EEG BRAIN SIGNAL CLASSIFICATION FOR EPILEPTIC SEIZURE DISORDER DETECTION EEG BRAIN SIGNAL CLASSIFICATION FOR EPILEPTIC SEIZURE DISORDER DETECTION SANDEEP KUMAR SATAPATHY SATCHIDANANDA DEHURI ALOK KUMAR JAGADEV SHRUTI MISHRA AcademicPressisanimprintofElsevier 125LondonWall,LondonEC2Y5AS,UnitedKingdom 525BStreet,Suite1650,SanDiego,CA92101,UnitedStates 50HampshireStreet,5thFloor,Cambridge,MA02139,UnitedStates TheBoulevard,LangfordLane,Kidlington,OxfordOX51GB,UnitedKingdom ©2019ElsevierInc.Allrightsreserved. Nopartofthispublicationmaybereproducedortransmittedinanyformorbyanymeans,electronicor mechanical,includingphotocopying,recording,oranyinformationstorageandretrievalsystem,without permissioninwritingfromthepublisher.Detailsonhowtoseekpermission,furtherinformationabout thePublisher’spermissionspoliciesandourarrangementswithorganizationssuchastheCopyright ClearanceCenterandtheCopyrightLicensingAgency,canbefoundatourwebsite:www.elsevier.com/ permissions. ThisbookandtheindividualcontributionscontainedinitareprotectedundercopyrightbythePublisher (otherthanasmaybenotedherein). Notices Knowledgeandbestpracticeinthisfieldareconstantlychanging.Asnewresearchandexperience broadenourunderstanding,changesinresearchmethods,professionalpractices,ormedicaltreatment maybecomenecessary. Practitionersandresearchersmustalwaysrelyontheirownexperienceandknowledgeinevaluatingand usinganyinformation,methods,compounds,orexperimentsdescribedherein.Inusingsuchinformation ormethodstheyshouldbemindfuloftheirownsafetyandthesafetyofothers,includingpartiesfor whomtheyhaveaprofessionalresponsibility. Tothefullestextentofthelaw,neitherthePublishernortheauthors,contributors,oreditors,assumeany liabilityforanyinjuryand/ordamagetopersonsorpropertyasamatterofproductsliability,negligenceor otherwise,orfromanyuseoroperationofanymethods,products,instructions,orideascontainedinthe materialherein. LibraryofCongressCataloging-in-PublicationData AcatalogrecordforthisbookisavailablefromtheLibraryofCongress BritishLibraryCataloguing-in-PublicationData AcataloguerecordforthisbookisavailablefromtheBritishLibrary ISBN978-0-12-817426-5 ForinformationonallAcademicPresspublications visitourwebsiteathttps://www.elsevier.com/books-and-journals Publisher:StacyMasucci AcquisitionEditor:RafaelE.Teixeira EditorialProjectManager:SamuelYoung ProductionProjectManager:MariaBernard CoverDesigner:MarkRogers TypesetbySPiGlobal,India PREFACE EEGsignalsaretheelectricwavesgeneratedinthehumanbrainwhichisthe main cause of different tasks performed by a person. These signals are very small in magnitude but collectively it defines the behavior of a person. Hence,carefulandcompleteanalysisofthesesignalscansolvedifferentdis- eases occurring in our brain. High volume and uncertainty of these signals make it very difficult for any medical person to diagnose a brain disorder diseasejustbylookingatthegraphicalrepresentationofthesesignals.Hence, there are plentiful opportunities for computer scientists to produce a computer-based model that can detect efficiently any specific neurological disease after proper analysis of these signals. The signals recorded from human brain contain several features, those needs to be extracted by some featureextractiontechnique.Afterthisanydiseasedetectionproblemcanbe solved by the application of different data mining techniques. The classifi- cation technique is basically used for solving these kinds of problems that takes the help of different supervised or unsupervised machine learning techniques. TheprimaryobjectiveofthisresearchistoclassifytheEEGbrainsignal for detection of epileptic seizures using machine learning techniques. Machine learning techniques have recently developed most efficient tech- niquesforsolvingdifferentkindsofproblems.Wearemainlyconcentrating on supervised machine learning techniques. These techniques require a prior knowledge about data that make the machine learn for recognizing newdataandperformclassificationoperation.Butitisnecessarytoprepro- cesstherawEEGdatacollectedfrompatientstobringthemintoaformfea- turesample-baseddataset.Thesignalsareanalyzedbydecompositionusing DiscreteWaveletTransform(DWT)withDaubechieswaveletoforder2up tolevel4.Afterthisdecompositiondifferentstatisticalfeaturesfordifferent coefficient are collected together to build the dataset. We performed an empirical survey of different machine learning techniques applied for designing classifier model for epilepsy identification. The different tech- niquesusedareMultilayerPerceptronNeuralNetwork(MLPNN)withdif- ferentlearningalgorithms(suchasback-propagation,resilient-propagation, and Manhattan update rule), different variations of neural network such as ProbabilisticNeuralNetwork(PNN),RecurrentNeuralNetwork(RNN), RadialBasisFunctionNeuralNetwork(RBFNN),andsoon,andSupport vii viii Preface Vector Machines (SVM) with different kernel functions (such as linear, polynomial, and RBF kernel). Further, we have enhanced the performance of RBFNN network by implementing different swarm intelligence-based optimization techniques as training algorithm. Swarm intelligence is one of the most efficient fields fortheoptimizationproblems.Theswarmintelligencemethodsusedinthis work are Particle Swarm Optimization (PSO) and Artificial Bee Colony (ABC) algorithms. This work proposes a novel hybrid technique to train RBFNNbyoptimizingtheparametersofthisnetworkusingtheimproved PSOalgorithmandABCalgorithm.PSOisoneofthemosteffectiveopti- mization technique used so far. But again we have modified the existing PSO algorithm to enhance the performance by adding a new technique forfindingthevalueoftheinertiaweightusedinthecalculationofvelocity update.Theexperimentalevaluationshaveprovedthattheperformanceof theproposedtechniqueishigherthanRBFNNtrainedwithgradientdecent approachaswellasRBFNNtrainedwiththeconventionalPSOalgorithm. WehavealsosuggestedachangetoexistingABCalgorithm,wherewehave usedtournamentselectioninplaceofroulettewheelselectioninonlookers beesphase.Besides,theoperationoftheproposedtechniquewasshownto bemoreefficientbydifferentexperimentalevaluations.Forthiswork,ABC algorithmwaschosenasmoreefficientalgorithmbecauseofthelessnumber ofdependableparametersandimprovedaccuracyofRBFNNascompared to PSO algorithm. Along with classification accuracy, specificity, sensitivity, and many other different parameters were considered for measuring the performance ofclassificationalgorithms.Also,theexperimentalevaluationsaresupported by a k-fold cross validation technique to validate the results obtained by different experiments. CHAPTER 1 Introduction Electroencephalogramorelectroencephalography(EEG)isatrialperformed onmentalcapacitytorecordtheelectricalactivityinbrain.Theneuralstruc- tureofthebrainconsistsofseveralneuronsintermsoflacs.Theseneurons communicate by colliding among themselves and passing information to each other. This collision leads to the generation of a very small amount of electricity. This is utterly different from the general electricity, which is very high in magnitude. This electric signal flow decides the behavior of a person. In a human brain the normal stream of electrical signal leads to a healthy person. And an abnormal electrical signal flow can pass to an unhealthyperson.Hence,thesesignscanberecordedandanalyzedtosolve manyneurologicaldisorderdiseases.Thetranscriptionoftheelectricalactiv- ity is essentially caused by putting electrodes on the scalp for 20–40min, which evaluates the potential fluctuations in the brain [1]. The nerve cells inthepsychearetheoriginofelectriccharge,andthentheyexchangeions withtheextracellularmilieu.Ionsofthesamechargerepeleachotherandin this manner they are forced out of the neurons when a number of ions are drivenoutatthesametimetheypromoteeachotherandformawayknown asvolumeconduction[2].Whenthiswavereachestheelectrodetheypush or force the ions along the air foil of the electrode which create potential differenceandthisvoltagedifferencerecordedovertimegivesEEGsignals. Thekeymotivationbehindthisresearchworkistherapidgrowthinvolume ofbiological andclinicaldata orrecords.Toextractknowledgefrom these datawhichcanbeservedtobeaclinicalapplication,therearedifferentdata analysisdifficultieswhichneedtobeovercome.Manyanalyticaltoolsbased on machine learning (ML) approaches have been invented to tackle with suchchallengingtaskofdataanalysisproblems.Around1%ofthetotalpop- ulation in the world are affected by a neurological disease called epilepsy. AcarefulanalysisoftheseEEGsignalscansolvemanyneurologicaldisorder diseases. EEGBrainSignalClassificationforEpilepticSeizureDisorderDetection ©2019ElsevierInc. 1 https://doi.org/10.1016/B978-0-12-817426-5.00001-6 Allrightsreserved. 2 EEGBrainSignalClassificationforEpilepticSeizureDisorderDetection 1.1 PROBLEM STATEMENT Nowadays,therecordingofEEGsignalcanbeeasilymanagedwiththeaid ofvarioushardwareandsoftwaretechniques.Bysimplyseeingatthesesig- nals with naked eyes one cannot make out any abnormality in the sign. Hence, the most important problem is to study these signals properly and extract thehiddenfeatures presentinside.Aneurological diseasecanoccur in a human brain due to abnormal EEG flow. This abnormality should be properlyanalyzedtospecifythepatternofthisdiseasethatcanhelpwithpre- dictionofanysuchtypeofdiseasesinhumanbrains.EEGrecordinggener- allyleadstothecollectionofahugeamountofnumericalinformation that consists of the state of electrical activity at different time. This recording is generallytakenfor10–15transactions.Thisdurationissufficienttounder- standthestateofahumanbrainwhichleadstocollectionofhugequantities of data. By plotting this information graphically, we can conclude some behaviorofthebrain,thoughnotcompletely.Asaresult,itismoresignif- icant to collect these data and pull information from this. In this research study,themainconcentrationisontheneurologicaldisorderdiseasecalled epilepsy.Thewholeproblemofthisresearchworkhasbeenbroadlyclassi- fied into two groups. First,is feature extraction and analysis of a very non- stationarysignallikeEEGsignal.Second,istheclassificationofEEGsignals to detect epileptic seizures. First,wehavedevelopedawell-definedandwell-structuredprocessfor extractingthehiddenfeaturesfromaverytransientandnonstationarysignal likeEEGsignal.Forthisasignaltransformationtechniquecalledasdiscrete wavelet transform (DWT) was used. To compare the significance of these extractedfeatures,otherfeaturesbasedonsomemathematicalcomputations were also extracted. Second, we have developed a well-defined and most efficient classifier modelthatcanidentifyanddistinguishepilepticseizuresfromnonepileptic ones. For this, we have considered mostly ML-based classification techniques likeANNbasedclassifiers,supportvectormachines(SVM),andevolution- ary theory-based classifiers. 1.2 GENERAL AND SPECIFIC GOALS AverycarefulanalysisofEEGsignalcanprovideresponseformanyneuro- logical disorder diseases. These signals are mainly responsible for different abnormalitiesgeneratedinthehumanbrain.Decadesagoitwasnotpossible Introduction 3 tostudythesesignalsduetohighpriceoftheequipmentandnonavailability of adequate engineering sciences. But nowadays these problems have been resolvedandmanyhardwareequipmentareavailablewithareasonableprice forrecordingoftheEEGsignal.Butstilltechnologicalimprovementinthe depthpsychologyofthesesignalsisgettingalong.Manyresearchersareput- ting their enormous efforts in this area. Recording of EEG signal is generally made for approximately 20–30min.Inthatrespectseveralelectrodesareplacedonthehumanscalp torecordtheseEEGsignals.Normallyaplentyofdataarebroughtforthdur- ingthisrecording,whichisnotpossibleforamedicalpersontoanalyzethem in naked eyes. Hence, there are several technologies developed to record thesesignalsanddirectlyplotthegraphsbyusingsoftware,whichareviewed byaspecialistandsomeobservationscanbemadeonthestateofthebrain. Routine EEG is a process which is applied for diagnosis in the following conditions. 1. To distinguish between epileptic seizures from nonepileptic seizures, syncope (fainting), subcortical movement disorders, and migraine variants. 2. To serve as an adjunct test of mind destruction. 3. To prognosticate in certain situations like patients with coma. 4. To determine whether to wean antiepileptic medications. IncertainsituationstheroutineEEGisnotsufficient,eventuallythepatients haveseizure.Inthiscase,thepatientsaretakentohospitalandEEGiscon- stantly recorded. This is mostly done to distinguish between epileptic sei- zures and other seizures. The most important advantage of EEG is its speed. A very complex neural activity can be read within a few minutes orintermsofseconds.EEGproducesaverylittlespatialresolutionascom- paredtomagneticresonanceimaging(MRI)andpositronemissiontomog- raphy(PET).ThussometimesEEGimagesarecombinedwithMRIscans. AccordingtoBickfordetal.[3]thedifferentclinicalapplicationofEEGin human and animals is used in following situations: 1. Monitor alertness, coma, and brain death. 2. Locateregionsofimpairmentfollowingheadtrauma,stroke,ortumor. 3. Monitor cognitive engagement. 4. Produce biofeedback situations. 5. Control anesthesia depth. 6. Investigate epilepsy and locate the seizure origin. 7. Test epilepsy drug effects. 8. Assist in experimental cortical excision of epileptic foci. 4 EEGBrainSignalClassificationforEpilepticSeizureDisorderDetection 9. Investigate sleep disorder and physiology. 10. Monitoring amobarbital effect during Wada test. Other than clinical uses there are different applications where EEGs are extensivelyused.Braincomputerinterface(BCI)isatypicalcommunication systemthatrecognizesuser’scommandonlythroughtheirbrainwavesand responds accordingly. It is frequently known as mind machine interface or brainmachineinterface.Itcanbeusedforpeoplewhoareunabletoexpress through speechor physicalactivities. The basicfunctionof thesedevicesis to intercept electrical signals that occur between nerve cells in the brain and render them into a signal that can be read by an external device. The different cases of BCI are invasive BCI, partially invasive BCI, and noninvasive BCI. Inthisresearchwork,theboundaryhasbeenlimitedonlytotheanalysis andclassificationofEEGsignalsforepilepticseizureidentification.Approx- imately 1% of the entire population is touched on by the disease known as epilepsy.Thehumanbrainisthemostcriticalpartofbodythatcontrolsthe coordination of human muscles. The transient and unexpected electrical interferencesinthebrainresultinanacutediseasecalledasepilepsyorepi- lepticseizure.Itisoneofthemostcommondiseasesaroundtheglobe,strik- ing more than 40 million people worldwide. 1.3 BASIC CONCEPTS OF EEG SIGNAL The EEG signal is normally employed for the purpose of recording down theelectricalactionsofthebrainsignalthattypicallyoriginatesinthehuman brain.AnEEGsignalismeasuredintermsofcurrentsthatflowduringexci- tationofsynapticactivitiesofmanyneuronsinthecerebralcortex[4].Dur- ing this excitation of brain cells, the small magnitude of currents are produced within the dendrites. As a result, these currents generate a mag- netic field which can be measured by electromyogram machines. DepthpsychologyandmeasurementofEEGsignalisappliedfortesting the stateandactivity of human brain.LikePETandMRI, EEGmethodis alsousedwidelyduetoitspowerofprovidingbesttemporalresolutionand lowprice.Humanbrainconsistsofvariousneurons.Whentherewillbeany flowofinformationintothebrain,theseneuronshiteachother.Duringthis procedureelectricityisgenerated,whichareverysmallinamountandtran- sient innature.Thisislikewisecalledasanonstationarysignal,becausethe frequency of this signal is not set up for a certain amount of time. Biolog- ically,itcanbeassumedthat,duringtheactivationofbraincells,thesynaptic currents are created within the dendrites [4].