Analysing Big Data to Build Knowledge Based System for Early Detection of Ovarian Cancer

Transcription

1 Indan Journal of Scence and Technology, Vol 8(4), DOI: /js/205/v84/65745, July 205 ISSN (Prn) : ISSN (Onlne) : Analysng Bg Daa o Buld Knowledge Based Sysem for Early Deecon of Ovaran Cancer P. Yasodha * and N. R. Anahanarayanan 2 Deparmen of Compuer Scence, Pachayappa s College for Women, Kanchpuram , Taml Nadu, Inda; yasodhaphd@gmal.com 2 Deparmen of CSA, SCSMV Unversy, Kanchpuram , Taml Nadu, Inda Absrac Bg daa analyss plays a crucal role n he healh care for early dagnoss of faal dsease. The daa mnng echnques are wdely used for daa analyss problem o dscover valuable knowledge from a large amoun of daa. Ths paper uses he daa mnng mehods such as feaure selecon and classfcaon o provde a predcve model for ovaran cancer deecon. A huge amoun of daase s gahered o buld knowledge based sysem. Rough se heory s ulzed o fnd he daa relance and reduce he feaure se conaned n he daa se. The Hybrd Parcle Genec Swarm Opmzaon (PGSO) s used o opmze he seleced feaures o effcenly classfy he ovaran cancer, eher normal or early or dfferen sages of ovaran cancer. Mul class SVM s adoped as he classfer o classfy normal or dfferen sages of ovaran cancer usng he opmzed feaure se. The epermen s done on dfferen ovaran cancer daase and he proposed sysem has obaned beer resuls for all daases. Keywords: Bg Daa Analyss, Genec Algorhm (GA), Mul Class Suppor vecor Machne (SVM), Parcle Swarm Opmzaon (PSO), Rough Se Theory, Ovaran Cancer. Inroducon Ovaran cancer s common dsease for women and ranks fourh among oher cancer for hgh deah rae. The deah rae of ovaran cancer s due o he mos of he women were unaware of he cancer unl he dsease has advanced o Sage III or Sage IV. Early deecon s a key o reduce he deah rae of Ovaran Cancer. An accurae and relable dagnoss process mgh followed by he physcan for early dagnoss. Bg daa analyss plays a crucal role n he healh care for early dagnoss of faal dsease 2. Analysng a large amoun of daa, whch s generaed from varous real me paen records produce a lo of poenal nformaon for creang qualy healh care a reduced coss 3. Tradonal daa analyss echnques have become nadequae for processng such huge volume of daa 4. Knowledge dscovery of daa s a new echnque ha encompasses a varey of paern recognon, sascal analyss and machne learnng echnques, whch eplo he knowledge from huge amoun of recorded daa 5,6. Daa Mnng s commonly alluded as paern eracons sraegy, whch assembles and formulaes he knowledge from remendous nformaon 7,8. Daa mnng sraeges have offcally demonsraed o wde range of medcne, ncludng prognoss, dagnoss and reamen. Ths paper uses he daa mnng echnque o oban a knowledge based sysem for early ovaran cancer deecon from an unorganzed daa se. The major commmen of he paper s as per he followng:. Inally, he rough se heory s appled o he daase o deermne he daa dependences and o mnmze he feaure se. 2. The Hybrd Parcle Genec Swarm Opmzaon (PGSO) s used o opmze he rough se feaure reducon o effecvely classfy he ovaran cancer umors eher normal or abnormal and 3. Mul Class Suppor Vecor Machne (SVM) s used for classfyng he dfferen sages of ovaran cancer and non-ovaran cancer. The res of he paper s organzed as follows: Secon II descrbes he recen relaed works abou he sysem. Secon III descrbes proposed knowledge base sysem. Epermenal resuls and analyss of he proposed sysem * Auhor for correspondence

2 Analysng Bg Daa o Buld Knowledge Based Sysem for Early Deecon of Ovaran Cancer s descrbed n Secon IV. Fnally, Secon V renders he conclusons. 2. Relaed Work A model s proposed n 9 for early deecon and correc dagnoss of lung cancer dsease o save he lfe of paen usng he daa mnng classfcaon mehods. The hsorcal lung cancer daabase s used o erac he concealed nformaon o predc he paens wh lung cancer or no. Three dfferen classfer are ulzed such as Nave Bayes ook afer by f-hen rule, Neural Nework and Decson rees and he performance of he Nave Bayes shows ha s obvously beer han he oher wo. The daa mnng echnques s used n 0 o mprove he breas cancer dagnoss and prognoss. The dfferen daa mnng echnques dscussed o fnd he effecve classfer for he breas cancer dagnoss. The UCI machne learnng daa se and SEER daa se used for he research and he decson ree aaned beer accuracy han oher classfers o predc he dsease. The daa mnng concep s used n o predc he head and neck cancer. The daa se s colleced from he varous dagnosc cenres wh boh he cancer and noncancer paen. A dfferen classfcaon echnque has been appled for he daa se and C4.5 obaned hgher accuracy rae han oher classfers. The performance of he Decson ree classfer- CART s analysed n 2 wh and whou feaure selecon on dfferen breas cancer daa ses. The epermenal resul shows ha classfcaon accuracy s mproved wh feaure selecon by removng he rrelevan feaures n he daases. The feaure selecon approach s analysed n 3 for classfcaon and a new approach s proposed for feaure selecon. The daabase conans a large number of feaures, so dmensonaly reducon s used usng he assocaon rules and correlaon arbues feaures selecon. Afer removal of unwaned arbues he accuracy of dfferen classfers s checked usng he reduced feaure se. A modfed REF-SVM based feaure selecon mehod s proposed n 4 for classfcaon problem. I s greedy mehod o fnd he bes possble combnaon for classfcaon o mprove he qualy of he fnal classfer. The of epermens s done on daase from UCI Machne Learnng Reposory and he promsng resuls has been obaned by conrbung local search on he classfcaon process. A new gene selecon echnque s proposed n 5, whch combnes Technques for Order Preference by Smlary o an Ideal Soluon and F-Score mehod o selec relevan gene. The seleced gene s fed n o four dfferen classfers such as K-Neares Neghbour (KNN), Decson Tree (DT), Suppor Vecor Machne (SVM) and Nave Bayes (NB). The SVM obaned beer classfcaon accuracy wh he reduced feaure se. A Mul objecve Frefly Algorhm echnque s proposed n 6 for Mulclass Gene Selecon. The mehod opmzes he mulple fre fles n he mulple class-specfc sascs o selec he genes. The mehod s compared wh he esng gene selecon mehods and he resuls show ha he mehod acheves hgh classfcaon accuracy wh less compley han he esng mehods. 3. Knowledge based Sysem Ulzng Daa Mnng Conceps for Early Deecon of Ovaran Cancer The goal of hs paper s o provde a knowledge base sysem o classfy he ovaran cancer and normal case. The daase conans N number of feaures, o deermne a relevan feaure o porray he daase. The rough se heory s well known mehod for feaure subse selecon. The Hybrd Parcle Genec Swarm Opmzaon (PGSO) s used o opmze he rough se feaure reducon o effecvely classfy he daa no normal or abnormal. The Mul class SVM s used n he classfcaon sage o classfy he daa no normal and abnormal case. The Fgure sows he proposed knowledge based sysem for early deecon of ovaran cancer. Fgure. Knowledge based sysem for early deecon of ovaran cancer. 2 Vol 8(4) July Indan Journal of Scence and Technology

3 P. Yasodha and N. R. Anahanarayanan 3. Rough Se Theory The rough se echnque s a mahemacal model nroduced by Pawlak for dealng wh uncerany 7,8. The rough se heory mnmzes he feaure se and reducs an sgnfcan feaure se, whch s capable of deec he vsble pons by he orgnal of Informaon Sysem (IS). Le IS = (U, A), where A( ) be he se of arbues and U( ) s ermed as Unverse and s a non-empy fne se of all objecs. A redac of A s a reduced se of arbue B A, such ha non empy subse of A. Indscernbly Relaon (IND (B)) s a relaon on U. Consder wo objecs and j U are vewed o be ndscernble by he se of arbues B n A, f and only, f a( ) = a( j ), a A-B..e. ( ) {(, ) j ( ) ( j) } IND B= U a Ba = a () A subse s obaned whou conanng any dspensable arbues and ha such subse s known as a reducs. The se of all reducs n IS s ndcaed as RED. 3.2 Parcle Genec Swarm Opmzaon (PGSO) based Rough Se Theory The GA and PSO s wdely used Evoluonary algorhm because of s smple process wh opmzed soluon. In hs paper, he hybrd PGSO procedure s followed by combnng he PSO and GA for opmzed feaure selecon. The GA s embedded whn he PSO o mprove he PSO by servng as a local opmzer a each eraon Parcle Swarm Opmzaon (PSO) In PSO, each parcle represen a canddae soluons of a populaon, smulaneously coes and evolve based on knowledge sharng wh neghbourng parcles 9. Each parcle fles on he problem search space and based on he dreced velocy vecor, wll generae a soluon. Each parcle changes s velocy o deermne he beer poson by usng s own flyng knowledge for he bes poson memory found n he earler flghs and eperence of neghbourng parcles as he bes deermned soluon of he populaon. The bes poson deermned by he parcles s represened as P bes and he endency o move forwards he prevous bes poson of he neghbourhood s g bes. The velocy of he parcle s updaed usng he followng equaon ( ) ( ) v = wv + cr p + cr p g (2) macoun 2 2 Where represens he curren poson of parcle, g p s he curren bes poson deermned by parcle, p s he global bes poson deermned among all parcles n he problem space up o eraon coun, c and c 2 represens he cognve and socal scalng parameers, r and r 2 are random numbers dsrbued unformly n he nerval (0,). w s he parcle nera, whch mnmze he search area dynamcally, macoun w = ( w w ) + w macoun ma mn mn (3) Where w ma and w mn ndcaes he mamum and mnmum of w respecvely ma coun represens he mamum eraon and represens he curren eraon number. The parcle poson updaed accordng o he followng equaon, {, rand () < sg v = (4) macoun 0, rand () sg v sg ( ) = + e ( ) + ( ) + Parcle fly owards he new poson accordng o equaon and 2. In hs way all parcles deermnes he new posons and updae her ndvdual bes poson g p and global bes poson p of he swarm. Ths process wll be connued unl mamum eraon reached. Consder an Informaon Sysem (IS) = (U,A) and A = (X Y )where X s a non-empy fne se of condon arbues and Y s a non-empy fne se of decson arbues, such ha RED (IS) X. The objecve funcon of parcle a poson s deermned by he followng equaon, ( ) ( ) (5) = + (6) f a y Y β Where y ( Y ) s he classfcaon qualy of parcle condon arbue se, whch conans he RED, and relave o decson able Y, and s shown n he followng equaon, y dred ( Y) d = (7) c Vol 8(4) July Indan Journal of Scence and Technology 3

4 Analysng Bg Daa o Buld Knowledge Based Sysem for Early Deecon of Ovaran Cancer Where d RED represens a dependency degree of RED on Y and d c represens he dependency degree of X on Y. s he number of lengh of seleced feaure subse for parcle, whle populaon of soluons P (number of parcles n he populaon) s a eraon coun. X s he oal number of condon arbues. The parameer α [0,] and β = -α represens mporance of classfcaon qualy and subse lengh Genec Algorhm Genec Algorhm (GA) s one of he compuaonal models used wdely because of s evoluon. GA opmzaon echnque conans selecon, crossover and muaon operaons o a populaon of compleng problem soluons. Afer he hree operaons are appled a new generaon of he populaons wll be generaed and a he same me he GA wll generae a se of chromosome randomly a he space. The fness value wll be calculaed for he chromosomes and he chromosome wh a hgher fness value wll be kep and he same operaon wll be performed unl a fed number of eraons are reached. To mprove he PSO performance, he GA s used as a local opmzer a each eraon 20,2. Afer he nal populaon are creaed, he operaon such as selecon, crossover and muaon wll be appled o he nally creaed parcles. Choose wo parcles randomly and deermne he relave dfferenced for hose wo parcles by he followng equaon, d = f f f 2 + f 2 Where f and f 2 are he fness value of parcle and parcle 2. Accordng o he relave dfference value he cross over operaon s chosen and s defned n he followng equaon, {If f < f2 and R < d, hen crossover operaon s done on parcle 2 If f and R > d, normal crossover operaon s choosen (9) If f > f 2 and R < d, hen crossover operaon s done on parcle f f > f 2 and R > r, normal crossover operaon s choosen Muaon operaon s done on he wo seleced parcles and s se o probably of /n (n= number of parcles). (8) The process wll be done unl he mamum eraon s reached. Then he PSO based opmzaon s performed for he parcles. The PGSO based rough se algorhm s gven n he Fgure 2. Inpu: C, he se of all condon arbues, ma coun Oupu: Reduc (RED) Sep : Inalzng poson, velocy, c,c2. Sep 2: For = o P Obanng he nera wegh for parcle by usng equaon (3) Calculae he fness (objecve) funcon for parcle by usng he equaon (6) End For g bes bes Sep 3: Inalze p = p and Gbes = f ( p ) bes Sep 4: For = o Macoun Selec wo random parcle and calculae he relave dfference usng equaon 8 Selec he cross over operaon accordng o he equaon (9) Muuae (parcle), Muuae (parcley) End For Sep 5: whle (< macoun) For = o P Compue he fness funcon for parcle usng eq (6) If pbes < f ( ) p = End f pbes ( ) < f bes 2 p Fnd he f ( p ) = ma { p, p, p } bes bes < f g bes bes bes = = f pbes p p and Gbes f p ( ) End f Evaluae he velocy for parcle by equaon (2) Updae parcle poson by usng equaon (4) End For End whle Sep 6: Redac RED Fgure 2. Algorhm for opmzed feaure se. 4 Vol 8(4) July Indan Journal of Scence and Technology

5 P. Yasodha and N. R. Anahanarayanan Classfcaon The effecve feaure se s bul by elmnang he nosy feaure and mproves he classfcaon accuracy. In hs paper, he Suppor Vecor Machne s used as a classfer and s orgnally desgned for bnary classfcaon. Currenly here es wo ypes of approaches for mulclass SVM. The oulne o llumnae mulclass SVM problems n one sep has parameers proporonal o he k classes. Therefore, for mulclass SVM echnques, eher a number of bnary classfers have o be consruced 22. There are four approaches for mulclass classfcaon based on bnary classfcaon: One Agans One (OAO), One Agans All (OAA), Fuzzy Decson Funcon (FDF) and Decson Dreced Acyclc Graph (DDAG) SVM. The One Agans All (OAA) mehod s used n hs paper o classfy dfferen sages of ovaran cancer cases and he normal cases. In OAA, k SVM models are consruced. The h SVM s learned wh all of he ranng eamples n he h class wh posve labels and all oher classes negave labels. The fnal oupu of he OAA s he class ha relaes o he SVM wh he hghes oupu value 23. Hence, by llumnang he opmzaon problem of SVM usng all he ranng eamples n he daase, h SVM decson funcon s D ( ) = w φ( ) + b (0) Where s he npu vecor, whch s assgned o he h class corresponds o he larges value of he decson funcon; w s he vecor n he feaure space,ϕ() s mappng funcon and b s a scalar. Sample s classfed no he class as defned n he followng equaon, = argma =,...k (D ()) () 4. Epermenal Resuls and Analyss The gene and blood es are normally used ndcaors for he ovaran cancer, so n hs paper wo daa ses are consdered. cga-daa.nc.nh.gov/). The daase composed of 2,042 genes whou any mssng values. The daase classfed 39 cases as Early (Sage I and Sage II) and 454 cases as lae sage (Sage III and Sage IV) Blood Assays The Daa se 2 has been aken from he Sngapore Naonal Unversy Hospal (NUH) (hp:// and he daase s based on he blood es resuls of 72 nsances and consss of 28 feaures. The daase conans 23 normal, 0 of borderlne, 9 of early (Sage I and Sage II) and 42 of lae sage (Sage I and Sage II). 4.2 Performance Mercs The sysem performance s compued n erms of mean absolue error, classfcaon accuracy, sensvy, roo mean square error and specfvy respecvely. Sensvy s he ably of he es o denfy he class correcly among oher classes belongng o he classes of he ovaran cancer daase. True Posve Sensvy = True Posve + False Negave (2) Where True posve s correcly denfed and False Negave s ncorrecly rejeced. Specfcy s ably of he ess o eclude a condon correcly. True Negave Specfcy = True Negave + False Posve (3) Where True negave s correcly rejeced and False posve s ncorrecly denfed. The classfcaon accuracy s drecly proporonal o he correcly classfed objecs and s deermned by he followng formula, Number of nsances classfed correcly Classfcaon Accuracy = Toal number of nsances (4) The Mean absolue error s he average of he dfference beween he predced and acual value n all es cases. 4. Daase 4.. Mcro Array Gene Epresson A normalzed gene epresson daa and consss of 493 nsances and s aken from he TCGA poral (hp:// MAE = a c + a c + + a c 2 2 n n Vol 8(4) July Indan Journal of Scence and Technology 5 n (5) The roo mean squared error yelds he error value he same dmensonaly as he acual and predced values.

6 Analysng Bg Daa o Buld Knowledge Based Sysem for Early Deecon of Ovaran Cancer RMSE = ( a c ) + ( a c ) + + ( a c ) n n n (6) Where he c s he compued value and a s he correspondng correced value. 4.3 Dscusson Fgure 3 shows he comparson of he Mulclass SVM, ANN and Nave Bayes n erms mean absolue error for wo ovaran daases. The mulclass SVM classfer ncurred a mnmum error rae by usng he opmzed feaure se for he wo daa se when compared o ANN and Nave Bayes. Fgure 4. Comparson of roo mean square error for orgnal feaure and reduced feaure se. (a) Daase I. (b) Daase II. Fgure 4 shows he comparson of he Mulclass SVM, ANN and Nave Bayes n erms of roo mean square error for wo daases. The mulclass SVM class obaned a mnmum error rae by usng he opmzed feaure se for he wo daa se when compared o ANN and Nave Bayes. Fgure 3. Comparson of mean absolue error for orgnal feaure and reduced feaures. (a) Daase I. (b) Daase II. Fgure 5. Classfcaon accuracy. Table. Classfcaon Accuracy for dfferen Ovaran cancer daase Classfer Daase I Daase II Sensvy Specfcy Accuracy Sensvy Specfcy Accuracy Mulclass SVM ANN Nave Bayes Vol 8(4) July Indan Journal of Scence and Technology

7 P. Yasodha and N. R. Anahanarayanan The classfcaon accuracy of he proposed sysem s compared wh he ANN and Nave Bayes and has acheved 96 percen accuracy for daase I and 98 percen accuracy for daase II and s graphcally represened n he Fgure 5. The Table gves he Sensvy, specfcy, and accuracy values for classfyng normal, early sage and lae sage for dfferen ovaran cancer daase and he proposed sysem acheves beer accuracy han he ANN and Nave Bayes. 5. Concluson Ovaran cancer dagnoss s an mporan research by consderng sgnfcan ncrease n deah rae, so early deecon and accurae sagng wll help he physcan for correc dagnoss procedure. The paper proposed a knowledge based sysem usng he daa mnng conceps such as feaure selecon and classfcaon. The opmzed feaure selecon process s acheved by he hybrd PGSO based rough se heory. The Mul class SVM s used for classfyng he normal, early sage and lae sage cases usng he opmzed arbue se. The sysem performance has been esed usng he wo dfferen ovaran cancer daa ses and he proposed Knowledge based sysem aan beer accuracy han oher classfers usng he reduced feaures. The sysem has endency o provde beer suggeson o he physcan for dagnosc process. 6. References. Tan TZ, Quek C, See Ng G, Razv K. Ovaran Cancer dagnoss wh complemenary learnng fuzzy neural nework. Arfcal Inellgence Sysem. 2008; 43: Pesker A, Dala S. Daa analycs for rural developmen. Indan Journal of Scence and Technology. 205; 8: Radhakrshnan A, Kalmad K. Bg daa medcal engne n he cloud (BDMEC): Your new healh docor. Infosys Lab Brefngs. 203; (). 4. Lm CS, Lee W, Joon YY, Shon K. Creang values from a nosy accumulaed conens based on daa analyss. Indan Journal of Scence and Technology. 205; 8: Mohammadzadeh N, Safdar R, Mohammadzadeh F. Usng nellgen daa analyss n cancer care: Benefs and challenges. Journal of Healh Informacs n Developng Counres. 204; 8(2): Noh K-S. Plan for valsaon of applcaon of bg daa for e-learnng n Souh Korea. Indan Journal of Scence and Technology. 205; 8: Ramesh Kumar KK, Anbuman A. Medcal mage segmenaon usng mulfracual analyss. Inernaonal Journal of Invenons n Compuer Scence and Engneerng. 204; (3): Sharma R, Sharma P. A novel approach owards bg daa challenges. Inernaonal Journal of Innovave Compuer Scence and Engneerng. 204; (2): Krshnaah V, Narsmha G, Subhash Chandra N. Dagnoss of lung cancer predcon sysem usng daa mnng classfcaon echnques. IJCSIT. 203; 4(): Kharya S. Usng daa mnng echnques for dagnoss and prognoss of cancer dsease. Inernaonal Journal of Compuer Scence, Engneerng and Informaon Technology (IJCSEIT). 202; 2(2): Suj RJ, Rajagopalan SP. An auomac oral cancer classfcaon usng daa mnng echnques. Inernaonal Journal of Advanced Research n Compuer and Communcaon Engneerng (IJARCCE). 203; 2(0): Lavanya D. Analyss of feaure selecon wh classfcaon: Breas cancer daases. Indan Journal of Compuer Scence and Engneerng (IJCSE). 20; 2(5): Rajeswar K, Vahyanahan V, Pede SV. Feaure selecon for classfcaon n medcal daa mnng. Inernaonal Journal of Emergng Trends and Technology n Compuer Scence (IJETTCS). 203; 2(2): Samb ML, Camara F, Ndaye S, Slman Y, Essegh MA. A novel RFE-SVM-based feaure selecon approach for classfcaon. Inernaonal Journal of Advanced Scence and Technology. 202; 43: Abd-El Faah M, Khedr WI, Sallam KM. A TOPSIS based mehod for gene selecon for cancer classfcaon. In J Compu Appl. 203; 67(7): Manoharan GV, Shanmugalakshm R. Mul-objecve frefly algorhm for mul-class gene selecon. Indan Journal of Scence and Technology. 205; 8: Nzar Banu PK, Hannah Inbaran H. Performance evaluaon of hybrdzed rough se based unsupervsed approaches for gene selecon. Inernaonal Journal of Compuaonal Inellgence and Informacs. 202; 2(2): Srman PK, Ko MS. Knowledge dscovery n medcal daa by usng rough se rule nducon algorhms. Indan Journal of Scence and Technology. 204; 7(7): Arafa H, Baraka S, Goweda AF. Usng nellgen echnques for breas cancer classfcaon. Inernaonal Journal of Emergng Trends and Technology n Compuer Scence (IJETTCS). 202; (3): Rouh F, Effanejad R. Un commmen n power sysem T by combnaon of Dynamc Programmng (DP), Genec Algorhm (GA) and Parcle Swarm Opmzaon (PSO). Indan Journal of Scence and Technology. 205; 8(2): Chaung L-Y, Yang C-H, L J-C, Yang C-H. A hybrd BP- SO-CGA approach for gene selecon and classfcaon of mcroarray daa. Journal of Compuaonal Bology. 202; 9(): Kohl N, Verma NK. Arrhyhma classfcaon usng SVM wh seleced feaures.ijest. 20; 3(8): Bha HF, Wan MA. Modfed one-agans-all algorhm based on suppor vecor machne. In J Adv Res Compu Sc Sofware Eng. 203; 3(2): Vol 8(4) July Indan Journal of Scence and Technology 7