Prediction of chronic kidney disease - A machine learning perspective

Recei ed Janua y 10, 2021, accep ed Janua y 15, 2021, da e o publica ion Janua y 22, 2021, da e o cu en e sion Feb ua y 1, 2021. Digi al Objec Iden i ie 10.1109/ACCESS.2021.3053763 P edic ion o Ch onic Kidney Disease - A Machine Lea ning Pe spec i e PANKAJ CHITTORA 1, SANDEEP CHAURASIA1, (Senio Membe , IEEE), PRASUN CHAKRABARTI2,3, (Senio Membe , IEEE), GAURAV KUMAWAT1, TULIKA CHAKRABARTI4, ZBIGNIEW LEONOWICZ 5, (Senio Membe , IEEE), MICHAŁ JASIŃSKI 5, (Membe , IEEE), ŁUKASZ JASIŃSKI 5, RADOMIR GONO 6, (Senio Membe , IEEE), ELŻBIETA JASIŃSKA 7, AND VADIM BOLSHEV 8 1Depa men o Compu e Science and Enginee ing, Manipal Uni e si y Jaipu , Jaipu 303007, India 2Depa men o Compu e Science Enginee ing, Techno India NJR Ins i u e o Technology, Udaipu 313003, India 3Da a Analy ics and A i icial In elligence Labo a o y, Enginee ing-Technology School, Thu Dau Mo Uni e si y, Thu Dau Mo 820000, Vie nam 4Depa men o Basic Science (Chemis y), Si Padampa Singhania Uni e si y, Udaipu 3136022, India 5Depa men o Elec ical Enginee ing Fundamen als, Facul y o Elec ical Enginee ing, W oclaw Uni e si y o Science and Technology, 50-370 W oclaw, Poland 6Depa men o Elec ical Powe Enginee ing, Facul y o Elec ical Enginee ing and Compu e Science, VSB–Technical Uni e si y o Os a a, 708 00 Os a a, Czech Republic 7Facul y o Law, Adminis a ion and Economics, Uni e si y o W oclaw, 50-145 W oclaw, Poland 8Labo a o y o Powe Supply and Hea Supply, Fede al Scien i ic Ag oenginee ing Cen e VIM, 109428 Moscow, Russia Co esponding au ho : Michał Jasiński (michal.jasinski@pw .edu.pl) This wo k was unded by he Chai o Elec ical Enginee ing Fundamen als (K38W05D02), W oclaw Uni e si y o Technology, W oclaw, Poland. ABSTRACT Ch onic Kidney Disease is one o he mos c i ical illness nowadays and p ope diagnosis is equi ed as soon as possible. Machine lea ning echnique has become eliable o medical ea men . Wi h he help o a machine lea ning classi ie algo i hms, he doc o can de ec he disease on ime. Fo his pe spec i e, Ch onic Kidney Disease p edic ion has been discussed in his a icle. Ch onic Kidney Disease da ase has been aken om he UCI eposi o y. Se en classi ie algo i hms ha e been applied in his esea ch such as a i icial neu al ne wo k, C5.0, Chi-squa e Au oma ic in e ac ion de ec o , logis ic eg ession, linea suppo ec o machine wi h penal y L1 & wi h penal y L2 and andom ee. The impo an ea u e selec ion echnique was also applied o he da ase . Fo each classi ie , he esul s ha e been compu ed based on (i) ull ea u es, (ii) co ela ion-based ea u e selec ion, (iii) W appe me hod ea u e selec ion, (i ) Leas absolu e sh inkage and selec ion ope a o eg ession, ( ) syn he ic mino i y o e -sampling echnique wi h leas absolu e sh inkage and selec ion ope a o eg ession selec ed ea u es, ( i) syn he ic mino i y o e - sampling echnique wi h ull ea u es. F om he esul s, i is ma ked ha LSVM wi h penal y L2 is gi ing he highes accu acy o 98.86% in syn he ic mino i y o e -sampling echnique wi h ull ea u es. Along wi h accu acy, p ecision, ecall, F-measu e, a ea unde he cu e and GINI coe icien ha e been compu ed and compa ed esul s o a ious algo i hms ha e been shown in he g aph. Leas absolu e sh inkage and selec ion ope a o eg ession selec ed ea u es wi h syn he ic mino i y o e -sampling echnique ga e he bes a e syn he ic mino i y o e -sampling echnique wi h ull ea u es. In he syn he ic mino i y o e -sampling echnique wi h leas absolu e sh inkage and selec ion ope a o selec ed ea u es, again linea suppo ec o machine ga e he highes accu acy o 98.46%. Along wi h machine lea ning models one deep neu al ne wo k has been applied on he same da ase and i has been no ed ha deep neu al ne wo k achie ed he highes accu acy o 99.6%. INDEX TERMS Ch onic kidney disease, machine lea ning, p edic ion. I. INTRODUCTION Ch onic kidney Disease (CKD) means you kidneys a e dam- aged and no il e ing you blood he way i should. The p ima y ole o kidneys is o il e ex a wa e and was e om The associa e edi o coo dina ing he e iew o his manusc ip and app o ing i o publica ion was Ha una Chi oma . you blood o p oduce u ine and i he pe son has su e ed om CKD, i means ha was es a e collec ed in he body. This disease is ch onic because o he damage g adually o e a long pe iod. I is la e ing a common disease wo ldwide [1]. Due o CKD may ha e some heal h oubles. The e a e many causes o CKD like diabe es, high blood p essu e, hea disease. Along wi h hese c i ical diseases, CKD also 17312 This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/ VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease depends on age and gende [2]. I you kidney is no wo king, hen you may no ice one o mo e symp oms like abdominal pain, back pain, dia hea, e e , nosebleeds, ash, omi ing. The e a e wo main diseases o CKD: (i) diabe es and (ii) high blood p essu e [3]. So ha con olling o hese wo diseases is he p e en ion o CKD. Usually, CKD does no gi e any sign ill kidney is damaged badly. CKD is being inc eased apidly as pe he s udies hospi aliza ion cases inc ease 6.23 pe cen pe yea bu he global mo ali y a e emains ixed [4]. The e a e ew diagnos ic es s o check he condi ion o CKD: (i) es ima ed glome ula il a ion a e(eGFR) (ii) u ine es (iii) blood p essu e. A. EGFR eGFR alue shows ha how you kidney cleaning he blood. I you eGFR alue is g ea e han 90, ha means he kidney is no mal. I eGFR alue is less han 60, ha means you ha e CKD [5]. B. URINE TEST The doc o also asks o u ine es o kidney unc ionali y because kidneys make u ine. I he u ine con ains blood and p o ein [6], ha means you kidney is no wo king p ope ly. C. BLOOD PRESSURE Doc o measu es blood p essu e as Blood p essu e ange shows how you hea is pumping blood. I eGFR alue eaches less han 15, ha means he pa ien has end-s age kid- ney disease. A his poin , he e a e only a ailable ea men s: (i) dialysis and (ii) kidney ansplan . Pa ien ’s li e a e dialysis depends on such ac o s as age, gende , equency and du a ion o dialysis, physical mo emen o he body and men al heal h [7]. I dialysis is no possible, he doc o has only one solu ion, i.e., kidney ansplan a ion. Howe e , i is ex emely expensi e [8]. The e o e, i is c i ical no ewo hiness in ea ly ecogni ion, moni o ing and handling o he disease. I is essen ial o p edic he s iding o CKD wi h app op ia e accu acy due o i s dynamic and sec e i e na u e in he ea ly s ages and pa ien abno mali y. Medical ea men o CKD is p esc ibed by he s age. Any hing o he han his, i is e y impe a i e o cha ac e ize he o ganiza ion o he in ec ion because i gi es a ew indica ions. I unde pins he assu ance o undamen al in e cessions and medica ions. Medical ea men is a e y signi ican applica ion a ea o in ellec ual in elligen sys ems [10]. A e wa ds, Da a mining can play a big ole o ind ou hidden in o ma ion om he huge pa ien medical and ea men da ase ha doc o s e- quen ly ob ain om pa ien s o ge pieces o knowledge abou he symp oma ic da a and o execu e p ecise ea men plans. Da a mining can be ca ego ized as he me hod o ex ac ing hidden in o ma ion om a huge da ase . Da a mining s a e- gies a e connec ed and u ilized b oadly in a ious con ex s and a eas. Using da a mining me hods, we may p edic , clas- si y, il e and clus e da a. The objec i e s a es he algo i hm p ocessing o a aining se con aining a se o a ibu es and a ge s. Da a mining is sui able o mining in da a i he da ase is huge bu we can also do i wi h he help o machine lea ning wi h a small da ase . The machine lea ning can also ind da a analysis and pa e n de ec ion [9]. A a ie y o heal h da ase is p esen so machine lea ning algo i hms a e bes i o imp o e he accu acy o diagnosis p edic ion [11]. As heal hca e elec onic da ase g ows apidly, machine lea ning algo i hms a e becoming mo e common in heal hca e. [12]. Qin e al. [13] p oposed da a asse ion and sample diagno- sis achie able in CKD diagnosis. KNN is used o da a asse - ion. Six classi ie s algo i hms used o accu acy o diagnosis: logis ic eg ession, andom o es , suppo ec o machine, K-nea es neighbo , nai e Bayes classi ie and eed- o wa d neu al ne wo k. In hese classi ie s andom o es gi es be e accu acy, i.e., 99.75%. Vasquez-Mo ales e al. [14] de eloped a neu al ne wo k model o isk p edic ion o Ch onic Kidney Disease de el- opmen on 40000 ins ances da ase and hei model accu acy was 95%. Chen e al. [15] applied h ee models on he da ase ha is p o ided by UCI. They used KNN, SVM and so inde- penden modelling o class analogy (SIMCA) o inding he isk calcula ion o pa ien using hese classi ie s. In which he SVM and KNN model a ained, he bes accu acy o 99.7% and SVM model has he g ea es capabili y o endu e noise dis u bance. Because CKD is in asi e, cos ly so ha many pa ien s eached a las s ages wi hou ea men s. So ha ea ly de ec- ion o his disease emains impo an . Besides, Ami galiye [16] ga e he expe imen al esul o SVM machine lea ning classi ie algo i hm wi h accu acy 93%. Padmanaban and Pa hiban [17] sugges ed ha he ea ly de ec ion o CKD o diabe ic pa ien s wi h he help o machine lea ning classi ie s algo i hms. They collec ed da a om Chennai based diabe es esea ch cen e and applied Nai e Bayes and Decision ee on he da ase . Fo inding he accu acy hey used Weka ool and concluded ha Naï e Bayes classi ie achie ed he highes accu acy o 91%. de Almeida e al. [18] in hei wo k applied Decision ee, Random Fo es , Suppo Vec o Machine (SVM) and also used SVM wi h linea , polynomial, sigmoid and RBF unc- ions. Fo hei esea ch, hey used he MIMIC-II da abase. They concluded ha andom o es and Decision ee go he bes esul in he o m o p edic ion accu acy o 80% and 87% espec i ely. Guna a hne e al. [19] buil a model o a ious machine lea ning classi ie s algo i hm and analysis o which algo i hm is bes sui ed o he da ase . They used da ase p o ided by UCI con aining 400 ins ances and 14 a ibu es. They con- cluded ha he Mul iclass decision o es algo i hm was bes i ed o he CKD da ase wi h an accu acy o 99.1%. Pola e al. [20] used SVM algo i hm o CKD p edic ion. Fo he accu a e esul , hey wo ked on an impo an ea u e. Fo selec ing he co ec ea u e, hey used wo-app oach W appe and il e wi h he SVM algo i hm. In he W appe , he e we e he g eedy s epwise sea ch engine o classi ie VOLUME 9, 2021 17313 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease subse e alua o and bes i s sea ch engine o W appe subse e alua o . In il e , he e we e he g eedy s epwise sea ch engine o co ela ion ea u e sec ion subse ea u e and bes i s sea ch engine o il e ed subse e alua o . The esul s o all echniques we e compa ed and i was ound ha SVM ga e he highes accu acy wi h il e ed subse e alua o , i.e. 98.5%. Suja a D all, Gu deep Singh D all, Sugandha Singh, Bha a D all e al. [21] wo ked on CKD da ase gi en by UCI wi h 400 ins ances and 25 a ibu es. Fi s ly, da a was p ep ocessed, he missing da a was ound, illed wi h 0, hen ans o med and applied on he da ase . A e p ep ocess- ing, au ho s applied algo i hm o impo an a ibu es and ound 5 mos impo an ea u es and hen he classi ica ion algo i hm: Naï e Bayes and K-Nea es Neighbo . The go en esul KNN achie ed he highes accu acy. Almasoud and Wa d [22] wo ked wi h CKD da ase o 400 ins ances and 25 a ibu es. They applied he il e ea u e selec ion me hod on a ibu es and ound ha haemoglobin, albumin and speci ic g a i y a e ea u e a ibu es in CKD da ase . A e ea u e selec ion, hey ained he da ase and alida ed wi h 10- old c oss- alida ion. The g adien boos - ing algo i hm achie ed he highes accu acy o 99.1%. Shanka e al. [23] applied h ee s eps on he same UCI da ase : (i) da a p ep ocessing & ea u e selec ion (ii), algo- i hms’ accu acy de e mina ion and (iii) die plan sugges ion. In he ea u e selec ion me hod, hey applied wo app oaches: one is he W appe and he o he is he LASSO me hod. A e he ea u e selec ion me hod, 4 classi ica ion algo i hms we e applied: Logis ic Reg ession, Random o es ee K-Nea es Neighbo s, Neu al Ne wo k and Wide and Deep Lea ning. Fo die plan sugges ion blood po assium le el was used. The blood po assium le el was di ided in o h ee g oups based on i s alue: Sa e Zone, Cau ion Zone and Dange zone. Vijaya ani and Dhayanand [24] collec ed kidney unc ion es (KFT) da ase om medical labs, esea ch cen es and hospi als. The da ase con ained 584 ins ances and 6 a ibu es and wo classi ie applied algo i hms: suppo ec o machine (SVM) and a i icial neu al ne wo k (ANN). I was ound ha ANN achie ed he highes accu acy o 87.7%. Xiao e al. [25] used he da a o 551 pa ien s and applied 9 machine lea ning algo i hm: XGBoos , logis ic eg ession, lasso eg ession, suppo ec o machine, andom o es , idge eg ession, neu al ne wo k, Elas ic Ne and K-nea es neighbo . They e alua ed accu acy, ROC cu e, p ecision and ecall and ound ha linea model ga e he highes accu acy. Reshma e al. [31] used he ea u e selec ion echnique on CKD Da ase . Fo ea u e selec ion, ACO me hod was applied. ACO is he me a heu is ic algo i hm o he ea u e selec ion. I is he ype o W appe me hod. In hei da ase , o al 24 a ibu es we e a ailable. A e applying ea u e selec ion algo i hm, 12 ea u es was used o making he model. The Suppo Vec o machine classi ie s algo i hm was used o building he model. Deepika e al. [32] buil a p ojec on p edic ion o Ch onic Kidney Disease based on old da ase o CKD. The da ase had 24 a ibu es and 1 a ge a iable. Fo building he model, hey applied KNN and Naï e Bayes supe ised machine lea ning algo i hm. KNN achie ed highes accu acy 97 % and Naï e Bayes achie ed 91% accu acy. Ma e al. [33] p oposed he deep lea ning algo i hm o p edic ing he Ch onic Kidney Disease s a ea ly s age. The deep neu al ne wo k was buil om He e ogeneous Modi ied a i icial neu al ne wo k algo i hm. Fo building he model, ul asound images we e used. Fo compa ing he esul , he e we e h ee di e en classi ie s: Suppo Vec o machine, a i- icial neu al ne wo k and mul ilaye pe cep on. UI Haq e al. [34] p oposed he machine lea ning model o p edic he diabe es disease a ea ly s age. They concluded ha machine lea ning can play i al ole in he heal hca e. Amin e al. [35] p oposed machine lea ning model o he p edic ion o Pa kinson’s disease a ea ly s age. Fo building he model, hey used SVM classi ie . Fea u e selec ion algo- i hms we e also applied o ex ac he impo an ea u es: Relie and ACO ea u e selec ion algo i hm. This esea ch a icle p ima ily aims o p edic whe he a pe son has Ch onic Kidney Disease o no . In his pe cep ion, se en di e en machine lea ning classi ie s we e applied on he da ase . All he algo i hms we e unning wi h bo h ull ea u es and selec ed ea u es. SMOTE was used o o e sam- pling and all he esul s we e eco ded. All he machine lea n- ing model esul s we e also compa ed wi h one deep neu al ne wo k algo i hm. Deep lea ning neu al ne wo k was used wi h wo hidden laye s. IBM SPSS Modele was applied o compu a ional pu pose. The con ibu ion e eals he accu acy es ima e o 99.6% when applying deep neu al ne wo k on he da ase . II. RESEARCH GAP Un il now, in majo i y o cases ull ea u es ha e been aken in o conside a ion. In his esea ch, ea u e op imiza ion was ca ied ou , whe ein h ee di e en ea u e selec ion algo- i hms we e applied o ind he algo i hm mos bene icial o ex ac he impo an ea u e o he p edic ion o Ch onic Kidney Disease. As many da ase s ha e imbalanced class, class balancing is needed o inc easing he pe o mance o classi ie model. In his esea ch SMOTE was used as a class balance . The highes accu acy o 99.6% was achie ed whe eas he a icle [22] p o ides an accu acy o 99.1% on he same da ase . Acco ding o [15] he highes accu acy o he model was 99.7%, bu hey wo ked on isk calcula ion o he pa ien whe eas he main aim o he a icle is o p edic Ch onic Kidney Disease. III. METHODS AND MATERIALS In his sec ion, he esea ch me hodology and a da ase will be discussed. A. DATASET Ch onic Kidney Disease da ase is used o his esea ch wo k. Many esea che s had also used his da ase [26]. This da ase is being p o ided by he UC I ine Machine 17314 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease Lea ning Reposi o y and i is a ailable on he UCI websi e. This da ase con ains 400 ins ances and 24 a ibu es wi h 1 a ge a ibu e. The a ge a ibu e has labelled in wo-class o ep esen CKD o non-CKD. The da ase was collec ed om a ious hospi als in 2015. I con ains also missing alue. The desc ip ion o all 24 a ibu es is ep esen ed in he able 1 below. B. METHODOLOGY In his esea ch, we ha e de eloped a model o p edic CKD disease in pa ien s. The pe o mance o he model was es ed on bo h all a ibu es and selec ed ea u es. Among ea u e selec ion me hods he e we e W appe , Fil e and Embedded [27] allowing o selec i al ea u es. Classi ie algo i hms pe o mance was es ed on he selec ed ea u es. IBM SPSS ool is used o p epa ing he model. The machine lea ning classi ie s such as a i icial neu al ne wo k (ANN), C5.0, logis ic eg ession, linea suppo ec o machine (LSVM), K- nea es neighbo s (KNN) and andom ee we e used o aining he model. Each classi ie alida ion and pe o - mance ma ix we e compu ed. The p ocedu e o his esea ch including i e s ages: (i) da ase p ep ocessing, (ii) ea u e selec ion, (iii) classi ie applica ion, (i ) SMOTE and ( ) ana- lyzing he pe o mance o he classi ie . Along wi h machine lea ning models, a deep neu al ne wo k was applied o com- pa ing he esul o machine lea ning models and deep neu al ne wo k. A i icial Neu al ne wo k classi ie was used o his pu pose. In his esea ch he signi icance o wo model we e checked by s a is ic es ing namely McNema ’s es . C. PREPROCESSING OF DATA Da a p ep ocessing could be a s a egy ha is u ilized o change o e he aw in o ma ion in o a clean da ase . I is a he basic s ep o ain e e y machine lea ning classi ie algo i hm. This echnique concludes such ac ions as handle missing alues, escaling o he da ase , ans o m in o bina y da a and s anda dize o he da ase . When he da ase included a ibu es wi h a ying scales, escaling is used o scale he da ase . The bina y ans o ma ion has been applied o con- e he alue in o 0 and 1. All alues o e e y a ibu e a e conside ed as 1 o abo e he h eshold and as 0 o below he h eshold. S anda dized me hod ensu es ha each a ibu e has mean 0 and s anda d de ia ion 1. D. FEATURE SELECTION Fea u e selec ion is needed o ained each machine lea ning classi ie because wi hou emo ing unnecessa y a ibu es om he da ase esul may be a ec ed. The classi ie algo- i hm wi h ea u e selec ion gi es be e pe o mance and educe he execu ion ime o he model. Fo his p ocess, h ee di e en ea u e selec ion me hods we e used in his esea ch. 1) FILTER METHOD The il e is one o he me hods o selec he app op ia e ea u e. I selec s he ea u e on hei in eg al ea u es wi hou in eg a ing any lea ning classi ie algo i hm. This me hod gi es esul as e as compa ed o he w appe me hod. The me hod assigns he sco e o e e y a ibu e based on hei s a is ical co ela o be ween a ibu es. The e a e many il e me hods a e a ailable, bu Co ela ion-based Fea u e Selec- ion (CFS) me hod has been used. CFS is he algo i hm o selec he ea u e-based on he a ibu e anks. I assigns he ank o a ibu e subse as based on he co ela ion heu is ic e alua ion unc ion [28]. The unc ion wo ks on he s a egy ha c ea es wo class labels, one is co ela ed o class and low co ela ed class and selec s only co ela ed label class a ibu es. 2) WRAPPER METHOD W appe me hod selec s he subse o ea u es based on a p ecise machine lea ning algo i hm [29]. I used he g eedy sea ch me hod o inding a possible subse o ea u es. The me hod can be implemen ed wi h using any o he ollow- ing algo i hms o wa d selec ion, backwa d elimina ion and ecu si e elimina ion. In he esea ch, we used he o wa d ea u e selec ion me hod. The o wa d ea u e selec ion i e - a i ely selec s he ea u e. This p ocedu e s a s wi h he null model and wo k i e a i ely and add he a ibu e in each s ep. The a ibu e is keeping add in he model un il he a ibu e does no imp o e model pe o mance. 3) EMBEDDED METHOD The embedded me hod is decision ee algo i hm o ea u e selec ion. I selec s he ea u e in each s ep wo ks ecu si ely while he ee is g owing and spli he sample se in o a smalle subse . The mos common decision ee algo i hm a e: ID3, C4.5 and CART. The e a e o he a ailable me hod s c ea ing linea models. The mos common me hods a e LASSO [30] wi h L1 penal y and Ridge wi h L2 penal y. In his esea ch LASSO (leas absolu e sh inkage and selec- ion ope a o ) algo i hm has been used. I pe o ms wo main asks: egula iza ion and ea u e selec ion. In egula iza ion, i sh inks some ea u e coe icien s o ze o ha means ea u es a e no impo an o he p edic o model. E. CLASSIFICATION ALGORITHMS Classi ica ion echnique is an impo an ea u e o supe ised lea ning. Classi ie s lea n om he aining da ase and apply on he es ing da ase o inding he a ge a ibu e. Below he e a e classi ica ion echniques used in esea ch. 1) ARTIFICIAL NEURAL NETWORK A i icial neu al ne wo k is a pa o a i icial in elligence. I is a ype o supe ised machine lea ning. I s s uc u e is he same as he human b ain. ANN also ha e neu ons and jus like in human all neu ons a e in e connec ed o one ano he , ANN neu ons a e connec ed o each o he in laye s o he ne wo k. Neu ons he e a e known as nodes. ANN can sol e he p oblem ha has been impossible o human o s a is ical s anda ds. ANN consis s o h ee laye s: inpu , hidden and ou pu laye s. The inpu laye akes inpu and weigh and passes o hidden laye o pe o ming calcula ion VOLUME 9, 2021 17315 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease TABLE 1. Desc ip ion o A ibu es in he Da ase . and inding he hidden s uc u es and pa e ns. The numbe o hidden laye s can be inc eased as equi ed. The ou pu laye compu es he ou pu . The weigh alues om he ou pu , i.e. p edic ed, and ac ual alue we e ecompu ed and he ne wo k again es a s o inding he class om he p e ious lea ning. The e o e, ANN wo ks based on backp opaga ion. 2) C5.0 C5.0 is a ype o decision ee because i c ea es he decision ee om he inpu . The ee has he numbe o b anches. I u ilizes he ee s uc u e o model he ela ionship be ween ea u es and po en ial ou comes. A each node o he ee, he a ibu e o he da ase is chosen. I can handle nominal and nume ic ea u es bo h. C5.0 is he ex ended e sion o he C4.5 classi ica ion algo i hm and uses in o ma ion en opy concep . En opy is used o inding he impu i y o ea u es. In o ma ion en opy is p oduced based on he calcula ion o pa en and child en opy alues. This p ocess is i e a i e and wo ks un il he e is no he u he spli . 3) LOGISTIC REGRESSION Logis ic eg ession is also a ype o supe ised lea ning algo- i hm. I is a s a is ical model. The p obabili y o a ge alue is p edic ed om logis ic eg ession. I is di ided he a ge a ibu e in o wo-classes: success o no success. Fo success, i e u ns 1 whe eas i e u ns 0 o no succeeding. Logis ic eg ession is ep esen ed by equa ion 1: P=1/(1+e^(−(b0+b1x+b2x^2)) (1) whe e P is he p edic ed alue, b0, b1, b2a e biases and x is is an a ibu e. I is used in a ious ield o machine lea ning applica ion in social sciences and medical a ena, o example, o spam de ec ion, diabe es de ec ion, cance de ec ion, e c. Logis ic eg ession is he ad anced e sion o linea eg ession. Th ough his echnique, we only conce n abou he p obabili y o he ou come a iable. 4) CHAID Chi-squa e au oma ic in e ac ion de ec ion (CHAID) is a ype o decision ee echnique. I is used o de e mine he ela- ionship be ween a iables. Nominal, o dinal and con inuous da a can be used in CHAID o inding he ou come. Fo each ca ego ical p edic o , all possible c oss- abula ion is c ea ed in he CHAID model and i p ocess wo ks un il he bes ou come is a ained. The a ge o dependen a iable becomes a oo node in he ee, he a ge a iable is spli in o 17316 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease wo o mo e pa s as pe he ca ego ies in a ge a iable and child o he oo node a e c ea ed using he s a is ical me hod and a iable ela ionship. Such a p ocess will be ill lea nodes o he ee. F es is used o he con inuous dependen a iable and he Chi-squa e es is used o he ca ego ical dependen a iable. 5) LINEAR SUPPORT VECTOR MACHINE (LSVM) linea suppo ec o machine (LSVM) is he mode n pa ic- ula ly as machine lea ning algo i hm o sol ing mul iclass classi ica ion p oblem o he la ge da ase based on a simple i e a i e app oach. I is c ea ed he SVM model in linea CPU ime o he da ase . LSVM can be used o he high dimensional da ase is he spa se and dense o ma . I is used o sol ing he la ge da ase machine lea ning p oblems in less expensi e compu ing esou ce. Suppo Vec o Machine is a supe ised classi ie algo i hm. I is used ke nel ick o sol ing he classi ica ion p oblem. Based on hese ans o - ma ions, ideal edge is ound be ween he possible ou pu s. SVM is used o he nonlinea ke nel, such as RBF. Fo he linea ke nel, LSVM is an app op ia e choice. LSVM classi ie is su icien o all linea p oblems. 6) K- NEAREST NEIGHBORS (KNN) KNN is a simple ype o supe ised algo i hm. I can be used o bo h classi ica ion and eg ession p oblems. Howe e , i is la gely used o classi ica ion p oblems. KNN does no use a pa icula aining s age and use all he da a o aining so ha i is a lazy lea ning algo i hm and also i does no conside any hing abou he unde lying da a, so ha is a non- pa ame ic lea ning algo i hm. KNN s o es he whole da ase because i has no model so ha he e is no lea ning equi ed. When he new da a en e o p edic ing he ou comes, i com- pa es K – neighbo s so ha selec ion o K’s alue is e y impo an . The dis ance is calcula ed be ween wo al eady label da a. The dis ance helps o ind he nea es neighbo o he new da a. A Euclidian me hod is used o inding he dis ance. 7) RANDOM TREE The andom ee is a ype o supe ised classi ie s. I p oduces lo s o dis inc lea ne s. The s ochas ic p ocess is used o o m he ee. I is a ype o ensemble lea ning echnique o classi ica ion. I wo ks he same as decision ee, bu a andom subse o a ibu es uses o each spli . This algo i hm uses o bo h classi ica ion p oblems and eg ession p oblems. A g oup o andom ees is known as a o es . The andom ees classi ie akes he inpu ea u e se and classi ies inpu o e e y ee in he o es . The ou pu o he andom ee selec s om he majo i y o o es. In he ee, e e y lea node holds a linea model. The bagging aining algo i hm is used o ain he model. F. VALIDATION METHOD OF CLASSIFIERS The da ase was di ided in o pa s: aining da ase and es ing da ase . IBM SPSS modelle was used o he pa i ion and TABLE 2. Con usion Ma ix. p edic ion o he esul . The aining da ase con ains 50% o he da a and emaining da a is conside ed as he es ing da a. The ype ool o IBM SPSS was applied o changing he ype o a ibu es. The pe o mance e alua ion ma ix was ecei ed o each classi ica ion algo i hm. G. PERFORMANCE EVALUATION MEASURE Va ious e alua ion ma ices we e used o checking he pe - o mance o he classi ie . Fo his pu pose, he con usion ma ix was used. I is a 2∗2 ma ix due o wo classes in he da ase . The con usion ma ix gi es wo ypes o co ec p e- dic ion o he classi ie and wo ypes o inco ec p edic ion o he classi ie . The con usion ma ix is p esen ed in Table 2. 1) CONFUSION MATRIX DESCRIPTION TP: T ue Posi i e means ou pu as posi i e such ha p edic ed esul is co ec ly classi ied. TN: T ue Nega i e means ou pu as nega i e such ha p edic ed esul is co ec ly classi ied. FP: False Posi i e means ou pu as posi i e such ha p e- dic ed esul is inco ec ly classi ied. FN: False Nega i e means ou pu as nega i e such ha p edic ed esul is inco ec ly classi ied. 2) CLASSIFICATION ACCURACY Classi ica ion accu acy shows he co ec a e o p edic ion esul s. I compu es om he con usion ma ix. The classi i- ca ion accu acy is ound by equa ion 2: accu acy =TP +TN TP +TN +FP +FN ∗100 (2) 3) CLASSIFICATION ERROR Classi ica ion e o shows he inco ec a e o p edic ion esul s. I compu es om he con usion ma ix. The classi- ica ion e o is ound by equa ion 3: E o =FP +FN TP +TN +FP +FN∗100 (3) 4) PRECISION P ecision is an impo an model pe o mance e alua ion ma ix. I is he ac ion o ela ed ins ances among he o al e ie ed ins ances. I is a posi i e p edic ed alue. The p e- cision is calcula ed as ollows in equa ion 4: P ecision =TP TP +FP ∗100 (4) VOLUME 9, 2021 17317 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease 5) RECALL Recall is also an impo an model pe o mance e alua ion ma ix. I is he ac ion o ela ed ins ances among he o al numbe o e ie ed ins ances. The ecall is calcula ed as ollows in equa ion 5: Recall =TP TP +FN ∗100 (5) 6) F-MEASURE I is also known as F Sco e. F-measu e is calcula ed so as o measu e he accu acy o es . I is calcula ed om he p ecision and ecall by equa ion 6: F−Measu e =2∗P ecision ∗Recall P ecision +Recall (6) 7) ROC AND AUC The pe o mance o he classi ica ion model is measu ed om he Recei e ope a ing a cha ac e is ic cu e (ROC). ROC is a g aph ha is c ea ed o ue posi i e a e s. alse posi i e a e a di e en classi ica ions h eshold. The en i e a ea unde he ROC cu e is known as a ea o he cu e (AOC). I gi es a collec i e measu e o pe o mance ac oss all achie able classi ica ion’s h eshold. 8) GINI COEFFICIENT I is also known as GINI index. I is a measu e o s a is ical dis ibu ion. I is used o measu e he inequali y amongs alues o a ibu es. I is also possible o say ha i calcu- la es he impu i y o a pa icula a ibu e in he o m o deg ee o p obabili y. H. SMOTE Syn he ic Mino i y O e sampling Technique (SMOTE) is used o o e sampling he mino i y class. I is also known as a balance . I akes he whole da ase as inpu bu wo ks only on mino i y class. I inc eases he pe cen age o mino i y class. SMOTE used KNN o inding new ins ances. I does no make any change in he majo i y cases. The new examples a e no simply duplica ing o exis ing mino i y cases. Ins ead, he calcula ion akes es s o he componen space o each a ge class and i s closes neighbo s and hen p oduces new models ha join a ibu es o he objec i e case wi h he highligh s o i s neighbo s. This me hodology builds he high- ligh s accessible o each class and makes es s p og essi ely b oad. The ma hema ical symbols used in his esea ch is shown in he able 3. I. STATISTICS TEST FOR MODEL COMPARISON Fo he pu pose o compa ing wo models, McNema ’s es was applied on he p edic ed ou pu o wo models. The McNema ’s es is used o de e mine whe he he e a e di - e ences on bipola dependen a iables be ween wo ela ed g oups. In his es 2∗2 con ingency ma ix is o med o wo g oups and p alue is calcula ed. Fo his pu pose, signi i- cance le el α=0.05 is conside ed. I p < α, we can ejec he null hypo hesis. I p > α, we ail o ejec he null hypo hesis. I p alue is less han α, i means bo h models show a signi ican di e ence as ega ds he hypo hesis. Howe e , i p > α, he di e ence would no be ega ded as s a is ically signi ican . IV. EXPERIMENTAL AND RESULT The esul o his esea ch including all ou comes and clas- si ica ion models om di e en pe cep ion will be discussed in his sec ion. IBM SPSS model is shown in igu e 1. Fi s , he pe o mance o di e en machine lea ning algo i hm iz. an a i icial neu al ne wo k, logis ic eg ession, C5.0, CHAID, andom ee, K-nea es neighbo s and linea suppo ec o machine ha e been checked on all ea u es. Second, pa ea u e selec ion algo i hm CFS, o wa d W appe and LASSO ha e been applied on he da ase o ind he impo - an ea u es. Thi d, he pe o mance o all abo e-men ioned classi ica ion algo i hm on impo an ea u es was checked. Fou h, SMOTE il e was also applied o he da ase and he esul o classi ie s we e checked. Va ious ools we e used. Weka ool was used o CFS and Fo wa d me hod. s udio was used o LASSO. IBM SPSS Modele was used o he pe o mance o classi ie s. Deep neu al ne wo k was buil in IBM SPSS modele . ANN was used wi h 2 hidden laye s o building deep neu al ne wo k. Twel e nodes we e used in hidden laye 1 and eigh nodes we e in hidden laye 2. A. RESULT WITHOUT FEATURES SELECTION In his subsec ion, he ull ea u es o he da ase we e used and he esul was es ed on all se en machine lea ning classi ica ion algo i hms wi h 50% o aining da a and 50% o es ing da a. The compa ison ma ix was c ea ed o all algo i hms. Wi h he esul an ma ix, h ee g aphs we e also c ea ed o checking he a ia ion o a ious classi ie s. The i s g aph p o ides a compa ison o all classi ie ’s accu acy, p ecision and ecall. The second one con ains he a ia ion o AUC and he hi d one includes he a ia ion o F-measu e. The compa ison o all classi ie s showed ha he C5.0 algo- i hm achie ed he highes accu acy, i.e. 96.10%. The alue o all pa ame e s o C5.0: p ecision was 92.40%, ecall was 97.30%, F-measu e was 94.80%, AUC was 97.80% and GINI index was 0.96. The a i icial neu al ne wo k was ained on 3 hidden laye s. The ANN achie ed accu acy o 94.63%, p ecision o 93.24% and ecall o 92%. The logis ic eg es- sion achie ed accu acy o 71.71%, p ecision o 56.48% and ecall o 98.6%. The CHAID algo i hm achie ed accu acy o 96%, p ecision o 93.50% and ecall o 92%. The LSVM wi h Penal y L1 and Lambda 0.5 achie ed accu acy o 92.2%, p ecision o 83.90% and ecall o 97.33%. The LSVM wi h Penal y L2 and Lambda 0.5 achie ed accu acy o 94.63%, p ecision o 90% and ecall o 96%. The KNN ga e he wo s esul o his da ase wi h a K alue o 5: accu acy o 64.39%, p ecision o 59.01% and ecall o 96%. The andom ee algo- i hm achie ed accu acy o 90.73%, p ecision o 83.34% and ecall o 93%. As a esul , ANN achie ed he highes AUC and C5.0 achie ed F-measu e. The esul o all classi ie s is 17318 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease TABLE 3. Desc ip ion o Used Ma hema ical Symbol. as ollows in able 4. The compa ison o p ecision, ecall and accu acy is desc ibed in igu e 2. The compa ison o AUC is desc ibed in igu e 3. The compa ison o F-Measu e is desc ibed in igu e 4. B. RESULT OF CORRELATION-BASED FEATURE SELECTION (CFS) In his subsec ion, he impo an ea u es we e selec ed by CFS algo i hm o pass in he classi ie algo i hms o p edic - ing he ou comes. Six mos impo an ea u es we e used o inding he ou comes such as bp, pc, pe, ane, pc and bc. As pe he CFS algo i hm, bp and pc a e he mos impo an ac o s o p edic ing Ch onic Kidney Disease. The esul o CFS algo i hm is shown in igu e 5. The pe o mance o all se en classi ie s was desc ibed in able 5. The LSVM wi h Penal y L2 and Lambda 0.5 achie ed he highes accu acy o selec ed ea u es om he CFS algo i hm wi h 95.12% accu acy, 93.34% p ecision and 93.34% ecall. C5.0 and CHAID achie ed an accu acy o 92.68%. The C5.0 algo i hm achie ed 85.71% p ecision and 96% ecall. The CHAID algo i hm achie ed 92.68% accu acy, 96.87% p ecision and 83% ecall. The ANN algo i hm achie ed 91.71% accu acy, 89.19% p ecision and 88% ecall. The logis ic eg ession algo i hm achie ed he lowes accu acy o 51.22%, 96.87% p ecision and 92.54% ecall. The LSVM wi h Penal y L1 and Lambda 0.5 achie ed 93.66% accu acy, 87.8% p ecision and 96% ecall. The KNN achie ed o his da ase wi h a K alue o 5 accu acy o 53.17%, p ecision o 97.05% and ecall o 100%. The andom ee algo i hm achie ed 87.80% accu acy,82.05% p ecision and 85% ecall. As om he esul , LSVM wi h penal y L1 achie ed he highes AUC. The compa ison o p ecision, ecall and accu acy is desc ibed in igu e 5. The compa ison o he GINI index is shown in igu e 6. The compa ison o AUC is desc ibed in igu e 7. C. RESULT OF WRAPPER FORWARD FEATURE SELECTION AND CLASSIFICATION In his subsec ion, he impo an ea u es we e selec ed by W appe o wa d ea u e selec ion algo i hm o pass in he classi ie algo i hms o p edic ing he ou comes. Six mos impo an ea u es we e used o inding ou comes such as hemo, h n, dm, cad, pe, al. As pe he CFS algo i hm, hemo and h n a e he mos impo an ac o s o p edic ing Ch onic Kidney Disease. The esul o he W appe algo i hm is shown in igu e 9. The esul o all classi ie algo i hm pe o - mance is desc ibed in able 6. The C5.0 achie ed he highes accu acy wi h he W appe algo i hm, namely 96.1% accu- acy, 98.55% p ecision and 90.67% ecall. ANN, CHAID and he andom ee also ga e a good esul . The ANN algo i hm achie ed 94.63% accu acy, 90% p ecision and 96% ecall. The CHAID algo i hm achie ed 94.63% accu acy, 93.24% p ecision and 92% ecall. The andom ee algo i hm achie ed 94.63% accu acy, 93.24% p ecision and 92% ecall. The logis ic eg ession algo i hm achie ed 78.54% accu acy, 98.55% p ecision and 100% ecall. The LSVM wi h Penal y L1 and Lambda 0.5 achie ed 94.15% accu acy, 88.89% p eci- sion and 96% ecall. The LSVM wi h Penal y L2 and Lambda 0.5 achie ed 93.66% accu acy, 87.80% p ecision and 96% ecall. The KNN ga e he wo s esul o his da ase wi h a K alue o 5 76: 10% accu acy, 95.58% p ecision and 95.58% ecall. As om he esul , LSVM achie ed he highes AUC. The compa ison o p ecision, ecall and accu acy is desc ibed in igu e 10. The compa ison o he GINI index is shown in igu e 11. The compa ison o AUC is desc ibed in igu e 12. D. RESULT OF LASSO FEATURE SELECTION In his subsec ion, he impo an ea u es we e selec ed by LASSO ea u e selec ion algo i hm o passe in he classi ie algo i hms o p edic ing he ou comes. Six mos impo an ea u es we e used o inding ou comes such as bc, pc, al, ba, su, pcc. As pe he LASSO FS algo i hm, bc and pc a e he mos impo an ac o s o p edic ing Ch onic Kidney Disease. The esul o he LASSO FS algo i hm is shown in igu e 13. The esul o algo i hm pe o mance o all se en classi ie s is desc ibed in able 7. LSVM and CHAID achie ed he highes accu acy o 97.07%. LSVM wi h bo h penal y L1 and L2 achie ed 97.07% accu acy, 98.59% p e- cision and 93.33% ecall. The CHAID algo i hm achie ed VOLUME 9, 2021 17319 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease FIGURE 1. IBM SPSS model o kidney disease p edic ion. 97.07% accu acy, 100% p ecision and 92% ecall. The ANN algo i hm achie ed 94.63% accu acy, 90% p ecision and 96% ecall. The andom ee algo i hm achie ed 90.24% accu acy, 80.90% p ecision and 96% ecall. The logis ic eg ession algo i hm achie ed 74.15% accu acy, 80.23% p ecision and 100% ecall. The andom ee algo i hm achie ed 88.78% accu acy, 78.26% p ecision and 96% ecall. The KNN ga e he wo s esul o his da ase wi h a K alue o 5: 56.59% accu acy, 92% p ecision and 100% ecall. As om he esul , LSVM achie ed he highes AUC. The compa ison o p e- cision, ecall and accu acy is desc ibed in igu e 13. The compa ison o he GINI index is shown in igu e 14. The compa ison o AUC is desc ibed in igu e 15 shows. E. RESULT OF SMOTE As he abo e esul , i was obse ed ha he highes accu acy achie ed on he selec ed ea u es was gi en by LASSO ea- u e selec ion me hod. Thus, SMOTE echnique was applied on ull ea u es and on selec ed ea u es gi en by LASSO eg ession me hod. The pe o mance o ANN, CHAID, LSVM and Random T ee was checked by SMOTE. These classi ica ion algo i hms we e pe o med e y well in all 17320 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease FIGURE 15. Pe o mance o GINI index o all classi ie s a e LASSO ea u e selec ion. FIGURE 16. Pe o mance o a ea unde he cu e o all classi ie s a e LASSO ea u e selec ion. FIGURE 17. Compa ison o p ecision, ecall and accu acy o all classi ie s wi h SMOTE and selec ed ea u es. p ecision and 91.38% ecall. LSVM wi h penal y L1 achie ed 96.53% accu acy, 91.04% p ecision and 100% ecall. ANN achie ed 96.47% accu acy, 98.14% p ecision and 91.38% ecall. C5.0 achie ed 96.45% accu acy, 96.61% p ecision and 93.44% ecall. The Random T ee achie ed 91.43% accu acy, 84.72% p ecision and 94% ecall. The compa ison g aph o VOLUME 9, 2021 17327 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease FIGURE 18. Pe o mance o AUC o all classi ie s wi h SMOTE and selec ed ea u es. FIGURE 19. Pe o mance o F-Measu e o all classi ie s a e LASSO ea u e selec ion and SMOTE. TABLE 8. Pe o mance o Classi ie s Wi h SMOTE and Selec ed Fea u es. p ecision, ecall and accu acy o all algo i hms is shown in igu e 20. The compa ison g aph o AUC o all algo i hms is shown in igu e 21. The compa ison g aph o F-Measu e o all algo i hms is shown in igu e 22. 17328 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease FIGURE 20. Compa ison o p ecision, ecall and accu acy o all classi ie s wi h SMOTE and ull ea u es. FIGURE 21. Pe o mance o AUC o all classi ie s wi h SMOTE and ull ea u es. FIGURE 22. Pe o mance o F-Measu e o all classi ie s a e LASSO ea u e selec ion and SMOTE. F. COMPARISION MATRIX OF ALL EXPERIMENTS I was obse ed ha ANN, C5.0, CHAID, LSVNM and Ran- dom ee pe o med well on he conside ed CKD da ase . The Logis ic eg ession and KNN ha e no gi en ou comes as expec ed. So, he compa ison able has been c ea ed o i e bes -pe o med algo i hms o all di e en echnique ype. VOLUME 9, 2021 17329 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease TABLE 9. Pe o mance o Classi ie s Wi h SMOTE and Full Fea u es. FIGURE 23. Compa ison o all classi ie models. TABLE 10. Classi ie Pe o mance in Va ious Models. The esul is desc ibed in able 10. The accu acy compa ison g aph is shown in igu e 23. G. PERFORMANCE OF LSVM IN ALL TECHNIQUES As he abo e esul , LSVM wi h penal y L2 ga e a be e esul in all echniques is i was discussed p e iously. In his sec ion, he pe o mance o LSVM will be discussed. The able 11 shows he esul o LSVM in all echniques. Along wi h he able, he e is he g aph on he able da a. Fig- u e 24 shows he compa ison o LSVM in all echniques. H. VALIDATE MACHINE LEARNING MODEL To alida e he indings abo e, he s udy includes esul s om ano he da a se ound a The Cance Imaging A chi e 17330 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease FIGURE 24. Compa ison o LSVM in di e en models. TABLE 11. Pe o mance o LSVM in all Machine Lea ning Models. (TCIA). The da ase has 210 ins ances o kidney disease pa ien . I con ains 48 a ibu es and 1 a ge a iable. The da ase was used on he same models applied ea lie . These indings a e gi en below and a e, in gene al, compa able o ea lie esul s wi h no signi ican di e ences obse ed. Though, he ou come o applying machine lea ning models la gely dependens on he speci ic da ase , he expe imen s abo e alida e ea lie indings, namely, SMOTE wi h ull ea u es esul . Table 12 shows he esul o bo h da ase s. I. PERFORMANCE OF DEEP NEURAL NETWORK In his esea ch wo k, a i icial neu al ne wo k was used o machine lea ning and deep neu al ne wo k-based analysis. Fo machine lea ning a i icial neu al ne wo k used only single hidden laye , bu o he usage o he a i icial neu al ne wo k as a deep neu al ne wo k mo e hidden laye s can be added. So as o es he pe o mance o machine lea ning classi ie algo i hms, one deep neu al ne wo k model was buil and he esul s we e no ed. In some cases, he deep neu al ne wo k ga e s ong esul and impo an ea u es we e ex ac ed by i sel , ha is, no ea u e selec ion algo i hm was equi ed. The same da ase was used o building a deep neu al ne wo k. I was no ed ha a deep neu al ne wo k achie ed he highes accu acy o 99.6% and i was be e han o he machine lea ning models. V. DISCUSSION OF MACHINE LEARNING MODELS All he machine lea ning models bu Logis ic and KNN classi ie s gi e sa is ac o y esul and ha e he negligible di e ence be ween p ecision and ecall alues. In compa ison wi h hem p ecision o Logis ic and KNN classi ie s is low whe eas ecall is high. I indica es ha hese wo classi ie s gi e many False posi i e esul s due o unbalanced da ase . Logis ic and KNN algo i hms ha e no enough capaci y o dis inguish be ween posi i e class and nega i e class as he ela ed AUC sco e is e y low. Along wi h AUC, he GINI coe icien is also no sa is ac o y. Hence, Logis ic and KNN a e no sui able o he p edic ion o CKD. In all cases LSVM wi h L1 and L2 penal y has he bes p ecision, ecall, AUC sco e and GINI coe icien and he model achie ed he highes accu acy in majo i y cases. VI. PERFORMANCE COMPARISON OF MACHINE LEARNING MODEL AND DEEP NEURAL NETWORK All he machine lea ning model esul s was discussed in he able 10 and acco ding o i LSVM wi h penal y L2 pe o med bes in SMOTE wi h ull ea u es and achie ed he highes accu acy o 98.46%. As discussed, he deep neu al ne wo k achie ed he highes accu acy om among all models wi h 99.6%. In o de o compa e he pe o mance o wo models, McNema ’s es was applied. Fo his es , he highes accu- acy was achie ed by machine lea ning model, i.e., LSVM VOLUME 9, 2021 17331 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease TABLE 12. Valida ion o Machine Lea ning Model. wi h SMOTE o all ea u es and a deep neu al ne wo k was aken and hei signi ican alue was no ed. The p alue o his es was 0.29 and i is g ea e han signi ican le el (α= 0.05) and, hence, we would ejec hypo hesis. VII. CONCLUSION This a icle objec s o p edic Ch onic Kidney Disease based on ull ea u es and impo an ea u es o CKD da ase . Fo ea u e selec ion h ee di e en echniques ha e been applied: co ela ion-based ea u e selec ion, W appe me hod and LASSO eg ession. In his pe cep ion, se en classi ie s algo i hm we e applied iz. a i icial neu al ne wo k, C5.0, logis ic eg ession, CHAID, linea suppo ec o machine (LSVM), K-Nea es neighbo s and andom ee. Fo each classi ie , he esul s we e compu ed based on ull ea- u es, selec ed ea u es by CFS, selec ed ea u es by W ap- pe , selec ed ea u es by LASSO eg ession, SMOTE wi h selec ed ea u es by LASSO, SMOTE wi h ull ea u es. I was obse ed ha LSVM achie ed he highes accu acy o 98.86% in SMOTE wi h ull ea u es. All classi ie s algo i hms pe o med well on ea u es selec ed by LASSO eg ession wi h SMOTE and wi hou SMOTE. SMOTE wi h ull ea u es ga e he bes esul o all 5 classi ie s. In his esea ch, a o al o 7 classi ie s we e used. Howe e , Logis ic and KNN did no gi e sui able esul s and i was why hey we e no used in SMOTE. As pe he esul , i is concluded ha SMOTE is a bes echnique o balancing a da ase . I is no ed ha SMOTE ga e be e esul s wi h selec ed ea- u es by LASSO eg ession as compa e o wi hou SMOTE on LASSO eg ession model. LSVM achie ed he highes accu acy in all expe imen s as compa ed o o he classi ie s algo i hms. REFERENCES [1] Q.-L. Zhang and D. Ro henbache , ‘‘P e alence o ch onic kidney disease in popula ion-based s udies: Sys ema ic e iew,’’ BMC Public Heal h, ol. 8, no. 1, p. 117, Dec. 2008. [2] W. M. McClellan, D. G. Wa nock, S. Judd, P. Mun ne , R. Kewal amani, M. Cushman, L. A. McClu e, B. B. Newsome, and G. Howa d, ‘‘Albumin- u ia and acial dispa i ies in he isk o ESRD,’’ J. Ame . Soc. Neph ol., ol. 22, no. 9, pp. 1721–1728, Aug. 2011. [3] M. K. Ha oun, ‘‘Risk ac o s o ch onic kidney disease: A p ospec i e s udy o 23,534 men and women in Washing on Coun y, Ma yland,’’ J. Ame . Soc. Neph ol., ol. 14, no. 11, pp. 2934–2941, No . 2003. [4] W. D. Souza, L. C. D. Ab eu, L. G. D. Sil aI, and I. M. P. Beze a, ‘‘Incidence o ch onic kidney disease hospi alisa ions and mo ali y in Espí i o San o be ween 1996 o 2017,’’ Wisi Cheungpasi po n, Uni . Mississippi Medical Cen e , Roches e , MN, USA, Tech. Rep., 2019, doi: 10.1371/jou nal.pone.0224889. [5] W. Mula-Abed, K. A. Rasadi, and D. Al-Riyami, ‘‘Es ima ed glome ula il a ion a e (eGFR): A se um c ea inine-based es o he de ec ion o ch onic kidney disease and i s impac on clinical p ac ice,’’ Oman Med. J., ol. 27, no. 4, pp. 339–340, 2012. [6] A. S. Le ey, D. Ca an, A. F iedman, W. G. Mille , J. Sedo , K. Tu le, B. Kasiske, and T. Hos e e , ‘‘P o einu ia as a su oga e ou come in CKD: Repo o a scien i ic wo kshop sponso ed by he na ional kidney ounda- ion and he US ood and d ug adminis a ion,’’ Ame . J. Kidney Diseases, ol. 54, no. 2, pp. 205–226, Aug. 2009. [7] S. Ge ogianni, ‘‘Conce ns o pa ien s on dialysis: A esea ch s udy,’’ Heal h Sci. J., ol. 8, no. 4, pp. 423–437, 2014. [8] J. R. Chapman, ‘‘Wha a e he key challenges we ace in kidney ansplan- a ion oday?’’ T ansplan a ion Res., ol. 2, no. S1, pp. 1–7, No . 2013. [9] T. Xiuyi and G. Yuxia, ‘‘Resea ch on applica ion o machine lea n- ing in da a mining,’’ in P oc. IOP Con ., Ma e . Sci. Eng., 2018, doi: 10.1088/1757-899X/392/6/06220. [10] B. Zupan, A. J. Hal e , and M. Bohanec, ‘‘Quali a i e model app oach o compu e assis ed easoning in physiology,’’ in P oc. In ell. Da a Anal. Med. Pha macol. (IDAMAP), B igh on, U.K., 2018, pp. 1–7. [11] A. Dhillon and A. Singh, ‘‘Machine lea ning in heal hca e da a analysis: A su ey,’’ J. Biol. Today’s Wo ld, ol. 8, no. 2, pp. 1–10, Jan. 2018. [12] M. Ghassemi, T. Naumann, P. Schulam, A. L. Beam, I. Y. Chen, and R. Rangana h, ‘‘A e iew o challenges and oppo uni ies in machine lea n- ing o heal h,’’ in P oc. AMIA Join Summi s T ansl. Sci., 2020, p. 191. [13] J. Qin, L. Chen, Y. Liu, C. Liu, C. Feng, and B. Chen, ‘‘A machine lea ning me hodology o diagnosing ch onic kidney disease,’’ IEEE Access, ol. 8, pp. 20991–21002, 2020. [14] G. R. Vasquez-Mo ales, S. M. Ma inez-Mon e ubio, P. Mo eno-Ge , and J. A. Recio-Ga cia, ‘‘Explainable p edic ion o ch onic enal disease in he colombian popula ion using neu al ne wo ks and case-based easoning,’’ IEEE Access, ol. 7, pp. 152900–152910, 2019. [15] Z. Chen, X. Zhang, and Z. Zhang, ‘‘Clinical isk assessmen o pa ien s wi h ch onic kidney disease by using clinical da a and mul i a ia e mod- els,’’ In . U ol. Neph ol., ol. 48, no. 12, pp. 2069–2075, Jun. 2016. [16] Y. Ami galiye , S. Shamiluulu, and A. Se ek, ‘‘Analysis o ch onic kidney disease da ase by applying machine lea ning me hods,’’ in P oc. IEEE 12 h In . Con . Appl. In . Commun. Technol. (AICT), Oc . 2018, pp. 1–4. [17] K. R. A. Padmanaban and G. Pa hiban, ‘‘Applying machine lea ning echniques o p edic ing he isk o ch onic kidney disease,’’ Indian J. Sci. Technol., ol. 9, no. 29, Aug. 2016. [18] L. Kil ia De Almeida, L. Lessa, A. Peixo o, R. Gomes, and J. Celes ino, ‘‘Kidney ailu e de ec ion using machine lea ning echniques,’’ in P oc. 8 h In . Wo kshop ADVANCEs ICT In as uc u es Se ices, 2020, pp. 1–8. [19] W. Guna a hne, K. D. M Pe e a, and K. A. D. C. P Kahandawaa achchi, ‘‘Pe o mance e alua ion on machine lea ning classi ica ion echniques o disease classi ica ion and o ecas ing h ough da a analy ics o ch onic kidney disease (CKD),’’ in P oc. IEEE 17 h In . Con . Bioin . Bioeng. (BIBE), Oc . 2017, pp. 291–296. [20] H. Pola , H. D. Meh , and A. Ce in, ‘‘Huseyin pola 1 & homay danaei meh 1 & aydin ce in,’’ J. Med. Sys ., ol. 41, no. 4, pp. 1–11, Ap . 2017, doi: 10.1007/s10916-017-0703-x. 17332 VOLUME 9, 2021 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease [21] S. D all, G. S. D all, S. Singh, and B. B. Naib, ‘‘Ch onic kidney disease p edic ion using machine lea ning: A new app oach,’’ In . J. Manage., Technol. Eng., ol. 8, pp. 278–287, May 2018. [22] M. Almasoud and T. E. Wa d, ‘‘De ec ion o ch onic kidney disease using machine lea ning algo i hms wi h leas numbe o p edic o s,’’ In . J. Ad . Compu . Sci. Appl., ol. 10, no. 8, pp. 89–96, 2019. [23] S. Shanka , S. Ve ma, S. Ela a hy, T. Ki an, and P. Ghuli, ‘‘Analysis and p edic ion o ch onic kidney disease,’’ In . Res. J. Eng. Technol., ol. 7, no. 5, May 2020, pp. 4536–4541. [24] S. Vijaya ani1 and S. Dhayanand, ‘‘Kidney disease p edic ion using SVM and ANN algo i hms,’’ In . J. Compu . Bus. Res., ol. 6, no. 2, pp. 1–12, Ma . 2015. [25] J. Xiao, R. Ding, X. Xu, H. Guan, X. Feng, T. Sun, S. Zhu, and Z. Ye, ‘‘Compa ison and de elopmen o machine lea ning ools in he p edic ion o ch onic kidney disease p og ession,’’ J. T ansl. Med., ol. 17, p. 119, Dec. 2019. [26] M. S. Gha ibdous i, K. Azimi, S. Ha hikal, and D. H. Won, ‘‘P edic ion o ch onic kidney disease using da a mining echniques,’’ in P oc. Ind. Sys . Eng. Con ., K. Cope ich, E. Cudney, H. Nembha d, Eds., 2017, pp. 2135–2140. [27] E. M. Ka abulu , S. A. Ozel, and T. Ib ikci, ‘‘A compa a i e s udy on he e ec o ea u e selec ion on classi ica ion accu acy,’’ P ocedia Technol., ol. 1, pp. 323–327, Jan. 2012. [28] A. Wosiak and D. Zak zewska, ‘‘In eg a ing co ela ion-based ea u e selec ion and clus e ing o imp o ed ca dio ascula disease diagnosis,’’ Complexi y, ol. 2018, Oc . 2018, A . no. 2520706. [29] N. A. Nnamoko, F. N. A shad, D. England, J. Vo a, and J. No man, ‘‘E al- ua ion o il e and w appe me hods o ea u e selec ion in supe ised machine lea ning,’’ in P oc. 15 h Annu. Pos g adua e Symp. Con e g. Telecommun., Ne w. B oadcas ., Li e pool, U.K., 2014, pp. 2–33. [30] J. M. Pe ei a, M. Bas o, and A. F. D. Sil a, ‘‘The logis ic lasso and idge eg ession in p edic ing co po a e ailu e,’’ P ocedia Econ. Finance, ol. 39, pp. 634–641, Jan. 2016. [31] P. G. Schola , ‘‘Ch onic kidney disease p edic ion using machine lea n- ing,’’ In . J. Eng. Res. Technol., ol. 9, no. 7, pp. 137–140, 2020. [32] B. Deepika, ‘‘Ea ly p edic ion o ch onic kidney disease by using machine lea ning echniques,’’ Ame . J. Compu . Sci. Eng. Su ey, ol. 8, no. 2, p. 7, 2020. [33] F. Ma, T. Sun, L. Liu, and H. Jing, ‘‘De ec ion and diagnosis o ch onic kidney disease using deep lea ning-based he e ogeneous modi ied a i i- cial neu al ne wo k,’’ Fu u e Gene . Compu . Sys ., ol. 111, pp. 17–26, Oc . 2020. [34] A. U. Haq, J. P. Li, J. Khan, M. H. Memon, S. Nazi , S. Ahmad, G. A. Khan, and A. Aliss, ‘‘In elligen machine lea ning app oach o e ec i e ecogni- ion o diabe es in E-heal hca e using clinical da a,’’ Senso s, ol. 20, no. 9, p. 2649, May 2020. [35] U. H. Amin, J. Li, Z. Ali, M. H. Memon, M. Abbas, and S. Nazi , ‘‘Recogni ion o he Pa kinson’s disease using a hyb id ea u e selec ion app oach,’’ J. In ell. Fuzzy Sys ., ol. 39, no. 1, pp. 1–21, Jul. 2020. PANKAJ CHITTORA ecei ed he B.Tech. deg ee in in o ma ion echnology om he Na ional Ins i- u e o Technology, Du gapu , in 2010, and he M.Tech. deg ee in compu e science and enginee - ing in 2012. He is cu en ly pu suing he Ph.D. deg ee om he Depa men o Compu e Sci- ence and Enginee ing, Manipal Uni e si y Jaipu , Jaipu . He has en yea s o expe ience as a Fac- ul y wi h he Compu e Science and Enginee ing Depa men . He is doing esea ch in he ield o heal hca e. His cu en esea ch in e es s include da a science, machine lea ning, da abase, p og amming, in o ma ion secu i y, and he analysis o algo i hms. He has deli e ed a ious subjec including da a s uc u e, algo- i hms analysis, heo y o compu a ion, compile cons uc ion, and o he s. SANDEEP CHAURASIA (Senio Membe , IEEE) is cu en ly a P o esso wi h he Depa men o CSE, School o Compu ing and I.T., Mani- pal Uni e si y Jaipu , Jaipu . He has mo e han 12 yea s o ich expe ience in academics and one yea in indus y. He has mo e han mo e han 30 publica ions in in e na ional/na ional jou - nals/con e ence p oceedings. His esea ch in e - es s include machine lea ning, so compu ing, algo i hms, and a i icial in elligence. He is asso- cia ed wi h machine lea ning o mo e han se en yea s. He is cu en ly wo king in he a ea o applica ion o machine/deep lea ning in na u al language p ocessing like seman ic analysis and lexical analysis. He is also guiding ou Ph.D. s uden s in he a ea o NLP, in usion de ec ion, and ood adul e a ion using AI echniques. He is also an ac i e membe o special in e es g oup and ini ia i e by MIR labs o connec he esea che s and p o essional ac oss he globe. He is a Senio Membe o LMCSI and MACM, and a membe o Machine In elligence Resea ch Labs, USA. He is also membe o e iewe boa d o a ious jou nals and echnical p og am commi ee o se e al epu ed con e ences. PRASUN CHAKRABARTI (Senio Membe , IEEE) ecei ed he Ph.D. (Engg.) deg ee om Jada pu Uni e si y, in 2009. He is cu en ly an Execu i e Dean (Resea ch and In e na ional Link- age) and also an Ins i u e Dis inguished Senio Chai P o esso wi h he Depa men o Compu e Science and Enginee ing, Techno India NJR Ins i- u e o Technology. He has se e al publica ions, books and 31 iled Indian pa en s in his c edi . He has supe ised en Ph.D. candida es success ully. On a ious esea ch assignmen s, he has isi ed Waseda Uni e si y, Japan, (2012 a ailing p es igious INSA-CICS a el g an ), Uni e si y o Mau i ius, in 2015, Nanyang Technological Uni e si y Singapo e, in 2015, 2016, and 2019, Lincoln Uni e si y College Malaysia, in 2018, he Na ional Uni e si y o Singapo e, in 2019, Asian Ins i u e o Technology, Bangkok, Thailand, in 2019, and ISI Delhi, in 2019. He is a Fellow o IETE, ISRD, U.K., IAER, London, AE(I), and CET(I). GAURAV KUMAWAT ecei ed he M.Tech. deg ee om Paci ic Uni e si y. He is cu en ly pu su- ing he Ph.D. deg ee om he Depa men o Compu e Science and Enginee ing, Manipal Uni- e si y Jaipu , Jaipu . He has a o al expe ience o 14 Yea s as a Facul y wi h he CSE Depa men . His esea ch in e es s include C p og amming, objec -o ien ed p og amming using C++, co e ja a, ad ance ja a, unix p og amming, web p o- g amming, ope a ing sys ems, compu e a chi ec- u e, in o ma ion secu i y, p og amming analysis, and da a analysis. TULIKA CHAKRABARTI ecei ed he Ph.D. (Sc.) deg ee om he Indian Ins i u e o Chemical Biol- ogy, Jada pu Uni e si y, in 2013. She is cu - en ly an Assis an P o esso (Senio G ade) wi h Si Padampa Singhania Uni e si y, Udaipu . She has se e al publica ions, books and iled pa en s o he c edi . She is a na ional me i schola - ship holde in bo h 10 h and 12 h g ade. She has isi ed NUS, NTU, Lincoln Uni e si y College, Malaysia, and AIT, Bangkok, on se e al academic assignmen s. VOLUME 9, 2021 17333 P. Chi o a e al.: P edic ion o Ch onic Kidney Disease ZBIGNIEW LEONOWICZ (Senio Membe , IEEE) ecei ed he M.S. and Ph.D. deg ees in elec- ical enginee ing om he W oclaw Uni e si y o Science and Technology, in 1997 and 2001, espec i ely, and he Habili a ion deg ee om he Bialys ok Uni e si y o Technology, in 2012. Since 1997, he has been wi h he Elec ical Engi- nee ing Facul y, W oclaw Uni e si y o Technol- ogy. He also ecei ed he wo i les o a Full P o esso om he P esiden o Poland, in 2019, and he P esiden o he Czech Republic. Since 2019, he has been a P o esso wi h he Depa men o Elec ical Enginee ing, whe e he is cu en ly he Head o he Chai o elec ical enginee ing undamen als. MICHAŁ JASIŃSKI (Membe , IEEE) ecei ed he M.S. and Ph.D. deg ees in elec ical enginee - ing om he W oclaw Uni e si y o Science and Technology, in 2016 and 2019, espec i ely. Since 2018, he has been wi h he Elec ical Enginee - ing Facul y, W oclaw Uni e si y o Technology, whe e he is cu en ly an Assis an P o esso . He has au ho ed o coau ho ed mo e han 60 scien i ic publica ions. His esea ch in e es s include using big da a in powe sys em especially in poin o powe quali y. ŁUKASZ JASIŃSKI ecei ed he deg ee in elec- ical enginee ing om he W oclaw Uni e si y o Technology, Poland, and he Ph.D. deg ee in dyscypine o au oma ion, elec onics and elec ical enginee ing om he W oclaw Uni e si y o Tech- nology. In 2020, he s a ed wo king as a designe o elec ical ins alla ions in he design o ice. His esea ch in e es includes big da a in powe sys- ems, especially in poin o powe quali y. RADOMIR GONO (Senio Membe , IEEE) ecei ed he M.Sc., Ph.D., Habili a e Ph.D., and P o esso deg ees in elec ical powe enginee - ing in 1995, 2000, 2008, and 2019, espec i ely. Since 1999, he has been wi h he Depa men o Elec ical Powe Enginee ing, VSB–Technical Uni e si y o Os a a, Czech Republic, whe e he is cu en ly a P o esso and he Vice Head o he depa men . His cu en esea ch in e es s include elec ic powe sys ems eliabili y, he op imiza ion o main enance, and enewable ene gy sou ces. ELŻBIETA JASIŃSKA ecei ed he Ph.D. deg ee om Poznań Uni e si y Technology, in 2013. Since 2019, she has been wi h he Facul y o Law, Adminis a ion and Economics, Uni e si y o W oclaw, whe e she is cu en ly an Assis an P o esso . She has au ho ed o coau ho ed mo e han 70 scien i ic publica ions. He esea ch in e - es s include sus ainable de elopmen , co po a e socially esponsible, enewable ene gy esou ces, and i ual powe plan s. VADIM BOLSHEV ecei ed he M.S. deg ee in elec ical enginee ing om O el S a e Ag a ian Uni e si y, in 2012, and he Ph.D. deg ee in elec- ical enginee ing in 2020. He gained expe ience in he indus y as an elec ician, an elec ical engi- nee , a chie enginee om 2010 o 2018. Since 2018, he has been a Resea che wi h he Labo- a o y o Powe and Hea Supply, Fede al Scien- i ic Ag oenginee ing Cen e VIM. His scien i ic ac i i y is o de elop me hods and ools aimed a imp o ing powe supply e iciency including he de elopmen o me hods and de ices o moni o ing powe quali y and he echnical s a e o powe supply sys em elemen s. 17334 VOLUME 9, 2021