Disponibleenlignesurwww.sciencedirect.comIRBM34 (2013)
244251Original articleComputer aided diagnostic problem solving:
Identication ofperipheral nerve disordersR. Kunhimangalama,, S.
Ovallathb, P.K. JosephaaNational Institute of Technology, NIT
Calicut (PO), Kozhikode, 673601 Kerala, IndiabDepartment of
Neurology, Kannur Medical College, Anjarakandy (PO), Kerala,
IndiaReceived 2 December 2012; received in revised form 6 April
2013; accepted 11 April 2013Available online 18 May
2013AbstractAim.Theaimwastodesignanddevelopadecisionsupportsystemwithagraphicaluserinterfaceforthepredictionofthecaseofperipheralnervedisorderandtobuildaclassierusingarticialneuralnetworksthatcandistinguishbetweencarpaltunnelsyndrome,neuropathyandnormalperipheralnerveconduction.Materialsandmethods.ThedatausedweretheNerveConductionStudydataobtainedfromKannurMedicalCollege,India.Arecurrentneuralnetworkandatwo-layerfeedforwardnetworktrainedwithscaledconjugate
gradientback-propagation
algorithmwereimplementedandresultswerecompared.Results.Boththenetworksprovidedfastconvergenceandgoodperformance,accuracybeing98.6%and97.4%fortherecurrentneuralnetworkandthefeedforwardnetworksrespectively,theconfusionmatrixineachcaseindicatedonlyafewmisclassications.Thedevelopeddecisionsupportsystemalsogaveaccurateresultsinagreementwiththespecialistsdiagnosisandwasalsousefulinstoringandviewingtheresults.Discussions.Intheeldofmedicine,programsarebeingdevelopedthataidsindiagnosticdecisionmakingbyemulatinghumanintelligencesuchaslogicalthinking,decisionmaking,learning,etc.Thesystemdevelopedprovesusefulincombinationwithothersystemsinprovidingdiagnosticandpredictivemedicalopinions.Itwasnotmeanttoreplacethespecialist,yetitcanbeusedtoassistageneralpractitionerorspecialistindiagnosingandpredictingpatientscondition.Conclusions.Thestudyprovesthatarticialneuralnetworksareindeedofvalueincombinationwithothersystemsinprovidingdiagnosticandpredictivemedicalopinions.Butthemajordrawbackofthesestudies,whichmakesuseofthenerveconductionstudydataaretheinherentshortcomingsoftheinterpretationoftheresults,whichincludelackofstandardizationandabsenceofpopulation-basedreferenceintervals.2013ElsevierMassonSAS.Allrightsreserved.1.IntroductionNeurologicaldisordersaffectingtheperipheralnervoussys-tem
consistsofaspectrumofdisorderswhichincludemorethan
100peripheralnervedisordersoutofwhichcarpaltunnelsyndrome(CTS)andsymmetricalperipheralneuropathypre-dominate.Thecontributionsofelectrophysiologicalstudiestothe
understandinganddiagnosisofperipheralnervedisordershave
beenreviewedextensively[1,2].CTSwhichis acom-mon
peripheralnervedisorderisanentrapmenttypeneuropathycaused
astheresultoftheentrapmentofthemediannerveCorresponding
author.E-mail addresses: [email protected],[email protected]
(R. Kunhimangalam).passingthroughthecarpaltunnel[3].
Thetestsforthediagno-sis includeelectromyography
(EMG)andthenerveconductionstudy
(NCS),wristX-raysshouldalsobedonetoruleoutotherproblemssuchaswristarthritis.However,NCSremainsthegoldstandard
fortheconrmationofthediagnosisofCTS[4,5].Peripheral
neuropathyreferstotheimpairmentofthenervesof
theperipheralnervoussystem,commonlyinducedeitherbydiseases
ortraumatothenerveorassecondary-effectsofsys-temic
illness.Differenttypesofperipheralneuropathyhavebeendescribed,
eachwithitsowncharacteristicsetofsymptoms,developmental
patternandmedicalprognosis.Theimpairedfunctions
andthesymptomsdependonthetypesofnervesthatare
damagedviz;themotor,sensoryorautonomic.Dependingonthe
patientsconditionmaybedescribedas
predominantlymotorneuropathy,predominantlysensoryneuropathy,sensory-motorneuropathy,autonomicneuropathy,etc.Thedifferentiationis1959-0318/$
see front matter 2013 Elsevier Masson SAS. All rights
reserved.http://dx.doi.org/10.1016/j.irbm.2013.04.003R.
Kunhimangalam et al. / IRBM34 (2013) 244251
245bestaccomplishedusingNCSandEMG[6].Thesetestscanconrmthepresenceofneuropathyandalsowhetheritis
motor,sensory orbothandalsothepathophysiology, i.e.whetheritisdue
todemyelinationoraxonalloss.NCSis acrucialcomponentof
theelectrodiagnosticevaluation,whichprovidesvaluablequantitative
andqualitativeinsightsintoneuromuscularfunc-tion,
particularly,theabilityofelectricalconductionofthemotorand
sensorynervesofthehumanbody[79].Itmustbeper-formed
withcarefulattentiontothetechniqueandmustbeinterpretedintheclinicalcontext.Mathematicalscienceandengineeringpreceptshavebeenwidelyemployedintheeldof
medicine[10].Therehasevolvedanumberoftechniques,which
aidinthemedicaldiagnosis,neuralnetworks(NN)beingone
amongthem[11].Theyhavebecomewellestablishedasexecutable,multipurpose,robustcomputationalmethodologieswith
rmtheoreticbackupandwithstrongpotentialtobeeffec-tive
inanydiscipline,especiallymedicine[12].Whateverbethecomputerlanguageortheunderlyingmethodologytheclinicaldecisionsupportsystemsdealswithmedicaldataandis
basedonthe
knowledgeofmedicinenecessarytointerpretsuchdata.Ingeneral
theyareemployedindeterminingthenatureofthedis-ease
buttheymayfurtherbeprogrammedto evenformulateanddevelop
aplanforreachingadiagnosisoradministeringtherapyappropriateforaspecicdiseaseorpatient[13].
Anycomputerprogram
designedtoassistinmakingaclinicaldecisioncanbecalled
aclinicaldecisionsupportsystem(CDSS)[13]oramedi-cal
decisionsupportsystem[14]. TherearemainlytwotypesofCDSS:
thersttypeisknowledgebasedwhichconsistsofthreeparts,
theknowledgebase,inferenceengine,andauserinter-face
whichformsamechanismforthecommunicationbetweenman
andmachine.Theknowledgebasecontainstherulesandassociationsofthecompileddata,whicharemostlyintheformof
IF-THENrules.Thesecondtypeisthenonknowledge-basedCDSSs
thatdonotuseaknowledgebasebutuseaformofarticial
intelligencecalledmachinelearning,whichallowcom-puters
tolearnfrompastexperiencesand/orndpatternsinclinical
data.Twotypesofnonknowledge-basedsystemsarearticialneuralnetworks(ANN)andgeneticalgorithms.In
thispaper,wehavedesignedanddevelopedknowledgebased
CDSSusingMATLABwithaGUIwhichconsistsofasimple
textorienteddisplayforthepredictionofthecaseofperipheralnervedisorderwhenprovidedwiththeNCSdataof
thepatient.Anonknowledge-basedsystem,usingANNin
theformofaclassierthatcandistinguishbetweenCTS,neuropathyandnormalperipheralnerveconductionwasalsodeveloped.Forthis,arecurrentneuralnetwork(RNN)andatwo-layer
feedforwardnetwork(FFN)trainedwithscaledconjugategradient
(SCG)back-propagation
algorithmwasimplemented.2.Methods2.1.DescriptionofdatabaseusedinANNclassicationInourstudy,wehaveusedtheelectronicmedicalrecordsof
KannurMedicalCollege,KeralaandselectedtheNCSdataof
254patientsoutofwhich90werenormalcases,i.e.thosewho
hadnormalNCSvaluesandhadnoelectrophysiologicalevidenceofCTSorneuropathy,100werepatientssufferingfrom
CTSandtheremainingwerehavingneuropathicsymp-toms.
TheNCSwasperformedusingthestandardtechniqueswith
surfaceelectroderecordingonbothhandsofeachsub-ject
usingconstantcurrentstimulator.Theethicalcommitteeapproval
wasobtained.Thefollowingcriteriawereappliedforidentifying
thedataforANNclassication[4]:For CTS
medianmotorlatencygreaterthan4.4ms,mediansensorylatency
greaterthan3.84ms,medianvelocitieslessthan50
m/s,normalulnarmotorvalues(latency:2.59.39ms,velocity:58.75.1m/s)andnormalulnarsensoryvalues(latency:
2.54.29ms,velocity:54.85.3m/s).For neuropathy
ulnarmotorlatencywillbegreaterthan3ms,theulnarsensorylatency
willbegreaterthan3.23ms,themedianmotorlatencyand
mediansensorylatencysameas theCTScase.Ulnarandmedian
velocitiesarelesserthan50m/s.The
slowingdownofthenerveconductionvelocity(NCV)and
prolongeddistallatenciesnormallysuggeststhereisdam-age
tothemyelinwhileareductioninthestrengthofimpulsesis
asignofaxonaldegeneration.Slowingofthemotorandsen-sory
latenciesofthemediannerveisanindicationofthefocalcompression
ofthemediannerveatthewrist,whichisanindi-cation
forCTS[4].Theslowingofallnerveconductionsinmore
thanonelimbindicatesgeneralizeddiseasednerves,i.e.it
indicatesgeneralizedperipheralneuropathy[15,16].Assess-ment
oftheperipheralneuropathyusingNCScandirectaphysician
towardstheappropriateautoimmunedisorder.Typ-ically,
demyelinatingneuropathiesdemonstrateslownerveconductionvelocities(NCV),oftenwithreducedamplitudesofsensory/motornerveconductionandprolongeddistallatencies.By
contrast,axonalneuropathiestypicallydemonstratenormalNCVs
withlowamplitudesofsensory/motornerveconduc-tion.
NeuropathiesmayalsohavemixedEMG/NCSresultsandexhibit
featuresofbothdemyelinationandaxonalloss[4].Atypical
NCSreportisshowninTable1, whichshowsthedatafor
anormalperson.2.2.Developingtheprogramandbuildingthegraphicaluser
interface(GUI)fordiagnosticpredictionComputeraidedinterpretationofmedicaldataiswidespreadbut
alotofphysicians arereluctantinrelyingonthecomputer,because
theadviceis deliveredbyacomputerprogramanditisnever
foolproof.Overthelastdecades,awiderangeofcomputersystemshasbeendevelopedintheareaofmedicinefordecisionsupportsystems,theyincludecomputertoolsforpatientspecicconsultationse.g.expertdiagnosticsystemsdesignedtoprovidedifferentialdiagnosisorexpertadvice.Thealgorithmsusedinthese
typesofdecisionsupportsystemsvarysubstantiallybutin
generalsuchsystemsdependupontheknowledgeandinfor-mationthatarecontainedinthesystem.Suchsystemsshould246
R. Kunhimangalam et al. / IRBM34 (2013) 244251Table 1Values of
nerve conduction study of a normal person.Site Latency (ms)
Amplitude Distance (mm) Interval (ms) NCV(m/s)Motor Nerve
Conduction StudyMedian, LWrist 3.85 9.25 mV 240 3.85 59.5Elbow 7.88
10 mV 4.03Median, RWrist 3.76 8.25 mV 250 3.76 58.2Elbow 8.05 8.5
mV 4.29Ulnar, RWrist 2.68 5.7 mV 220 2.68 57.6Elbow 6.5 6.25 mV
3.82Ulnar, LWrist 2.67 6 mV 225 2.67 55.83Elbow 6.7 6.5 mV
4.03Sensory Nerve Conduction StudyMedian, LWrist 2.88 28.6 V 140
2.88 48.6Median, RWrist 2.75 32.6 V 140 2.75 51Ulnar, RWrist 2.25
25.4 V 120 2.25 53.33Ulnar, LWrist 2.29 20.3 V 120 2.29 52.4NCV:
nerve conduction
velocity.ideallybeabletokeepupwiththehumandecisionmakingpro-cess.
Thedecisionsupportsystemdevelopedconsistsofthreemain
parts:theinput,therulesetandtheoutput.Diagnosisofa
diseaseisdonebyaspecialistusingasetofrulesandthedesignedsysteminvolvesthecollectionoftheserulesandeval-uation
oftherulebaseforagivensetofinputs.Fordevelopingdiagnostic
toolforCTSandneuropathy,dataisrequiredthatiscapable
ofrepresentingthediseases.Byconsultingthespecial-ist
andbyanalysingthedataofthepatients,eightNCSvaluewere nalizedas
theinputsfordiagnosisviz.motormedianlatency,
motormedianNCV,motorulnarlatency,motorulnarNCV,
sensorymedianlatency,sensorymedianNCV,sensoryulnar
latencyandsensoryulnarNCV.Theinputis givenontothe
graphicaluserinterface,thevaluesarecheckedandcom-pared
totherulesetandnallytheoutputisobtainedastheresult. Theaimis
todevelopaMATLABGUIsystem,whichmodels
thereasoningprocessoftheconsultantsintheparticularmedical
scenariounderconsideration.Therule-baseddecisionmakingprogramdevelopedutilisesanapproachwherealltheknowledge,informationandtheconcerneddataiscontainedina
grouporsetofrules.TherulebasedevelopedconsistedofIF-THEN-ELSErules,whichwereformulatedusingthesamecriteria
appliedforidentifyingthedataforANNclassication.Every
rulecontainsmultiplehypotheses
andconclusions,whichrepresentthelogical,thoughtprocessesofthespecialists.Thegiven
setsofrulesareactivatedwhentheinformationorthedatais
inputintotheuserinterface.Theknowledgebasecanbeeasilymodiedorchanged.Theproposedsystemreducesthediagno-sis
timeofaphysician andadditionallyincreasestheaccuracyof
thediagnosis.Theproposedsystemisnotonlyusedfordiag-nosis,
butalsousedtostoreandreadtheresultsofthediagnosisfor
futurereference.TheuserhastoentertheNCSvaluesintotheinterfacetogetherwiththepatientdetailsandwhenclickedon
theGETRESULTbuttonthediagnosisisoutputbythepro-gram.
Theprogramcandistinguishbetweenandgivetheresultas
normal,suggestiveofbilateralCTS,suggestiveofleftCTS,
suggestiveofrightCTS,predominantlymotorneu-ropathy,
predominantlysensoryneuropathy, sensory-motorneuropathy,
orautonomicneuropathy. Theprogramhasalsoprovisions
forstoringtheresultonaspreadsheetbyclickingon
theSAVERESULTbutton.Thisstoreddatacanbeeasilyretrieved
forfurtherreferenceas andwhenneeded.Provisionisalso
providedintheinterfaceforthedoctororthetechniciantoinclude
theircommentsforeachpatientwhichcanalsobestored.Involvementofthedistalpartoftheperipheralnervealoneisincluded
inthissoftwareprogram.Theprogramcanbefurtherexpanded
toincludetheinvolvementoftheproximalpartofthenerves
byincludingvaluesofstimulationattheelbowlevelanderbs
pointintheupperlimb.2.3.TrainingofthearticialneuralnetworkANNisamathematical/computationalmodelinspiredbythe
structuralandfunctionalaspectsofthebiologicalneuralnetworks.Itconsistsofnodescalledneuronsandweightedcon-nections
thattransmitsignalsbetweentheneuronsinaforwardor
loopedfashion;theyprocesstheinformationusingacon-nectionistapproachtocomputing.TheFFNcanapproximateaspatially
nitefunctionwithalargesetofhiddennodesbyoper-ating
ontheinputspace.ThebasicdifferencefoundintheRNNis
thattheyoperatenotonlyonaninputspacebutalsoonaninternal
statespace,whichrepresentssomeinformationonwhatalready
hasbeenprocessedbythenetwork.Thedifferenceintheoperating
principleofthetwocanbeunderstoodfromFig.1.R. Kunhimangalam et al. /
IRBM34 (2013) 244251 247Fig. 1. (a) A schematic representation
showing the difference in the basic principle of operation of the
recurrent neural networks and feed forward networks. (b)Feed
forward network implementation in MATLAB. (c) Recurrent neural
network implementation in
MATLAB.InthecaseofRNNeverytimeitreceivesapattern,theunitcomputesits
activationjustlikeafeedforwardnetwork.Butits
netinputwillcontainatermreectingthestateofthenet-work
beforethepatternwasseen.Inthesubsequentpatterns,the
hiddenandoutputstateswillbeafunctionofeverythingthenetwork
hasseensofar,i.e.thenetworkbehaviourofRNNisbased
onitshistory[17].Consideratwo-layerednetwork,i.e.a
networkwithtwolayersofnodes;aninputlayer,ahiddenorstate
layer,andanoutputlayer.Inafeedforwardnetwork,theinput vector,x,is
propagated throughaweightlayer,Uandwehave thefollowingEqs.14:yk(t)
= f (netk(t)) (1)netk(t) =n
ixi(t) uki+k(2)yj(t) = g
netj(t)
(3)netj(t) =l
kyk(t)
wjk+j(4)withthefollowingindexvariables:nfortheinputnodes,kforthe
hidden,j fortheoutputnodes,sarethebiasesandf andgare
outputfunctions.Wis thesetofweightsintheoutputlayerand
yistheoutputvector.Itconsistsofthreemainlayers:theinput
layer;whichreceivesdata,i.e.clinicalndings,theoutputlayer
whichgivestheresultsandhiddenlayerthatprocessesthedata
andarrivesattheconclusion.Thestructureofthenetworkchanges
basedontheexternalorinternalinformationthatowsthroughitduringthelearningphase.Attheentranceofeacharticial
neuron,whichis thebasicbuildingblockofeveryANNthe
inputsareweighted,i.e.everyinputsignalismultipliedwithan
individualweight.Inthesucceedingsectionisthesummingfunction
thataddstogetheralltheweightedinputsandbias.Atthe
exitthesumofpreviouslyweightedinputsandbiasespassesthrough
theactivationfunctionalsocalledthetransferfunctionto
getthenaloutput.Fig.2showstheANNbasicstructure.In
asimpleRNN,theinputvectorissimilarlypropagatedthrough
aweightlayer,andisalsocombinedwiththepreviousstate
activationthroughanadditionalrecurrentweightlayer,V,[18]andwehaveEqs.5and6.yk(t)
= f (netk(t)) (5)netk(t) =n
ixi(t) uki+l
mym(t1)
vkm+k(6)wherelisthenumberofstatenodes.Eqs.3and4areappli-cable
fortheoutputlayeroftheRNNalso.Thegeneralneuralnetwork
designprocessconsistsofthefollowingsteps:thecol-lection
ofdata,thecreationofthenetwork,itsconguration,theinitializationoftheweightsandbiases,thetrainingofthenet-work,
thevalidationofthenetworkandnallyusingthenetwork[19].MATLABsoftwarepackage(MATLABversion7.9.0withneural
networkstoolbox)wasusedforimplementationoftheclassiers
usingneuralnetworks.TheNCSdatacollectedwasseparated
intoinputsandtargets.Thesignicantfeatureswereidentiedfromthedataandtheyactsastheinputstotheneuralnetwork.Eightinputswereidentiedwhicharemotormedian248
R. Kunhimangalam et al. / IRBM34 (2013) 244251Fig. 2. The basic
structure of articial neural networks (ANN).Table 2The attributes
of the nerve conduction study datasets.Attribute No. Attribute
description Attribute range Mean Standard deviation1 Motor median
latency (ms) 29 4.27 1.322 Motor median nerve conduction velocity
(m/s) 3065 52.35 8.383 Motor ulnar latency (ms) 29 2.87 0.694 Motor
ulnar nerve conduction velocity (m/s) 3065 53.76 8.95 Sensory
median latency (ms) 29 3.6 1.256 Sensory median nerve conduction
velocity (m/s) 3065 49.8 5.797 Sensory ulnar latency (ms) 29 3.3
0.6338 Sensory ulnar nerve conduction velocity (m/s) 3065 52.76
6.78latency,motormedianNCV,motorulnarlatency,motorulnarNCV,
sensorymedianlatency,sensorymedianNCV,sensoryulnar
latency,sensoryulnarNCV.Theattributesofthedatasetsare
giveninTable2.Thetargetsfortheneuralnetworkwerethelogicalindicesofthe
diseasesamples.Thenormalsampleswereidentiedwitha1
00,theCTSwitha010andtheneuropathywith001.Thesamples
weredividedintotraining,validationandtestsets.Thetraining
setteachesthenetworkandtrainingcontinuesaslongas
thenetworkcontinuesimprovingonthevalidationset.Thetrainingstopsautomaticallywhengeneralizationstopsimprov-ing
whichisindicatedbyanincreaseintheMeanSquareError(MSE)ofthevalidationsamples.MSEis
theaveragesquareddifference
betweentheoutputsandtargets,lowervaluesarebet-ter,
zeromeansnoerror.TheerrorgoalintrainingwasxedasE