Optimal Fuzzy Min-Max Neural Network (FMMNN) for Medical Data Classification Using Modified Group Search Optimizer Algorithm

Transcription

1 1 Opmal Fuzzy Mn-Max Neural Nework (FMMNN) for Medcal Daa Classfcaon Usng Modfed Group Search Opmzer Algorhm D. Mahammad Raf 1 * Chear Ramachandra Bharah 2 1 Vvekananda Insue of Engneerng & Technology, JNTU Unversy, Hyderabad, Inda 2 Vel Tech Unversy, Chenna, Tamlnadu, Inda *Correspondng auhor s Emal: dmahammadraf0780@gmal.com Absrac: The man nenson of he research s o classfy he medcal daa wh hgh accuracy value. In order o acheve promsng resuls, we have desgned o ulze orhogonal local preservng proecon and opmal classfer. Inally, pre-processng wll be appled o exrac useful daa and o conver suable sample from raw medcal daases. In he proposed mehod, npu daase wll be hgh dmensonal or hgh feaures; so he hgh number of feaure s a grea obsrucon for predcon. Therefore, feaure dmenson reducon mehod s used n our proposed echnque. Here, orhogonal local preservng proecon wll be used o reduce he feaure dmenson. Once he feaure reducon s formed, he predcon wll be done by opmal classfer. Here modfed group search opmzer algorhm combned wh Fuzzy Mn-Max neural nework. The mplemenaon wll be done n MATLAB. The performance of he proposed echnque s evaluaed usng accuracy, sensvy and specfcy. Keywords: orhogonal local preservng proecon; group search opmzer algorhm; Fuzzy Mn-Max neural nework; opmal classfer. 1. Inroducon Daa mnng s an essenal sep n he process of knowledge dscovery n daabases n whch nellgen mehods are appled n order o exrac paerns [1]. I s he process of analyzng daa from dfferen perspecves and summarzng no useful nformaon. The man goal of daa mnng s o dscover new paerns for he users and o nerpre he daa paerns o provde meanngful and useful nformaon for he users. I appled o fnd useful paerns o help n he mporan asks of medcal dagnoss and reamen [2]. The algorhms, when appropraely used, are capable of mprovng he qualy of predcon, dagnoss and dsease classfcaon [3]. Wh daa echnque such knowledge can exraced and accessed ransformng he daa base asks form sorng and rereval o learnng and exracng knowledge [4]. Classfcaon of medcal daa for accurae dagnoss s a growng feld of applcaon n daa mnng [5]. I s a process of fndng he class model accordng o her arbues. There are wo dfferen caegory of learnng process namely supervsed learnng and unsupervsed learnng. Decson ree s a common classfcaon algorhm and wdely used n many applcaons. Decson ree algorhm ncludes Classfcaon and Regresson Tree [6]. The classfy panel enables he user o apply classfcaon and regresson algorhms o he resulng daase, o esmae he accuracy of he resulng predcve model, and o vsualze erroneous predcons, or he model self [7]. An ensemble model s defned as a se of ndvdually raned classfer whose predcons are combned when classfyng a new daa. Ensemble combnes he oupu of several classfers produced by weak learner no a sngle compose classfcaon. I can be used o reduce he error of any weak learnng algorhm [8]. Classfcaon model could make predcaon of caegorcal label nclude dscree or unordered varables [9]. A daa se s mbalanced f he classfcaon caegores are no equally represened [4]. Classfcaon dvdes daa samples no arge classes. The classfcaon echnque predcs he arge class for each daa pons. Bnary and mullevel are he wo mehods of classfcaon. In bnary classfcaon he performance and qualy of classfcaon algorhms s usually evaluaed usng predcve accuracy [10]. No wh sandng, Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

2 2 hs s no suable when he nformaon s unequvocally mbalanced as dspares n he quany of ems beween he classes may promp serous weakenng of he groupng exacness [11]. Daa mnng and knowledge dscovery n daabases have been aracng sgnfcan amoun of research n felds lke ndusry medcal, commerce, scence whch s aracng he meda aenon of lae [12]. Afer, classfer was bul, hs model would be evaluaed accordng o he answers accuracy ha model wll gve hrough esng. Whereas, anoher daa, unlabeled examples ha are known es daa, provde he model [13]. Many successful applcaons ha whch are based on assocaon rule mnng algorhms are used o produce very large numbers of rules. In mos of he decson suppor sysems he accuracy of classfer s measured on he bass of all arbues [14]. Varous algorhms and echnques are used n he medcal daa mnng lke Classfcaon, Cluserng, Regresson, Arfcal Inellgence, Neural Neworks, Assocaon Rules, Decson Trees, Genec Algorhm, Neares Neghbor mehod [15]. 2. Relaed Work In hs secon, we have dscussed some recen papers abou medcal daa classfcaon. M. A. Jabbar e al [16] have proposed a mehod ha o dscover assocaon rules n medcal daa o predc hear dsease for Andhra Pradesh. Tha approach was expeced o help physcans o make accurae decson. Moraly daa from he regsrar general of Inda shows ha he coronary hear dsease (CHD) were a maor cause of deah n Inda. They deermne he precse cause of deah n rural areas of Andhra Pradesh have revealed ha CHD cause abou 30% deah n rural areas. Indu San e al [17] have proposed algorhm was evaluaed on wo manually annoaed sandard daabases such as CSE and MIT-BIH Arrhyhma daabase. Tha work ells abou dgal band-pass fler was used o reduce false deecon caused by nerference presen n ECG sgnal and furher graden of he sgnal was used as a feaure for QRSdeecon. They also found an addon he accuracy of KNN based classfer was largely dependen on he value of K and ype of dsance merc. The deecon raes of 99.89% and 99.81% were acheved for CSE and MIT-BIH daabases respecvely. Sna Khanmohammad and Mandana Rezaeahar [18] have proposed o choose a machne learnng classfcaon ha has been used for developng clncal decson suppor sysem. They Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016 presened en sample medcal daases. They sugges model SVM as he mos desrable classfcaon algorhm for developng CDSS. The research was no o denfy a classfcaon algorhm ha has been performng bes n all medcal daases. The relably of he model has been mproved usng more sample daases. R. Chra and V. Seenvasagam [19] have proposed a growng research on hear dsease predcng sysem, ha has been become mporan o caegores he research oucomes and provdes readers wh an overvew of he exsng hear dsease predcon echnques n each caegory. Neural Neworks were one of many daa mnng analycal ools ha have been ulzed o make predcons for medcal daa. R. Bhuvaneswar and K. Kalaselv [20] have proposed he use of Nave Baye s classfer n medcal applcaons. Qualy servce mples dagnosng paens correcly and admnserng reamens ha are effecve. Decson Suppor n Hear Dsease Predcon Sysem was also developed usng Nave Bayesan Classfcaon echnque. Naïve Bayes classfcaons have been used as a bes decson suppor sysem. P K. Anoo [21] have proposed a weghed fuzzy rule-based clncal decson suppor sysem (CDSS) was presened for he dagnoss of hear dsease, auomacally obanng knowledge from he paen s clncal daa. They conss wo phases namely: (1) Auomaed approach for he generaon of weghed fuzzy rules and (2) Developng a fuzzy rule-based decson suppor sysem. M. Akhl abbar e al [22] have proposed a combnes approach of KNN and genec algorhm have o mprove he classfcaon accuracy of hear dsease daa se. They used genec search as a goodness measure o prune redundan and rrelevan arbues, and o rank he arbues whch have been conrbue more owards classfcaon. Leas ranked arbues were removed, and classfcaon algorhm s bul based on evaluaed arbues. A. Sudha, e al [23] have proposed a prncple componen analyss algorhm was used for reducng he dmensons and ha have been deermnes he arbues nvolvng more owards he predcon of sroke dsease and predcs wheher he paen sufferng from sroke dsease or no. Tha algorhm deals abou sroke was a lfe hreaenng dsease ha has been ranked hrd leadng cause of deah n saes and n developng counres. The sroke was a leadng cause of serous, long erm dsably n US.

3 3 Mar Juhola e al [24] have proposed complcaed varable dsrbuon of he daa alhough here were only wo classes. In addon o a sraghforward daa cleanng mehod, hey used an effcen way called neghborhood cleanng ha solved he problem and mproved her classfcaon accuraces 5 10%, a her bes, up o 95% of all es cases. Tha shows mporan ha was frs very carefully o sudy dsrbuons of daa ses have been classfed and use dfferen cleanng echnques n order o oban bes classfcaon resuls. B. Denns and S. Muhukrshnan [25] have proposed an effcen medcal daa classfcaon sysem based on Adapve Genec Fuzzy Sysem (AGFS). In her research 1) Generang rules from daa as well as for he opmzed rules selecon, adapng of genec algorhm s done and o explan he exploraon problem n genec algorhm, nroducon of new operaor, called sysemac addon s done, 2) Proposng a smple echnque for schemng of membershp funcon and Dscrezaon, and 3) Desgnng a fness funcon by allowng he frequency of occurrence of he rules n he ranng daa. Adeny e al. [26] has presened a sudy of auomac web usage daa mnng and recommendaon sysem based on curren user behavor hrough hs/her clck sream daa on he newly developed Really Smple Syndcaon (RSS) reader webse, n order o provde relevan nformaon o he ndvdual whou explcly askng for. The K-Neares-Neghbor (KNN) classfcaon mehod has been raned o be used onlne and n Real-Tme o denfy clens/vsors clck sream daa, machng o a parcular user group and recommend a alored browsng opon ha mee he need of he specfc user a a parcular me. Ther resul shows ha he K-Neares Neghbor classfer was ransparen, conssen, sraghforward, smple o undersand, hgh endency o possess desrable quales and easy o mplemen han mos oher machne learnng echnques specfcally when here s lle or no pror knowledge abou daa dsrbuon. 3. Problem Defnon In hs secon dscuses he problem defnon of my research work The daa conans redundan and rrelevan arbues classfcaon produce less accurae resul [22]. The sraghforward cleanng of medcal daa se mpared s classfcaon resul consderably wh some machne learnng mehods, bu no all Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016 of hem unexpecedly and agans nuon compare o he orgnal suaon whou any daa cleanng [24]. The avalably of huge amoun of medal daa leads o he need for powerful daa analyss ools o exrac useful knowledge. Dagnoss of mos of he dsease s expensve as many ess are requred o predc dsease. The cos dagnoss by avodng many ess by selecon of hose arbue. Number of work carred ou for predcon varous dsease comparng he performance of predcve daa mnng. The physcs dagnose represened by human experse can be ncurrence o fal. In conras he daa mnng can be recru he exrac knowledge from huge of clncal daa hrough daa mnng and produce a predcve model use he classfcaon ask o acheve he dagnosc [13]. The analyss of daa mnng process requred for medcal daa mnng especally o dscover locally frequency dsease such as hear algnmen lung cancer, breas cancer and so on. Classfcaon accuracy s mproved by removng mos rrelevan feaures he daase. Ensemble model s used for mprovng classfcaon accuracy by combnng he predcon of mulple classfers. The healhcare ndusry gahers enormous of hear dsease daa whch s unforunaely o dscover hdden nformaon for effecve decson The sgnfcan progress made n he dagnoss and reamen of hear dsease, furher nvesgaon s sll needed [16]. Medcal professonals need a relable predcon mehodology o dagnose Dabees. 4. Proposed Mehodology In hs research we have nend o propose an effcen mehod o classfy he medcal daa. In medcal daa classfcaon, n order o acheve beer resul orhogonal local preservng proecon and opmal classfer wll be used n he proposed echnque. A frs he npu daa se s seleced from he medcal daabase. Then preprocessng wll be appled n he npu medcal daa se. In preprocessng sage we have o exrac he useful daa from he raw medcal daase. Afer preprocessng he npu daase wll be hgh dmensonal or hgh feaures; so he hgh number of feaure s a grea obsrucon for predcon. Therefore, feaure dmenson reducon mehod wll

4 4 be appled o reduce he feaures space whou losng he accuracy of predcon. Here, orhogonal local preservng proecon (OLPP) wll be used o reduce he feaure dmenson. Once he feaure reducon s formed, he predcon wll be done based on he opmal classfer. In he opmal classfer, group search opmzer algorhm wll be used wh Fuzzy Mn-Max neural nework. The dealed process of our proposed mehod s shown n Fg.1. frs sep of he algorhm s PCA whch helps n dmensonaly reducon. An adacency graph s bul by OLPP and he class relaonshp beween he sample pons s bes refleced by. I s no easy o reconsruc he daa snce Localy Preservng Proecons (LPP) s non-orhogonal normally. By means of Orhogonal Localy Preservng Proecon mehod, hs problem s overcome whch produces orhogonal bass funcons and can have more localy preservng power han LPP. The Orhogonal exenson of LPP s called as he Orhogonal Localy Preservng Proecons (OLPP). The overall seps are shown n he flowchar: Fgure.1 Block dagram of proposed mehod The overall process of he proposed framework s dvded no hree sages, Sage1: Preprocessng Sage2: Feaure reducon usng OLPP Sage3: Classfcaon usng GFMMNN Sage1: Preprocessng In preprocessng sage he npu daase s gven as he npu. Here he npu medcal daa has raw daa. Ths raw daa s hghly suscepble o nose, mssng values and nconssency. The qualy of raw daa affecs he resuls of he mplemened mehod. In order o mprove he qualy of he medcal daa and consequenly, of he resuls raw daa s pre-processed so as o mprove he effcency and ease of mnng process. Daa pre-processng s one of he mos crcal seps n a daa mnng process whch deals wh he preparaon and ransformaon of he nal daase. In he paper, pre-processng s appled o he daase for geng he numercal daa from he non-numercal daa. In he sage, he non-numercal daa are removed and obaned he numercal daase for proceedng furher. The preprocessed oupu s fed o he furher process. Sage2: Feaure Reducon usng OLPP OLPP algorhm dffers from Prncpal Componen Analyss (PCA) and Lnear Dscrmnan Analyss (LDA). The am of boh algorhms s dmensonaly reducon, snce he Fgure.2 The flowchar for OLPP mehod The seps occuped n OLPP: PCA Proecon: Prncpal Componens Analyss s a mehod ha reduces daa dmensonaly by performng a covarance analyss beween facors. The PCA proecon nvolves he followng seps: () Oban a se of feaures from he npu daabase () Calculae he mean value () Compue covarance marx and hen calculae egen vecor and egen value of covarance marx (v) The egen value and egenvecors are ordered and pared Consrucng he Adacency Graph: Le D d1, d2,..., d m be a se of npu daa. Consder G denoes a graph wh n nodes. The h node corresponds o he npu daa d. An edge s pu beween nodes and, f d andd are close,.e. d s among p neares neghbors of d or d s among p neares neghbors of d. If he Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

5 5 class nformaon s avalable n any wo nodes we smply pu an edge beween ha wo nodes belongng o he same class. Choosng he Weghs: f he node and are conneced, he weghw s calculaed usng he followng equaon, W e d d e (1) Where Consan If he node and are no conneced means we pu W 0. The wegh marx We of e graph G models havng he local srucure of varous npu daa. Compung he Orhogonal Bass Funcons: Afer fndng he wegh marx W e we calculae he dagonal marx M. A dagonal marx M s defned as, whose enres are column (or row) sums ofw e. M W e (2) Afer ha we calculae he Laplacan marx L usng dagonal marx M and wegh marxw e. L M W e (3) Le O r1, Or 2,..., Or m be orhogonal bass vecors, we defne T 1 A m1 O r1, Or 2,..., O m1, r 1 T B m Am1 Z Am1 (4) Where; 1 T Z DMD The orhogonal bass vecors O r, Or, , Or m can be compued as follows () Compue O 1 as he egenvecor of T Z 1 DLD assocaed wh he smalles egenvalue. () Compue Om as he egenvecor of assocaed wh he smalles egenvalue of J m 1 T 1 T I Z A B Z DLD Jm m1 m1 (5) OLPP Embeddng : Le T OLPP O 1 O2 O3.... O l embeddng s follow, Y DT (6) T T PCA T OLPP (7) Where; T Transformaon marx Y One dmensonal represenaon of D T Ths ransformaon marx reduces he dmensonaly of he feaure vecors of he npu daa. Ths dmensonaly reduced feaures, gven o he classfcaon process. Sage3: Classfcaon usng GFMMNN Fuzzy Mn Max Neural Nework FMNN learnng algorhm comprses of hree acons: 1) expanson, 2) overlap es, and 3) conracon. Is prncple s o locae a suable hyperbox for each npu paern. If he suable hyperbox exss, s sze canno surpass he mnmum and maxmum lms. Afer expanson, all hyperboxes ha have a place wh dsncve classes mus be checked by overlap es o fgure ou wheher any overlap exss. So a dmenson by dmenson comparson beween hyperboxes of dfferen class s performed. FMNN nends four es cases, a leas one of he four cases s sasfed, and hen overlap exss beween he wo hyperboxes. Oherwse, a new hyperbox needs o be added o he nework. If no overlaps occur, he hyperboxes are solaed and no conracon s requred. Oherwse, a conracon process s needed o elmnae he confuson n overlapped areas. The ranng se conss of se of ordered pars {X, I}, where X= {X 1, X 2, X n} s he npu daa and Iϵ(1,2, m) s he ndex of one of he class. The learnng process begns by selecng an ordered par and fndng a hyper box for he same class ha can expand (f necessary) o nclude he npu. If a hyper box canno be found ha mees he expanson crera, a new hyper box s formed and added o he neural nework. The membershp funcon s defned wh respec o he mnmum and maxmum pons of a hyper box. I descrbes he degree o whch a paern fs n he hyper box. The hyper boxes have a range from 0 o 1 along each dmenson. A paern whch s conaned n he hyper box has a uny membershp funcon. Mahemacally, he defnon of each hyper box fuzzy se H s defned by, H X, V, W, F( X, V, W )} (8) { mn max mn max Where, X-Inpu daa, V mn (V mn1, V mn2,.v mnn) s he mnmum pons of H W max (W max1, W max2,.w maxn) s he maxmum pons of H F (X, V mn, W max) s he membershp funcon The membershp funcon for he h hyperbox (H) s gven below, Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

6 6 1 H 2n n 1 [max (0,1 max (0, mn (1, X w ))) max (0,1 max (0, mn (1, v X )))] (9) Where, γ s a sensvy parameer ha regulaes how fas he membershp value decreases as he dsance beween X and H ncreases. The archecure of fuzzy mn max neural nework consss of hree layers of node. Frs layer represen he npu layer ha conans npu daa. Las layer represen he oupu layer ha conan he number of classes. The mddle or hdden layer s called hyper box layer. The overall srucure of fuzzy mn max neural nework s shown n Fg.3. Fgure.3 The srucure of fuzzy mn max neural nework MGSO Algorhm Group search opmzaon algorhm (GSO) s proposed o opmze he wegh n he neural nework. The Group search opmzaon algorhm s developed wh he movaon from he searchng acvy of anmals. The searchng acvy of anmals s manly done wh he nenon of dscoverng resources ha nclude food and sheler. Here he radonal GSO algorhm s mproved wh he help of velocy updaon nsead of ranger performance o randomly selec he resource. In hs MGSO algorhm, he populaon s ermed as a group and he ndvduals resdng whn he group are known as members. The members whn a group are of hree knds, namely, he producers, he scroungers and he rangers. The acvy of he producers as well as he scroungers reles on he PS model. The rangers move n an arbrary manner. Producers: These members go n search of resources. Scroungers: These members lnk he resources, whch he producer dscovers. Rangers: These are he members ha make movemens n an arbrary manner and perform searchng n an organzed way, so ha effcen fndng of resources could be acheved. Sep by sep procedure of Modfed GSO:- Sep 1: Inalze he search soluon as well as he head angle The head angle can be saed as, 1... ( n1) (10) The members drecon of search reles on he head angle L ) l... l (11 ) ( 1 ( n) Polar and Caresan coordnae ransformaon s employed o assess he drecon of search L L n 1 1 p 1 sn( cos( p n1 ( 1) p ) cos( p ); (12) Where (=2.n-1) (13) Ln sn( ( n1) ) (14) Sep 2: Fness funcon s calculaed fness mn ( MSE) (15) Sep 3: Fnd he producer (Z p) of he group The member wh he bes fness s called as he producer Producer performance Durng he execuon of he MGSO algorhm, he acvy of he producer Z p a eraon can be elucdaed as follows, () Scannng operaon a zero degree Z Z d L (16 z p 1 max p ) Where, d max denoes he maxmum search dsance. () Scannng operaon a he rgh hand sde Z max r Z p 1d max L p 2 (17) 2 () Scannng operaon a he lef hand sde Z max l Z p 1d max L p 2 (18) 2 Where, ε 1 pons o a normally dsrbued random number wh zero mean and uny sandard Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

7 7 devaon. And ε 2 sands for a unformly dsrbued random sequence ha akes value beween zero and one. The compuaon of maxmum search dsance d max Maxmum search angle Φ max max (19) 2 c The consan c can be saed as: C round ( n 1) (20) Where, n denoes he dmenson of he search space. max (21) n 1 The presen bes locaon would ake a new bes locaon, f s resource s found as no beer han ha n he new locaon. Else, he producer wll manan s locaon and urn s head accordng o he head angle drecon ha s arbrarly generaed. Scrounger performance In all eraons, several members ha exclude he producer are also chosen and hey are called as scroungers. The scroungng acon of GSO usually nvolves he area copyng acvy. 1 Z Z ( Z Z ) (22) 3 Where, ᴏ specfes he Hadamard produc ha compues he produc of he wo vecors n an enrywse manner and ε 3 denoes a unform random sequence lyng n he nerval of (0, 1). Ranger performance The rangers are he remanng members of he group, whch have been dsplaced from her presen locaon. The rangers can also fnd he resources effecvely hrough random walks or an organzed searchng procedure. Random walks are preferred n cases, where he resources are found o be dsrbued. In our Modfed GSO here nsead of ranger performance we go for he velocy Updaon. V new V c2. r2.( IGbes S ) (23) C1 and C2 consans conanng value of 2 r1 and r2 random numbers equally produced n he range [0, 1] S nal poson of h parcle Once he enre process ges compleed, he fness of he updaed soluon s evaluaed. The bes soluon wll be ganed, f he process s repeaed for number of eraons. Based on hese he weghs are opmzed. Fnally we classfy he medcal daa wh hgh predcon value. p c. r.( IP 1 1 bes S ) 5. Resuls and Dscusson The proposed sysem s mplemened usng MATLAB 2014 and he expermenaon s performed wh 5 processor of 3GB RAM. Daase descrpon The proposed mehod s expermened wh he four daase namely Kdney chronc, Cleveland, Hungaran and Swzerland. These daases are aken from he UCI machne learnng reposory. () Mammographc Mass Daa Se The mammographc mass daase used here has been colleced a he Insue of Radology of he Unversy Erlangen-Nuremberg beween 2003 and The daa se s avalable by hp access of he Unversy of Calforna a Irvne (UCI) machne learnng reposory. Dgal Daabase for Screenng Mammography (DDSM) has been used o evaluae he proposed sysem. The daabase conans approxmaely 2,620 cases. () Pma Indans Dabees Daa Se The source of Pma Indan dabees daa se s he UCI machne learnng reposory. The daa source uses 768 samples wh wo class problems o es wheher he paen would es posve or negave for dabees. All he paens n hs daabase are Pma Indan women a leas 21 years old and lvng near Phoenx Arzona, USA. () Cleveland daa Ths daa base conans 76 characerscs, however all dsrbued ess refer o ulzng a subse of 14 of hem. Specally, ML researchers use only he Cleveland daabase ll oday. The "goal" feld refers o he presence of hear dsease n he paen. I s neger valued from 0 (no presence) o 4. Expermens wh he Cleveland daabase have concenraed on smply aempng o dsngush presence (values 1, 2, 3, 4) from absence (value 0). The names and socal secury numbers of he paens were recenly removed from he daabase, replaced wh dummy values. (v) Hungaran daa Owng o a vas percenage of mssng values hree of he characerscs have been reeced however he forma of he daa s precsely he smlar as ha of he Cleveland daa. Thry-four examples of he daabase were reeced on accoun of mssng values and 261 examples were presen. Class dsrbuons are 62.5% hear dsease no presen and 37.5% hear dsease presen. (v) Swzerland daa More number of mssng values s n Swzerland daa. I encloses 123 daa nsances and Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

8 8 14 feaures. Class dsrbuons are 6.5% hear dsease no presen and 93.5% hear dsease presen. Evaluaon mercs In order o assess he effcency of he proposed sysem an evaluaon merc s employed. I conans a se of measures ha pursue a general underlyng evaluaon mehodology some of he mercs ha we have selec for our evaluaon purpose are True Posve, True Negave, False Posve and False Negave, Sensvy, Specfcy and Accuracy. TP Sensv y (29) TP FN TN Specfc y (30) FP TN TP TN Accuracy (31) TP FN FP TN Performance Analyss The resuls of proposed work help o analyze he effcency of he predcon process. The subsequen able.1 abulaes he resuls. Here he resuls of four daases are gven n able.1. Comparave Analyss The leraure revew works are compared n hs secon wh he proposed work o show ha our proposed work s beer han he sae-of-ar works. We can esablsh ha our proposed work helps o aan very good accuracy for he medcal daa classfcaon. In our proposed mehod we use fuzzy mn max neural nework wh modfed group search algorhm for classfcaon. And also we can esablsh hs predcon accuracy oucome by comparng oher classfers. Here he proposed classfer s compared wh he radonal neural nework. Below specfed Fg.4 explans he comparson oucomes of he Sensvy measures. Table 1. Performance of he proposed mehod usng varous daase Daase Accuracy Sensvy Specfcy Mammographc Mass daa Pma Indans Dabees daa Fgure.4 he comparson oucomes of he Sensvy measures Cleveland daa Hungaran daa Swzerland daa From able.1, he evaluaon mercs are analyzed for he fve numbers of daases, by whch we can observe he effcency of proposed medcal daa classfcaon sysem. The accuracy values of fve daase are 94.2%, 95.30%, 85.14%, 83.33% and 84.01%. The sensvy values for he fve daases are 0.974, , 0.851, and The specfcy values for he fve daases are 0.915, 0.933, 0.851, 0.69 and Fgure.5 The comparson oucomes of he Specfcy measures The sensvy values for he exsng mehods are 0.96, 0.958, 0.81, 0.88, and 0.87 whch s low when compared wh our opmal classfer, he Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016

9 9 sensvy values of our opmal classfer are 0.974, 0.970, 0.851, 0.952, and Below specfed Fg.5 explans he comparson oucomes of he Specfcy measures. The specfcy values for he exsng mehods are 0.872, 0.89, 0.81, and 0.650, whch s low when compared wh our opmal classfer, he specfcy values of our opmal classfer are 0.915, 0.93, 0.85, 0.69 and Below specfed Fg.6 explans he comparson oucomes of he Accuracy. done by opmal classfer whch combned MGSO and Fuzzy Mn-Max Neural Nework (GFMMNN).The mplemenaon of he proposed mehod was done n MATLAB. For expermenaon, he daase gven n he UCI machne learnng reposory such as, Mammographc Mass daa, Pma Indans Dabees daa, Cleveland, Hungaran and Swzerland ec., wll be subeced o analyze he performance of he proposed echnque n class mbalance problem ulzng accuracy, sensvy and specfcy. The resuls of our proposed mehod have shown ha, our opmal classfer acheves beer resul when compared o oher mehod. Our mehod acheves he maxmum accuracy value for Mammographc Mass medcal daase and Pma Indans Dabees daa. Boh daase acheves 92.65% of accuracy value for medcal daa classfcaon. Fgure.6 The comparson oucomes of he Accuracy measures The accuracy values for he exsng mehods are 92.65%, 92.65%, 81.51%, 81.97% and 81.6%, whch s low when compared wh our opmal classfer, he accuracy values of our opmal classfer are 94.21%, 95.30%, 85.14%, 83.33% and 84%. When compared o he exsng mehod he proposed mehod acheves beer resul. The drawbacks of oher exsng echnque conan boh redundan and rrelevan arbues; leads o less classfcaon accuracy. In order o overcome hs drawback, nally he proposed echnque removes he rrelevan daa and also reduces he dmenson wh he help of OLPP echnque. Then classfcaon s done by opmal classfer. So ha, he mplemened echnque reaches he maxmum accuracy value compared o he oher exsng echnques. 6. Concluson In hs paper we have proposed a mehod orhogonal local preservng proecon and opmal classfer. Inally, he pre-processng wll be appled o exrac useful daa and o conver suable sample from raw medcal daases. Feaure reducon s done wh he help of OLPP and he classfcaon Reference Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016 [1] S. Kharya, Usng Daa Mnng Technques for Dagnoss and Prognoss of Cancer Dsease, Inernaonal Journal of Compuer Scence, Engneerng and Informaon Technology, Vol. 2, No. 2, pp , Aprl [2] K. Raesh, and V. Sangeeha, Applcaon of Daa Mnng Mehods and Technques for Dabees Dagnoss, Inernaonal Journal of Engneerng and Innovave Technology, Vol. 2, No. 3, pp , Sep [3] M. A. Khaleel, S. K. Pradham and G. N. Dash, A Survey of Daa Mnng Technques on Medcal Daa for Fndng Locally Frequen Dseases, Inernaonal Journal of Advanced Research n Compuer Scence and Sofware Engneerng, Vol. 3, No. 8, pp , Augus [4] Q. A. A. Radadeh and E. A. Nag, Usng Daa Mnng Technques o Buld a Classfcaon Model for Predcng Employees Performance, Inernaonal Journal of Advanced Compuer Scence and Applcaons, Vol. 3, No.2, pp , [5] S. TR and N. B. K. AR, Hybrd Feaure Reducon and Selecon for Enhanced Classfcaon of Hgh Dmensonal Medcal Daa, In proceedng of IEEE Inernaonal Conference on Compuaonal Inellgence and Compung Research (ICCIC), pp.1-4, Dec [6] H. I. Elshazly, A. M. Elkorany, A. E. Hassanen and A. T. Azar, Ensemble classfers for bomedcal daa: performance evaluaon, In proceedng of 8h Inernaonal Conference on Compuer Engneerng & Sysems (ICCES), pp , Nov [7] V. Chaurasa and S. Pal, Daa Mnng Approach o Deec Hear Deses, Inernaonal Journal of Advanced Compuer Scence and Informaon Technology, Vol. 2, No. 4, pp , [8] P. Puar and J. B. Gupa, Improvng Classfcaon Accuracy by Usng Feaure Selecon and Ensemble

10 10 Model, Inernaonal Journal of Sof Compung and Engneerng, Vol. 2, No. 2, pp , May [9] S. Mualb, R. A. Razak, S. N. S. A. Rahman and A. Mohamed, Inellgen Classfcaon n Medcal Daa, In proceedng of IEEE EMBS Inernaonal Conference on Bomedcal Engneerng and Scences, pp , Dec [10] D. Tomar and S. Agarwal, A survey on Daa Mnng approaches for Healhcare, Inernaonal Journal of Bo-Scence and Bo-Technology, Vol. 5, No. 5, pp , [11] B. Krawczyk and G. Schaefer, Ensemble fuson mehods for medcal daa and classfcaon, In proceedng of 11h symposum on neural nework applcaon n elecrcal engneerng, pp , [12] V. S. Laha, P.Y.L. Sweha, M. Bhavya, G. Geeha and D. K.Suhasn, Combned Mehodology of he Classfcaon Rules For medcal Daa-Ses, Inernaonal Journal of Engneerng Trends and Technology, Vol. 3, No. 1, pp , [13] L. A. N. Muhammed, Usng Daa Mnng echnque o dagnoss hear dsease, In proceedng of sasc n scence, busness and engneerng, pp. 1-3, Sep [14] K. Raeswar, V. Vahyanahan and S. V. Pede Feaure Selecon for Classfcaon n Medcal Daa Mnng, Inernaonal conference of emergng Tren and echnology n compuer scence, Vol. 2, No. 2, pp , Aprl [15] A. A. and S. A. Hannan, Daa Mnng Technques o Fnd Ou Hear Dseases: An Overvew, Inernaonal Journal of Innovave Technology and Explorng Engneerng, Vol. 1, No. 4, pp , Sep 2012, [16] M. A. Jabbar, P. Chandra and B. L Deekshaulu, Knowledge Dscovery From Mnng Assocaon Rules For Hear Dsease Predcon, Journal of Theorecal and Appled Informaon Technology, Vol. 41, No. 2, pp , July 2012 [17] I. San, D. Sngh and A. Khosla, QRS deecon usng K-Neares Neghbor algorhm (KNN) and evaluaon on sandard ECG daabases, Elsever Journal of Advanced Research, Vol. 4, No. 4, pp , July 2013 [18] S. Khanmohammad and M. Rezaeahar, AHP based classfcaon algorhm selecon for clncal decson suppor developmen, Elsever Proceda Compuer Scence, Vol. 36, pp , [19] R. Chra and V. Seenvasagam, Revew Of Hear Dsease Predcon Sysem Usng Daa Mnng And Hybrd Inellgen Technques, Icac Journal On Sof Compung, Vol. 03, No. 04, pp , July [20] R. Bhuvaneswar and K. Kalaselv, Nave Bayesan Classfcaon Approach n Healhcare Applcaons, Inernaonal Journal of Compuer Scence and Telecommuncaons, Vol. 3, no. 1, pp , January [21] P. K. Anoo, Clncal decson suppor sysem: Rsk level predcon of hear dsease usng weghed fuzzy rules, Elsever Compuer and Informaon Scences, Vol. 24, No. 1, pp , January [22] M. A. Jabbar, B.L. Deekshaulu and P. Chandra, Classfcaon of Hear Dsease Usng K- Neares Neghbor and Genec Algorhm, Elsever Proceda Technology, Vol. 10, pp , [23] A. Sudha, P. Gayahr and N. Jasankar, Effecve Analyss and Predcve Model of Sroke Dsease usng Classfcaon Mehods, IEEE Inernaonal Journal of Compuer Applcaons, Vol. 43, No.14, pp , Aprl [24] M. Juhola, H. Jousok, H. Aalo and T. P. Hrvonen, On classfcaon n he case of a medcal daa se wh a complcaed dsrbuon, Elsever Appled Compung and Informacs, Vol. 10, No. 2, pp , January [25] B. Denns and S. Muhukrshnan, "AGFS: Adapve Genec Fuzzy Sysem for medcal daa classfcaon", Appled Sof compung, Vol. 24, pp , [26] D. A. Adeny, Z. Wa, Y. Yongquan, Auomaed web usage daa mnng and recommendaon sysem usng K-Neares Neghbor (KNN) classfcaon mehod, Appled Compung and Informacs, Vol. 3, Inernaonal Journal of Inellgen Engneerng and Sysems, Vol.9, No.3, 2016