employedto distinguish whether the candidate peak is a true peak ortrue non-peak from the extracted features Each feature hasa corresponding threshold value in the classification processGiven a set of features a true peak only can be identified ifall the feature values are greater than or equal to the decisionthreshold values Otherwise the candidate peak belongs totrue non-peak The form of the rule is

IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873

THEN Candidate Peak is a True Peak(7)

where 119891119894is denoted as a one of sixteen peak features th

119894is

denoted as one of the decision threshold values of this peakfeature and true peak is predicted peak at a particular peakpoint location

3 Experimental Setup

In this section two experiments are conducted for peakdetection of EEG signal For first experiment the frameworkis executed without feature selection For second experimentthe experiment is executed with feature selection The exper-imental protocols are discussed in the next subsection Thetraining and testing EEG signal are prepared to evaluate theperformance of the proposed framework Then the resultsare discussed and analyzed

Each experiment is conducted in 10 independent runsFor each run 30 particles are used to perform featureselection and parameters estimation For each particle thetotal number of dimensions is depending on the number offeatures in a feature set The maximum iteration was set to1000 For the initial value of PSO parameters the maximuminertia weight 120596max is 09 and the minimum inertia weight120596min is 04 The cognitive component 119888

1 and the social

Table 5 Parameters setting of standard PSO and RA-PSO algo-rithms

Initial PSO parametersParameters ValueDecrease inertia weight 120596 09sim04Cognitive component 119888

12

Social component 1198882

2Random value 119903

1and 1199032

Random number [0 1]Velocity vector for each particle 0Initial pbest score for each particle 0Initial gbest score 0Range of search space for 119899119891 + 1 to 119899119891 + 5 [0 30]

Range of search space for 119899119891 + 6 to 119899119891 + 12 [0 78125]

Range of search space for 119899119891 + 13 to 119899119891 times 2 [0 2416]

component 1198882 are set to 2 These values are proposed by

Shi and Eberhart in 1999 [41] The random values 1199031and

1199032 are randomly distributed values from 0 to 1 The velocity

clamping Vmax for binary version is set to 6 [39]The velocityvector for each particle the pbest score for each particle andgbest score is set to 0The parameters setting of standard PSOand RA-PSO algorithms are tabulated in Table 5

31 Experimental Protocols This study uses the eye move-ment EEG signal as a case study to evaluate the proposedframeworkThe observation of the eyemovement EEG signalindicates that the most observable signal pattern is the peakpoint which signifies the brain response on eye movementsThe known peak point locations through the response ofthe brain can be translated into an output for examplewheelchair movement

The experimental protocol to acquire this EEG signalwas reviewed and approved by the Medical Ethic Committee(MEC) in theUniversity ofMalayaMedical Centre (UMMC)The subject gave a written consent prior to the data collectionsessionThis EEG signal was acquired in the Applied Controland Robotic (ACR) Laboratory Department of ElectricalEngineering Faculty of Engineering University of MalayaMalaysia A healthy subject was involved voluntarily in thisdata collection session who is a postgraduate student in theFaculty of Engineering

The EEG signal recording was conducted using thegMOBIlab portable signal acquisition system The EEGsignal was recorded from C4 channel The EEG signal ofchannel CZ was used as a reference The ground electrodewas located on the forehead The electrode was placed usingthe 10ndash20 international electrode placement system Thesampling frequency was set to 256Hz

Before the session begins the subject was advised to getgood rest Thus he can give full focus during the sessionThe subject was also advised to wash his hair During thedata collection session the subject was required to be readywithin 0 to 4 seconds for waiting for an external cue The cueis a command for a subject to move their eyes to the rightposition Within the standby time the subject is required notto move their eyes into a frontal position


0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)

Figure 3 Filtered EEG signal

Table 6 Signal specifications

Specification Channel C4Total sampling point 10240Total length signal (second) 40Number of peak points in the signal 40Sampling frequency (Hz) 256

When the time is exactly 5 seconds the external cueappears on the screen monitor The instruction allows thesubject to move back their eyes in a frontal position Theexternal cue appears for 40 times The total length of EEGrecording is 40 seconds As a cleanliness procedure theelectrodes and head-cap that are used in the session werewashed The filtered EEG signal is shown in Figure 3 Fortylocations of definite peak points are highlighted in the redcircle The next process is to prepare the training and testingdata

From the data collection 40 definite peak point locationshave been identified by EEG expert In 40-second signal thereare 10240 sampling points 119909(119868)There are only 40 peak pointsand the remaining of 10200 sampling points are the non-peak points For preparing the training and testing signal thetraining signal is selected from 1 to 5120 sampling pointswhilethe remaining EEG signal is used for testing signalThe signalspecification is summarized in Table 6

4 Results and Discussions

To evaluate the proposed framework for training and testingphase four different measures are used including the averageGmean themaximumGmean theminimumGmean and thestandard deviation (STDEV)

41 Results of Peak Detection Algorithm without FeatureSelection Four peak models are employed for evaluatingthe peak models performance in the proposed frameworkThe training and testing performance based on those four

different measures for each model is shown in Table 7 Thestandard PSO algorithm is used to find the optimal thresholdvalues for each peak model The obtained results for eachpeak model are compared with the results of peak detectionalgorithm and the feature selection framework based onstandard PSO Notably in this section only standard PSO isconsidered in the peak detection algorithm without featureselection framework

Referring to Table 7 the training performance for aver-age maximum minimum and STDEV is 8401 89158058 and 443 for Dumpala et alrsquos peak model 7448059 6708 and 371 for Acir et alrsquos peak model and9098 9476 8366 and 551 for Dingle et alrsquos peakmodel respectively The testing performance for averagemaximum minimum and STDEV is 8122 9183 7415and 913 for Dumpala et alrsquos peak model 6859 77435477 and 697 for Acir et alrsquos peak model and 88789475 7744 and 798 for Dingle et alrsquos peak modelrespectively

Overall the average performance of the training phase forDumpala et alrsquos peakmodel Acir et alrsquos peakmodel andDin-gle et alrsquos peakmodel is greater than the average performanceof their testing phase However for the peakmodel Liu et alrsquospeak model will give zero percent performance for trainingand testing phase This result indicates the limitation of rule-based classifier when dealing with both feature sets Duringthe training process on the feature sets the particles in thePSO algorithm do not meet the optimum decision thresholdvalues and the particlesmight also be trapped at local optimaBased on the preceding rule a true peak only can be identifiedif all the feature values are greater than or equal to the decisionthreshold values So if one of the feature values does notsatisfy the decision threshold value the classifier will decidethe peak candidate as a non-peak point When this happensto all peak candidates the TP is equal to zeroGmeanwill givezero percent performance even if TN is equal to some valuesThe end results indicate the employment of the presented ruleis only valid for Dumpala et alrsquos peak model Acir et alrsquos peakmodel and Dingle et alrsquos peak model

Compared to the test average performance of the peakmodels the highest test performance is obtained by Dingle etalrsquos peakmodel which is 8878 then follows by Dumpala etalrsquos peak model which is 8122 The worst test performanceis obtained by Acir et alrsquos peak model which is 6859 Itcan be concluded from the findings of experimental resultsthe finest peak model for the filtered EEG signal is Dingle etalrsquos peak model and the worst peak model for the filteredEEG signal is Acir et alrsquos peak model True peak rate andtrue non-peak rate of test performance are shown in Table 8It can be concluded that from the finding experimentalresults the chosen peakmodels limit the designed frameworkto obtain the best accuracy Therefore the feature selectiontechnique using standard PSO is employed into the designedframework

42 Results of PeakDetectionAlgorithmwith Feature SelectionThe results of peak detection algorithmwith feature selectionare categorized into two subsections which are the results of


Table 7 Training and testing performance of peak detection for each peak model (without feature selection)

Peak model Training () Testing ()Average Max Min STDEV Average Max Min STDEV

Dumpala et al (1982) [8] 8401 8915 8058 443 8122 9183 7415 913Acir et al (2005) [7 11 26] 744 8059 6708 371 6859worst 7743 5477 697Liu et al (2002) [10] 0 0 0 0 0 0 0 0Dingle et al (1993) [9] 9098 9476 8366 51 8878best 9475 7744 798

Table 8 TPR and TNR test results for EEG signal (without featureselection)

Peak model TPR () TNR ()Dumpala et al (1982) [8] 650 997Acir et al (2005) [7 11 26] 500 999Liu et al (2002) [10] 00 00Dingle et al (1993) [9] 800 993

feature selection using standard PSO and the results of featureselection using RA-PSO Also the results from the two PSOalgorithms in the proposed framework are discussed

421 Feature Selection Using Standard PSO The feature setsof 10 runs using the standard PSO algorithm are shownin Table 9 The result shows the variety of the optimalcombination of features that give the higher classificationperformance mostly higher than 9969 The maximumtraining accuracy is 9998Themost significant peak featureis the feature 119891

5because all the 10 runs appear as a selected

feature by PSO Feature 1198915is the amplitude that is calculated

from the difference between peak points (PP) and movingaverage curve (MAC) Another most significant feature isfeature 119891

2 which is the calculated amplitude between a peak

point and valley point at the second haft wave The feature1198916is chosen 4 times The feature 119891

6is chosen 4 times The

features 1198914and 119891

9are chosen 2 times The feature 119891

10is only

selected at 9th runBased on the results in Table 9 the combination of peak

features (1198912 1198915 and 119891

6) appears 4 times the combination

of peak features (1198912 1198915 and 119891

9) appears 2 times and the

combination of peak features (1198912and 119891

5) appears 2 times

Therefore there are 3 optimal combinations of features thatcan be chosen

Table 10 has the optimal threshold values for the optimalcombination of the featuresThe threshold values are selectedbased on the selected peak features that are highlighted in thetable

The average of training and testing results of 10 runs usingstandard PSO algorithm is tabulated in Table 11The results ofstandard PSO show the average training accuracy is 9991The maximum training accuracy is 9998 The minimumtraining accuracy is 9969 and the standard deviation is807 On the other hand the testing accuracy is 9373Themaximum testing accuracy is 9992 The minimum testingaccuracy is 7741

In terms of peak and the non-peak rate (TP and TN) fortraining results the classifier accurately predicted all 20 peakpoints and 5113 non-peak points The results also show thatthe classifiermisclassified 27 non-peak pointsThemaximumof the true peak point is 20 and true non-peak point is 5118The minimum of true peak point is 20 and true non-peakpoint is 5109

For testing results the classifier accurately predicted 18peak points and 5110 non-peak points The maximum of thetrue peak point is 20 and true non-peak point is 5114 Theminimum of true peak point is 12 and true non-peak point is5106 In general the average testing result that correspondsto the selected peak features using the proposed featureselection framework is greater than the average testing resultof Dinglersquos peak model which is 9373 and 8878 Thefeature set of the Dinglersquos peak model is 119891

5 1198916 11989111 and 119891

12

while the feature set that gives a higher training performancein this experiment is 119891

2and 119891

5

However the proposed framework based on standardPSO produces slightly high variance model as it measuresfrom the STDEV index The STDEV is evaluated for mea-suring the algorithm consistency where lowest STDEV valueindicates a good generalization algorithm Based on theresults of the STDEV in Table 13 the STDEV values ofthe standard PSO are 807 and 718 for training andtesting respectively This results show that the high standarddeviation of the accuracy is recorded between maximumand minimum of classification rate The experimental resultsare reasonable due to the limitation of the standard PSOalgorithm

422 Feature Selection Using RA-PSO Table 12 shows thefeature selection results of 10 runs based on the RA-PSOalgorithm The feature set was highlighted of each run Thethreshold values for all selected features are also given inTable 13The highestGmean value of training phase is 9991The significant peak features are119891

5and1198918The corresponding

threshold values are 920 and 4 Note that feature 1198915is the

amplitude that is calculated from the difference between peakpoints (PP) andmoving average curve (MAC) Another mostsignificant feature is feature 119891

8 which is the width between

peak point and valley point of second half wave The features1198912 1198914 and 119891


12is only

selected at second runThree similar results were obtained out of ten runs Other

significant feature sets that are obtained in this result are thecombination of peak features (119891

2and 119891

5) and (119891

4and 119891

5)

These feature sets also appear 3 times


Table 9 Training results the feature sets of 10 runs using standard PSO

RunPeak features

Amplitudes Widths Slopes Gmean ()11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998

Table 10 Training results the optimal decision threshold values of 10 runs using standard PSO

Run Optimal threshold valuesth1 th2 th3 th4 th5 th6 th7 th8 th9 th10 th11 th12 th13-th14

1 mdash 040 mdash mdash 907 mdash mdash mdash 9 mdash mdash mdash mdash2 mdash mdash mdash 027 920 mdash mdash mdash mdash mdash mdash mdash mdash3 mdash 124 mdash mdash 927 mdash mdash mdash 17 mdash mdash mdash mdash4 mdash 037 mdash mdash 893 12 mdash mdash mdash mdash mdash mdash mdash5 mdash 043 mdash mdash 918 12 mdash mdash mdash mdash mdash mdash mdash6 mdash 125 mdash mdash 1134 mdash mdash mdash mdash mdash mdash mdash mdash7 mdash 093 mdash mdash 907 11 mdash mdash mdash mdash mdash mdash mdash8 mdash 038 mdash mdash 910 8 mdash mdash mdash mdash mdash mdash mdash9 mdash mdash mdash 043 913 mdash mdash mdash mdash 8 mdash mdash mdash10 mdash 090 mdash mdash 1007 mdash mdash mdash mdash mdash mdash mdash mdash

Table 11 Average training and testing results of 10 runs with feature selection using standard PSO

Algorithm Results Training TestingGmean () TN FP TP FN Gmean () TN FP TP FN

Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718

Table 12 Training results the feature sets of 10 runs using RA-PSO

RA-PSO Peak featuresAmplitudes Widths Slopes Gmean ()

Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990


Table 13 Training results the optimal decision threshold values of 10 runs using RA-PSO

Run Optimal threshold values using RA-PSOth1 th2 th3 th4 th5 th6 th7 th8 th9 th10 th11 th12 th13-th14

1 mdash mdash mdash mdash 920 mdash mdash 4 mdash mdash mdash mdash mdash2 mdash mdash mdash mdash 921 mdash mdash mdash mdash mdash mdash 06 mdash3 mdash mdash mdash 038 921 mdash mdash mdash mdash mdash mdash mdash mdash4 mdash mdash mdash 022 920 mdash mdash mdash mdash mdash mdash mdash mdash5 mdash mdash mdash mdash 920 mdash mdash 4 mdash mdash mdash mdash mdash6 mdash 036 mdash mdash 904 mdash mdash mdash mdash mdash mdash mdash mdash7 mdash 039 mdash mdash 922 mdash mdash mdash mdash mdash mdash mdash mdash8 mdash mdash mdash 025 920 mdash mdash mdash mdash mdash mdash mdash mdash9 mdash 027 mdash mdash 919 mdash mdash mdash mdash mdash mdash mdash mdash10 mdash mdash mdash mdash 920 mdash mdash 4 mdash mdash mdash mdash mdash

Table 14 Average training and testing results of 10 runs with feature selection using RA-PSO


RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133

Table 14 shows the average training and testing results of10 runs with feature selection using RA-PSO algorithm TheaverageGmean value of the RA-PSO algorithm is 9990 and9859 for training and testing respectively The maximumGmean value of the RA-PSO algorithm is 9991 and 9986for training and testing respectively The minimum Gmeanvalue of the RA-PSO algorithm is 9987 and 9733 fortraining and testing respectively


For testing results the classifier accurately predicted 19peak points and 5106 non-peak points The maximum of thetrue peak point is 20 and true non-peak point is 5107 Theminimum of true peak point is 19 and true non-peak point is5103

As compared to the framework using standard PSO RA-PSO is found to offer lower variance model The recordedSTDEV values of the RA-PSO are 115 and 133 for trainingand testing respectively Therefore the RA-PSO may offer areliable and reasonable model as compared to standard PSOwith consistent classification rate

5 Conclusions

In this study the framework of feature selection and param-eters estimation is proposed for EEG signal peak detectionalgorithm The proposed framework involves peak candidate

detection feature extraction feature selection and classifica-tion The framework is developed based on PSO algorithmand a rule-based classifier In general the binary PSO basedalgorithm was utilized for selecting the peak features whilethe continuous PSO based algorithm was utilized for opti-mizing the classifier parameters Two PSO based algorithmsare employed in the proposed framework (1) standard PSOand (2) RA-PSO Fourteen peak features were employedin this study All these peak features were taken from theexisting peak models in the time domain approach Theavailable peak features are then automatically selected incombinatorial form using the proposed framework Based onthe experiment results of peak detection algorithm withoutfeature selection the best peak model is Dingle et alrsquos [9]peak model where the highest performance obtained is8878 Meanwhile the experimental results with featureselection show the proposed framework with standard PSOcan further improve the Dingle et alrsquos model However therecorded results are inconsistent due to high variances of theclassification accuracy The unreliability of the standard PSOcan be further improved based on the proposed frameworkusing RA-PSO In general the proposed feature selectiontechnique offers a better performance as compared to anypeak models without feature selection For future work theproposed framework will be employed in more case studiesand will invent more classification methods

Conflict of Interests

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


Acknowledgments

This project is funded by the Ministry of Education Malaysiafor High Impact Research Grant (UM-D000016-16001) Uni-versity of Malaya Research Acculturation Grant Scheme(RDU121403) Universiti Malaysia Pahang FundamentalResearch Grant Scheme (VOT 4F331) Universiti TeknologiMalaysia and MyPhD scholarship from Ministry of Educa-tion Malaysia The authors would like to thank the Facultyof Engineering the University of Malaya for supporting thisresearch The authors also would like to acknowledge theEditor and anonymous reviewers for their valuable commentsand suggestions

References

[1] 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

[2] H Zeng and A G Song ldquoRemoval of EOG artifacts from EEGrecordings using stationary subspace analysisrdquo The ScientificWorld Journal vol 2014 Article ID 259121 9 pages 2014

[3] R Tafreshi A Jaleel J Lim andL Tafreshi ldquoAutomated analysisof ECG waveforms with atypical QRS complex morphologiesrdquoBiomedical Signal Processing andControl vol 10 pp 41ndash49 2014

[4] N Xu X Gao B Hong X Miao S Gao and F Yang ldquoBCIcompetition 2003mdashdata set IIb enhancing P300wave detectionusing ICA-based subspace projections for BCI applicationsrdquoIEEE Transactions on Biomedical Engineering vol 51 no 6 pp1067ndash1072 2004

[5] Q G Ma and Q Shang ldquoThe influence of negative emotionon the Simon effect as reflected by P300rdquo The Scientific WorldJournal vol 2013 Article ID 516906 6 pages 2013

[6] K P Indiradevi E Elias P S Sathidevi S DineshNayak and KRadhakrishnan ldquoA multi-level wavelet approach for automaticdetection of epileptic spikes in the electroencephalogramrdquoComputers in Biology and Medicine vol 38 no 7 pp 805ndash8162008

[7] N Acir and C Guzelis ldquoAutomatic spike detection in EEGby a two-stage procedure based on support vector machinesrdquoComputers in Biology and Medicine vol 34 no 7 pp 561ndash5752004

[8] S R Dumpala S Narasimha Reddy and S K Sarna ldquoAnalgorithm for the detection of peaks in biological signalsrdquoComputer Programs in Biomedicine vol 14 no 3 pp 249ndash2561982

[9] A A Dingle R D Jones G J Carroll and W R Fright ldquoAmultistage system to detect epileptiform activity in the EEGrdquoIEEE Transactions on Biomedical Engineering vol 40 no 12 pp1260ndash1268 1993

[10] H S Liu T Zhang and F S Yang ldquoA multistage multimethodapproach for automatic detection and classification of epilepti-form EEGrdquo IEEE Transactions on Biomedical Engineering vol49 no 12 I pp 1557ndash1566 2002

[11] N Acir I Oztura M Kuntalp B Baklan and C GuzelisldquoAutomatic detection of epileptiform events in EEG by a three-stage procedure based on artificial neural networksrdquo IEEETransactions on Biomedical Engineering vol 52 no 1 pp 30ndash40 2005

[12] W Lu M M Nystrom P J Parikh et al ldquoA semi-automaticmethod for peak and valley detection in free-breathing respira-tory waveformsrdquoMedical Physics vol 33 no 10 pp 3634ndash36362006

[13] L XuMQ-HMeng R Liu andKWang ldquoRobust peak detec-tion of pulse waveform using height ratiordquo in Proceedings of the30th IEEE Annual International Conference of the Engineering inMedicine and Biology Society pp 2856ndash3859 British ColumbiaCanada 2008

[14] R Barea L Boquete S Ortega E Lopez and J M Rodrıguez-Ascariz ldquoEOG-based eye movements codification for humancomputer interactionrdquo Expert Systems with Applications vol 39no 3 pp 2677ndash2683 2012

[15] M SManikandan andK P Soman ldquoA novelmethod for detect-ing R-peaks in electrocardiogram (ECG) signalrdquo BiomedicalSignal Processing and Control vol 7 no 2 pp 118ndash128 2012

[16] A Juozapavicius G Bacevicius D Bugelskis and R Samai-tiene ldquoEEG analysismdashautomatic spike detectionrdquo Journal ofNonlinear Analysis Modelling and Control vol 16 no 4 pp375ndash386 2011

[17] L Senhadji and F Wendling ldquoEpileptic transient detectionwavelets and time-frequency approachesrdquo NeurophysiologieClinique vol 32 no 3 pp 175ndash192 2002

[18] M Putignano A Intermite and P Welsch ldquoA non-linearalgorithm for current signal filtering and peak detection inSiPMrdquo Journal of Instrumentation vol 7 pp 1ndash19 2012

[19] R E Bonner L Crevasse M IreneFerrer and J C GreenfieldJr ldquoA new computer program for analysis of scalar electrocar-diogramsrdquoComputers and Biomedical Research vol 5 no 6 pp629ndash653 1972

[20] V P Oikonomou A T Tzallas andD I Fotiadis ldquoA Kalman fil-ter based methodology for EEG spike enhancementrdquo ComputerMethods and Programs in Biomedicine vol 85 no 2 pp 101ndash1082007

[21] Y-C Liu C-C K Lin J-J Tsai and Y-N Sun ldquoModel-basedspike detection of epileptic EEG datardquo Sensors vol 13 pp12536ndash12547 2013

[22] N Sinno and K Tout ldquoAnalysis of epileptic events usingwavelet packetsrdquo The International Arab Journal of InformationTechnology vol 5 no 4 pp 165ndash169 2008

[23] Z Ji X Wang T Sugi S Goto and M Nakamura ldquoAutomaticspike detection based on real-time multi-channel templaterdquo inProceedings of the 4th International Conference on BiomedicalEngineering and Informatics (BMEI rsquo11) pp 648ndash652 IEEEOctober 2011

[24] T P Exarchos A T Tzallas D I Fotiadis S Konitsiotisand S Giannopoulos ldquoEEG transient event detection andclassification using association rulesrdquo IEEE Transactions onInformation Technology in Biomedicine vol 10 no 3 pp 451ndash457 2006

[25] C J James R D Jones P J Bones and G J Carroll ldquoDetectionof epileptiform discharges in the EEG by a hybrid systemcomprising mimetic self-organized artificial neural networkand fuzzy logic stagesrdquo Clinical Neurophysiology vol 110 no 12pp 2049ndash2063 1999

[26] N Acir ldquoAutomated system for detection of epileptiformpatterns in EEG by using a modified RBFN classifierrdquo ExpertSystems with Applications vol 29 no 2 pp 455ndash462 2005

[27] J F Gao Y Yang P Lin P Wang and C X Zheng ldquoAutomaticremoval of eye-movement and blink artifacts from EEG sig-nalsrdquo Brain Topography vol 23 no 1 pp 105ndash114 2010


[28] S B Wilson and R Emerson ldquoSpike detection a review andcomparison of algorithmsrdquo Clinical Neurophysiology vol 113no 12 pp 1873ndash1881 2002

[29] C-L Huang and J-F Dun ldquoA distributed PSO-SVM hybridsystem with feature selection and parameter optimizationrdquoApplied Soft Computing vol 8 no 4 pp 1381ndash1391 2008

[30] J Kennedy andRC Eberhart ldquoParticle swarmoptimizationrdquo inProceedings of the IEEE International Conference on Neural Net-works (ICW 95) pp 1942ndash1948 Perth Australia November-December 1995

[31] K S Lim Z Ibrahim S Buyamin et al ldquoImproving vectorevaluated particle swarm optimisation by incorporating non-dominated solutionsrdquo The Scientific World Journal vol 2013Article ID 510763 19 pages 2013

[32] M S Mohamad S Omatu S Deris M Yoshioka A Abdullahand Z Ibrahim ldquoAn enhancement of binary particle swarmoptimization for gene selection in classifying cancer classesrdquoAlgorithms for Molecular Biology vol 8 article 15 2013

[33] Z Ibrahim N K Khalid J A A Mukred et al ldquoA DNAsequence design for DNA computation based on binary vectorevaluated particle swarm optimizationrdquo International Journal ofUnconventional Computing vol 8 no 2 pp 119ndash137 2012

[34] A Adam A F Zainal Abidin Z Ibrahim A R Husain ZMd Yusof and I Ibrahim ldquoA particle swarm optimizationapproach to Robotic Drill route optimizationrdquo in Proceedingsof the 4th International Conference on Mathematical Modellingand Computer Simulation (AMS rsquo10) pp 60ndash64 May 2010

[35] M N Ayob Z M Yusof A Adam et al ldquoA particle swarmoptimization approach for routing in VLSIrdquo in Proceedings ofthe 2nd International Conference on Computational IntelligenceCommunication Systems and Networks (CICSyN rsquo10) pp 49ndash53July 2010

[36] Y Shi and R Eberhart ldquoModified particle swarm optimizerrdquoin Proceedings of the IEEE International Conference on Evolu-tionary Computation pp 69ndash73 Anchorage Alaska USA May1998

[37] Y Shi and R C Eberhart ldquoParameter selection in particleswarm optimizationrdquo in Proceedings of the 7th Annual Con-ference on Evolutionary Programming pp 591ndash601 San DiegoCalif USA 1998

[38] J Kennedy and R C Eberhart ldquoDiscrete binary version ofthe particle swarm algorithmrdquo in Proceedings of the IEEEInternational Conference on Computational Cybernetics andSimulation pp 4104ndash4108 IEEE Orlando Fla USA October1997

[39] S Mirjalili and A Lewis ldquoS-shaped versus V-shaped transferfunctions for binary Particle Swarm Optimizationrdquo Swarm andEvolutionary Computation vol 9 pp 1ndash14 2013

[40] J Rada-Vilela M Zhang and W Seah ldquoA performance studyon synchronicity and neighborhood size in particle swarmoptimizationrdquo Soft Computing vol 17 no 6 pp 1019ndash1030 2013

[41] Y Shi and R Eberhart ldquoEmpirical study of particle swarm opti-mizationrdquo inProceedings of the IEEEConference on EvolutionaryComputation pp 1945ndash1950 Washington DC USA 1999

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



frequency domain [16] time-frequency domain [10 17] andnonlinear [18] In time domain approach the peaks are ana-lyzed in time In frequency domain approach the peaks areanalyzed in frequency In time-frequency domain approachthe peaks are analyzed in both time and frequency domainIn nonlinear approach some statistical parameters of thepeaks are analyzed The general framework of peak detectionalgorithm usually involves several processes which are signalpreprocessing peak candidate detection feature extractionand classification Various signal preprocessing methodshave been employed such as data compression [19] wavelettransform [6] Kalman filter [20] and Hilbert transform [15]Two methods for peak candidate detection have been usedwhich are three point sliding windowmethod [8] and k-pointnonlinear energy operator (k-NEO) method [21] Variousfeature extraction techniques have been proposed which aremodel-based [21] wavelet analysis [22] template matching[23] and power spectra analysis [24] Several classifiers havebeen used which are rule-based [8 24] artificial neuralnetwork (ANN) [10 11 25 26] support vector machine(SVM) [7 27] and expert system [10] The highlightedpurposes in designing the framework are to achieve thehighest performance and to reduce the computational timeAlmost all studies in the EEG peak detection literature focuson the problem of detecting peaks in epileptic EEG signals Areview of peak detection algorithms that is employed to theepileptic EEG signal is presented in [28]Thepeak detection isjust a first step in epileptic event detectionThemain goal is todetermine the epileptic spikes not thewhole peaksThereforefor an epileptic event detection system the epileptic spikedetection performance not the peak detection performanceis the performance of interest

In time domain approach fourteen different peak featuresare recognized from different peak models [7ndash10] The peakmodel is a set of peak features that represents a peak by itsamplitude width and slope Most algorithms [7ndash13 21] intime domain approach consider different peak models andthe different styles of framework The peak model is chosenbased on the experiences of EEG expert To date there isno any peak detection framework that automatically findsthe finest existing peak model The use of the finest peakmodel will give a chance for the algorithm to achieve a goodperformance On the other hand the chosen peak model isnot necessarily suitable for different types of biological signalMoreover the finest peak model represents somemeaningfulinformation on the signal to be evaluated Therefore theadaptation of feature selection technique is important inthis study to automatically find the finest peak model Theutilization of feature selection on peak detection algorithmwill also reduce the computational time

In this study feature selection and classifier parame-ters estimation method based on standard particle swarmoptimization (PSO) and random asynchronous PSO (RA-PSO) algorithm are employed The process to find the finestpeak model and classifier parameter estimation is executedsimultaneously The peak features will be evaluated by a rule-based classifier The role of the classifier is to distinguishbetween peak point and non-peak point Rule-based classifieris employed due to the ability to provide an outstanding

interpretation for the obtained decisions [24] In addition theparameter values are tricky to be estimated manually A PSOalgorithm is considered to be appropriate for addressing theproblem based on the reason in which the feature selectionis a binary search problem and determination of classifierparameter is a continuous search problem [29]

11 Peak Model in Time Domain Analysis Peak model is a setof peak features that represents a peak by its amplitude widthand slope In time domain analysis fourteen different peakfeatures are recognized from different peak models [8ndash10]The earliest peak model was introduced by Dumpala et al in1982 [8] The peak model comprises four features which are(1) the amplitude between the magnitude of peak point andthe magnitude of valley point at the first half wave (2) thewidth between valley point of first half point and valley pointat second half wave (3) and (4) two slopes between a peakpoint and valley point in the first half wave and second halfwave A similar definition of the peak amplitude and slopesare also been used in [7 11 13]

An additional feature of peak amplitude and two featuresof peak width have been introduced by Acir et al [7 11]The additional peak amplitude is the amplitude between themagnitude of peak point and themagnitude of valley point ofthe second half wave The peak widths are the width betweenpeak point and valley point of first half wave and secondhalf wave The total features that are introduced by Acir etal are six features Acir et al did not use the width featurethat was introduced by Dumpala et al A similar definition ofthe peak amplitudes widths and slopes has also been used in[21] In [21] an additional peak feature is added with a set offeatures that is introduced in [7 11] which is the area of peakHowever the definition of area integration is not presented inthe paper

In addition Liu et al [10] have introduced eleven peakfeatures The proposed peak model consists of four ampli-tudes (1) the amplitude between themagnitude of peak pointand the magnitude of valley point at the first half wave (2)the amplitude between the magnitude of peak point and themagnitude of valley point of the second half wave (3) theamplitude between themagnitude of peak and themagnitudeof turning point at the first half wave and (4) the amplitudebetween themagnitude of peak and themagnitude of turningpoint at the second half wave The turning point is definedas the point where the slope decreases more than 50 ascompared to the slope of the preceding pointThemodel alsoconsists of three widths (1) the width between valley point atfirst half point and valley point at second half wave (2) thewidth between turning point at first half wave and turningpoint at second half wave and (3) the width between halfpoint at first half wave and half point at second half waveThere are four slopes that are also measured (1) and (2) twoslopes between a peak point and valley point in the first halfwave and second half wave (3) and (4) two slopes betweenpeak point and turning point at first half wave and secondhalf wave

Another peak model consists of four features which hasbeen proposed by Dingle et al [9] The peak amplitude is




















Rule-based

classifier

Rule-based

classifier


Featureextraction



Featureextraction

FilteredEEG

signalPeak features

FilteredEEG

signal

Training

Testing

PSOalgorithm

Thresholds


Finest peakmodel




2 Methodology




119894 as shown


119894minus 1) lt 119909(PP

119894) gt 119909(PP

119894+ 1) and two




PPi

HP1i

TP1i

VP1i

VP2i

HP2i

TP2i

MAC (PPi)


1) gt 119909(VP1119894) lt 119909(VP1

119894+ 1) and 119909(VP2

119894minus 1) gt 119909(VP2

119894) lt

119909(VP2119894+ 1)






119894) the








Amplitudes

1198911=1003816100381610038161003816119909(PP119894) minus 119909(VP1

119894)1003816100381610038161003816


1198912=1003816100381610038161003816119909(PP119894) minus 119909(VP2

119894)1003816100381610038161003816


1198913=1003816100381610038161003816119909(PP119894) minus 119909(TP1

119894)1003816100381610038161003816


1198914=1003816100381610038161003816119909(PP119894) minus 119909(TP2

119894)1003816100381610038161003816


1198915=1003816100381610038161003816119909(PP119894) minusMAC(PP


Widths

1198916=1003816100381610038161003816VP1119894minus VP2

119894


1198917=1003816100381610038161003816PP119894minus VP1

119894


1198918=1003816100381610038161003816PP119894minus VP2

119894


1198919=1003816100381610038161003816TP1119894minus TP2

119894

1003816100381610038161003816


11989110=1003816100381610038161003816HP1119894minusHP2

119894


Slopes

11989111=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP1

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989112=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP2

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989113=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP1

119894)

PP119894minus TP1

119894

10038161003816100381610038161003816100381610038161003816


11989114=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP2

119894)

PP119894minus TP2

119894

10038161003816100381610038161003816100381610038161003816




1 1198916 11989111 11989112

4Acir et al (2005) [7 11 26] 119891

1 1198912 1198917 1198918 11989113 11989114

6Liu et al (2002) [10] 119891

1 1198912 1198913 1198914 1198916 1198919 11989110 11989111 11989112 11989113 11989114

11Dingle et al (1993) [9] 119891

5 1198916 11989111 11989112

4













119904119896

119894119909119896

1198941119909119896


119896

119894119889119909119896

1198941119909119896


119896

119894119863


1198941 119909119896

1198942 119909119896

1198943 119909

119896

119894119889 while


V1198961198941 V1198961198942 V1198961198943 V119896


119901119896

119894= 119901119896

1198941 119901119896

1198942 119901119896

1198943 119901

119896


119892=

119901119896

1198921 119901119896

1198922 119901119896

1198923 119901

119896


particle 119904119896+1119894


V119896+1119894= 120596V119896119894+ 11988811199031(119901119896

119894minus 119909119896

119894) + 11988821199032(119901119896

119892minus 119909119896

119894) (1)



1




) times 119896 (2)


119904119896+1

119894= 119904119896

119894+ V119896+1119894 (3)


119879 (V119896+1119894) =

10038161003816100381610038161003816tanh (V119896+1

119894)

10038161003816100381610038161003816 (4)

119904119896+1

119894=

(119904119896

119894)

minus1

if rand lt 119879 (V119896+1119894)

119904119896

119894rand ge 119879 (V119896+1

119894)

(5)



119894and


iteration 119896 and (119904119896119894)

minus1







119894= 119909

119896

1198941 119909119896

1198942 119909

119896

119894119889 119909119896

1198941 119909119896

1198942 119909

119896

119894119863










TPR = TPTP + FN

TNR = TNTN + FP


(6)



IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873



119894is





1 and the social



12



1and 1199032













0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






















Rule-based

classifier

Rule-based

classifier


Featureextraction



Featureextraction

FilteredEEG

signalPeak features

FilteredEEG

signal

Training

Testing

PSOalgorithm

Thresholds


Finest peakmodel




2 Methodology




119894 as shown


119894minus 1) lt 119909(PP

119894) gt 119909(PP

119894+ 1) and two




PPi

HP1i

TP1i

VP1i

VP2i

HP2i

TP2i

MAC (PPi)


1) gt 119909(VP1119894) lt 119909(VP1

119894+ 1) and 119909(VP2

119894minus 1) gt 119909(VP2

119894) lt

119909(VP2119894+ 1)






119894) the








Amplitudes

1198911=1003816100381610038161003816119909(PP119894) minus 119909(VP1

119894)1003816100381610038161003816


1198912=1003816100381610038161003816119909(PP119894) minus 119909(VP2

119894)1003816100381610038161003816


1198913=1003816100381610038161003816119909(PP119894) minus 119909(TP1

119894)1003816100381610038161003816


1198914=1003816100381610038161003816119909(PP119894) minus 119909(TP2

119894)1003816100381610038161003816


1198915=1003816100381610038161003816119909(PP119894) minusMAC(PP


Widths

1198916=1003816100381610038161003816VP1119894minus VP2

119894


1198917=1003816100381610038161003816PP119894minus VP1

119894


1198918=1003816100381610038161003816PP119894minus VP2

119894


1198919=1003816100381610038161003816TP1119894minus TP2

119894

1003816100381610038161003816


11989110=1003816100381610038161003816HP1119894minusHP2

119894


Slopes

11989111=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP1

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989112=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP2

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989113=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP1

119894)

PP119894minus TP1

119894

10038161003816100381610038161003816100381610038161003816


11989114=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP2

119894)

PP119894minus TP2

119894

10038161003816100381610038161003816100381610038161003816




1 1198916 11989111 11989112

4Acir et al (2005) [7 11 26] 119891

1 1198912 1198917 1198918 11989113 11989114

6Liu et al (2002) [10] 119891

1 1198912 1198913 1198914 1198916 1198919 11989110 11989111 11989112 11989113 11989114

11Dingle et al (1993) [9] 119891

5 1198916 11989111 11989112

4













119904119896

119894119909119896

1198941119909119896


119896

119894119889119909119896

1198941119909119896


119896

119894119863


1198941 119909119896

1198942 119909119896

1198943 119909

119896

119894119889 while


V1198961198941 V1198961198942 V1198961198943 V119896


119901119896

119894= 119901119896

1198941 119901119896

1198942 119901119896

1198943 119901

119896


119892=

119901119896

1198921 119901119896

1198922 119901119896

1198923 119901

119896


particle 119904119896+1119894


V119896+1119894= 120596V119896119894+ 11988811199031(119901119896

119894minus 119909119896

119894) + 11988821199032(119901119896

119892minus 119909119896

119894) (1)



1




) times 119896 (2)


119904119896+1

119894= 119904119896

119894+ V119896+1119894 (3)


119879 (V119896+1119894) =

10038161003816100381610038161003816tanh (V119896+1

119894)

10038161003816100381610038161003816 (4)

119904119896+1

119894=

(119904119896

119894)

minus1

if rand lt 119879 (V119896+1119894)

119904119896

119894rand ge 119879 (V119896+1

119894)

(5)



119894and


iteration 119896 and (119904119896119894)

minus1







119894= 119909

119896

1198941 119909119896

1198942 119909

119896

119894119889 119909119896

1198941 119909119896

1198942 119909

119896

119894119863










TPR = TPTP + FN

TNR = TNTN + FP


(6)



IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873



119894is





1 and the social



12



1and 1199032













0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Amplitudes

1198911=1003816100381610038161003816119909(PP119894) minus 119909(VP1

119894)1003816100381610038161003816


1198912=1003816100381610038161003816119909(PP119894) minus 119909(VP2

119894)1003816100381610038161003816


1198913=1003816100381610038161003816119909(PP119894) minus 119909(TP1

119894)1003816100381610038161003816


1198914=1003816100381610038161003816119909(PP119894) minus 119909(TP2

119894)1003816100381610038161003816


1198915=1003816100381610038161003816119909(PP119894) minusMAC(PP


Widths

1198916=1003816100381610038161003816VP1119894minus VP2

119894


1198917=1003816100381610038161003816PP119894minus VP1

119894


1198918=1003816100381610038161003816PP119894minus VP2

119894


1198919=1003816100381610038161003816TP1119894minus TP2

119894

1003816100381610038161003816


11989110=1003816100381610038161003816HP1119894minusHP2

119894


Slopes

11989111=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP1

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989112=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(VP2

119894)

PP119894minus VP1

119894

10038161003816100381610038161003816100381610038161003816


11989113=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP1

119894)

PP119894minus TP1

119894

10038161003816100381610038161003816100381610038161003816


11989114=

10038161003816100381610038161003816100381610038161003816

119909(PP119894) minus 119909(TP2

119894)

PP119894minus TP2

119894

10038161003816100381610038161003816100381610038161003816




1 1198916 11989111 11989112

4Acir et al (2005) [7 11 26] 119891

1 1198912 1198917 1198918 11989113 11989114

6Liu et al (2002) [10] 119891

1 1198912 1198913 1198914 1198916 1198919 11989110 11989111 11989112 11989113 11989114

11Dingle et al (1993) [9] 119891

5 1198916 11989111 11989112

4













119904119896

119894119909119896

1198941119909119896


119896

119894119889119909119896

1198941119909119896


119896

119894119863


1198941 119909119896

1198942 119909119896

1198943 119909

119896

119894119889 while


V1198961198941 V1198961198942 V1198961198943 V119896


119901119896

119894= 119901119896

1198941 119901119896

1198942 119901119896

1198943 119901

119896


119892=

119901119896

1198921 119901119896

1198922 119901119896

1198923 119901

119896


particle 119904119896+1119894


V119896+1119894= 120596V119896119894+ 11988811199031(119901119896

119894minus 119909119896

119894) + 11988821199032(119901119896

119892minus 119909119896

119894) (1)



1




) times 119896 (2)


119904119896+1

119894= 119904119896

119894+ V119896+1119894 (3)


119879 (V119896+1119894) =

10038161003816100381610038161003816tanh (V119896+1

119894)

10038161003816100381610038161003816 (4)

119904119896+1

119894=

(119904119896

119894)

minus1

if rand lt 119879 (V119896+1119894)

119904119896

119894rand ge 119879 (V119896+1

119894)

(5)



119894and


iteration 119896 and (119904119896119894)

minus1







119894= 119909

119896

1198941 119909119896

1198942 119909

119896

119894119889 119909119896

1198941 119909119896

1198942 119909

119896

119894119863










TPR = TPTP + FN

TNR = TNTN + FP


(6)



IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873



119894is





1 and the social



12



1and 1199032













0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








119904119896

119894119909119896

1198941119909119896


119896

119894119889119909119896

1198941119909119896


119896

119894119863


1198941 119909119896

1198942 119909119896

1198943 119909

119896

119894119889 while


V1198961198941 V1198961198942 V1198961198943 V119896


119901119896

119894= 119901119896

1198941 119901119896

1198942 119901119896

1198943 119901

119896


119892=

119901119896

1198921 119901119896

1198922 119901119896

1198923 119901

119896


particle 119904119896+1119894


V119896+1119894= 120596V119896119894+ 11988811199031(119901119896

119894minus 119909119896

119894) + 11988821199032(119901119896

119892minus 119909119896

119894) (1)



1




) times 119896 (2)


119904119896+1

119894= 119904119896

119894+ V119896+1119894 (3)


119879 (V119896+1119894) =

10038161003816100381610038161003816tanh (V119896+1

119894)

10038161003816100381610038161003816 (4)

119904119896+1

119894=

(119904119896

119894)

minus1

if rand lt 119879 (V119896+1119894)

119904119896

119894rand ge 119879 (V119896+1

119894)

(5)



119894and


iteration 119896 and (119904119896119894)

minus1







119894= 119909

119896

1198941 119909119896

1198942 119909

119896

119894119889 119909119896

1198941 119909119896

1198942 119909

119896

119894119863










TPR = TPTP + FN

TNR = TNTN + FP


(6)



IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873



119894is





1 and the social



12



1and 1199032













0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










TPR = TPTP + FN

TNR = TNTN + FP


(6)



IF 1198911ge th1AND 119891

2ge th2AND AND

119891119872ge th119873



119894is





1 and the social



12



1and 1199032













0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




0 2000 4000 6000 8000 10000 12000

0

10

20

Sampling points

minus40

minus30

minus20

minus10

(120583V

)






























10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



















10is only


features (1198912 1198915 and 119891





5) appears 2 times






5 1198916 11989111 and 119891

12


2and 119891

5









12is only



2and 119891

5) and (119891

4and 119891

5)




RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





RunPeak features


1 0 1 0 0 1 0 0 0 1 0 0 0 0 99892 0 0 0 1 1 0 0 0 0 0 0 0 0 99913 0 1 0 0 1 0 0 0 1 0 0 0 0 99694 0 1 0 0 1 1 0 0 0 0 0 0 0 99925 0 1 0 0 1 1 0 0 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99957 0 1 0 0 1 1 0 0 0 0 0 0 0 99948 0 1 0 0 1 1 0 0 0 0 0 0 0 99919 0 0 0 1 1 0 0 0 0 1 0 0 0 999610 0 1 0 0 1 0 0 0 0 0 0 0 0 9998






Standard PSO

AVG 9991 5113 27 20 0 9373 5110 30 18 2MAX 9998 5118 22 20 0 9992 5114 26 20 0MIN 9969 5109 31 20 0 7741 5106 34 12 8

STDEV 807 718



Run 11989111198912119891311989141198915119891611989171198918119891911989110119891111198911211989113-11989114

1 0 0 0 0 1 0 0 1 0 0 0 0 0 99912 0 0 0 0 1 0 0 0 0 0 0 1 0 99873 0 0 0 1 1 0 0 0 0 0 0 0 0 99904 0 0 0 1 1 0 0 0 0 0 0 0 0 99905 0 0 0 0 1 0 0 1 0 0 0 0 0 99916 0 1 0 0 1 0 0 0 0 0 0 0 0 99907 0 1 0 0 1 0 0 0 0 0 0 0 0 99908 0 0 0 1 1 0 0 0 0 0 0 0 0 99909 0 1 0 0 1 0 0 0 0 0 0 0 0 999010 0 0 0 0 1 0 0 1 0 0 0 0 0 9991

AVERAGE Gmean 9990







RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in









RA-PSO

AVG 9990 5110 30 20 0 9859 5106 34 19 1MAX 9991 5111 29 20 0 9986 5107 33 20 0MIN 9987 5107 33 20 0 9733 5103 37 19 1

STDEV 115 133





5 Conclusions






Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Acknowledgments


References


















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in

























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



feature selection and classifier parameters estimation for

Documents