Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Intelligent Systems Developed for the Early Detection of Chronic Kidney Disease

Intelligent Systems Developed for the Early Detection of Chronic Kidney Disease HindawiPublishingCorporation AdvancesinArti�cialNeural�ystems Volume2013,ArticleID539570,7pages http://dx.doi.org/10.1155/2013/539570 ResearchArticle IntelligentSystemsDevelopedfortheEarlyDetectionof ChronicKidneyDisease 1 1 1 2 3 RueyKeiChiu, ReneeY.Chen, Shin-AnWang, Yen-ChunChang, andLi-ChienChen DepartmentofInformationManagement,FuJenCatholicUniversity,XinzhuangDistrict,NewTaipeiCity24205,Taiwan OfficeofComputerProcessing,EnChuKongHospital,SanxiaDistrict,NewTaipeiCity23702,Taiwan OfficeofInformationProcessing,CardinalTienHospital,XindianDistrict,NewTaipeiCity231,Taiwan CorrespondenceshouldbeaddressedtoRueyKeiChiu;rkchiu@mail.u.edu.tw Received10August2012;Revised5November2012;Accepted5November2012 AcademicEditor:PingFengPai Copyright©2013RueyKeiChiuetal.isisanopenaccessarticledistributedundertheCreativeCommonsAttributionLicense, whichpermitsunrestricteduse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited. is paper aims to construct intelligence models by applying the technologies of arti�cial neural networks including back- propagation network (BPN), generalized feedforward neural networks (GRNN), and modular neural network (MNN) that are developed, respectively, for the early detection of chronic kidney disease (CKD). e comparison of accuracy, sensitivity, and speci�cityamongthreemodelsissubse�uentlyperformed.emodelofbestperformanceischosen.Byleveragingtheaidofthis system,CKDphysicianscanhaveanalternativewaytodetectchronickidneydiseasesinearlystageofapatient.Meanwhile,itmay alsobeusedbythepublicforself-detectingtheriskofcontractingCKD. 1.Introduction In this paper, we aim to develop a feasible intelligent model for detecting CKD for evaluating the severity of a According to the statistical data announced by the Depart- patientwithorwithoutCKD.einputdataformodeldevel- ment of Health of Taiwan’s government in 2010 [1], the opmentandtestingiscollectedfromthehealthexamination mortality caused by kidney disease has been ranked in the whichisperiodicallycarriedoutbythecollaborativeteaching 10th place in all causes of death in Taiwan and thousands of hospitalofthisresearch. othersareatincreasedrisk.emortalitycausedfromkidney disease is estimated as 12.5 in every 100,000 people. As a 2.TheMajorMethodsforMeasuring result,itcostsashighas35percentofhealthinsurancebudget ChronicKidneyDisease to treat the chronic kidney disease (CKD) patients with the ageover65yearsoldandend-stagekidneydiseasepatientsin As it is mentioned in prior section, the GFR is the most allages.Itoccupiesahugeamountofexpendituresinnational common method used to measure kidney health function. insurancebudget. It refers to the water �lterability of glomerular of peo- ple’s kidney. e normal value should be between 90 and Regarding the measurement of serious levels of CKD, presently glomerular �ltration rate (GFR) is the most com- 120mL/min/1.73m (i.e., measured by mL per minute per monly measuring indicator used in health institutions to 1.73m ). ere are three common computation methods of estimate kidney health function. e physician in the health GFR,whichare(1)removingrateof24-hoururinecreatinine institutioncancalculateGFRfrompatient’sbloodcreatinine, (i.e., Creatinine Clearance Rate, Ccr), (2) Cockcro-Gault age,race,gender,andotherfactorsdependinguponthetype formula (also known as C-G formula), and (3) Modi�cation offormal-recognizedcomputationformulas[2,3]employed. of Diet in Renal Disease formula (also known as MDRD) eGFRmayindicatethehealthofapatent’skidneyandcan [4]. e CKD is categorized into �ve stages by making use also be taken to determine the stage of severity of a patient of GFR to measure kidney function. Although the course of withorwithoutkidneydisease. thechangefromstageoneto�vemayusuallylastforyears,it 2 AdvancesinArti�cialNeuralSystems T1:Classi�cationofCKDde�nedbyKDOQI. the desired (targeted) value to develop our neural network models.Onehas Classi�cation Stage Stateofkidneyfunction byseverityby mL/min −1.154 −0.203 GFR GFR = 186×creatinine ×age 󶀥󶀥 󶀵󶀵 . (1) 1.73m Kidneydamagewithnormalor GFR≥90 increasinginGFR Note:Forfemaletheresultshouldbemultipliedbyafactorof Kidneydamagewithmilddecreasingin GFRof60–89 GFR 0.742. 3 Itisspecially worthyofnote thattherecent studyfrom ModeratewithdecreasinginGFR GFRof30–59 Matsushitaetal.[7]indicatesthatalthoughtheModi�cation SeverewithdecreasinginGFR GFRof15–29 of Diet in Renal Disease (MDRD) study equation is recom- GFR<15(or Kidneyfailure mended for estimating GFR, the Chronic Kidney Disease dialysis) Epidemiology Collaboration (CKD-EPI) has proposed an alternativeequation,whichisknownasCKD-EPI.eCKD- EPI applies different coefficients to the same 4 variables sometimes may enter into ��h stage pretty soon resulted in which include age, sex, race, and serum creatinine level, thenecessityofdialysisorkidneytransplant. used in the MDRD study equation. e study takes the Again, National Kidney Foundation’s Kidney Disease data from more than one million participant cases residing Outcomes Quality Initiative (KDOQI) [3] provides a in 40 countries or regions. ey �nd that the CKD-EPI conceptual framework for the diagnosis of the severity equation estimates measured GFR more accurately than the stages of CKD based on the different function levels MDRD study equation in most of the study areas. It shows of glomerular �ltration rate (GFR). e new system approximately one-fourth of cases were reclassi�ed to a represented a signi�cant conceptual change, since higher estimated GFR category by the CKD-EPI equation kidney disease historically had been categorized mainly comparedwiththeMDRDstudyequation.Inthisone-fourth bycauses.ediagnosisofCKDreliesonmarkersofkidney of cases are reclassi�ed upward in GFR �gure by CKD-EPI, damage and/or a reduction in GFR. Stages 1 and 2 de�ne 24.4% in the general population cohorts, 15.4% in the high conditions of kidney damage in the presence of a GFR of riskcohorts,and6.6%intheCKDcohorts.isimprovement 2 2 at least 90mL/min/1.73m or 60 to 89mL/min/1.73m , byCKD-EPIclassi�cationmaylowertheprevalenceofCKD. respectively,andstages3to5de�neconditionsofmoderately Participantcaseswhoarereclassi�edupwardhadlowerrisks and severely reduced GFR irrespective of markers of kidney of mortality and end-stage renal disease (ESRD) compared damage. e summary of this guideline is shown in Table withthosenotreclassi�ed[7]. 1 [3]. However, Levey et al. [5] especially mentioned that although this guideline was endorsed by the Kidney Disease: Improving Global Outcomes (KDIGO) in 2004 S"UJêDJBMVSFBM/PSLFUX/ and this framework was constantly promoted to increase the attention to chronic kidney disease in clinical practice, e arti�cial neural network (ANN), usually simply called research, and public health, it had also generated debate. It neural network (NN), is a mathematical model or com- is the position of KDIGO and KDOQI that the de�nition putational model that is inspired by the structure and/or and classi�cation should re�ect patient prognosis and functional aspects of biological neural networks. Kriesel [6] that an analysis of outcomes would answer key questions indicated neural networks were a bioinspired mechanism of underlying the debate. e common de�nition of CKD has data processing that enables computers to learn technically facilitated comparisons between studies. Nevertheless, there similar to human-being brain. A neural network consists of are limitations to this classi�cation system, which is by its aninterconnectedgroupofarti�cialneurons,anditprocesses naturesimpleandnecessarilyarbitraryintermsofspecifying informationusingaconnectionistapproachtocomputation. the thresholds for de�nition and different stages. When the Inmostcases,anANNisanadaptivesystemthatchangesits classi�cationsystemwasdevelopedin2002[3],theevidence structure based on external (input) or internal information base used for the development of this guideline was much that�owsthroughthenetworkduringthelearningphase.It smaller than the CKD evidence base today. erefore, this is properly the most prestigious and adoptable model in all guidelinehasbeenconstantlyrevisedfromthenon. applicationmodelsinthe�eldofarti�cialintelligence. InTaiwan,theTaiwanSocietyofNephrology(TSN)also ere are many different types of neural network mod- presentedtheself-detectingmethodofkidneyforthepublic. els derived from the generic structure of ANN. In this Atpresent,theMDRDformula[3,6]isrecognizedasamostly paper’sstudy,threeneuralnetworkmodelsemployedforthe common method adopted by kidney physicians to estimate experiment and comparison include the back-propagation GFR from serum creatinine level. erefore, in this paper neural network (BPNN) [6, 8], the generalized feedforward we take MDRD to calculate the GFR for the detection. e neural network (GFNN) [9, 10], and the modular neural method of computational formula is shown in formula (1). network(MNN)[11].egenericarchitecturesofthree-layer e input data for the GFR calculation of each individual back-propagation neural network, generalized feedforward caseinhealthexaminationareprovidedbythecollaborative neural network and modular neural network are illustrated teaching hospital. We also used the calculated results as inFigures1,2,and3,respectively. ··· ··· ··· ·· · · ·· · ·· AdvancesinArti�cialNeuralSystems 3 Expert 1 Input signals ॶ Expert 2 ॷ ॵ ॵ २ २३ ॶ १२ ३ ॷ &YQFSU ॗ ० ० Gate Input Hidden Output layer layer layer F3:Agenericarchitectureofmodularneuralnetwork. Error signals T 2: e in�uence factors derived from the computation F 1: A generic model of three-layer back-propagation neural formulaforGFR. network. Factors Mean Max. Mini. Description Generalized 0→Female shunting Output neurons Sex — —— Input neurons 1→Male neurons (perceptrons) Age Integer 77.38 93 67 Creatinine 0.88 4.28 0.37 Real Weight 60.719 98 32 Real 0→Negative GFRlevels — —— 1,2,3,4,5→Positive ड़ ै ५ measuredinlevels २५ ै २५ einputdatasetfordevelopingneuralnetworksarecol- lected from the cases of health examination provided by the collaborative hospital of this study. 1161 heath examination cases covering past few years are selected. Before they are ३२ inputfortrainingandtestingnetworkmodels,adataprepro- cessing is processed to remove duplication and correct the ड़ ै २१ २१ error, inconsistency, and missing �elds in each case record. Among these cases, there exit some unknown duplication cases which have unknown causes, they are identi�ed and Excitatory synapses removed. Some cases have error in certain �elds such as the Inhibitory synapses �gurebeingbeyonditsreasonablerangeorina�uestionable F2:Agenericarchitectureofgeneralizedfeedforwardneural high or low level. is is oen seen in the �gures for some network. physiological test. We try to con�rm these �gures with the authority of health examination center of hospital and try to correct it, otherwise they are removed. Some cases show 4.ResearchMaterialsandMethods postal code error or appear inconsistent with the personal Based on the literature reviews and expert interviews, this contact address or vice versa, using different representation paper takes three neural network models including back- toreferthesamemeaning,whichiscommonlyseeningender propagation neural network, feedforward neural network, �led and name �eld. More signi�cantly, because not every and modular neural network, respectively, to generate the subject conducts a complete health examination, certain detectionmodelforchronickidneydisease.ekeyin�uence �elds used for risk measurement are missing. ese cases input factors for each model are determined and veri�ed are simply removedas wellbecause theycan not beused from professional experts and physicians of kidney diseases. for model development. By this process, we can ensure the e set of in�uence factors derived from GFR computation accuracy,completeness,andintegrityofinputdata.Aerdata preprocessing, only most accurate 430 patient cases remain formulasisshowninTable2andanothersetofkeyin�uence factors selected and determined in this paper as the input forthedevelopmentofintelligentmodels.Amongthesecases, of neural networks is shown in Table 3, respectively. e 145 cases are prediagnosed as negative with CKD, and 285 classi�cation performances of two input sets for model cases are prediagnosed as positive. e details are shown in developmentarecompared. Table4. 4 AdvancesinArti�cialNeuralSystems T3:ein�uencefactorsfortheclassi�cationofneuralnetworksforC�Ddetection. Factors Description Mean Max. Min. Creatinine Real 0.88 4.28 0.37 Glucose(GLU) Real 109.63 271 82 Systolicpressure(SP) Real 135.78 195 88 0→Negative 1→Trace Proteininurine(UP)orcalledproteinuria 2→+ — — — 3→++ 4→+++ Hematuria — — — Bloodureanitrogen(BUN) Real 15.80 80.1 5.4 0→Negative GFR-level — — — 1,2,3,4,5→Positivemeasuredinlevels T 4: Data distribution for model training, testing, and cross- T 5: e best settings of network parameters for back- veri�cation. propagationnetworkmodel. Numberof ModelFA ModelFB Networkparameters Inputdataformodeling Numberofrecords desired Learningrule Levenberg-MarquarLevenbergMarquar output TanhAxon SigmoidAxon Transformationfunctions 101Negative 2 2 No.ofhidinglayer 199Positive Trainingdata 300 invarious No.ofneuronsinhiding 5 8 levels layer MSE :0.0001 MSE :0.0001 34Negative min min Testingdata 100 Criterionfortermination 66Positive Epochs :3000 Epochs :3000 max max 10Negative Batch Batch Methodsforweightupdate Crossveri�cation 30 20Positive 5.1. e Model Development of Back-Propagation Neural Network. rough a series of training and testing with the 5.ModelDevelopment differentsetsofparametercombinationselected,wegainthe edevelopmentforaspeci�cnetworkmodelcanbeviewed best model for back-propagation neural network in terms as a series of modeling and simulation. We attempt to of its respective network parameters with respect to two pursuethefeasiblemodelthroughtheexperimentofdifferent different sets of input factors. One set is by adopting the key combinations of modeling parameters selected for different factorsusedincomputationformula(simplycalledModelPA network models. e parameters selected for experiment hereaer).eothersetisbyadoptingthekeyfactorsselected include the learning rules, the transform functions, the in this paper (simply called Model FB hereaer). Table 5 number of hiding layers, the number of neurons in each showsthesetwobestsettingsofparameters.eclassi�cation hidden layer, the weight update methods, and the criteria accuracy, sensitivity, and speci�city with respect to Model for the determination in model training and testing. e FA and FB, respectively, are shown in Table 6. In Table 6, well-known NBuilder of neural solution is the tool adopted themetricof“accuracy”isusedtomeasuretheclassi�cation for model development. Among 430 cases, 300 cases are of accuracy with the proportion of the sum of the number selected for training, 100 cases are used for testing, and of true positives and the number of true negatives. e the rest of 30 cases are taken for cross-veri�cation during metric of “sensitivity” is used to measure the proportion of the model development. is data distribution is shown in actual positives which are correctly classi�ed as such. e Table 4. e desired output of classi�cation for each case is metric of “speci�city” is used to measure the proportion of formallyidenti�edbeforehandbytheprofessionalphysicians negatives which are correctly identi�ed. From the results of chronic kidney disease from the collaborative teaching shown in Table 6, a pretty good classi�cation result is hospital by using GFR computation formula. e hybrid gained with BPN both for Model FA and FB, while it shows model of combining each individual neural network and a signi�cant drop in classi�cation performance in model genetic algorithm (GA) is also conducted in this research. testing stage both in BPN and BPN plus GA. By the results In other words, GA will be taken to combine with back- shown in Table 6, pure BPN model may gain better results propagation neural network, feedforward neural network, in accuracy measure, while BPN plus GA may gain better and modular neural network, respectively, in model devel- resultsinsensitivitymeasure.eresultalsoshowsthemodel opmentandcomparisonaswell. with the adoption of the key factors used in computation AdvancesinArti�cialNeuralSystems 5 T6:eperformancegainedfromback-propagationneuralnetworkmodels. BPN ModelFA ModelFB PureBPN Training Test Training Test Accuracyofclassi�cation 100% 94.75% 100% 72.42% Sensitivity 100% 97.06% 100% 70.59% Speci�city 100% 92.42% 100% 74.24% BPN+GA Test Test Accuracyofclassi�cation 91.71% 77.68% Sensitivity 97.06% 81.82% Speci�city 86.36% 73.53% T7:ebestsettingsofnetworkparametersforgeneralizedfeedforwardneuralnetwork. Networkparameters ModelFA ModelFB Learningrule LevenbergMarquar LevenbergMarquar Transformationfunctions SigmoidAxon TanhAxon Numberofhidinglayer 2 2 Numberofneuronsinhidinglayer 8 8 MSE :0.0001 MSE :0.0001 min min Criterionfortermination Epochs :3000 Epochs :3000 max max Methodsforweightupdate Batch Batch formula gains better results in all three measures. ese inTables11and12,respectively.Table11showsthedetection results show that a hybrid model with the combination of accuracyinthreepureneuralnetworkmodelswhileTable12 BPNandGAseamsdoesnotimprovethemodelperformance shows the detection accuracy with GA model embedded in in accuracy measure but it is helpful to improve sensitivity three respective pure models. We found BPN may gain the measure. best accuracy regardless of in both Model FA and Model FB which are measured with 94.75% and 72.42% accuracy 5.2. e Model Development of Generalized Feedforward shown in Table 11, respectively in model testing. However, NeuralNetwork. Table7showsthebestnetworkparameters ModelFA,whichisthetestbyadoptingthekeyfactorsused settingsforthedevelopmentofgeneralizedfeedforwardneu- incomputationformula,maygainmuchbetterperformance ral network (GFNN�. e classi�cation accuracy, sensitivity, than Model FB, which is the test by adopting the key factors andspeci�citywithrespecttoModelFAandFB,respectively, selectedinthispaper.esameresultisgainedwiththeGA are shown in Table 8. As the results in Table 8 show, pure embedded in each fundamental model in test. As the results GFNN obtains a perfect classi�cation percentage in three observedfromTable12,bothBPNplusGAandGFNNplus measures including accuracy, sensitivity, and speci�city in GA gain close results in accuracy measure which is much training stage while GFNN plus GA apparently may gain betterthanMNNplusGA. better results in testing. However, they also show the results By further observation from the results of model test gained from GFNN are not so good in all three measures as shown in Section 4, it is found that almost all models thosegainedfromBPNandBPNplusGA,respectively. employed in test may gain near 100% accuracy in CKD detection in the training stage regardless of which sets of 5.3. e Model Development of Modular Neural Network. in�uencefactorsusedinmodeltraining.However,iffurther Table 9 shows the parameters for the test of modular neural modeltestingisconducted,itisfoundthatthenetwork network (MNN�. e classi�cation accuracy, sensitivity, and models with the input of in�uence factors of CKD used by speci�city with respect to Model FA and FB, respectively, physicians employed in the computational formula always are shown in Table 10. As the testing results in Table 10 show better detection performance in all three aspects of show,pureMNNisthesameaspriortwomodelswhichmay measure including accuracy, sensitivity, and speci�city mea- obtainaperfectclassi�cationpercentageinthreemeasuresin sures.AswecanseefromTable11,theBPNgainsthehighest trainingstage,butMNNplusGAapparentlymaygainworse 94.75% accuracy measure in the testing stage among three results in testing stage, which is different from the results fundamentalneuralnetworkmodelswhileGFNNgainsonly showninpriortwomodels. 86.63%inaccuracymeasurewhichisthelowestperformance inthreemodels. rough further observations from test results as it is 6.PerformanceComparisonofModels shown in Table 12, we �nd the hybrid network model of eperformancesofthreeneuralnetworkmodelsdeveloped GFNNwithGAembeddedmaysigni�cantlyshowimprove- mentindetectionperformanceinallthreemeasuresfromits for comparison in this paper in terms of detection (i.e., classi�cation�accuracymeasurearesummarizedandshown fundamentalmodelinthetestingstagealthoughthereversed 6 AdvancesinArti�cialNeuralSystems T8:eperformancegainedfromgeneralizedfeedforwardneuralnetworkmodels. GFNN ModelFA ModelFB PureGFNN Training Test Training Test Accuracyofclassi�cation 100% 86.63% 100% 86.63% Sensitivity 100% 82.35% 100% 82.35% Speci�city 100% 90.91% 100% 90.91% GFNN+GA Test Test Accuracyofclassi�cation 91.09% 78.39% Sensitivity 88.24% 76.47% Speci�city 93.94% 80.30% T9:ebestsettingsofnetworkparametersformodularneuralnetwork. Networkparameters ModelFA ModelFB Learningrule LevenbergMarquar LevenbergMarquar Transformationfunctions SigmoidAxon SigmoidAxon Numberofhidinglayer 2 2 Numberofneuronsinhidinglayer 4 8 MSE :0.0001 MSE :0.0001 min min Criterionfortermination Epochs :3000 Epochs :3000 max max Methodsforweightupdate Batch Batch T10:eperformancegainedfrommodularneuralnetworkmodels. MNN ModelFA ModelFB PureMNN Training Test PureMNN Training Accuracyofclassi�cation 100% 93.23% Accuracyofclassi�cation 100% Sensitivity 100% 97.06% Sensitivity 100% Speci�city 100% 89.39% Speci�city 100% MNN+GA Test Test Accuracyofclassi�cation 88.82% 78.34% Sensitivity 89.39% 79.41% Speci�city 88.24% 77.27% T11:Detectionaccuracyinthreepureneuralnetworkmodels. Byadoptingthekeyfactorsusedincomputationformula Networkmodels Byadoptingthekeyfactorsselectedinthispaper(ModelFB) (ModelFA) BPN 94.75% 72.42% GFNN 86.63% 71.03% MNN 93.23% 70.99% enhancementisshowninBPNandMNN.Again,bythetest ModelFAamongthreefundamentalneuralnetworkmodels �gures shown in Section 4, although pure neural network employed in detecting CKD while the models BPN and models with GA show degraded performance improvement GFNN with GA embedded, respectively, might be the best for Model FA in testing stage in terms of three measures hybrid models for CKD detection but only GFNN plus GA in accuracy, sensitivity, and speci�city, it shows signi�cant cangainenhancementindetectionperformanceinbothsets improvement for Model FB. As a result, we conclude a of in�uence factors as input. ese conclusions should be hybrid model indeed may improve detection performance. furtherveri�edandcomparedwithothermodelsselectedfor is result may conclude that GA provides no bene�t in the testinlaterstudiesbeforetheycanbeassured. yield of better detection performance with the adoption of the key factors used in computation formula as the input to 7.Conclusion allmodelsintesting. AccordingtothedetectionperformanceshowninTables WeconcludethatneuralnetworkmodelsdevelopedforCKD 11 and 12, we conclude that BPN might be the best��tting detectionmayeffectivelyandfeasiblyequipmedicalstaffwith AdvancesinArti�cialNeuralSystems 7 T12:Detectionaccuracyinthreeneuralnetworkmodelswith [10] G. Arulampalam and A. Bouzerdoum, “A generalized feedfor- GA. ward neural network classi�er,” inProceedings of the Interna- tionalJointConferenceonNeuralNetworks,pp.1429–1434,July Networkmodels ModelFA ModelFB BPN+GA 91.71% 77.68% [11] R. Jacobs, M. Jordan, S. Nowlan, and G. Hinton, “Adaptive GFNN+GA 91.09% 78.39% mixturesoflocalexperts,”NeuralComputation,vol.3,no.1,pp. MNN+GA 88.82% 78.34% 79–87,1991. the ability to make precise diagnosis and treatment to the patient. In the future study, further model modi�cation and testing for the intelligence models developed in this paper should be conducted to enhance the accuracy in detection performance and to ensure they are sufficiently good for being truly employed in medical practice. In the meantime, different intelligence models can be widely and persistently applied for system development in this domain application aswellinordertosearchforabestonemodeltobeadopted. Inthefuturesystemdevelopment,itisworthwhiletodeploy thesystemtothecloudplatformsothatthepublicuserscan alsousethissystemtoconductaself-detectionofhavinghad CKD. References [1] Department of Health (DOH), “e statistical data of the causes of civilian mortality,” Taiwan, 2010, http://www .doh.gov.tw/CHT2006/DM/DM2_2.aspx?now_fod_list_no= 11122&class_no=440&level_no=3. [2] A.S.Levey,P.E.DeJong,J.Coreshetal.,“ede�nition,clas- si�cation, and prognosis of chronic kidney disease: a KDIGO controversies conference report,” Kidney International, vol. 80, no.1,pp.17–28,2011. [3] NationalKidneyFoundation(NKF),“KDOQIclinicalpractice guidelinesforchronickidneydisease:evaluation,classi�cation andstrati�cation,”AmericanJournalofKidneyDiseases,vol.39, pp.S1–S266,2002. [4] A. S. Levey, T. Greene, J. W. Kusek, and G. L. Beck, “A simpli�ed equation to predict glomerular �ltration rate from serumcreatinine,”JournalAmericanSocietyforNephrology,vol. 11,article155A,2000. [5] A.S.Levey,K.�.Eckardt,Y.Tsukamotoetal.,“De�nitionand classi�cation of chronic kidney disease: a position statement fromKIidneyDisease:ImprovingGlobalOutcomes(KDIGO),” KidneyInternational,vol.67,no.6,pp.2089–2100,2005. [6] D. Kriesel, “A brief introduction to neural networks,” 2011, http://www.dkriesel.com. [7] K. Matsushita, B. K. Mahmoodi, M. Woodward et al., “Com- parisonofriskpredictionusingtheCKD-EPIequationandthe MDRDstudyequationforestimatedGlomerular�ltrationrate,” e Journal of the American Medical Association, vol. 307, no. 18,pp.1941–1951,2012. [8] R.K.Chiu,C.I.Huang,and Y. C.Chang, “e study onthe construction of intelligence-based decision support system for diabetes diagnosis and risk evaluation,” Journal of Medical Informatics,vol.19,no.3,pp.1–24,2010. [9] G. Arulampalam and A. Bouzerdoum, “Training shunting in- hibitoryarti�cialneuralnetworksasclassi�ers,”NeuralNetwork World,vol.10,no.3,pp.333–350,2000. Journal of Advances in Industrial Engineering Multimedia Applied Computational Intelligence and Soft Computing International Journal of The Scientific Distributed World Journal Sensor Networks Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 Advances in Fuzzy Systems Modelling & Simulation in Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Submit your manuscripts at http://www.hindawi.com Journal of Computer Networks and Communications  Advances in  Artic fi ial Intelligence Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 International Journal of Advances in Biomedical Imaging Artificial Neural Systems International Journal of Computer Games Advances in Advances in Computer Engineering Technology Software Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 International Journal of Reconfigurable Computing Computational Advances in Journal of Journal of Intelligence and Human-Computer Electrical and Computer Robotics Interaction Neuroscience Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Advances in Artificial Neural Systems Hindawi Publishing Corporation

Intelligent Systems Developed for the Early Detection of Chronic Kidney Disease

Download PDF
8 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/intelligent-systems-developed-for-the-early-detection-of-chronic-w4FuXzA71T
Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2013 Ruey Kei Chiu et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
1687-7594
DOI
10.1155/2013/539570
Publisher site
See Article on Publisher Site

Abstract

HindawiPublishingCorporation AdvancesinArti�cialNeural�ystems Volume2013,ArticleID539570,7pages http://dx.doi.org/10.1155/2013/539570 ResearchArticle IntelligentSystemsDevelopedfortheEarlyDetectionof ChronicKidneyDisease 1 1 1 2 3 RueyKeiChiu, ReneeY.Chen, Shin-AnWang, Yen-ChunChang, andLi-ChienChen DepartmentofInformationManagement,FuJenCatholicUniversity,XinzhuangDistrict,NewTaipeiCity24205,Taiwan OfficeofComputerProcessing,EnChuKongHospital,SanxiaDistrict,NewTaipeiCity23702,Taiwan OfficeofInformationProcessing,CardinalTienHospital,XindianDistrict,NewTaipeiCity231,Taiwan CorrespondenceshouldbeaddressedtoRueyKeiChiu;rkchiu@mail.u.edu.tw Received10August2012;Revised5November2012;Accepted5November2012 AcademicEditor:PingFengPai Copyright©2013RueyKeiChiuetal.isisanopenaccessarticledistributedundertheCreativeCommonsAttributionLicense, whichpermitsunrestricteduse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited. is paper aims to construct intelligence models by applying the technologies of arti�cial neural networks including back- propagation network (BPN), generalized feedforward neural networks (GRNN), and modular neural network (MNN) that are developed, respectively, for the early detection of chronic kidney disease (CKD). e comparison of accuracy, sensitivity, and speci�cityamongthreemodelsissubse�uentlyperformed.emodelofbestperformanceischosen.Byleveragingtheaidofthis system,CKDphysicianscanhaveanalternativewaytodetectchronickidneydiseasesinearlystageofapatient.Meanwhile,itmay alsobeusedbythepublicforself-detectingtheriskofcontractingCKD. 1.Introduction In this paper, we aim to develop a feasible intelligent model for detecting CKD for evaluating the severity of a According to the statistical data announced by the Depart- patientwithorwithoutCKD.einputdataformodeldevel- ment of Health of Taiwan’s government in 2010 [1], the opmentandtestingiscollectedfromthehealthexamination mortality caused by kidney disease has been ranked in the whichisperiodicallycarriedoutbythecollaborativeteaching 10th place in all causes of death in Taiwan and thousands of hospitalofthisresearch. othersareatincreasedrisk.emortalitycausedfromkidney disease is estimated as 12.5 in every 100,000 people. As a 2.TheMajorMethodsforMeasuring result,itcostsashighas35percentofhealthinsurancebudget ChronicKidneyDisease to treat the chronic kidney disease (CKD) patients with the ageover65yearsoldandend-stagekidneydiseasepatientsin As it is mentioned in prior section, the GFR is the most allages.Itoccupiesahugeamountofexpendituresinnational common method used to measure kidney health function. insurancebudget. It refers to the water �lterability of glomerular of peo- ple’s kidney. e normal value should be between 90 and Regarding the measurement of serious levels of CKD, presently glomerular �ltration rate (GFR) is the most com- 120mL/min/1.73m (i.e., measured by mL per minute per monly measuring indicator used in health institutions to 1.73m ). ere are three common computation methods of estimate kidney health function. e physician in the health GFR,whichare(1)removingrateof24-hoururinecreatinine institutioncancalculateGFRfrompatient’sbloodcreatinine, (i.e., Creatinine Clearance Rate, Ccr), (2) Cockcro-Gault age,race,gender,andotherfactorsdependinguponthetype formula (also known as C-G formula), and (3) Modi�cation offormal-recognizedcomputationformulas[2,3]employed. of Diet in Renal Disease formula (also known as MDRD) eGFRmayindicatethehealthofapatent’skidneyandcan [4]. e CKD is categorized into �ve stages by making use also be taken to determine the stage of severity of a patient of GFR to measure kidney function. Although the course of withorwithoutkidneydisease. thechangefromstageoneto�vemayusuallylastforyears,it 2 AdvancesinArti�cialNeuralSystems T1:Classi�cationofCKDde�nedbyKDOQI. the desired (targeted) value to develop our neural network models.Onehas Classi�cation Stage Stateofkidneyfunction byseverityby mL/min −1.154 −0.203 GFR GFR = 186×creatinine ×age 󶀥󶀥 󶀵󶀵 . (1) 1.73m Kidneydamagewithnormalor GFR≥90 increasinginGFR Note:Forfemaletheresultshouldbemultipliedbyafactorof Kidneydamagewithmilddecreasingin GFRof60–89 GFR 0.742. 3 Itisspecially worthyofnote thattherecent studyfrom ModeratewithdecreasinginGFR GFRof30–59 Matsushitaetal.[7]indicatesthatalthoughtheModi�cation SeverewithdecreasinginGFR GFRof15–29 of Diet in Renal Disease (MDRD) study equation is recom- GFR<15(or Kidneyfailure mended for estimating GFR, the Chronic Kidney Disease dialysis) Epidemiology Collaboration (CKD-EPI) has proposed an alternativeequation,whichisknownasCKD-EPI.eCKD- EPI applies different coefficients to the same 4 variables sometimes may enter into ��h stage pretty soon resulted in which include age, sex, race, and serum creatinine level, thenecessityofdialysisorkidneytransplant. used in the MDRD study equation. e study takes the Again, National Kidney Foundation’s Kidney Disease data from more than one million participant cases residing Outcomes Quality Initiative (KDOQI) [3] provides a in 40 countries or regions. ey �nd that the CKD-EPI conceptual framework for the diagnosis of the severity equation estimates measured GFR more accurately than the stages of CKD based on the different function levels MDRD study equation in most of the study areas. It shows of glomerular �ltration rate (GFR). e new system approximately one-fourth of cases were reclassi�ed to a represented a signi�cant conceptual change, since higher estimated GFR category by the CKD-EPI equation kidney disease historically had been categorized mainly comparedwiththeMDRDstudyequation.Inthisone-fourth bycauses.ediagnosisofCKDreliesonmarkersofkidney of cases are reclassi�ed upward in GFR �gure by CKD-EPI, damage and/or a reduction in GFR. Stages 1 and 2 de�ne 24.4% in the general population cohorts, 15.4% in the high conditions of kidney damage in the presence of a GFR of riskcohorts,and6.6%intheCKDcohorts.isimprovement 2 2 at least 90mL/min/1.73m or 60 to 89mL/min/1.73m , byCKD-EPIclassi�cationmaylowertheprevalenceofCKD. respectively,andstages3to5de�neconditionsofmoderately Participantcaseswhoarereclassi�edupwardhadlowerrisks and severely reduced GFR irrespective of markers of kidney of mortality and end-stage renal disease (ESRD) compared damage. e summary of this guideline is shown in Table withthosenotreclassi�ed[7]. 1 [3]. However, Levey et al. [5] especially mentioned that although this guideline was endorsed by the Kidney Disease: Improving Global Outcomes (KDIGO) in 2004 S"UJêDJBMVSFBM/PSLFUX/ and this framework was constantly promoted to increase the attention to chronic kidney disease in clinical practice, e arti�cial neural network (ANN), usually simply called research, and public health, it had also generated debate. It neural network (NN), is a mathematical model or com- is the position of KDIGO and KDOQI that the de�nition putational model that is inspired by the structure and/or and classi�cation should re�ect patient prognosis and functional aspects of biological neural networks. Kriesel [6] that an analysis of outcomes would answer key questions indicated neural networks were a bioinspired mechanism of underlying the debate. e common de�nition of CKD has data processing that enables computers to learn technically facilitated comparisons between studies. Nevertheless, there similar to human-being brain. A neural network consists of are limitations to this classi�cation system, which is by its aninterconnectedgroupofarti�cialneurons,anditprocesses naturesimpleandnecessarilyarbitraryintermsofspecifying informationusingaconnectionistapproachtocomputation. the thresholds for de�nition and different stages. When the Inmostcases,anANNisanadaptivesystemthatchangesits classi�cationsystemwasdevelopedin2002[3],theevidence structure based on external (input) or internal information base used for the development of this guideline was much that�owsthroughthenetworkduringthelearningphase.It smaller than the CKD evidence base today. erefore, this is properly the most prestigious and adoptable model in all guidelinehasbeenconstantlyrevisedfromthenon. applicationmodelsinthe�eldofarti�cialintelligence. InTaiwan,theTaiwanSocietyofNephrology(TSN)also ere are many different types of neural network mod- presentedtheself-detectingmethodofkidneyforthepublic. els derived from the generic structure of ANN. In this Atpresent,theMDRDformula[3,6]isrecognizedasamostly paper’sstudy,threeneuralnetworkmodelsemployedforthe common method adopted by kidney physicians to estimate experiment and comparison include the back-propagation GFR from serum creatinine level. erefore, in this paper neural network (BPNN) [6, 8], the generalized feedforward we take MDRD to calculate the GFR for the detection. e neural network (GFNN) [9, 10], and the modular neural method of computational formula is shown in formula (1). network(MNN)[11].egenericarchitecturesofthree-layer e input data for the GFR calculation of each individual back-propagation neural network, generalized feedforward caseinhealthexaminationareprovidedbythecollaborative neural network and modular neural network are illustrated teaching hospital. We also used the calculated results as inFigures1,2,and3,respectively. ··· ··· ··· ·· · · ·· · ·· AdvancesinArti�cialNeuralSystems 3 Expert 1 Input signals ॶ Expert 2 ॷ ॵ ॵ २ २३ ॶ १२ ३ ॷ &YQFSU ॗ ० ० Gate Input Hidden Output layer layer layer F3:Agenericarchitectureofmodularneuralnetwork. Error signals T 2: e in�uence factors derived from the computation F 1: A generic model of three-layer back-propagation neural formulaforGFR. network. Factors Mean Max. Mini. Description Generalized 0→Female shunting Output neurons Sex — —— Input neurons 1→Male neurons (perceptrons) Age Integer 77.38 93 67 Creatinine 0.88 4.28 0.37 Real Weight 60.719 98 32 Real 0→Negative GFRlevels — —— 1,2,3,4,5→Positive ड़ ै ५ measuredinlevels २५ ै २५ einputdatasetfordevelopingneuralnetworksarecol- lected from the cases of health examination provided by the collaborative hospital of this study. 1161 heath examination cases covering past few years are selected. Before they are ३२ inputfortrainingandtestingnetworkmodels,adataprepro- cessing is processed to remove duplication and correct the ड़ ै २१ २१ error, inconsistency, and missing �elds in each case record. Among these cases, there exit some unknown duplication cases which have unknown causes, they are identi�ed and Excitatory synapses removed. Some cases have error in certain �elds such as the Inhibitory synapses �gurebeingbeyonditsreasonablerangeorina�uestionable F2:Agenericarchitectureofgeneralizedfeedforwardneural high or low level. is is oen seen in the �gures for some network. physiological test. We try to con�rm these �gures with the authority of health examination center of hospital and try to correct it, otherwise they are removed. Some cases show 4.ResearchMaterialsandMethods postal code error or appear inconsistent with the personal Based on the literature reviews and expert interviews, this contact address or vice versa, using different representation paper takes three neural network models including back- toreferthesamemeaning,whichiscommonlyseeningender propagation neural network, feedforward neural network, �led and name �eld. More signi�cantly, because not every and modular neural network, respectively, to generate the subject conducts a complete health examination, certain detectionmodelforchronickidneydisease.ekeyin�uence �elds used for risk measurement are missing. ese cases input factors for each model are determined and veri�ed are simply removedas wellbecause theycan not beused from professional experts and physicians of kidney diseases. for model development. By this process, we can ensure the e set of in�uence factors derived from GFR computation accuracy,completeness,andintegrityofinputdata.Aerdata preprocessing, only most accurate 430 patient cases remain formulasisshowninTable2andanothersetofkeyin�uence factors selected and determined in this paper as the input forthedevelopmentofintelligentmodels.Amongthesecases, of neural networks is shown in Table 3, respectively. e 145 cases are prediagnosed as negative with CKD, and 285 classi�cation performances of two input sets for model cases are prediagnosed as positive. e details are shown in developmentarecompared. Table4. 4 AdvancesinArti�cialNeuralSystems T3:ein�uencefactorsfortheclassi�cationofneuralnetworksforC�Ddetection. Factors Description Mean Max. Min. Creatinine Real 0.88 4.28 0.37 Glucose(GLU) Real 109.63 271 82 Systolicpressure(SP) Real 135.78 195 88 0→Negative 1→Trace Proteininurine(UP)orcalledproteinuria 2→+ — — — 3→++ 4→+++ Hematuria — — — Bloodureanitrogen(BUN) Real 15.80 80.1 5.4 0→Negative GFR-level — — — 1,2,3,4,5→Positivemeasuredinlevels T 4: Data distribution for model training, testing, and cross- T 5: e best settings of network parameters for back- veri�cation. propagationnetworkmodel. Numberof ModelFA ModelFB Networkparameters Inputdataformodeling Numberofrecords desired Learningrule Levenberg-MarquarLevenbergMarquar output TanhAxon SigmoidAxon Transformationfunctions 101Negative 2 2 No.ofhidinglayer 199Positive Trainingdata 300 invarious No.ofneuronsinhiding 5 8 levels layer MSE :0.0001 MSE :0.0001 34Negative min min Testingdata 100 Criterionfortermination 66Positive Epochs :3000 Epochs :3000 max max 10Negative Batch Batch Methodsforweightupdate Crossveri�cation 30 20Positive 5.1. e Model Development of Back-Propagation Neural Network. rough a series of training and testing with the 5.ModelDevelopment differentsetsofparametercombinationselected,wegainthe edevelopmentforaspeci�cnetworkmodelcanbeviewed best model for back-propagation neural network in terms as a series of modeling and simulation. We attempt to of its respective network parameters with respect to two pursuethefeasiblemodelthroughtheexperimentofdifferent different sets of input factors. One set is by adopting the key combinations of modeling parameters selected for different factorsusedincomputationformula(simplycalledModelPA network models. e parameters selected for experiment hereaer).eothersetisbyadoptingthekeyfactorsselected include the learning rules, the transform functions, the in this paper (simply called Model FB hereaer). Table 5 number of hiding layers, the number of neurons in each showsthesetwobestsettingsofparameters.eclassi�cation hidden layer, the weight update methods, and the criteria accuracy, sensitivity, and speci�city with respect to Model for the determination in model training and testing. e FA and FB, respectively, are shown in Table 6. In Table 6, well-known NBuilder of neural solution is the tool adopted themetricof“accuracy”isusedtomeasuretheclassi�cation for model development. Among 430 cases, 300 cases are of accuracy with the proportion of the sum of the number selected for training, 100 cases are used for testing, and of true positives and the number of true negatives. e the rest of 30 cases are taken for cross-veri�cation during metric of “sensitivity” is used to measure the proportion of the model development. is data distribution is shown in actual positives which are correctly classi�ed as such. e Table 4. e desired output of classi�cation for each case is metric of “speci�city” is used to measure the proportion of formallyidenti�edbeforehandbytheprofessionalphysicians negatives which are correctly identi�ed. From the results of chronic kidney disease from the collaborative teaching shown in Table 6, a pretty good classi�cation result is hospital by using GFR computation formula. e hybrid gained with BPN both for Model FA and FB, while it shows model of combining each individual neural network and a signi�cant drop in classi�cation performance in model genetic algorithm (GA) is also conducted in this research. testing stage both in BPN and BPN plus GA. By the results In other words, GA will be taken to combine with back- shown in Table 6, pure BPN model may gain better results propagation neural network, feedforward neural network, in accuracy measure, while BPN plus GA may gain better and modular neural network, respectively, in model devel- resultsinsensitivitymeasure.eresultalsoshowsthemodel opmentandcomparisonaswell. with the adoption of the key factors used in computation AdvancesinArti�cialNeuralSystems 5 T6:eperformancegainedfromback-propagationneuralnetworkmodels. BPN ModelFA ModelFB PureBPN Training Test Training Test Accuracyofclassi�cation 100% 94.75% 100% 72.42% Sensitivity 100% 97.06% 100% 70.59% Speci�city 100% 92.42% 100% 74.24% BPN+GA Test Test Accuracyofclassi�cation 91.71% 77.68% Sensitivity 97.06% 81.82% Speci�city 86.36% 73.53% T7:ebestsettingsofnetworkparametersforgeneralizedfeedforwardneuralnetwork. Networkparameters ModelFA ModelFB Learningrule LevenbergMarquar LevenbergMarquar Transformationfunctions SigmoidAxon TanhAxon Numberofhidinglayer 2 2 Numberofneuronsinhidinglayer 8 8 MSE :0.0001 MSE :0.0001 min min Criterionfortermination Epochs :3000 Epochs :3000 max max Methodsforweightupdate Batch Batch formula gains better results in all three measures. ese inTables11and12,respectively.Table11showsthedetection results show that a hybrid model with the combination of accuracyinthreepureneuralnetworkmodelswhileTable12 BPNandGAseamsdoesnotimprovethemodelperformance shows the detection accuracy with GA model embedded in in accuracy measure but it is helpful to improve sensitivity three respective pure models. We found BPN may gain the measure. best accuracy regardless of in both Model FA and Model FB which are measured with 94.75% and 72.42% accuracy 5.2. e Model Development of Generalized Feedforward shown in Table 11, respectively in model testing. However, NeuralNetwork. Table7showsthebestnetworkparameters ModelFA,whichisthetestbyadoptingthekeyfactorsused settingsforthedevelopmentofgeneralizedfeedforwardneu- incomputationformula,maygainmuchbetterperformance ral network (GFNN�. e classi�cation accuracy, sensitivity, than Model FB, which is the test by adopting the key factors andspeci�citywithrespecttoModelFAandFB,respectively, selectedinthispaper.esameresultisgainedwiththeGA are shown in Table 8. As the results in Table 8 show, pure embedded in each fundamental model in test. As the results GFNN obtains a perfect classi�cation percentage in three observedfromTable12,bothBPNplusGAandGFNNplus measures including accuracy, sensitivity, and speci�city in GA gain close results in accuracy measure which is much training stage while GFNN plus GA apparently may gain betterthanMNNplusGA. better results in testing. However, they also show the results By further observation from the results of model test gained from GFNN are not so good in all three measures as shown in Section 4, it is found that almost all models thosegainedfromBPNandBPNplusGA,respectively. employed in test may gain near 100% accuracy in CKD detection in the training stage regardless of which sets of 5.3. e Model Development of Modular Neural Network. in�uencefactorsusedinmodeltraining.However,iffurther Table 9 shows the parameters for the test of modular neural modeltestingisconducted,itisfoundthatthenetwork network (MNN�. e classi�cation accuracy, sensitivity, and models with the input of in�uence factors of CKD used by speci�city with respect to Model FA and FB, respectively, physicians employed in the computational formula always are shown in Table 10. As the testing results in Table 10 show better detection performance in all three aspects of show,pureMNNisthesameaspriortwomodelswhichmay measure including accuracy, sensitivity, and speci�city mea- obtainaperfectclassi�cationpercentageinthreemeasuresin sures.AswecanseefromTable11,theBPNgainsthehighest trainingstage,butMNNplusGAapparentlymaygainworse 94.75% accuracy measure in the testing stage among three results in testing stage, which is different from the results fundamentalneuralnetworkmodelswhileGFNNgainsonly showninpriortwomodels. 86.63%inaccuracymeasurewhichisthelowestperformance inthreemodels. rough further observations from test results as it is 6.PerformanceComparisonofModels shown in Table 12, we �nd the hybrid network model of eperformancesofthreeneuralnetworkmodelsdeveloped GFNNwithGAembeddedmaysigni�cantlyshowimprove- mentindetectionperformanceinallthreemeasuresfromits for comparison in this paper in terms of detection (i.e., classi�cation�accuracymeasurearesummarizedandshown fundamentalmodelinthetestingstagealthoughthereversed 6 AdvancesinArti�cialNeuralSystems T8:eperformancegainedfromgeneralizedfeedforwardneuralnetworkmodels. GFNN ModelFA ModelFB PureGFNN Training Test Training Test Accuracyofclassi�cation 100% 86.63% 100% 86.63% Sensitivity 100% 82.35% 100% 82.35% Speci�city 100% 90.91% 100% 90.91% GFNN+GA Test Test Accuracyofclassi�cation 91.09% 78.39% Sensitivity 88.24% 76.47% Speci�city 93.94% 80.30% T9:ebestsettingsofnetworkparametersformodularneuralnetwork. Networkparameters ModelFA ModelFB Learningrule LevenbergMarquar LevenbergMarquar Transformationfunctions SigmoidAxon SigmoidAxon Numberofhidinglayer 2 2 Numberofneuronsinhidinglayer 4 8 MSE :0.0001 MSE :0.0001 min min Criterionfortermination Epochs :3000 Epochs :3000 max max Methodsforweightupdate Batch Batch T10:eperformancegainedfrommodularneuralnetworkmodels. MNN ModelFA ModelFB PureMNN Training Test PureMNN Training Accuracyofclassi�cation 100% 93.23% Accuracyofclassi�cation 100% Sensitivity 100% 97.06% Sensitivity 100% Speci�city 100% 89.39% Speci�city 100% MNN+GA Test Test Accuracyofclassi�cation 88.82% 78.34% Sensitivity 89.39% 79.41% Speci�city 88.24% 77.27% T11:Detectionaccuracyinthreepureneuralnetworkmodels. Byadoptingthekeyfactorsusedincomputationformula Networkmodels Byadoptingthekeyfactorsselectedinthispaper(ModelFB) (ModelFA) BPN 94.75% 72.42% GFNN 86.63% 71.03% MNN 93.23% 70.99% enhancementisshowninBPNandMNN.Again,bythetest ModelFAamongthreefundamentalneuralnetworkmodels �gures shown in Section 4, although pure neural network employed in detecting CKD while the models BPN and models with GA show degraded performance improvement GFNN with GA embedded, respectively, might be the best for Model FA in testing stage in terms of three measures hybrid models for CKD detection but only GFNN plus GA in accuracy, sensitivity, and speci�city, it shows signi�cant cangainenhancementindetectionperformanceinbothsets improvement for Model FB. As a result, we conclude a of in�uence factors as input. ese conclusions should be hybrid model indeed may improve detection performance. furtherveri�edandcomparedwithothermodelsselectedfor is result may conclude that GA provides no bene�t in the testinlaterstudiesbeforetheycanbeassured. yield of better detection performance with the adoption of the key factors used in computation formula as the input to 7.Conclusion allmodelsintesting. AccordingtothedetectionperformanceshowninTables WeconcludethatneuralnetworkmodelsdevelopedforCKD 11 and 12, we conclude that BPN might be the best��tting detectionmayeffectivelyandfeasiblyequipmedicalstaffwith AdvancesinArti�cialNeuralSystems 7 T12:Detectionaccuracyinthreeneuralnetworkmodelswith [10] G. Arulampalam and A. Bouzerdoum, “A generalized feedfor- GA. ward neural network classi�er,” inProceedings of the Interna- tionalJointConferenceonNeuralNetworks,pp.1429–1434,July Networkmodels ModelFA ModelFB BPN+GA 91.71% 77.68% [11] R. Jacobs, M. Jordan, S. Nowlan, and G. Hinton, “Adaptive GFNN+GA 91.09% 78.39% mixturesoflocalexperts,”NeuralComputation,vol.3,no.1,pp. MNN+GA 88.82% 78.34% 79–87,1991. the ability to make precise diagnosis and treatment to the patient. In the future study, further model modi�cation and testing for the intelligence models developed in this paper should be conducted to enhance the accuracy in detection performance and to ensure they are sufficiently good for being truly employed in medical practice. In the meantime, different intelligence models can be widely and persistently applied for system development in this domain application aswellinordertosearchforabestonemodeltobeadopted. Inthefuturesystemdevelopment,itisworthwhiletodeploy thesystemtothecloudplatformsothatthepublicuserscan alsousethissystemtoconductaself-detectionofhavinghad CKD. References [1] Department of Health (DOH), “e statistical data of the causes of civilian mortality,” Taiwan, 2010, http://www .doh.gov.tw/CHT2006/DM/DM2_2.aspx?now_fod_list_no= 11122&class_no=440&level_no=3. [2] A.S.Levey,P.E.DeJong,J.Coreshetal.,“ede�nition,clas- si�cation, and prognosis of chronic kidney disease: a KDIGO controversies conference report,” Kidney International, vol. 80, no.1,pp.17–28,2011. [3] NationalKidneyFoundation(NKF),“KDOQIclinicalpractice guidelinesforchronickidneydisease:evaluation,classi�cation andstrati�cation,”AmericanJournalofKidneyDiseases,vol.39, pp.S1–S266,2002. [4] A. S. Levey, T. Greene, J. W. Kusek, and G. L. Beck, “A simpli�ed equation to predict glomerular �ltration rate from serumcreatinine,”JournalAmericanSocietyforNephrology,vol. 11,article155A,2000. [5] A.S.Levey,K.�.Eckardt,Y.Tsukamotoetal.,“De�nitionand classi�cation of chronic kidney disease: a position statement fromKIidneyDisease:ImprovingGlobalOutcomes(KDIGO),” KidneyInternational,vol.67,no.6,pp.2089–2100,2005. [6] D. Kriesel, “A brief introduction to neural networks,” 2011, http://www.dkriesel.com. [7] K. Matsushita, B. K. Mahmoodi, M. Woodward et al., “Com- parisonofriskpredictionusingtheCKD-EPIequationandthe MDRDstudyequationforestimatedGlomerular�ltrationrate,” e Journal of the American Medical Association, vol. 307, no. 18,pp.1941–1951,2012. [8] R.K.Chiu,C.I.Huang,and Y. C.Chang, “e study onthe construction of intelligence-based decision support system for diabetes diagnosis and risk evaluation,” Journal of Medical Informatics,vol.19,no.3,pp.1–24,2010. [9] G. Arulampalam and A. Bouzerdoum, “Training shunting in- hibitoryarti�cialneuralnetworksasclassi�ers,”NeuralNetwork World,vol.10,no.3,pp.333–350,2000. Journal of Advances in Industrial Engineering Multimedia Applied Computational Intelligence and Soft Computing International Journal of The Scientific Distributed World Journal Sensor Networks Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 Advances in Fuzzy Systems Modelling & Simulation in Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Submit your manuscripts at http://www.hindawi.com Journal of Computer Networks and Communications  Advances in  Artic fi ial Intelligence Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 International Journal of Advances in Biomedical Imaging Artificial Neural Systems International Journal of Computer Games Advances in Advances in Computer Engineering Technology Software Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 International Journal of Reconfigurable Computing Computational Advances in Journal of Journal of Intelligence and Human-Computer Electrical and Computer Robotics Interaction Neuroscience Engineering Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal

Advances in Artificial Neural Systems – Hindawi Publishing Corporation

Published: Jan 9, 2013

References

  • The statistical data of the causes of civilian mortality,
    Department of Health (DOH)
  • The definition, classification, and prognosis of chronic kidney disease: a KDIGO controversies conference report,
    A. S. Levey, P. E. De Jong, J. Coresh et al.
  • KDOQI clinical practice guidelines for chronic kidney disease: evaluation, classification and stratification,
    National Kidney Foundation (NKF)
  • A simplified equation to predict glomerular filtration rate from serum creatinine,
    A. S. Levey, T. Greene, J. W. Kusek, and G. L. Beck
  • Definition and classification of chronic kidney disease: a position statement from KIidney Disease: Improving Global Outcomes (KDIGO),
    A. S. Levey, K. U. Eckardt, Y. Tsukamoto et al.
  • A brief introduction to neural networks,
    D. Kriesel
  • Comparison of risk prediction using the CKD-EPI equation and the MDRD study equation for estimated Glomerular filtration rate,
    K. Matsushita, B. K. Mahmoodi, M. Woodward et al.
  • The study on the construction of intelligence-based decision support system for diabetes diagnosis and risk evaluation,
    R. K. Chiu, C. I. Huang, and Y. C. Chang
  • Training shunting inhibitory artificial neural networks as classifiers,
    G. Arulampalam and A. Bouzerdoum
  • A generalized feedforward neural network classifier,
    G. Arulampalam and A. Bouzerdoum
  • Adaptive mixtures of local experts,
    R. Jacobs, M. Jordan, S. Nowlan, and G. Hinton