schizophreniaeegsignalclassificationbasedonswarm … · 2020. 8. 12. ·...

Research ArticleSchizophrenia EEG Signal Classification Based on SwarmIntelligence Computing

Sunil Kumar Prabhakar 1 Harikumar Rajaguru2 and Sun-Hee Kim 1

1Department of Brain and Cognitive Engineering Korea University Anam-dong Seongbuk-gu Seoul 02841 Republic of Korea2Department of Electronics and Communication Engineering Bannari Amman Institute of TechnologySathyamangalam 638402 India

Correspondence should be addressed to Sunil Kumar Prabhakar sunilprabhakar22gmailcom and Sun-Hee Kim sunheekimkoreaackr

Received 12 August 2020 Revised 23 October 2020 Accepted 26 October 2020 Published 30 November 2020

Academic Editor Antonio Dourado

Copyright copy 2020 Sunil Kumar Prabhakar et al +is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

One of the serious mental disorders where people interpret reality in an abnormal state is schizophrenia A combination ofextremely disordered thinking delusion and hallucination is caused due to schizophrenia and the daily functions of a person areseverely disturbed because of this disorder A wide range of problems are caused due to schizophrenia such as disturbed thinkingand behaviour In the field of human neuroscience the analysis of brain activity is quite an important research area For generalcognitive activity analysis electroencephalography (EEG) signals are widely used as a low-resolution diagnosis tool +e EEGsignals are a great boon to understand the abnormality of the brain disorders especially schizophrenia In this work schizophreniaEEG signal classification is performed wherein initially features such as Detrend Fluctuation Analysis (DFA) Hurst ExponentRecurrence Quantification Analysis (RQA) Sample Entropy Fractal Dimension (FD) Kolmogorov Complexity Hjorth ex-ponent Lempel Ziv Complexity (LZC) and Largest Lyapunov Exponent (LLE) are extracted initially +e extracted features arethen optimized for selecting the best features through four types of optimization algorithms here such as Artificial Flora (AF)optimization Glowworm Search (GS) optimization Black Hole (BH) optimization and Monkey Search (MS) optimization andfinally it is classified through certain classifiers +e best results show that for normal cases a classification accuracy of 8754 isobtained when BH optimization is utilized with Support Vector Machine-Radial Basis Function (SVM-RBF) kernel and forschizophrenia cases a classification accuracy of 9217 is obtained when BH optimization is utilized with SVM-RBF kernel

1 Introduction

+e reason for schizophrenia occurrence is generally notknown but researchers believe that it is a combination of theenvironment and brain genetics which contributes a lot tothe development of this disorder [1] +e signs and symp-toms usually involve disorganized speech delusions im-paired functions of organs and hallucinations [2]Symptoms generally vary with type and severity dependingon time sometimes remission of symptoms can occur andsometimes the existing symptom can worsen to a greatextent +e symptom of schizophrenia in teenagers is moreor less the same as those in adults as they experience a dropin performance at school lack of motivation irritabilitydepressed mindset withdrawal from friends and family and

also suffer from trouble sleeping [3] People affected withschizophrenia generally lack awareness that their difficultyoriginates from a mental disorder which requires carefulmedical screening [4] Suicidal thoughts and behaviour tooare very common symptoms of schizophrenia Certainnaturally occurring brain channels such as neurotransmit-ters when altered or disturbed may contribute to schizo-phrenia [5]+ough the precise cause of schizophrenia is notknown the common risk factors are having a family historyof schizophrenia birth complications exposure to toxicelements malnutrition and psychotropic drugs which alterthe state of mind [6] If schizophrenia is left untreated thenit can result in a plethora of problems such as anxietydisorder depression alcohol abuse social isolation ag-gressive behaviour leading to victimization social isolation

HindawiComputational Intelligence and NeuroscienceVolume 2020 Article ID 8853835 14 pageshttpsdoiorg10115520208853835

and financial issues followed due to various health andmedical problems [7] +ere is definitely no sure method toprevent schizophrenia but considering and taking thetreatment plan effectively can help in preventing relapses [8]Characterized by relapsing episodes of psychosis it is aserious disordermental illness To determine the neuraldynamics of human cognition EEG recording acts as asensitive tool as it can provide a millisecond-level resolution[9] EEG data are complex and at the same time they aredimensional too as they are dependent on an event of timeseries [10] As EEG signals provide electrical activity of thebrain it is easy to analyze the schizophrenia patient data

Some of the most common works related to schizo-phrenia EEG signal analysis and classification reported inthe literature are as follows An EEG-dependent nonlinearityanalysis technique for schizophrenia diagnosis was con-ducted by Zhao et al [11] +e abnormal EEG complexity inpatients with schizophrenia and depression was performedby Li et al [12] Fractal dimension was used by Raghavendraet al [13] to analyze the complexity of EEG in schizophreniapatients +e complexity measures and entropy for EEGsignal classification of both schizophrenia and controlparticipants were performed by Sabeti et al [14] A pre-liminary report on the reduced nonlinear complexity andchaos during sleep in the first episode schizophrenia wasgiven by Keshavan et al [15] A machine learning-baseddiagnosis of schizophrenia using combined sensor-level andsource level EEG features was proposed by Shim et al [16] Apreliminary data analysis of comparing the EEG nonline-arity in deficit and nondeficit schizophrenia patients wasconducted by Cerquera et al [17] +e nonlinear analysis ofEEG in schizophrenia patients with persistent auditoryhallucination was performed by Lee et al [18] For aschizophrenia patient the estimate of the first positiveLyapunov exponent of the EEG signal was performed byKim et al [19] A magnetoencephalography (MEG) study onthe LZC in schizophrenia patients was conducted by Fer-nandez et al [20] A multiscale entropy analysis for theabnormal EEG complexity signals was performed byTakahashi et al [21] As a diagnostic test for schizophreniathe spectral EEG abnormality was analyzed by Boutros et al[22] An in-depth analysis of the utility of quantitative EEGin unmedicated schizophrenia was conducted by Kim et al[23] For schizophrenic and healthy adults the machinelearning identification of EEG features which helps inpredicting working memory performance was done byJohannesen et al [24] A data-driven methodology forresting state EEG signal classification of schizophrenia withcontrol participants using random matrix theory was de-veloped by Liu et al [25] Deep-learning methods along withrandom forest and voting classifiers was performed by Chuet al for the individual recognition in schizophrenia usingresting-state EEG streams [26] Convolution Neural Net-works (CNNs) along with the Pearson Correlation Coeffi-cient (PCC) to represent the EEG channel relationships wereused to classify the schizophrenic and healthy patients usingEEG signals by Naira and Alamo [27] Random ForestMachine learning algorithm was used to identify and di-agnose the schizophrenia EEG signals by Zhang [28] A

higher order pattern discovery was used to classify theschizophrenia EEG by Sheng et al [29] +e complexity ofthe EEG signals in schizophrenia syndromes was analyzed byKutepov et al [30] A fractal-based classification of EEGsignals for both schizophrenic and healthy patients wasperformed by Namazi et al [31] A new approach for EEGsignal classification using Linear Discriminant Analysis(LDA) and Adaboost was found for schizophrenic andcontrol participants by Sabeti et al [32] In this work withthe advent of some features optimization techniques andclassifiers the schizophrenia EEG signal classification isperformed and this attempt is the first of its kind inschizophrenia EEG signal classification +e organization ofthe work is as follows Section 2 explains the materials andmethods Section 3 explains about the feature extractiontechniques Section 4 explains about the swarm intelligencecomputing techniques Section 5 explains about the classi-fication techniques Section 6 explains the results and dis-cussion followed by the conclusion in Section 7

2 Materials and Methods

+e EEG dataset for 14 healthy subjects and 14 Schizo-phrenic subjects was collected from the Institute of Psy-chiatry and Neurology Warsaw Poland and the details areexplained clearly in [33] For the subjects 7 males with anaverage age of 279 + 33 years and 7 females with an averageage of 283 + 41 years were selected +e standard 10ndash20International system was used for the acquisition of the EEGdata +e patientsrsquo data were obtained in a relaxed state andwith their eyes closed+e segmentation of the acquired EEGsignals was performed where it is considered to be sta-tionary A very simplified pictorial representation of thework is given in Figure 1

Over the duration of 15 minutes 19 channel EEG signalsare obtained Every channel of EEG signals comprises of225000 samples which are then divided into groups of5000 sample segments +erefore the data matrix of[5000times 45] is framed per channel For the preprocessing ofEEG signals Independent Component Analysis (ICA) wasutilized in this work

3 Feature Extraction Techniques

To describe a very large amount of data feature extraction isnecessary as it involves in mitigating the number of re-sources to a certain limit Once the preprocessing of the EEGsignals is conducted the following features are extractedfrom the EEG signals as follows

(i) DFA to trace the self-similarity characteristics ofthe EEG signal DFA [34] is used

(ii) Hurst exponent the self-similarity and the pre-dictability estimation of the EEG signal areexpressed by the Hurst exponent [35] If themagnitude value of the Hurst exponent is greaterthen it denotes that the EEG signal is pretty smoothand less complicated

2 Computational Intelligence and Neuroscience

(iii) RQA to measure the complexity of the EEG signalsthe total number of times of recurrences is evaluatedby RQA [36]

(iv) Entropy to evaluate the irregularity and uncertaintypresent in the EEG signal entropy features are usedWhen the complexity and the variability of the EEGsignal increases then the entropy of the EEG signalsis higher Sample entropy [37] Shannon entropy[38] approximate entropy [39] Renyi entropy [40]permutation entropy [41] KolmogorovndashSinai en-tropy [42] and fuzzy entropy [43] are the types ofentropy usually used for the analysis and in thiswork only sample entropy has been extracted

(v) FD to compare and analyze the complexity ofdetails in the EEG FD is used thereby enabling thedetection of EEG signal patterns [44] +e FractalDimension of a signal is expressed through the Katzmethod as follows

D log10(H)

log10(d) (1)

where between the successive points H representsthe sum of distances and d represents the diameterestimated

(vi) Kolmogorov complexity the characteristics of theEEG signal are easily explained by this parameter[45] If the signals are more random then thedescription length is also longer

(vii) Hjorth to quantify the morphology of the EEGsignal the important Hjorth parameters utilizedhere are complexity mobility and activity [46]

(viii) LZC to assess the repetitiveness of the EEG signalsLZC is used [47] If the LZC values are higher thenit shows that the signal is more repetitive

(ix) LLE to assess the degree of chaos present in theEEG signals an estimate of it is located by the LLE[48] If the complexity of the signals is high thenthe value of LLE is also high

+e feature extraction is initiated using DFA at firstamong the nine features +e attained feature matrix permethod per channel is in the form of [5000times10] +en fourtypes of optimization producer such as AF optimization GSOptimization BH optimization and MS optimization areutilized to further extract a better represented feature col-umn matrix as [5000times1] +is procedure is repeated for allthe channels among the subjects +is feature extractionmethod is repeated for all the other eight features such as theHurst exponent RQA entropy fractal dimension Kol-mogorov complexity Hjorth and LLE and thus the featureextraction is performed for each data segment as such

4 Swarm Intelligence Computing Techniques

To determine and understand a certain subset of initialfeatures feature selection is required +e features which areselected will usually have the useful and most relevant in-formation from the original data so that using the reducedform of representation the desired task can be performedeasily +e following four optimization techniques are uti-lized in this work

41 Artificial Flora Algorithm Developed by Cheng et al[49] four basic elements are comprised in this algorithmsuch as original plant offspring plant location of the plantand the distance of propagation initially +e plants that areused to spread seeds are called original plants +e seeds oforiginal plants are called offspring plants and at that mo-ment they cannot spread seeds Plant location is the specificlocation of the plant and the distance of propagation refersto how long a seed can spread +e three major patterns arepresent here such as evolution behaviour spread behaviourand select behaviour +e probability that the evolvement ofthe plant adapts to the environment behaviour is calledevolution behaviour +e movement of seeds is referred bythe spreading behaviour and select behaviour refers to thesurvival or death of the flora due to the environment +emain rules here are as follows

Rule 1 Due to the environmental or any external factors arandom distribution of a species in a region is performed inthat region no such species were found earlier and so itbecomes the primitive original plant

Rule 2 As the environment changes the plants will adopt tolive in the main environment +erefore a complete in-heritance to the parent plant does not depend on the properdistance of offspring plants

Rule 3 When the seeds are spread around the original plantautonomously the range is a circle where radius is themaximum propagation distance Anywhere in the circle theoffspring plants can be distributed

EEG signals

Preprocessing of EEG signals

GSoptimization

Various feature extraction methods

AFoptimization

BSoptimization

Classification techniques

MHoptimization

Figure 1 Pictorial representation of the work

Computational Intelligence and Neuroscience 3

Rule 4 Plants have varied survival probability because ofenvironmental factors such as climate and temperature +esurvival probability is related to the fitness of the plantsHere the adaptability of the plants to the environment canbe referred by fitness Or in other words the survivalprobability of a plant in a specific position is termed asfitness If the fitness is higher then the probability of survivalis greater

Rule 5 +e survival probability will be lower if the distanceis further from the original plants because the basic dif-ference between the current environment and the previousenvironment will be much greater

Rule 6 If the seeds are spread in an external margin thenthe spread distance cannot cross the maximum area of limitbecause of other constraints

411 Evolution Behaviour +e seeds are spread by theoriginal plant around in a circle which is nothing but thepropagation distance and is evolved for the respectivepropagation distances of the parent plant and the grand-parent plant and is represented as

dh d1h times rand(0 1) times c1 + d2h times rand(0 1) times c2 (2)

where d1h is the propagation distance of the grandparentplant d2h is the propagation distance of the parent plant c1and c2 are the learning coefficients and rand(0 1) denotesthe independent uniformly distributed number in (0 1)

+e normal grandparent distance of propagation isexpressed as

d1hprime d2h (3)

+e standard deviation between the respective positionof the original and offspring plant is the new parentpropagation distance and is given as

d2hprime

1113936Ni1 Lih minus Lih

prime1113872 11138732

N

1113971

(4)

412 Spreading Behaviour +e original flora with N so-lutions is randomly generated by the AF algorithm if thereare N plants in the flora+ematrix Lih is used to express theoriginal plant position where i is the dimension and h is thetotal number of plants in the flora

Lih rand(0 1) times d times 2 minus d (5)

where the maximum limit area is represented by d andrand(0 1) is an array of random numbers that have auniform distribution between (0 1) With the help of thefollowing propagation function the position of the offspringplant is generated as follows

Lihtimesmprime Dihtimesm + Lih (6)

where the number of seeds that one plant can propagate isexpressed as m L ihtimesm

prime denotes the offspring position Lih is

the original plant position and Dihtimesm is random numberwith the Gaussian distribution with zero mean and variancedh If the survival of the offspring plant is not guaranteedthen a new plant is generated as shown in the above-mentioned equation

413 Select Behaviour +e survival probability is used toassess whether the offsprings are alive or not and is rep-resented as

l

F Lihtimesmprime1113872 1113873

Fmax

1113974111397211138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868

times P(htimesmminus1)y (7)

where P(htimesmminus1)y is Py to the power of (h times m minus 1) and Py is

the selective problem +is value has to be between 0 and 1and in our experiment it is 05 When the offspring plant isfarther from the original plant then fitness is lower Py canassess the exploration ability of this algorithm To get a localoptima solution Py should be larger +e maximum fitnessin the flora is determined by the Fmax and the fitness of thehth solution is determined by F(Lihtimesm

prime ) To decide whetherthe offspring plant is alive or not the roulette wheel selectionmethod is used in our work +e procedure is explained inPseudocode 1

Pseudocode 1

Input times maximum run timeB maximum branching numberN number of original plantsl survival probability of offspring plantst 0 population Initialization

Define related parametersEvaluate the fitness value of N individuals and computethe best solutionWhile (tlt times)

For i 1 NlowastB

Propagation distance is evolved by new plantsOriginal plants spread their offspringIf rand(0 1)gt l

Offspring is aliveElseOffspring is deadEnd if

End forEvaluate novel solutions and randomly select N

plants as new original plantsIf new solution is better then old plant is replaced bynew plantFind the current best solutiont t + 1

End whileOutput optimized solution


42 Glowworm Swarm Optimization Algorithm In this al-gorithm initially the position location and information ondata exchange can be carried out by most glowworms bymeans of sending out rhythmed short beams [50] Around aparticular searching scope the main intention of GS opti-mization is to find out the flaring neighbor +e glowwormsalways move from the first position to a best position andfinally into a more extreme value point +e attraction of theglowworm individuals is highly related to its brightness +eattractiveness of a particular individual glow worm is in-versely proportional to the distance between the two indi-viduals so it implies that it has a direct proportion tobrightness Each position of the individual glow worm ac-counts for the objective function value+e individual searchscope is defined by the dynamic decision domain +e in-dividual movement is later updated step by step in theprocedure given below

421 Procedure

(i) Parameter initialization primeiprime individuals are initiallyplaced in a random fashion around a feasible regionf0 denotes the fluorescein value q0 indicates thedynamic decision domain st indicates step in ex-presses domain threshold the update coefficient ofdomain is expressed as β qst denotes the maximumsearch radius and the iteration number is expressedas n

(ii) +e objective function value H(yj(n)) is trans-formed to fj(n) as

fj(n) (1 minus ρ)fj(n minus 1) + cH yj(n)1113872 1113873 (8)

where yj(n) expresses for the individual position j

at primenprime instant of time(iii) In each q

j

d(n) the higher fluorescein value is se-lected thereby forming a set of neighborhood Ij(n)+erefore

Ij(n) h yh(n) minus yj(n)

lt qj

d(n) fj(n)lefh(n)1113882 1113883 (9)

(iv) +e probability of a particular individual j mayprogress forward as h and is expressed as

pjh(n) fh(n) minus fj(n)

1113936misinIj(n)fm(n) minus fj(n) (10)

where h is chosen by pjh(n)(v) +e updation of the position of individual j is

expressed as

yj(n + 1) yj(n) + styh(n) minus yj(n)

yh(n) minus yj(n)

⎛⎝ ⎞⎠ (11)

(vi) +e updation of the dynamic decision domain isexpressed as

qj

d(n + 1) min qst max 0 qj

d(n) β in minus Ij(n)11138681113868111386811138681113868

111386811138681113868111386811138681113874 11138751113882 11138831113882 1113883 (12)

43 Black Hole Algorithm +e algorithm is inspired fromthe black hole phenomenon [51] In this method at eachiteration the best candidate is chosen as the black hole andthe other normal candidates are chosen as normal stars +eblack hole creation is one of the original candidates of theentire population and it is not random Depending on onersquoscurrent location and a randomly generated number themovement of the candidate towards the black hole isascertained +e step-by-step details of the algorithm is asfollows

(i) It is a very famous population-based algorithm+erefore in the search space of some functionsome randomly generated population of candidatesolutions (stars) is placed

(ii) +e fitness value evaluation of the population isconducted after the initialization process

(iii) +e best candidate in the population is the onewhich has the highest fitness value and it is as-sumed to be the initial black hole

(iv) While this process is going on the rest of thecandidates do the normal stars

(v) +e specialty of the black hole is that it has thecapacity to absorb the surrounding stars around it

(vi) Once the stars and the black hole are initializedthe absorption of the stars by the black holetakes place and therefore the movement of allother stars is now towards or around the blackhole

(vii) +e formulation of the capability of the ab-sorption of stars by the black hole is ascertainedas follows

Yj(t + 1) Yj(t) + rand Ycl minus Yj(t)1113872 1113873 (13)

Here Yj(t) and Yj(t + 1) are the locations of thestar j in iteration t and t + 1 Ycl is the location ofthe black hole in the entire search space Randomnumber in the interval of [0 1] is determined byrand

(viii) +e black hole absorbs the candidate solution(each star representing it) that crosses the horizonof the black hole Using the following equation theradius of the horizon in the BH algorithm iscomputed as follows


R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)

where Gcl represents the black hole fitness value Gj is thefitness value of the star j S denotes the number of starswhich indicates the size of populationcandidate solutionsand R denotes the black hole radius

+e candidate is usually collapsed when the distancebetween the best candidate (black hole) and the candidatesolution is less than R In such a case the value of a newcandidate would be carried out and it is randomly dis-tributed in the search space giving the optimized and the bestvalues

44 Monkey Search Algorithm Based on the mountainclimbing procedure of monkeys one of the recently de-veloped metaheuristic algorithms is MS algorithm [52]Initially for the monkeys the population size is definedEach monkeyrsquos specific position is denoted by a vector aiand it is performed randomly With the help of a step-by-step climbing process the monkey positions are changedOne of the famous recursive optimization algorithm calledSimultaneous Perturbation Stochastic Approximation(SPSA) was used to design the climb process in MS algo-rithm +e objective function value is improved by theclimbing process Once a monkey arrives on top of amountain after the climbing process then it will search forhigher points than its current position A monkey with itssharp and keen eyesight would easily shift its base to a higherpoint by jumping +e maximum distance watched by amonkey is determined by the eyesight of a monkey +eupdation of the position is carried out +en the monkeysstart to search the novel search domains by utilizing thecurrent position as a pivot +is process is termed as asomersault procedure and it helps the monkeys to get a newposition +e evaluation of the objective function values isconducted and if the number of iterations is satisfactorythen the process will be stopped +e main process of thisalgorithm is as follows

(1) Representation of a solution the population size ofmonkey M is defined initially For any monkeyj isin 1 2 M defines its positionaj (aj1 aj2 ajn) which gives the optimizationproblem solutions with the n dimensions

(2) Initialization in a random manner the initial so-lution of the monkeys is generated

(3) Climbing process

(i) A vector is generated randomly Δaj (Δaj1

Δaj2 Δajn) where Δajk is set as s where s

denotes the step length of climb process

g jkprime aj1113872 1113873

g aj + Δaj1113872 1113873 minus g aj minus Δaj1113872 1113873

2Δajk

(15)

k 1 2 n and at point aj the vector is termedas the pseudogradient of the objective function

(ii) Assign zk ajk + α sign(gjkprime(aj)) k 1 2 n

and z (z1 z2 zn)

(iii) Assume ajlarrz has given z is feasible or else keepas such aj

(iv) Unless the maximum number of allowed itera-tions Nq has reached or there is a little change inthe objective function of the iteration the steps (i)to (iii) of the climbing process are repeated

(4) Watch jump process(i) +e real number zj is generated randomly from

(ajk minus e ajk + e) k 1 2 n where e repre-sents the eyesight of the monkey which denotes themaximum distance that can be witnessed by amonkey

(ii) Assume aj⟵ z given that g(z)gtg(aj) and z isfeasible Unless a certain number of watch timeshas been obtained or until an appropriate point z

is found out the steps are repeated(iii) +e climb process is repeated by employing z as

an initial position(5) Somersault process(i) From the somersault interval [q d] a real number

θ is generated randomly(ii) Set zk ajk + θ(vk minus ajk) where

vk (1M) 1113936Mj1 ajk k 1 2 n

v (v1 v2 vn) is termed somersault pivot and(vk minus ajk) is the somersault direction of monkey j

(iii) Set ajlarrz if z (z1 z2 zn) is quite feasibleUnless a feasible solution z is found out steps (i)and (iii) of this somersault process are repeated

(6) Termination

Unless the stopping criterion is met the above-mentioned steps are repeated As the stopping criteria thenumber of iterations are used

+us using these metaheuristic algorithms the optimaland best solutions are explored and exploited from the initialrandom solutions

5 Classification Techniques

+e optimized values or the best selected feature valuesthrough swarm intelligence computing techniques are finallyfed to the classifiers for classification

51 ANN Here a three-layer perceptron comprising of aninput layer hidden layer and an output layer was utilized[53]+e total number of hidden neurons is expressed by thefollowing equation

H [(A + B)d] (16)

where the number of input neurons is expressed as A thenumber of output neurons is expressed as B and d is a


constant and its range is from d isin [0 1] For the ANNclassifier the total number of hidden neurons used here is50

52 QDA +e ratio of the between class variance is max-imized and the ratio of the within class variance is mini-mized by QDA [54] Between classes it also allows quadraticdecision boundaries so that the classifier can perform theclassification more accurately providing good classificationaccuracy In this QDA no shrinkage or regularization isused

53 SVM Due to its good scalability and its high classifi-cation performance SVM is used [55] Creating a hyper-plane to maximize the margins between the classes that canbe obtained by mitigating the cost function so that themaximum classification accuracy is attained is the main ideaof SVM +e hyperplanes which are represented by thevectors are known as support vectors By minimizing thecost function the optimal solution that maximizes thedistance between the hyperplane and the nearest trainingpoint is obtained by the SVM as follows

Minimize12w

2+ C 1113944

n

j1ξj

subject to zj wTxj + f1113872 1113873

3ge 1 minus ξj ξj ge 0

(17)

where wT xj isin R2 and f isin Rprime w2 wT w

+e tradeoff between the margin and the error is denotedas C +e measure of the training data is expressed as ξj +eclass label for the jth sample is denoted as zj SVM can beutilized as a both linear and nonlinear classifier Varioustypes of kernel functions are utilized to make SVM as anonlinear classifier +e types of kernels generally used arePolynomial Radial Basis Function (RBF) and Sigmoidkernels Here in our work only SVM-RBF kernel is usedand this nonlinear SVM is used to get higher classificationaccuracy

54 Logistic Regression Between an independent variableand a response variable the relationship is assessed byLogistic Regression (LR) [56] An observation is classifiedinto one of the two classes using Logistic Regression in a verysimple manner When the variables are nominal or binary itcan be used Similar to the Bayesian group the data arecomprehensively analyzed after the discretization processfor the continuous variables is performed

55 FLDA +e main intention of the Fischers LinearDiscriminant Analysis (FLDA) is to trace a direction in thefeature space along which the specific distance of the meansrelative to the within-class scatter explained by the within-class scatter matrix SW reaches a maximum [57] When thewithin-class scatter matrix reaches a maximum the classseparability is maximized By maximizing the following

criterion with the between-class scatter matrix this goal canbe achieved

J(W) W

TSBW

WTSWW

(18)

Tomaximize the criterion the directionw is expressed asW Sminus1

W(m1 minus m2) where m1 and m2 are the means for thetwo classes For the two classes the FLDA acts as a sub-optimal classifier when their respective distributions areGaussian in nature

56 KNN One of the famous nonparametric algorithmsutilized for both classification and regression is KNN [58]On the underlying data distribution KNN does not makeany assumption +ere is also no explicit training phaseavailable here During the testing phase the utilization of thetraining data is carried out where the measurement betweenthe training instance and test instance is performed +eprediction of the class of the test instance is performed byutilizing the majority voting of the any of the K-nearesttraining instances +e value of K is assumed to be 4 in ourwork

6 Results and Discussion

It is classified with 70 of training and 30 of testingmethodology and the performance of it is computed +eexperiment was repeated five times to check whether we getthe similar results every time when the analysis is done +emathematical formulae for computing the PerformanceIndex (PI) Sensitivity Specificity Accuracy Good Detec-tion Rate (GDR) and Mean Square Error (MSE) rate ismentioned in literature and using the same the values arecomputed and exhibited PC indicates Perfect ClassificationFA indicates False Alarm and MC indicates Missed Clas-sification respectively

+e sensitivity is expressed as

Sensitivity PC

PC + FAtimes 100 (19)

Specificity is expressed as

Specificity PC

PC + MCtimes 100 (20)

Accuracy is expressed as

Accuracy Sensitivity + Specificity

2 (21)

Performance Index is expressed as

PI PC minus MC minus FA

PC1113874 1113875 times 100 (22)

Good Detection Rate (GDR) is expressed as

GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)

+e Mean Square Error (MSE) is expressed as follows


MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)

where Oi indicates the observed value at a specific time Tj

denotes the target value at model j j 1 to 19 and N is thetotal number of observations per channel for a single op-timization method and in our case it is 45000 +e trainingof the classifiers was implemented with a zero-training errorof MSE

Table 1 shows the average statistical parameters suchas mean variance skewness and kurtosis of the nineextracted features through four-optimization process forthe normal cases +e higher value of mean indicates thepeaked position of the feature selected Lower value ofmean indicates that there exists the peak and valleypoints in the features As in the case of AF optimizationin Kolmogorov complexity features peaked value ofmean is attained +e variance parameter shows theenergy component of the feature Here also in Kolmo-gorov complexity in AF optimization method arrivedhigher value Skewness depicts the skewed features of thedata points and all the features in Table 1 indicates thesame +e flatness is indicated by the higher kurtosisvalues In the case of Hurst exponent Fractal dimensionand Hjorth features at AF optimization show highervalue of kurtosis

Table 2 demonstrates the average statistical parame-ters such as mean variance skewness and kurtosis of thenine extracted features through four-optimization processfor the schizophrenia cases +e higher value of meanindicates the peaked position of the feature selectedLower value of mean indicates that there exist the peakand valley points in the features as all the optimizationmethods made the mean parameter as a smaller oneamong the nine features Skewness depicts the skewedfeatures of the data points and all the features in Table 2indicate the same +e flatness is indicated by the higherkurtosis values In the case of Kolmogorov complexity atAF optimization and BH optimization it shows highervalue of Kurtosis

+e correlation among the normal and schizophreniacases can be established by calculating the CanonicalCorrelation Analysis (CCA) as shown in Table 3 If theCCA value is greater than 05 that indicates the twoclasses are highly correlated and for lower value it is viceversa As shown in Table 3 the CCA is calculated for thenine features among the normal and schizophrenia casesAs observed from Table 3 the lower value of CCA in-dicates the Nil correlation among the features across theclasses

Table 4 exhibits the average PCC with different featuresfor normal and schizophrenia cases +e values in Table 2indicates the nonlinear relation among the features in thesame class of the data +erefore the uncorrelated andnonlinear features have to be optimized by the optimizationprocess

Table 5 shows the CCA of various optimization tech-niques with different features for normal and schizophrenia

cases It is observed from Table 5 that the low value of CCA isdefinitely indicating the Nil correlation among the featuresof the two classes

Table 6 depicts the Average PCC at various opti-mization techniques with nine different features fornormal and schizophrenia cases As shown in Table 6low value of CCA in the normal cases indicates thepresence of nonlinearity in the features +e negativevalue of PCC in the schizophrenia cases mention aboutthe inverse relation among the features as well as theoptimization methods

Table 7 shows the consolidated results of accuracyamong the classifiers at various optimization techniqueswith different features of normal cases As indicated inTable 7 ANN classifier is ebbed at low value of accuracyin the three types of optimization methods such as AFGS and BH +e poor performance of ANN is due toover learning and the exhibition of false alarm in theclassifier outputs FLDA classifier arrived at low accu-racy in the case of MS optimization method SVMclassifier is outperforming all the classifiers in terms ofhigher accuracy value of 8754 in BH optimizationmethod

Table 8 denotes the consolidated results of accuracyamong the classifiers at various optimization techniqueswith different features of schizophrenia cases As ob-served in Table 8 LR classifier is ebbed at low value ofaccuracy of 7864 in the black hole optimizationmethod +e poor performance of LR is due to the rigidparametric values and the exhibition of missed classi-fication in the classifier outputs SVM classifier is out-performing all the classifiers in terms of higher accuracyvalue of 9217 in BH optimization method Evenlynature of the performance shows higher accuracy valuesamong the classifiers for MS method

Table 9 represents the average perfect classificationamong the classifiers at various optimization techniqueswith different features of normal cases As observed in Table9 LR classifier is ebbed at low value of perfect classificationof 543 in the AF optimization method +e poor per-formance of LR is due to the nonadaptive parametric valuesand the exhibition of false alarm in the classifier outputsSVM classifier is outperforming all the classifiers in terms ofhigher perfect classification value of 75093 in BH opti-mization method Evenly nature of the performance showshigher perfect classification values among the classifiers forMS method

Table 10 denotes the average perfect classificationamong the classifiers at various optimization techniqueswith different features of schizophrenia cases As shown inTable 10 LR classifier is reached at low value of perfectclassification of 5729 in the BH optimization method+e poor performance of LR is due to the nonadaptiveparametric values and the exhibition of missed classifica-tion in the classifier outputs SVM classifier is out-performing all the classifiers in terms of higher perfectclassification value of 8434 in BH optimization methodEvenly nature of the performance shows higher perfectclassification values among the classifiers for MS method


Table 11 depicts the average Performance Index amongthe classifiers at various optimization techniques with dif-ferent features of normal cases As shown in Table 11 LRclassifier is reached at low value of PI of 1563 in AF

optimization method +e poor performance of LR is due tothe exhibition of missed classification in the classifier out-puts SVM classifier is outperforming all the classifiers interms of higher PI value of 646 in BH optimization

Table 1 Average statistical parameters at various optimization techniques with different features for normal cases

Optimizationmethods Parameters DFA Hurst RQA Sample

entropyFractal

dimensionKolmogorovcomplexity Hjorth LZC LLE

Artificial flora

Mean 526965 0413804 3689182 0869676 041564 2875062 04173837 0416007 04172Variance 2207468 0000125 0961903 1100939 8993Eminus06 3118672 216155Eminus10 117142Eminus07 3939Eminus08Skewness 1724139 minus792829 minus098766 2580366 minus4544 1179709 minus672855 15590 minus37801Kurtosis 5223521 7760378 0927814 9156234 300779 2794157 49372894 26237 20561

Glowwormswarm

Mean 2320835 0733608 0396612 1731146 074723 0609228 0196184 0269188 01472Variance 103798 0021621 0001249 0049742 0019817 0033745 214Eminus06 515Eminus06 000035Skewness 1903065 0269845 1104594 minus020551 0240516 229958 0698693 0408151 102453Kurtosis 4714625 0122304 2464601 0171725 0032322 8864067 0245379 8982281 243118

Black hole

Mean 2710718 0351737 1363744 2227318 0524266 3234247 175846 161927 159492Variance 0123383 0044921 0011921 0524723 0055867 0173108 0000821 0005577 0110766Skewness 055226 minus02302 minus041571 1004853 minus045211 0861662 1221792 minus112091 0540982Kurtosis 0090556 0133614 minus156857 4072836 0456164 0082189 minus046675 2260795 1804519

Monkey search

Mean 0444369 0000265 0044637 0066802 0000593 0786268 0556733 0740261 0708153Variance 0008841 647Eminus08 764Eminus05 0000456 351Eminus07 0005208 0001272 0003981 6704619Skewness 00065 2394641 0614119 0430518 2170684 minus033499 0572398 1990603 16253Kurtosis minus013621 8868095 0241653 0532993 6750117 0005905 0058357 6873791 282403

Table 2 Average parameters at various optimization techniques with different features for schizophrenia cases


entropyFractal


Artificial flora

Mean 1947 0797 0895 10434 09192 09182 0922 0922 072722Variance 61955 000198 8215Eminus05 09472 7331Eminus05 000011 482Eminus09 3393Eminus08 000382Skewness 27591 minus00952 01699 34882 minus79156063 minus316135 48058 1162 066921Kurtosis 10378 minus03245 minus000067 165549 1399124 9987822 25231 2795 034590

Glowwormswarm

Mean 360521 045140 0528535 320544 1109916 2762614 0775 0874 043363Variance 140291 000050 0000336 018103 0019184 0846133 000022 385Eminus06 0000213Skewness 183474 018256 minus00874 minus002972 021627 1943948 053419 minus0373 0538756Kurtosis 437384 minus012708 minus009477 minus032559 minus017723 4561639 00462 1151 05087

Black hole

Mean 004486 164093 038062 008666 0471683 0193926 0246 0209 1409073Variance 001454 0827833 0000476 0020484 0054866 000054 109Eminus05 67Eminus06 0102673Skewness minus0911996 2338573 0602759 minus025059 2264549 3166492 minus2561 9701 1243274Kurtosis 1689095 8655974 0923601 minus138165 8043591 1003234 5236 15625 3155236

Monkeysearch

Mean 0360178 064032 0142771 0139355 0000866 0540915 0933 0000151 2272099Variance 001044 0449776 0000665 0001762 215Eminus07 0006902 000703 19Eminus12 854282Skewness 0232422 2867251 0441075 0154336 0980327 minus020562 05348 0799874 1161515Kurtosis minus006264 1354124 0384905 minus046198 11591 minus030885 00581 1016941 167628

Table 3 Average CCA with different features for normal and schizophrenia cases

Parameters DFA Hurst RQA Sample entropy Fractal dimension Kolmogorov complexity Hjorth LZC LLECCA 014231 005738 00876 014356 006889 015469 014003 006534 012817

Table 4 Average PCC with different features for normal and schizophrenia cases

Parameters PCC DFA Hurst RQA Sampleentropy

Fractaldimension

Kolmogorovcomplexity Hjorth LZC LLE

Normal 0021077 0014018 0006344 0047908 0016744 0023395 000045 000376 0062845Schizophreniacases 0053808 0012278 0068795 0030514 0032101 0107226 0003109 000216 0043985


Table 5 CCA at various optimization techniques with different features for normal and schizophrenia cases

Optimization methods CCAArtificial flora 0047944Glowworm swarm 0088456Black hole 0060256Monkey search 0089556

Table 6 Average PCC at various optimization techniques with different features for normal and schizophrenia cases

Optimization methodsPCC

Normal Schizophrenia casesArtificial flora 0005745 minus001573Glowworm swarm 004604 minus007745Black hole 0040539 minus001178Monkey search 008175 minus01422

Table 7 Consolidated results of accuracy () among the classifiers at various optimization techniques with different features for normalcases

Optimization methods ANN QDA SVM LR FLDA KNNArtificial flora 7742581 7908989 8540703 7715035 8007678 8261267Glowworm swarm 7735069 7903598 8714997 833216 8018607 8187646Black hole 7725071 7882409 8754716 773041 8027763 793533Monkey search 8147733 7993699 8508729 8448008 7967025 8107509

Table 8 Consolidated results of accuracy () among the classifiers at various optimization techniques with different features forschizophrenia cases


Table 9 Average perfect classification () among the classifiers at various optimization techniques with different features for normal cases


Table 10 Average perfect classification () among the classifiers at various optimization techniques with different features for schizo-phrenia cases



method +e performance of PI parameter for the ANN andQDA classifier among the four-optimizationmethod is poorand this is due to the more missed classification and falsealarms of the classifier outputs

Table 12 signifies the average Performance Index amongthe classifiers at various optimization techniques with dif-ferent features of schizophrenia cases As shown in Table 12LR classifier is reached at low value of PI of 241 in the BHoptimization method +e poor performance of LR is due tothe exhibition of missed classification in the classifier out-puts SVM classifier is outperforming all the classifiers interms of higher PI value of 8159 in BH optimizationmethod Once again MS optimization evenly handled the PIparameter among the classifiers

Table 13 exhibits the average performance of parametersamong the classifiers at various optimization techniqueswith different features of normal cases As shown in Table 13

ANN classifier is reached at low value of PI of 2201 and PCof 567 FLDA denotes low GDR value of 3767 SVMclassifier is outperforming all the classifiers in terms ofhigher PC PI and GDR values and lower error rate value of274 For a significant accuracy and error rate KNNclassifier closely follows the performance of the SVMclassifier

Table 14 depicts the average performance of parametersamong the classifiers at various optimization techniqueswith different features of schizophrenia cases As shown inTable 14 LR classifier is reached at low value of PI of 315and PC of 6067 LR denotes low GDR value of 419 SVMclassifier is outperforming all the classifiers in terms ofhigher PC PI and GDR values and lower error rate value of1811 For a significant accuracy and error rate ANNclassifier closely follows the performance of the SVMclassifier

Table 11 Average performance index () among the classifiers at various optimization techniques with different features for normal cases


Table 12 Average performance index () among the classifiers at various optimization techniques with different features for schizophreniacases


Table 13 Average performance of parameters among the classifiers at various optimization techniques with different features for normalcases

Performance parameters ANN QDA SVM LR FLDA KNNPerfect classification 5675227 5844348 7259401 6112578 601035 6245814Performance index 220149 2730796 5976428 3275924 3079819 3585038Accuracy 7837614 7922174 8629786 8056403 8005268 8122938GDR 5451164 5844344 7243905 5911251 376745 5596857Error rate 4324759 4155656 274062 3887433 3989642 3754199

Table 14 Average performance of parameters among the classifiers at various optimization techniques with different features forschizophrenia cases



7 Conclusions and Future Work

+e disorders in the areas of the lobes of the brain can lead toschizophrenia As the lobes of the brain are important forinformation processing and memory management activitiesa huge damage occurs to it due to schizophrenia +e di-agnosis classification and analysis of schizophrenia spec-trum disorders are quite challenging To incorporate thelatest scientific techniques to clinical diagnosis the scientificcommunity is working very hard For the brain state in-terpretation and diagnosis EEG is emerged as a highly usefuland beneficial tool +e proposed method in this workexplores and utilizes a plethora of features with four differenttypes of optimization techniques before proceeding toclassification +e best results show that for normal cases aclassification accuracy of 8754 is obtained when BHoptimization is utilized with SVM-RBF Kernel and forschizophrenia cases a classification accuracy of 9217 isobtained when BH optimization is utilized with SVM-RBFkernel We plan to incorporate other optimization mecha-nisms with deep learning techniques for schizophrenia EEGsignal classification in future work

Data Availability

Data will be provided to genuine researchers upon request

Conflicts of Interest

+e authors declare that there are no conflicts of interest

Acknowledgments

+is work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Korea gov-ernment (MSIT) (no NRF-2018R1A2B6006046)

References

[1] P Lysaker M Bell and J Beam-Goulet ldquoWisconsin cardsorting test and work performance in schizophreniardquo Psy-chiatry Research vol 56 no 1 pp 45ndash51 1995

[2] M D Bell P H Lysaker R M Milstein and J L Beam-Goulet ldquoConcurrent validity of the cognitive component ofschizophrenia relationship of PANSS scores to neuro-psychological assessmentsrdquo Psychiatry Research vol 54 no 1pp 51ndash58 1994

[3] N C Andreasen and S Olsen ldquoNegative v positive schizo-phreniardquo Archives of General Psychiatry vol 39 no 7pp 789ndash794 1982

[4] A S Bellack J J Blanchard P Murphy and K PodellldquoGeneralization effects of training on the Wisconsin CardSorting Test for schizophrenia patientsrdquo Schizophrenia Re-search vol 19 no 2-3 pp 189ndash194 1996

[5] D Koren L J Seidman M J Lyons et al ldquoFactor structure ofthe Wisconsin card sorting test dimensions of deficit inschizophreniardquo Neuropsychology vol 12 no 2 pp 289ndash3021998

[6] X Zhu H-I Suk S-W Lee and D Shen ldquoCanonical featureselection for joint regression and multi-class identification in

Alzheimerrsquos disease diagnosisrdquo Brain Imaging and Behaviorvol 10 no 3 pp 818ndash828 2016

[7] X Chen H Zhang L Zhang C Shen S W Lee and D ShenldquoExtraction of dynamic functional connectivity from braingrey matter and white matter for MCI classificationrdquo HumanBrain Mapping vol 38 no 10 pp 5019ndash5034 2017

[8] X Ding and S-W Lee ldquoChanges of functional and effectiveconnectivity in smoking replenishment on deprived heavysmokers a resting-state FMRI studyrdquo PLoS One vol 8 no 3Article ID e59331 2013

[9] S K Prabhakar H Rajaguru and S-W Lee ldquoA frameworkfor schizophrenia EEG signal classification with nature in-spired optimization algorithmsrdquo IEEE Access vol 8 no 12020

[10] S K Prabhakar H Rajaguru and S-W Lee ldquoMetaheuristic-based dimensionality reduction and classification analysis ofPPG signals for interpreting cardiovascular diseaserdquo IEEEAccess vol 7 pp 165181ndash165206 2019

[11] Q Zhao B Hu and L Liu ldquoAn EEG based nonlinearityanalysis method for schizophrenia diagnosisrdquo BiomedicalEngineering vol 9 p 136 2012

[12] Y Li S Tong D Liu et al ldquoAbnormal EEG complexity inpatients with schizophrenia and depressionrdquo Clinical Neu-rophysiology vol 119 no 6 pp 1232ndash1241 2008

[13] B S Raghavendra D N Dutt H N Halahalli and J P JohnldquoComplexity analysis of EEG in patients with schizophreniausing fractal dimensionrdquo Physiological Measurement vol 30no 8 pp 795ndash808 2009

[14] M Sabeti S Katebi and R Boostani ldquoEntropy and com-plexity measures for EEG signal classification of schizophrenicand control participantsrdquo Artificial Intelligence in Medicinevol 47 no 3 pp 263ndash274 2009

[15] M S Keshavan J D Cashmere J Miewald andV K Yeragani ldquoDecreased nonlinear complexity and chaosduring sleep in first episode schizophrenia a preliminaryreportrdquo Schizophrenia Research vol 71 no 2-3 pp 263ndash2722004

[16] M Shim H-J Hwang D-W Kim S-H Lee and C-H ImldquoMachine-learning-based diagnosis of schizophrenia usingcombined sensor-level and source-level EEG featuresrdquoSchizophrenia Research vol 176 no 2-3 pp 314ndash319 2016

[17] A Cerquera K Gjini S M Bowyer and N BoutrosldquoComparing EEG nonlinearity in deficit and nondeficitschizophrenia patients preliminary datardquo Clinical EEG andNeuroscience vol 48 no 6 pp 376ndash382 2017

[18] S-H Lee J-S Choo W-Y Im and J-H Chae ldquoNonlinearanalysis of electroencephalogram in schizophrenia patientswith persistent auditory hallucinationrdquo Psychiatry Investi-gation vol 5 no 2 pp 115ndash120 2008

[19] D-J Kim J Jeong J-H Chae et al ldquoAn estimation of the firstpositive Lyapunov exponent of the EEG in patients withschizophreniardquo Psychiatry Research Neuroimaging vol 98no 3 pp 177ndash189 2000

[20] A Fernandez M-I Lopez-Ibor A Turrero et al ldquoLempel-Zivcomplexity in schizophrenia a MEG studyrdquo Clinical Neu-rophysiology vol 122 no 11 pp 2227ndash2235 2011

[21] T Takahashi R Y Cho T Mizuno et al ldquoAntipsychoticsreverse abnormal EEG complexity in drug-naive schizo-phrenia a multiscale entropy analysisrdquo Neuroimage vol 51no 1 pp 173ndash182 2010

[22] N N Boutros C Arfken S Galderisi JWarrick G Pratt andW Iacono ldquo+e status of spectral EEG abnormality as adiagnostic test for schizophreniardquo Schizophrenia Researchvol 99 no 1-3 pp 225ndash237 2008


[23] J W Kim Y S Lee D H Han K J Min J Lee and K LeeldquoDiagnostic utility of quantitative EEG in un-medicatedschizophreniardquo Neuroscience Letters vol 589 pp 126ndash1312015

[24] J K Johannesen J Bi R Jiang J G Kenney andC M A Chen ldquoMachine Learning Identification of EEGfeatures predicting working memory performance inschizophrenia and healthy adultsrdquo Neuropsychiatric Electro-physiology vol 2 no 3 pp 1ndash21 2016

[25] H Liu T Zhang Y Ye et al ldquoA data driven approach forresting-state eeg signal classification of schizophrenia withcontrol participants using random matrix theoryrdquo 2017httpsarxivorgabs171205289

[26] L Chu R Qiu H Liu Z Ling and X Shi ldquoIndividualrecognition in schizophrenia using deep learning methodswith random forest and voting classifiers insights fromresting state EEG streamsrdquo pp 1ndash7 2017 httpsarxivorgabs170703467

[27] C A T Naira and C J L D Alamo ldquoclassification of peoplewho suffer schizophrenia and healthy people by eeg signalsusing deep learningrdquo International Journal of AdvancedComputer Science and Applications vol 10 no 10 pp 511ndash516 2019

[28] L Zhang ldquoEEG signals classification using machine learningfor the identification and diagnosis of schizophreniardquo inProceedings of the 2019 41st Annual International Conferenceof the IEEE Engineering in Medicine and Biology Society(EMBC) pp 4521ndash4524 Berlin Germany July 2019

[29] Z Sheng Q Shini and W Wei ldquoClassification of schizo-phreniarsquos EEG based on high order pattern discoveryrdquo inProceedings of the 2010 IEEE Fifth International Conference onBio-Inspired Computing Ceories and Applications (BIC-TA)pp 1147ndash1149 Changsha China September 2010

[30] I E Kutepov V A Krysko A V Krysko et al ldquoComplexityof EEG Signals in Schizophrenia Syndromesrdquo in Proceedingsof the 29th International Conference on Computer Graphicsand Vision pp 1ndash4 Bryansk Russia 2019

[31] H Namazi E Aghasian and T S Ala ldquoFractal-based classi-fication of electroencephalography (EEG) signals in healthyadolescents and adolescents with symptoms of schizophreniardquoTechnology and Health Care vol 27 no 3 pp 233ndash241 2019

[32] M Sabeti S D Katebi R Boostani and G W Price ldquoA newapproach for EEG signal classification of schizophrenic andcontrol participantsrdquo Expert Systems with Applicationsvol 38 no 3 pp 2063ndash2071 2011

[33] E Olejarczyk and W Jernajczyk ldquoGraph-based analysis ofbrain connectivity in schizophreniardquo PLoS One vol 12Article ID e0188629 2017

[34] K Hu P C Ivanov Z Chen P Carpena and H E StanleyldquoEffect of trends on detrended fluctuation analysisrdquo PhysicalReview E vol 64 no 1 19 pages Article ID 011114 2001

[35] Y-C Chang L-C Lai L-H Chen C-M Chang andC-C Chueh ldquoA Hurst exponent estimator based on autor-egressive power spectrum estimation with order selectionrdquoBio-Medical Materials and Engineering vol 24 no 1pp 1041ndash1051 2014

[36] R Fusaroli I Konvalinka and S Wallot ldquoAnalyzing socialinteractions the promises and challenges of using cross re-currence quantification analysisrdquo in Translational Recur-rences From Mathematical Ceory to Real-WorldApplications N Marwan M Riley A Giuliani andC Webber Eds Springer International Publishing LondonUK pp 137ndash155 2014

[37] J S Richman and J R Moorman ldquoPhysiological time-seriesanalysis using approximate entropy and sample entropyrdquoAmerican Journal of Physiology-Heart and Circulatory Phys-iology vol 278 no 6 pp H2039ndashH2049 2000

[38] J A Tenreiro Machado ldquoShannon entropy analysis of thegenome coderdquo Mathematical Problems in Engineeringvol 2012 p 12 Article ID 132625 2012

[39] S M Pincus ldquoApproximate entropy as a measure of systemcomplexityrdquo Proceedings of the National Academy of Sciencesvol 88 no 6 pp 2297ndash2301 1991

[40] C-Y Kee S G Ponnambalam and C-K Loo ldquoBinary andmulti-class motor imagery using Renyi entropy for featureextractionrdquo Neural Computing and Applications vol 28no 8 pp 2051ndash2062 2017

[41] M Staniek and K Lehnertz ldquoParameter selection for per-mutation entropy measurementsrdquo International Journal ofBifurcation and Chaos vol 17 no 10 pp 3729ndash3733 2007

[42] A N Kolmogorov ldquo+ree approaches to the quantitativedefinition of informationrdquo International Journal of ComputerMathematics vol 2 no 1-4 pp 157ndash168 1968

[43] F Zhao Q S Wang and W L Hao ldquoImprovement ofconstraint conditions and new constructional method forintuitionistic fuzzy entropyrdquo Journal of Computer Applica-tions vol 35 no 12 pp 3461ndash3464 2015

[44] M K Biswas T Ghose S Guha and P K Biswas ldquoFractaldimension estimation for texture images a parallel approachrdquoPattern Recognition Letters vol 19 no 3-4 pp 309ndash313 1998

[45] E Mayordomo ldquoA Kolmogorov complexity characterizationof constructive Hausdorff dimensionrdquo Information ProcessingLetters vol 84 no 1 pp 1ndash3 2002

[46] B Hjorth ldquoEEG analysis based on time domain propertiesrdquoElectroencephalography and Clinical Neurophysiology vol 29no 3 pp 306ndash310 1970

[47] J Ziv and A Lempel ldquoA universal algorithm for sequentialdata compressionrdquo IEEE Transactions on Information Ceoryvol 23 no 3 pp 337ndash343 1977

[48] Y Pan S S Ge A Al Mamun and F R Tang ldquoDetection ofseizures in EEG signal using weighted locally linear embed-ding and SVM classifierrdquo in Proceedings of the IEEE Inter-national Conference on Cybernetics and Intelligent Systems(CIS rsquo08) pp 358ndash363 Chengdu China September 2008

[49] L Cheng X H Wu and Y Wang ldquoArtificial flora (AF)optimization algorithmrdquo Applied Sciences vol 8 no 239pp 1ndash22 2018

[50] S Mannar and S N Omkar ldquoSpace suit puncture repair usinga wireless sensor network of micro-robots optimized byGlowworm Swarm Optimizationrdquo Journal of Micro-nanoMechatronics vol 6 no 3-4 pp 47ndash58 2011

[51] R Soto B Crawford R Olivares et al ldquoOnline control ofenumeration strategies via bat algorithm and black holeoptimizationrdquo Natural Computing vol 16 no 2 pp 241ndash2572017

[52] T-H Yi H-N Li and X-D Zhang ldquoA modified monkeyalgorithm for optimal sensor placement in structural healthmonitoringrdquo Smart Materials and Structures vol 21 no 10pp 65ndash69 2012

[53] S K Prabhakar and S-W Lee ldquoTransformation based tri-level feature selection approach using wavelets and swarmcomputing for prostate cancer classificationrdquo IEEE Accessvol 8 pp 127462ndash127476 2020

[54] A K Jain P W Duin and J Jianchang Mao ldquoStatisticalpattern recognition a reviewrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 22 no 1 pp 4ndash372000


[55] S K Prabhakar and S-W Lee ldquoAn integrated approach forovarian cancer classification with the application of stochasticoptimizationrdquo IEEE Access vol 8 pp 127866ndash127882 2020

[56] S K Prabhakar H Rajaguru and S-H Kim ldquoFuzzy-inspiredphotoplethysmography signal classification with bio-inspiredoptimization for analyzing cardiovascular disordersrdquo Diag-nostics vol 10 no 10 p 763 2020

[57] C L Jiang ldquoLinear discriminatory analysis on coal and gasoutburst danger and its critical valuerdquo Journal of ChinaUniversity of Mining amp Technology vol 29 no 1 pp 63ndash662000

[58] Y P Mack ldquoLocal properties of k-NN regression estimatesrdquoSIAM Journal on Algebraic Discrete Methods vol 2 no 3pp 311ndash323 1981


and financial issues followed due to various health andmedical problems [7] +ere is definitely no sure method toprevent schizophrenia but considering and taking thetreatment plan effectively can help in preventing relapses [8]Characterized by relapsing episodes of psychosis it is aserious disordermental illness To determine the neuraldynamics of human cognition EEG recording acts as asensitive tool as it can provide a millisecond-level resolution[9] EEG data are complex and at the same time they aredimensional too as they are dependent on an event of timeseries [10] As EEG signals provide electrical activity of thebrain it is easy to analyze the schizophrenia patient data

Some of the most common works related to schizo-phrenia EEG signal analysis and classification reported inthe literature are as follows An EEG-dependent nonlinearityanalysis technique for schizophrenia diagnosis was con-ducted by Zhao et al [11] +e abnormal EEG complexity inpatients with schizophrenia and depression was performedby Li et al [12] Fractal dimension was used by Raghavendraet al [13] to analyze the complexity of EEG in schizophreniapatients +e complexity measures and entropy for EEGsignal classification of both schizophrenia and controlparticipants were performed by Sabeti et al [14] A pre-liminary report on the reduced nonlinear complexity andchaos during sleep in the first episode schizophrenia wasgiven by Keshavan et al [15] A machine learning-baseddiagnosis of schizophrenia using combined sensor-level andsource level EEG features was proposed by Shim et al [16] Apreliminary data analysis of comparing the EEG nonline-arity in deficit and nondeficit schizophrenia patients wasconducted by Cerquera et al [17] +e nonlinear analysis ofEEG in schizophrenia patients with persistent auditoryhallucination was performed by Lee et al [18] For aschizophrenia patient the estimate of the first positiveLyapunov exponent of the EEG signal was performed byKim et al [19] A magnetoencephalography (MEG) study onthe LZC in schizophrenia patients was conducted by Fer-nandez et al [20] A multiscale entropy analysis for theabnormal EEG complexity signals was performed byTakahashi et al [21] As a diagnostic test for schizophreniathe spectral EEG abnormality was analyzed by Boutros et al[22] An in-depth analysis of the utility of quantitative EEGin unmedicated schizophrenia was conducted by Kim et al[23] For schizophrenic and healthy adults the machinelearning identification of EEG features which helps inpredicting working memory performance was done byJohannesen et al [24] A data-driven methodology forresting state EEG signal classification of schizophrenia withcontrol participants using random matrix theory was de-veloped by Liu et al [25] Deep-learning methods along withrandom forest and voting classifiers was performed by Chuet al for the individual recognition in schizophrenia usingresting-state EEG streams [26] Convolution Neural Net-works (CNNs) along with the Pearson Correlation Coeffi-cient (PCC) to represent the EEG channel relationships wereused to classify the schizophrenic and healthy patients usingEEG signals by Naira and Alamo [27] Random ForestMachine learning algorithm was used to identify and di-agnose the schizophrenia EEG signals by Zhang [28] A

higher order pattern discovery was used to classify theschizophrenia EEG by Sheng et al [29] +e complexity ofthe EEG signals in schizophrenia syndromes was analyzed byKutepov et al [30] A fractal-based classification of EEGsignals for both schizophrenic and healthy patients wasperformed by Namazi et al [31] A new approach for EEGsignal classification using Linear Discriminant Analysis(LDA) and Adaboost was found for schizophrenic andcontrol participants by Sabeti et al [32] In this work withthe advent of some features optimization techniques andclassifiers the schizophrenia EEG signal classification isperformed and this attempt is the first of its kind inschizophrenia EEG signal classification +e organization ofthe work is as follows Section 2 explains the materials andmethods Section 3 explains about the feature extractiontechniques Section 4 explains about the swarm intelligencecomputing techniques Section 5 explains about the classi-fication techniques Section 6 explains the results and dis-cussion followed by the conclusion in Section 7

2 Materials and Methods

+e EEG dataset for 14 healthy subjects and 14 Schizo-phrenic subjects was collected from the Institute of Psy-chiatry and Neurology Warsaw Poland and the details areexplained clearly in [33] For the subjects 7 males with anaverage age of 279 + 33 years and 7 females with an averageage of 283 + 41 years were selected +e standard 10ndash20International system was used for the acquisition of the EEGdata +e patientsrsquo data were obtained in a relaxed state andwith their eyes closed+e segmentation of the acquired EEGsignals was performed where it is considered to be sta-tionary A very simplified pictorial representation of thework is given in Figure 1

Over the duration of 15 minutes 19 channel EEG signalsare obtained Every channel of EEG signals comprises of225000 samples which are then divided into groups of5000 sample segments +erefore the data matrix of[5000times 45] is framed per channel For the preprocessing ofEEG signals Independent Component Analysis (ICA) wasutilized in this work

3 Feature Extraction Techniques

To describe a very large amount of data feature extraction isnecessary as it involves in mitigating the number of re-sources to a certain limit Once the preprocessing of the EEGsignals is conducted the following features are extractedfrom the EEG signals as follows

(i) DFA to trace the self-similarity characteristics ofthe EEG signal DFA [34] is used

(ii) Hurst exponent the self-similarity and the pre-dictability estimation of the EEG signal areexpressed by the Hurst exponent [35] If themagnitude value of the Hurst exponent is greaterthen it denotes that the EEG signal is pretty smoothand less complicated





D log10(H)

log10(d) (1)













EEG signals


GSoptimization


AFoptimization

BSoptimization


MHoptimization










d1hprime d2h (3)


d2hprime


prime1113872 11138732

N

1113971

(4)









l


Fmax

1113974111397211138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868





Pseudocode 1



For i 1 NlowastB








421 Procedure






j



lt qj

d(n) fj(n)lefh(n)1113882 1113883 (9)





expressed as


yh(n) minus yj(n)

⎛⎝ ⎞⎠ (11)


qj


d(n) β in minus Ij(n)11138681113868111386811138681113868

111386811138681113868111386811138681113874 11138751113882 11138831113882 1113883 (12)













R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)










g jkprime aj1113872 1113873


2Δajk

(15)



and z (z1 z2 zn)









vk (1M) 1113936Mj1 ajk k 1 2 n



(6) Termination






H [(A + B)d] (16)






Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References


































































D log10(H)

log10(d) (1)













EEG signals


GSoptimization


AFoptimization

BSoptimization


MHoptimization










d1hprime d2h (3)


d2hprime


prime1113872 11138732

N

1113971

(4)









l


Fmax

1113974111397211138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868





Pseudocode 1



For i 1 NlowastB








421 Procedure






j



lt qj

d(n) fj(n)lefh(n)1113882 1113883 (9)





expressed as


yh(n) minus yj(n)

⎛⎝ ⎞⎠ (11)


qj


d(n) β in minus Ij(n)11138681113868111386811138681113868

111386811138681113868111386811138681113874 11138751113882 11138831113882 1113883 (12)













R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)










g jkprime aj1113872 1113873


2Δajk

(15)



and z (z1 z2 zn)









vk (1M) 1113936Mj1 ajk k 1 2 n



(6) Termination






H [(A + B)d] (16)






Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References






































































d1hprime d2h (3)


d2hprime


prime1113872 11138732

N

1113971

(4)









l


Fmax

1113974111397211138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868111386811138681113868111386811138681113868





Pseudocode 1



For i 1 NlowastB








421 Procedure






j



lt qj

d(n) fj(n)lefh(n)1113882 1113883 (9)





expressed as


yh(n) minus yj(n)

⎛⎝ ⎞⎠ (11)


qj


d(n) β in minus Ij(n)11138681113868111386811138681113868

111386811138681113868111386811138681113874 11138751113882 11138831113882 1113883 (12)













R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)










g jkprime aj1113872 1113873


2Δajk

(15)



and z (z1 z2 zn)









vk (1M) 1113936Mj1 ajk k 1 2 n



(6) Termination






H [(A + B)d] (16)






Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References
































































421 Procedure






j



lt qj

d(n) fj(n)lefh(n)1113882 1113883 (9)





expressed as


yh(n) minus yj(n)

⎛⎝ ⎞⎠ (11)


qj


d(n) β in minus Ij(n)11138681113868111386811138681113868

111386811138681113868111386811138681113874 11138751113882 11138831113882 1113883 (12)













R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)










g jkprime aj1113872 1113873


2Δajk

(15)



and z (z1 z2 zn)









vk (1M) 1113936Mj1 ajk k 1 2 n



(6) Termination






H [(A + B)d] (16)






Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References































































R Gcl

1113936Sj1 Gj

1113868111386811138681113868111386811138681113868111386811138681113868

1113868111386811138681113868111386811138681113868111386811138681113868 (14)










g jkprime aj1113872 1113873


2Δajk

(15)



and z (z1 z2 zn)









vk (1M) 1113936Mj1 ajk k 1 2 n



(6) Termination






H [(A + B)d] (16)






Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References


































































Minimize12w

2+ C 1113944

n

j1ξj



(17)






J(W) W

TSBW

WTSWW

(18)







Sensitivity PC



Specificity PC




2 (21)



PC1113874 1113875 times 100 (22)


GDR [PC minus MC]

[PC + FA]1113888 1113889 times 100 (23)



MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References































































MSE 1N

1113944

N

i1Oi minus Tj1113872 1113873

2 (24)



















entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References



































































entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkey search




entropyFractal


Artificial flora


Glowwormswarm


Black hole


Monkeysearch






Fractaldimension


































Data Availability




Acknowledgments


References





























































































Data Availability




Acknowledgments


References































































schizophreniaeegsignalclassificationbasedonswarm … · 2020. 8. 12. ·...

Documents