An algorithm for automatic detection of repeater F-waves and MUNE studies

(1)

ContentslistsavailableatScienceDirect

Biomedical

Signal

Processing

and

Control

j ou rn a l h o m e pa g e :w w w . e l s e v i e r . c o m / l o c a t e / b s p c

An

algorithm

for

automatic

detection

of

repeater

F-waves

and

MUNE

studies

N. Tu˘grul

Artu˘g

a,∗

_,

_N.

_Görkem

_S¸

_irin

b

_,

_Emel

_O˘guz

_Akarsu

b

_,

_M.

_Baris

_Baslo

b

_,

_A.

_Emre

_Öge

b a_Electrical_and_Electronics_Engineering,_Istanbul_Arel_University,_Tepekent,_{Buyukcekmece,}_Istanbul,_Turkey

b_Istanbul_Medical_Faculty,_Istanbul_University,_Fatih,_Capa,_Istanbul,_Turkey

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received9July2018

Receivedinrevisedform5February2019 Accepted26February2019

Availableonline7March2019 Keywords:

Electromyography RepeaterF-waves MUNE

ALS

AutomatedF-responseanalysis

a

b

s

t

r

a

c

t

ThepresentstudyaimstodevelopanalgorithmandsoftwarethatautomaticallydetectsrepeaterF-waves whichareverydifﬁculttoanalyzewhenelicitedashighnumberofrecordingsinmotorunitnumber estimationstudies.ThemainstrategyofthestudywastotaketherepeaterFwavesdiscriminatedbythe neurologist,fromlimitednumberofrecordings,asthegoldstandardandtotesttheconformityofthe resultsofthenewautomatedmethod.

TenpatientswithALSandtenhealthycontrolswereevaluated.90F-waveswithsupramaximalstimuli and300F-waveswithsubmaximalstimuliwererecorded.Supramaximalrecordingswereevaluatedboth manuallybyanexpertneurologistandautomaticallybythedevelopedsoftwaretotesttheperformance ofthealgorithm.Theresultsbothacquiredfromtheneurologistandfromthesoftwarewerefound compatible.Therefore,themainexpectedimpactofthepresentstudyistomaketheanalysisofrepeater Fwaveseasierprimarilyinmotorunitnumberestimationstudies,sincethereiscurrentlyacontinuing needforsuchautomatedprogramsinclinicalneurophysiology.

Submaximalrecordingswereexaminedonlybythedevelopedsoftware.Theextractedfeatureswere: maximumMresponseamplitude,meanpowerofMresponse,meanofsMUPvalues,MUNEvalue,number ofbaskets,persistenceofF-waves,persistenceofrepeaterF-waves,meanofF-waves’powers,medianof F-waves’powers.Featureselectionmethodswerealsoappliedtodeterminethemostvaluablefeatures. Variousclassifierssuchasmulti-layerperceptron(MLP),radialbasisfunctionnetwork(RBF),support vectormachines(SVM)andknearestneighbors(k-NN)weretestedtodifferentiatetwoclasses. Ini-tiallyallfeatures,thendecreasednumbersoffeaturesafterfeatureselectionprocesswereappliedtothe aforementionedclassifiers.Theclassificationperformanceusuallyincreasedwhendecreasedfeatures wereappliedtointelligentsystems.Ulnarrecordingsundersubmaximalstimulationshowedbetter per-formancewhencomparedwithsupramaximalequivalentsormediannerveequivalents.Thehighest performancewasobtainedas90%withk-NNalgorithmwhichwasacommitteedecisionbasedclassifier. Thisresultwasachievedwithonlytwofeatures,namelymeanofsMUPamplitudeandMUNEvalue.

1. Introduction

F-wavesareoneofthelateresponsesacquiredduringroutine

EMGstudies.Theletter“F”comesfromtheinitialof“foot”whichis

theﬁrstrecordinglocationofthatsignal[1].F-wavesappearafter

theMresponsefollowingthenervestimulation[2].F-waves

con-sistoftheactionpotentialswhicharedebouncedfromthelower

motorneurondendritesandreturnedtomusclewheretheyare

beingrecorded[3].TheactionpotentialscontributedtoF-waves

∗ Correspondingauthor.

E-mailaddresses:tugrulartug@arel.edu.tr(N.T.Artu˘g),

gorkemsirin@yahoo.com.tr(N.G.S¸irin),emeloguz@yahoo.com(E.O.Akarsu), mbbaslo@istanbul.edu.tr(M.B.Baslo),aemreoge@istanbul.edu.tr(A.E.Öge).

arealsopresentinMresponse[4–6].However,allcomponentsthat

formMresponsecan’tdebouncefromspinalcord.F-waveschange

fromstimulustostimulusbecauseanF-wavemaybeaproduced

signalfromamotorunitoritcanbeacombinedsignalfrom

sev-eralmotorunits[3,7].HoweverthemorphologyoftheF-wavefor

asinglemotorunitisthesame[3].

F-waveshavelatenciesafterthelatencyofMresponsebecause

ofthelongpathwaytheyhavetotravel.Thelengthofthis

path-waydeterminesthelatencyofF-wave[8].Itisseenadecrement

inthenumberofF-waveproducingmotorneuronsindenervation

whichisdepictedbytheirdecreasedpersistencewhiletheratio

ofrepeaterF-wavesincreases[8–11].F-wavesarelowamplitude

signalsinhealthystatebutitcanbeobservedthatsomesignals

mayreachupto700␮Vamplitude[12,13].TheabilityforF-wave

generationchangesfrommuscletomusclesuchastherearemore

(2)

stimulatingthemedian,peroneal,ulnarandtibialnerves.Theyalso

questionedtheeffectofheight,ageandgenderinthisstudy.They

foundthat10cmincrementinheightcauses1.6to3ms

prolonga-tionofminimumF-wavelatencyaccordingtothelocalizationof

recordingmuscle.Agealsohasaneffectonlatencyprolongation

butlesspronouncedcomparingtotheheight.Genderwasfound

ineffective.

RepeaterF-wavesareknownastheF-wavesthathavethesame

latency,sameamplitudeandsamemorphology[3].Theideal

pop-ulationtodoastudyonrepeaterF-wavesisobtainedastheresult

ofatleast90stimuli[3,27].Inthestudythatwasconductedby

Stashuketal.300stimuluswereapplied[15].

Indailypracticethemeasureableparametersof F-wavesare

evaluatedmanuallybyneurologist.Therearestudiesaboutdoing

the investigations automatically. The developed algorithm by

Stashuketal.evaluatesF-wavesautomatically[15]andcalculates

maximumMresponse,sumofF-waves,meanofS-MUAPaswellas

theestimationofthenumberofmotorunits.Eveniftheprogram

doesthesecalculationsautomaticallytheselectionofS-MUAPis

donethebyoperatormanually.For motorunitnumber

estima-tion(MUNE)300F-wavesarerecordedwithsupramaximalstimuli.

Thedevelopedalgorithmcanbeappliedtoeverymusclethathas

F-wavelatencymorethan20ms.

Hachisukaetal.conductedastudy[28]aboutcalculatingthe

MUNEvalue.TheyrecordedF-wavesofmedianandtibialnerves

from43Poliopatientsand20healthyindividualsinresponseto

100stimuli.TheyobserveddecreasedpersistenceoftheF-waves

alongwithanincrementinthenumberofrepeaterF-wavesinthe

patients.Theyalsohavefoundanegativecorrelationbetweenthe

numberofrepeaterF-wavesandtheMUNEvalue.

AstudyconductedbyChronietal.[29]determinedthe

char-acteristicsofrepeaterF-wavesbyexcitinglowandhighthreshold

motorﬁbersinhealthyindividuals.Collisiontechniqueand

sub-maximalstimuluswereusedforthispurposetoexamineifmotor

neuronsubgroupswereresponsibleofrepeaterF-waves.Collision

techniquewaspreferredtoeliminatetheF-wavesgeneratedfrom

lowthresholdneurons.Therecordingswereacquiredfromulnar

nerves(ADMmuscle)of12healthyparticipantsinﬁvedifferent

sessionsbyusingsupramaximalandsubmaximalstimuli.Anew

softwarewasdevelopedandintroducedfordetectingrepeater

F-waves.TheirstudyshowedthatthepresenceofrepeaterF-waves

weremorerelatedtohighpersistenceratherthantheneuron

sub-groupsgeneratingthem.

Kamel etal.’s study[30] wasabout combiningF-wave with

singleﬁberconductionvelocity.Theyincluded16healthy

partici-pantsand16patientswithmildneuropathy.F-wavemeasurement

wasdonefromﬁbularnerve (EDBmuscle) with20

supramaxi-malstimuliusingasurfaceelectrode. Theextractedparameters

wereminimumandmaximumF-wavelatency,F-wavedispersion,

thedifferencebetweenFmaxandFmin,F-wavepersistence.

Sin-gleﬁberF-waverecordingsweredonefromthesamemuscle.In

wavesautomatically,

bTesttheperformanceofthemethodbycomparingtheresults

ofF-waveanalysismanuallydonebyanexpertwiththeonesof

developedsoftware

cExtract some features from F-waves to classify patients and

healthyindividualswithintelligentsystems

ThepapercontinueswithmaterialsandmethodsinSection2.

Theory/Calculationpartcontainsthealgorithmandﬂowchartfor

detectingF-waveswhichisinSection3.Resultsarepresentedin

Section4.InSection5,Discussionpartarguestheprosandconsfor

theproposedmethodandcomparesthemwithpreviousstudies.

ThispaperendswiththeConclusion.

2. Materialandmethods

Inthisstudyatwoclassdatasetisformedthatcontains

neuro-genicpatientsandhealthyindividuals.F-waveswererecordedin

responsetosupramaximalandsubmaximalstimuli.Interpretation

wasdonebothwithmanuallybyanexpertclinical

neurophysiolo-gistandbythedevelopedalgorithm.Resultsofbothinterpretations

werecompared.

Thestudyhasbeenapprovedbythelocalethicscommitteeof

IstanbulUniversity,IstanbulMedicalFaculty(2016/162).

2.1. Subjects

Ten patientswithALSand tenhealthy controlswere

evalu-atedforpreliminaryanalysisofautomatedF-wave.Themeanage

was53.4±10.2inpatients(ranged36 to64)and 51.5±13.6in

healthycontrols(ranged30to67).Eightofthepatientswere

clas-sifiedasdefiniteALSandtwowereclassifiedaspossibleALSdue

toAwajicriteria[31].Inthepatientgroup,meandurationof

symp-tomswas15.2±20.2months(2–72)andmeanALS-FRSscorewas

41.4±5.3(34–47)[32].Patientswhohadsensorimotor

polyneu-ropathy,carpaltunnelsyndrome,ulnarentrapmentneuropathy,

diabetes,uremia,chronicalcoholuse,ormalignancywereexcluded

fromthestudy.Patientshavingcompoundmuscleaction

poten-tial(CMAP)amplitudes less than 1mV in eitherof the studied

muscleswereexcludedfromthestudy.Healthycontrolshad

nor-malneurologicalexamination,normalEMGstudiesandnoclinical

symptomsofcarpaltunnelsyndrome,ulnarentrapment

neuropa-thyorpolyneuropathy.

2.2. Electrophysiologicalevaluation

CompoundmuscleactionpotentialsandF-waveswererecorded

fromthelessaffectedupperextremityinthepatients.Ifbothupper

extremities wereaffectedequally,then non-dominantside was

(3)

chosen.MedianandulnarCMAPswereelicitedbystimulatingthe

nervesatthe wristandrecording fromabductor pollicisbrevis

(APB)andabductordigitiminimi(ADM)musclesrespectively,with

aMedelecSynergyEMGmachine.Disposablerecordingelectrodes

wereplacedovermusclesaccordingtothe“belly-tendon

record-ing”principleasdescribedpreviously[33].CMAPwasrecordedin

bothmusclesbysupramaximalstimulus.Theﬁltercut-offsettings

werearrangedas20Hz-10kHz.Then,cathodeofthestimulatorwas

placedproximaltoanodeforF-waverecording.Forthe

prelimi-naryanalysis,90F-waveswereelicitedbysupramaximalstimuliat

afrequencyof0.5Hzforeachmuscle.ForF-waveMUNEanalysis,

stimulusintensitywasdecreasedtoalevelof10–50%ofthebaseto

peakamplitudeofmaximumCMAPwhichisknownassubmaximal

stimulus.Then,300F-waveswererecordedbysubmaximalstimuli

atafrequencyof0.5Hzforeachmuscle[15,17].

2.3. Manualanalysis

NinetyF-waveswereanalyzedvisuallybythesameexperienced

clinicalneurophysiologist(EOA) fromprint-outsmadewiththe

sensitivityof500␮V-1mV/divisionandsweepdurationof100ms

withoutsplittingthescreen.RepeatingF-wavesmorethanonce

withsamelatency,amplitudeandshapewereconsideredassingle

motorunitpotential(sMUP).Peaktopeakamplitudesofrepeating

F-wavesweremeasuredinordertocalculatethemeanamplitude

ofsMUPs.

TheMUNEvaluewascalculatedbydividingtheamplitudeofthe

CMAPtothatofthemeansMUPforbothmethods.Aftercalculating

sMUPandMUNEvalues,oneoftheresearchersperformedthe

sta-tisticalanalysisforpreliminaryanalysis.CalculatingMUNEvalues

from300F-waveselicitedbysubmaximalstimuliwereperformed

onlybyautomatedanalysis.

2.4. Automatedanalysis

AnalgorithmwasdevelopedforextractingF-wavesfromeach

signalrecordandgroupingeachrepeaterF-waveindifferent

“bas-kets”,calculatingthenumberofrepeaterF-wavesandthenumber

ofrepeatsforeachrepeaterF-waveinthebaskets.Thesoftware

isalsoabletodisplaythemostsimilarF-wavepairs, calculates

theMresponsemaximumamplitudevalue,F-wavepeaktopeak

amplitudevalue,powervalueforeachF-waveandMUNEvalue.

SPSSv21wasusedtoperformstatisticalanalysis.MUNEand

sMUPwerecalculatedvisuallybytheneurophysiologistand

auto-maticallybytheproposedmethod.Theresultswerecomparedwith

Wilcoxonsingleranktest.Spearman’srhotestwasusedtomeasure

thecorrelationbetweenthetwomethods.

3. Theory/calculation

ThemethodforextractingF-wavesanddeterminingrepeater

F-wavesaredescribedasfollows.Firsttherecordedsignalsare

ﬁl-teredfromnoisebyusingwavelettransformbasednoisereduction

method.TheDaubechieswaveletispreferredfornoisereduction.

Thethresholdfunctionischosenashardthreshold.Inthemulti

res-olutionanalysis,decompositionlevelwasselectedas3.Thegraphic

forthe15recordedsamplesignalswhichbelongstoapatientis

giveninFig.1.

Twomillisecondpartfromthebeginningofeverysignalis

dis-carded.Thispartcontainsstimulusartifact soitis unnecessary

forcalculations.ThenthemaximumamplitudeofMresponseis

determinedforeachrecordedsignalandmeanofitiscalculated

(MGloMax).

AftertheF-wavesarecutfromthebeginningandtheend

loca-tions,themaximum(Fmax)andtheminimum(Fmin)amplitude

valuesarecalculated.ThelocationsofFmaxandFminarerecorded.

ThegraphicfortheF-wavesaftercuttingfromtherawsignalsis

showninFig.2.

Ifasignal’speaktopeakamplitudevalueVpp<=40␮V,itis

acknowledgedasnoise, ﬂooredtolevel 0and is notevaluated,

accordingtotherecommendationsfortheclinical

neurophysiol-ogystudies[3,14].Thesignalsthathaveamplitudesgreaterthan

40␮VareapprovedasF-waves.Moreover,ifasignaldoesnotgoes

down40␮Vtotheleftandrightin3msfromtheFmaxlocation,

thissignalisevaluatedasnoiseandisﬂooredtolevel0too.

Aftertheseprocesses,allsignalsarealignedaccordingtotheir

Fmaxlocations.BecausetherepeaterF-wavesmusthavethesame

amplitude,samelatencyandsamemorphology;thesignalpairs

thatarecloserthan0.5msuptotheFmaxandFminlocationsare

determinedasrepeaterF-wavecandidates.

IfthedifferencevalueforFmaxbetweencandidatesislower

than10%andthedifferenceofindividualpowervaluesbetween

themislowerthan20%,theykeeptheircandidacy.Besides,the

correlationcoefﬁcientbetweencandidatesisinspected.Ifthe

coef-ﬁcientislowerthan0.9thatpairisdiscardedfromcandidacy.

Thedifferenceofamplitudesforcandidatesignalpairsis

calcu-latedandthedifferencesignalsarerectiﬁedtocalculatepowers.A

“similaritycoefﬁcient”iscalculatedaccordingtothesecandidates’

amplitudedifferenceandpowerdifference.Thethresholdvaluefor

similaritycoefﬁcientisdeterminedas0.6,duringthepreparation

periodofthepresentstudyafterexaminingamultitudeofdifferent

signals.Ifanycandidate’ssimilaritycoefﬁcientislowerthan0.6it

becomesarepeaterF-wave.IfanyotheridenticalrepeaterF-wave

pairispresent,theyarecombinedinthesamebasket.

Theuniquesignalsthatarenotinanybasketarealignedupto

theFminlocationthistime.Ifsignalpairsarecloserthan0.5ms

uptotheFminlocations,theyaredeterminedasrepeaterF-wave

candidates.

IfthedifferencevalueforFminbetweencandidatesislowerthan

5%andthedifferenceofindividualpowervaluesbetweenthemis

lowerthan10%,theykeeptheircandidacy.

Thedifferenceofamplitudesforcandidatesignalpairsis

calcu-latedandthedifferencesignalsarerectiﬁedtocalculatepowers.A

“similaritycoefﬁcient”iscalculateduptothesecandidates’

ampli-tudedifferenceandpowerdifference.Ifanycandidate’ssimilarity

coefﬁcientislowerthan0.6itbecomesarepeaterF-wave.Ifany

otheridenticalrepeaterF-wavepairispresent,theyarecombined

inthesamebasket.

Themeanofthepeaktopeakamplitudes(sMUP)ofallsignalsin

abasketiscalculated.Themeanvalueforallbaskets’sMUPvalueis

calculated.MUNEvalueiscalculatedwiththeformulagivenbelow:

MUNE=

MGloMax_j

k=1 sMUPk

/j (1)

ThenumberofF-wavesisdisplayed.Howmanyofthemarein

abasketandhowmanyofthemareuniquecanbeseenonthe

monitor.LastlyMUNEvalueisdisplayed.Themostsimilarsignal

pairscanbeplottedoneundertheotheroroverlappedwith

simi-laritycoefﬁcientvalue.ThemostsimilarF-waverepeatersamong

therecordedsamplesignalscanbeseeninFig.3.

ThealgorithmforthedevelopedsoftwareisgiveninFig.4.

3.1. Featureextractionfromdataset

Nine features were extracted from this dataset.These were

mean of maximum M response amplitude, mean power of M

response,meanofsMUPvalues,MUNEvalue,numberofbaskets,

persistenceofF-waves,persistenceofrepeaterF-waves,meanof

(4)

Fig.1. M-ResponsesandF-Wavesin15RecordedSampleSignals.

(5)

Fig.3.MostSimilarRepeaterF-WavePairAmongtheRecordedSampleSignals.

FirstthemaximumamplitudeforallMresponsesfrom

base-linewasdeterminedandtheirmeanwascalculated.Thenpower

valueforallMresponsesfrombaselinetopositivepeakvaluein

eachrecordwascalculated.MeanofpowervaluesforM-responses

wascalculated.Thirdfeaturewascalculatedasthemeanofeach

F-waverepeaterbasketsMUPvalues.MUNEvaluewascalculatedas

thefourthfeatureasdescribedbefore.NumberofdifferentF-wave

repeaterswasthenumberofbaskets.Persistencewastheratioof

F-wavesignalstotheallsignalrecords.Seventhfeaturewas

per-sistenceofrepeaterF-waves.Itcanbecalculatedasgivenbelow.

PersistanceofrepeaterF−waves

=(AllRecords_All−_RecordsNoiseSignals)₋_Noise−_SignalsUniqueSignals (2)

NextfeaturewasmeanofallindividualF-wavepowervalues

andthelastfeaturewasthemedianvalueofallindividualF-wave

powervalues.

3.2. Featureselection

Infeature selectionprocess ReliefFalgorithm waspreferred.

Reliefalgorithmwasdevelopedformulti-featureddatasetson

fea-tureselectionprocess.Inthealgorithmtwonearestneighborsare

determinedfromeachclassforeachsample.Neighborinthesame

classwiththesampleismarkedasnearesthit(H),other

neigh-borismarkedasnearestmiss(M)[34].Foreachfeatureaweightis

calculatedandfeaturesareevaluatedtobeselectedbytheir

impor-tance(weights).ReliefFalgorithmwasdevelopedtoovercomethe

drawbackofgenericReliefalgorithmwhichcanonlyseparatetwo

classes[35].In ReliefFalgorithmknearestneighborsare

deter-minedinsteadoftwoneighborsforeachsample.

3.3. Classiﬁcationofdata

Afterthefeatureextractionprocesssupramaximalrecordings

had10instancesinmediannerveand8instancesinulnarnerve

foreachclasswhich werehealthyindividualsand ALSpatients.

Table1

SpeciﬁcationsforDataSet.

StimulationType #instances #features Median Ulnar Median Ulnar

Supramaximal 10Healthy 8Healthy 9

10ALS 8ALS

Submaximal 5Healthy₅_ALS 8

Thereasonforlowernumberofinstancesinulnarnervewasno repeaterF-wavewasdetectedbythesoftware,andso2instances fromeachclasswerediscarded.Afterthisprocess5-foldcross vali-dationformediannerveand4-foldcrossvalidationforulnarnerve couldbeapplied.Submaximalrecordingshad5instancesforeach classbuttheyhad8featuresbyexcludingMresponsemeanpower. AsummaryfordataspeciﬁcationsweregiveninTable1.

Fourdifferentclassiﬁersweretestedforseparationofthetwo

classes.Firstoneismulti-layerperceptron(MLP).Ithastwo

hid-den layers and there are 18 and 9 neurons for supramaximal

records,16and 8neuronsforsubmaximalrecordsinthose

lay-ers.Levenberg–Marquardtwasselectedasthetrainingalgorithm

forMLPnetwork.

Secondclassiﬁerwasradialbasisfunctionnetwork(RBF).Spread

parameterwas0.2formediannerverecordingsand0.6forulnar

nerverecordings.RBFnetworkshaveonlyonehiddenlayerandthe

algorithmaddsneuronstoitformeetingtheperformancegoal.The

algorithmwasadded25neuronstoitshiddenlayerforeachtrial

inthisclassiﬁcationtask.

Supportvectormachines(SVM)weresuitableforthis

classiﬁ-cationproblembecauseitwassufﬁcientforrecognitionofthetwo

classes.

Lastclassiﬁcationalgorithmwask-NNandneighborhoodvalue

wasselectedas1,3and5.

4. Results

IfnorepeaterF-waveisobservedinanervebytheexpert

(6)

Fig.4. AutomatedF-WaveRepeaterDetectionAlgorithm.

analysis.Afterthisprocess,atotalof18medianand16ulnarnerve

recordingsareevaluatedforsupramaximalstimuli.For

submaxi-malstimulation,5recordingsforeachnerveacquiredbothfrom

healthyparticipantsandALSpatientswereincluded.

4.1. Statisticalanalysis

Maininterestfortherecordingswithsupramaximalstimulation

(7)

Table2

MeansMUPandMUNEValuesAcquiredbySupramaximalStimulationforHealthyParticipants.

HealthyControls8Median,7Ulnar Software[mean±SD](min-max) Neurologist[mean±SD](min-max)

MeansMUPMedian[␮V] 347.45±158.21(145.9-568) 437.34±296.99(132.5–903.6)

MeansMUPUlnar[␮V] 283.26±149.37(148.3–552) 257.82±282(78.3–873.5)

MUNEMedian 28.08±14.1(16.3–53.4) 40.43±35.28(12.1–110.2)

MUNEUlnar 31.26±12.74(15.6–50.2) 70.29±40.95(13.6–141.8)

Table3

MeansMUPandMUNEValuesAcquiredbySupramaximalStimulationforALSPatients.

Patient(ALS)10Median,9Ulnar Software[mean±SD](min-max) Neurologist[mean±SD](min-max)

MeansMUPMedian[␮V] 466.37±292.84(184.4–1083) 509.26±321.85(186.8–1120.5)

MeansMUPUlnar[␮V] 373.38±165.14(157.5–619.5) 486.14±220.52(238–849.4)

MUNEMedian 17.34±11.14(3.2–40.9) 19.81±10.5(4.4–35.2)

MUNEUlnar 15.62±7.96(5.8–28.4) 17.49±8.77(5.1–30.3)

comparisonsweredonebetweentheresultsofthesoftwareand theexpertneurologist.However,fortherecordingsofsubmaximal stimulation,itwasaimedtotest theperformance ofthe devel-opedsoftwareindifferentiationbetweenthehealthyonesandALS patients.

For supramaximal stimulation, 8 median and 7 ulnar nerve recordingswereevaluatedforhealthyparticipants.sMUPvalues formedian nerve arecalculated as437.34␮Vand 347.45␮Vby anexpertneurologistandsoftware,respectively.Thesamevalues forulnarnerveare257.82␮Vand283.26␮V.MeanMUNEvalue formediannerve iscalculatedas40.43and28.08bytheexpert neurologistandthesoftwarerespectively,thisvalueisobtainedas 70.29and31.26fortheulnarnerve.Descriptivestatisticalvaluesfor healthyparticipantswhichwereacquiredfrommedianandulnar nervesaregiveninTable2.

sMUPvalueof ulnar nervein healthy controlswastheonly

parameterwhichhasasigniﬁcantcorrelationbetweentheexpert

neurophysiologistandthesoftware(p<0.05).

Ten median and 9 ulnar nerve records were evaluated for

patients.sMUPvaluesformediannervearecalculatedas509.26␮V

and466.37␮V.Thesamevaluesforulnarnervearecalculatedas

486.14␮Vand373.38␮V.MeanMUNEvalueformediannerveis

calculatedas19.81and17.34;thisvalueisobtainedas17.49and

15.62fortheulnarnerve.

DescriptivestatisticalvaluesforALSpatients’datawhichwere

acquiredfrommedianandulnarnervesaregiveninTable3.

InALSpatients’data,therewassigniﬁcantcorrelations(p<0.05)

betweenthecalculatedsMUPandMUNEvalueswhichweredone

bytheexpertneurologistandthesoftware.

Fig.5showsthecorrelationofthevaluesbythesoftwareand

theexpertneurologistwhichdepictsthesMUPandMUNEvalues

frommedianandulnarnervesoftheALSpatients.

Thefeaturesotherthan meansMUPand MUNEvalues were

unnecessarytobepresentedinthispaperbecausetherewasno

signiﬁcantcorrelationbetweenthevaluesprovidedbytheexpert

neurologistandthedevelopedsoftware.

Fivemedian and ulnarnerve recordingsin response to

sub-maximalstimulationinhealthyparticipantsandALSpatientswere

evaluatedbythedevelopedsoftware.

DescriptivestatisticalvaluesforbothhealthycontrolsandALS

patientdatawhichwereacquiredfrommedianandulnarnerves

aregiveninTable4.

MeanofsMUPvaluesforhealthycontrolsandALSpatientsin

mediannervearecalculatedas184.74␮Vand200.12␮V

respec-tively. The same values for the ulnar nerve are calculated as

154.17␮Vand241.73␮V.MeanMUNEvalueformediannervein

thehealthycontrolsandALSpatientsarecalculatedas56.2and

33.09;thisvalueisobtainedas63.87and32.62fortheulnarnerve.

The MUNE values acquired from median and ulnar nerves

weresigniﬁcantlylowerinpatientswithALScomparingtothose

acquiredfromhealthyparticipants(p<0.05).Theintergroup

dif-ferencebetweenthemeansMUPamplitudeswasonlysigniﬁcant

fortheulnarnerverecordings(p<0.05).Fortheotherfeatures

cal-culated,therewasnosigniﬁcantdifferencebetweenthehealthy

participantsandALSpatients.

4.2. PowerspectraldensityofF-wavesignals

Signalsfromapatientandahealthyvolunteerwerepresented

assuperimposedtracesinFig.6.

Thepersistence ofF-waves inhealthy volunteer was81.11%

whileitwaslowinthepatient(21.11%).Sumoftherectiﬁedtraces

werecalculatedandthepowerspectraldensity(PSD)wasobtained.

Thepointsusedin thediscreteFourier transformweresameas

thoseoftheinputsignal.Thesamplingfrequencywas20KHzfor

therecordings.ThePSDgraphforsignalsfromahealthyvolunteer

andapatientwasgiveninFig.7.

Inthehealthyvolunteer,thepowerofthesignalwashighlike

thepersistenceoftheF-waves.Frequencyspectrumyieldedawider

bandcomparingtopatient’sdata,probablyrelatedtothehigher

variability of F-wave in healthy situation. Conversely, patient’s

spectrum revealedless powerthat wasin accordancewiththe

lowpersistenceoftheF-waves.Likewisepatient’sdatashoweda

narrowbandreﬂectingthepresenceofrepeaterF-waves.

4.3. Classiﬁcationresults

Theclassiﬁcationaccuraciesforallfourdifferentclassiﬁersare

giveninTable5.

Thehighestperformance(90%)wasobtainedintheulnarnerve

withsubmaximalstimulationbyusing1-NNalgorithm.Bothofthe

mediannervewithsupramaximalstimulationbyusing5-NNand

theulnarnervewithsubmaximalstimulationbyusingMLPand

RBFgave80%classiﬁcationaccuracy.

AfterfeatureselectionwithReliefFalgorithmtheclassiﬁcation

accuracythatwasacquiredforallpreviousclassiﬁersisgivenin

Table6.

Thehighestperformance(90%)wasobtainedintheulnarnerve

withsubmaximalstimulationbyusingk-NNalgorithmforallk

val-uesrelyingonmeansMUPamplitudeandMUNEvalues.Median

andulnarnerveswithsubmaximalstimulationbyMLPnetwork

performed80%accuracy.

5. Discussion

Underthetermof“lateresponses”,F-wavesareproducedbythe

(8)

stim-Fig.5.CorrelationGraphsofsMUPAmplitudesandMUNEValuesofMedian(M)andUlnar(U)NervesforALSPatients.

Table4

CalculatedFeaturesAcquiredbySubmaximalStimulationforHealthyParticipantsandALSPatients.

(5Median,5Ulnar) HealthyControls Patient(ALS) [mean±SD](min-max) [mean±SD](min-max)

MeansMUP(M)[␮V] 184.74±45.22(138.8–238.18) 200.12±57.36(120.02–278.57)

MeansMUP(U)[␮V] 154.17±37.29(91.87–181.18) 241.73±55.43(151.92–288.86)

MUNE(M) 56.2±11.02(47.74–75.37) 33.09±17.99(9.83–53.60)

MUNE(U) 63.87±18.53(49.12–94.7) 32.62±8.25(21.46–44.76)

#baskets(M) 24±13,27(14–47) 19.6±10.43(9–35)

#baskets(U) 14.2±4.27(10–19) 17±8.46(7–30)

MeanofMaxMAmp(M)[␮V] 10,200±2185.18(7600-12500) 6920±4613.24(2100-11700)

MeanofMaxMAmp(U)[␮V] 9360±1388.52(8100–11700) 7560±1180.25(6200–9100)

Persistence(M) 0.54±0.17(0.3–0.75) 0.52±0.15(0.3-0.7)

Persistence(U) 0.4±0.15(0.24–0.62) 0.45±0.16(0.21–0.64)

PersistenceF-Rep(M) 0.52±0.11(0.34–0.65) 0.51±0.16(0.27–0.67)

PersistenceF-Rep(U) 0.53±0.23(0.26-0.78) 0.62±0.22(0.25-0.78)

MeanofF-waves’Powers(M) 3.92±2.51(1.68–7.91) 2.6±1.24(0.84–3.91)

MeanofF-waves’Powers(U) 3.91±2.18(1.1–6.13) 4.82±3.4(1.9-8.91)

MedianofF-waves’Powers(M) 3.55±2.28(1.45–7.17) 1.88±1.26(0.47–3.75)

MedianofF-waves’Powers(U) 3.62±2.03(1.03-5.96) 3.85±3.43(0.72-8.23)

Table5

ClassiﬁcationAccuracyforClassiﬁers.

AllFeatures MLP RBF SVM 1-NN 3-NN 5-NN

Supramaximal Median 55 50 45 50 70 80

Ulnar 75 56.25 50 50 50 68.75

(9)

Fig.6.SuperimposedF-WaveTracesofaHealthyVolunteer(a)andanALSPatient(b).

Table6

ClassiﬁcationAccuracyAfterFeatureSelectionforClassiﬁers.

ReliefF MLP RBF SVM 1-NN 3-NN 5-NN

Supramaximal Median 55(4) 35(3) 70(3) 60(4) 75(4) 60(4)

Ulnar 75(5) 56.25(4) 56.25(5) 56.25(5) 50(5) 50(5)

Submaximal Median_Ulnar 80₈₀(3)₍₄₎ 50₇₀(3)₍₄₎ 60₇₀₍₄₎(2) ₉₀50₍₂₎(3) ₉₀60₍₂₎(3) ₉₀50(3)₍₂₎

(10)

Fig.7.PowerSpectralDensityofF-wavesCalculatedfromaHealthyVolunteer(BlueLine)andaPatient(RedLine).

ulation.Althoughsupramaximal stimulationrecruitsmore than onelowermotorneuron,itisalsopossibletorecordsingleaxon’s responsewithsubmaximalstimulation.Theshape,latencyandthe persistenceofF-wavesaretheindicatorsofdiseasestatesaffecting thelowermotorneuronanditsaxon[2,3,19].F-wavesareuseful

fortheestimationoffunctioninglowermotorneuroncountaswell

[15,17].

InordertogetmeaningfulresultsinF-wavestudies,itis

essen-tialtodeliveratleast20stimuliforsupramaximalandsometimes

hundredsofstimuliforsubmaximalstimulation[16,36].Manuel

analysisofthetracesharboringF-wavesiscumbersomeandthere

isalwaysapossibilityofmisrecognitionandfaultyinterpretation.

Automatedanalysissoftware,suchastheonewhichisdeveloped

inthepresentstudy,isaneedtoovercomethetroubledsideof

naked-eyeevaluation.

ScientistswerecuriousabouttheautomatedanalysisofF-wave

parameterssince1990s.Intheirstudy,Stashuketal[15]reported

analgorithmin1994,whichwasdevelopedforcalculating

maxi-mumMresponseamplitude,summingF-waves,calculatingmean

ofS-MUAPandestimatingthenumberofmotorunitsina

mus-cle.Theydemonstratedthattheestimatednumberofmotorunits

inthenarmusclescalculatedbytheirdevelopedsoftwarewhich

selectsS-MUAPsfromF-responsesautomaticallywasalmost

iden-ticaltotheonecalculatedbymanually.Theestimatednumberof

motorunitsbyusingF-wavesyieldedsimilarvalueswithanother

validated MUNE method, namely “multiple point stimulation”.

Althoughthealgorithm performs perfectly well,someoperator

controlmightbeneededforrecruitingtheS-MUAPsincasethey

couldnotbeselectedbythesoftwareforthecalculationofmean

S-MUAPvalue.

Fouryearslater,Felice,KJ[17]calculatedMUNEvalueofthenar

muscles in patientswith ALS byusing the software developed

by Stashuk et al [15] and demonstrated that the patients had

signiﬁcantly low number of motor units comparingto healthy

participants.Sotheycarriedastepforwardbyshowingthatthe

Stashuk’salgorithmworksforthepatientssufferingfrommotor

neurondisease.

Morris A. Fisher questioned the accuracy of an automated

methodforthemeasurementofF-wavelatency.Inhispaper[37]

publishedin2005,hestudiedtheF-wavesof80peronealnerves

andshowedthatthecomputersoftware(NEUROMetrix)picksthe

correctsitesforlatencymeasurementofF-waveswhichwerein

concordancewiththeonesselectedmanually.

Inastudywherethereproducibilityofnerveconductionstudies

wasquestioned,Kongetal.[38]performedserialmeasurements

ofconduction parametersand developedsoftwarewhich

calcu-latesF-wavelatencyautomaticallyincombination withvarious

other features. Among them, automaticallycalculated mean

F-wavelatencywasthemostreproducibleone.

InanotherstudyofKongetal.[39],theauthorscomparedthe

F-waveparametersofperonealnervebothacquiredinresponseto

supraandsubmaximalstimulationforthesakeofpatient’s

com-fort.TheyusedautomatedalgorithmsforthecalculationofF-wave

parameterssuchastheamplitude,persistence,latency,duration

andchronodispersion.Itisclearthatthesubmaximalstimulation

is welltolerated.Althoughtheamplitudeand persistence were

highwith supramaximal stimulation,the latency,duration and

chronodispersionofF-waveswerealmostsimilar.

Chronietal.[14]conductedaretrospectivestudydealingwith

the quantiﬁed parameters of F-waves in 2012. They formed a

datasetwith50 healthyparticipantsand patientswithdiabetic

polyneuropathy, ALS,carpal tunnel syndrome, ulnar

mononeu-ropathyand L5 rootlesion.Each groupofpatientsconsistedof

50participants.Theystimulatedmedian,ulnarandﬁbularnerves

supramaximally20timesandrecordedMandF-wavesfrom

abduc-torpollicisbrevis,abductordigitiminimiandextensordigitorum

brevismusclesrespectively.Amongtheparameterscalculatedfrom

theirdataset,thosedealingwiththeF-wavesweremeanlatency,

maximumamplitude,persistence,repeaterneuronindex,total

F-repeaterindex,totalF-repeaterpersistence,non-repeaterF-wave

persistence,repeaterneuron meanlatencyand repeaterneuron

maximumamplitude.Comparingtohealthyparticipants,patients’

datarevealedsigniﬁcantlyhigherpercentagesofrepeaterneuron

(11)

decreasednumbersofmotorneuronscontributingtoF-waves.The

maximumamplitudeofrepeaterneuronswaslowerthanthe

F-wavemaximumamplitudeformedianandﬁbularnervesinhealthy

participantsandforallthreenervesinpatients.Thisﬁndingwasin

contrasttothestudyconductedbyGuiloffandModarres-Sadeghiin

1991[40].Theauthorsdiscussedthisﬁndingasthesmaller

ampli-tudeofrepeaterneuronsreﬂectsthebackﬁringofsingleneurons

ratherthanbeingacombinedactivity.

In2017Chronietal.[29]developedacustomdesignedsoftware.

Theabilitiesofthissoftwarewereplottingtherecordings,detecting

F-waves,determiningtherepeatersandextractingtheirfeatures.It

alsoallowedsettingthepointsmanuallyinwhichthesignalleaves

andreturnstobaseline.However,theiralgorithmhassome

dif-ferencesfromours.Theyusedlowpassﬁlteringtosuppresshigh

frequencycomponentsofthesignal.Moreover,theyused

deriva-tivewithalowpassﬁltertoremovethespikeswhichwerecaused

bytheoutliersofthesignals.Ontheotherhand,ouralgorithmuses

waveletbasednoisereductiontosmooththerecordingsandto

dis-cardnoisycomponents.Theiralgorithmfordeterminingrepeater

F-wavesisstricterthanours.Theydeterminedatoleranceband

of±0.05msforallnegativeandpositivepeaks,onsetandreturn

tobaseline ofthesignals.In additiontothose rules,their

algo-rithmhadanareabasedcriteria.Theyalsoseta25␮Vdifference

limittominimumandmaximumamplitudesforsignalpairs.Inour

study,weaimedtodevelopanalgorithmthatdecidesby

consider-ingminorvariabilityamongthesubjectsintoaccount.Firstofall,

ouralgorithmcalculatesthesignal’speaktopeakamplitude.Ifthis

valueislowerthan40␮V,thesignalismarkedasnoiseandisﬂoored

tolevel0.ThenitcheckstheFminandFmaxlocationsiftheyare

closetoeachotherina0.5mswindow.Afterthisstep,ifthe

ampli-tudedifferenceforFmaxvaluebetweensignalpairsislowerthan

10%anddifferenceofindividualpowervaluesbetweensignalpairs

islowerthan20%andlastlythecorrelationcoefﬁcientbetween

themisgreaterthan0.9theyaremarkedascandidates.We

deter-minedaparameternamedassimilaritycoefﬁcientwhichconsists

ofamplitudedifferenceandpowerdifferenceofsignalpairs.Iftheir

sumisbelowtheempiricallydeterminedthreshold,thesignalsare

markedasrepeaters.Afterthat,asecondinspectionisdoneandit

isgivenelaboratelyinPart3“Theory/Calculation”andin

summa-rizedformatFig.4.ThealgorithmofChronietal.andourstudy

grouprepeaterssimilartoeachother.Theiralgorithmadditionally

assignsacolorforeachrepeaterF-wave.

Inthepresentstudy,theresultsofthedevelopedsoftware

con-cerningsMUPandMUNEvaluesacquiredbysupramaximalstimuli

werefoundtobecompatiblewiththeobservationsoftheexpert

neurologist.Interestingly,thisconcordancewasmoreprominent

inALSpatientsthatcanbeexplainedbyreducedavailable

num-bersoflowermotorneuronsforF-wavegeneration.Supramaximal

stimulationmightalsocontributetolowernumbersofMUNE

val-uesbothinhealthyvolunteersandthepatients,byincreasingthe

sMUPamplitudeaswell.Nevertheless,meansMUPamplitudesfor

medianandulnarnerveswereclosetothevaluesofChroni’sstudy

[14].Theycalculatedthemaximumamplitudeofrepeaterneurons

as0.4and0.2mVformedianandulnarnervesinhealthy

individ-uals,respectively.Thesevalueswerefound0.35and0.28mVby

thedevelopedsoftwareformedianandulnarnervesinthepresent

study.ForpatientswithALS,Chroniatal[14].measuredthe

maxi-mumamplitudeoftherepeaterneuronsas0.5mVformedianand

0.4mVforulnar nerve,whileweautomaticallycalculatedmean

sMUPamplitudes as0.47mVand 0.37mVformedianandulnar

nervesrespectively.

TheMUNEvaluescalculatedfromsubmaximalstimulusdataset

revealedthattheALSpatientshadsigniﬁcantlydecreasednumbers

ofmotorunitsinbothoftheirmuscles,namelyabductorpollicis

brevisandabductordigitiminimi.F-waveswhichareproducedin

responsetosubmaximalstimulusarethoughtthattheyare

com-ingfromasinglemotorneuroniftheyrepeatatthesamelatency

andinthesameshape[15].MeansMUPamplitudewascalculated

fromtheserepeaterF-waveswhichwereautomaticallychosenby

thedevelopedsoftware.ThecalculatedmeansMUPamplitudewas

higherforboth musclesinALSpatientsalthoughthedifference

betweenhealthyvolunteersandpatientswasonlystatistically

sig-niﬁcantforulnar nerve. Theincrease inmean sMUPamplitude

mightbeareﬂectionofreinnervationbycollateral sproutingin

affectedmuscles.MUNEvalueissimplycalculatedbydividingthe

M-responseamplitudebymeansMUPamplitude,sothata low

M-responseamplitudeincombinationwithhighsMUPamplitude

yieldeddecreasedMUNEvalueinpatients’groupaﬁndingwhich

mightbeexpected.

FindingsoflowerMUNEvalue,decreased M-response

ampli-tudeandhighsMUPamplitudeinpatientswithALSisnotsurprising

becauseofALSischaracterizedbyprogressivelossoflowermotor

neuronsandreinnervationbycollateralsproutingfromthe

surviv-ingones.However,comparingtopreviousstudies,MUNEvalues

calculatedinthepresentstudywerealsolowforbothmusclesof

thehealthysubjectsaswell,aﬁndingthatisnoteworthyto

dis-cuss.Previousstudiesreportedtheestimatednumbersofmotor

unitsaremorethantwohundredinhandmuscles[15,17,41].The

estimatedvaluesofmotorunitsinthepresentstudywerewithin

48to75forthemediannerveand49to95fortheulnarnerve.The

underlyingmechanismoflowMUNEvaluesinthisstudymightbe

relatedtotherecordingtechniqueoftheM-responseand/or

prefer-entiallyselectingthehigh-amplitudesMUPsforcalculation[42,43].

Nevertheless,techniquesdealingwithmotorneuroncountingare

indirectandtheygiveonlyestimatedvalues[44–47].MUNEis

valu-ablefordifferentiatingpatientsfromhealthyindividualsbutalso

itmakesa greatcontributionforpatient’sfollowupbytracking

thechangesintheMUNEvaluesoverthetime[48–51].Following

thepatientbyusingthesamemethodallowsforaccurate

predic-tionofthedeclineinlowermotorneuroncount,althoughitisjust

estimationratherthanrevealingtherealnumber.

A classiﬁcation studywas also donetodetermine themost

effective features for differentiating ALS patients from healthy

volunteers.Extractedfeatureswereappliedtovariousintelligent

systemssuchasMLP,RBF,SVMandk-NN.First,alloftheextracted

features andthen, decreased numbersof featureswereapplied

whichweredepictedbyfeatureselection algorithms.Theulnar

nerve recordings with submaximal stimulation yielded highest

performance.Thedatasetcomposedbysupramaximalstimulation

didnotshowagreataccuracy(maximum75%).Howeverfor

sub-maximalstimulation,MLPnetworkrevealed80%accuracywhen3

or4featuresselectedformedianandulnarrecordingsrespectively.

Featureselectionincreasedtheclassiﬁcationperformancenearly

forallclassiﬁers.Acommitteedecisionbasedclassiﬁerk-NNwas

superiortoallothersafterfeatureselectionatulnarnerve

record-ingsundersubmaximalstimulation.Thehighestperformancewas

obtainedas90%withonlytwofeatures.Themostvaluable

fea-tureswhichhelpdifferentiationoftwoclassesweremeanofsMUP

amplitudeandMUNEvalue.

Thesuperioraspectsofthepresentstudyare(1)algorithmwas

testedonbothhealthyindividuals andpatients,(2) comparison

withtheresultsofaneurologist(asforthegoldstandard)wasdone

totesttheperformanceofthedevelopedalgorithm,(3)original

ﬁl-tersettingsfromtheEMGinstrumentwerekeptasthesamesothe

frequencycomponentsoftherecordedsignalswerenotdistorted.

Ontheotherhandtherearesomedrawbackssuchas(1)numbers

ofthecasesandthevarietyofdiseasegroupsareneededtobe

increased,(2)thealgorithmisplannedtobefullyautomaticfor

F-waveanalysisbutthereisanalgorithminastudy[15]whichallows

usertoselectrepeaterF-wavesmanuallyaswell.Thedeveloped

algorithmhasacomprehensiveviewersoanoptionmaybeadded

(12)

softwarewastested withsupramaximalrecordings.The results

wereinconcordancewiththeexpertneurologistformeansMUP

andMUNEvalue.Adatasetwasformedwithsupramaximaland

submaximal recordings. Some features were extracted from

F-wavestoclassifypatientsandhealthyindividualswithintelligent

systems.Featureselectionalgorithmswereappliedtodetermine

themost effective features. The classiﬁcationperformance was

increasedinmanyclassiﬁerswhenthefeaturesweredecreased.

Theulnarnerverecordingsshowedhighestperformanceas90%

withk-NNalgorithm whenthestimulusstrengthwas

submaxi-mal.ThemostvaluablefeaturesweremeansMUPamplitudeand

MUNEvalue.

Acknowledgements

Theauthorswouldliketoacknowledgethevaluable

contribu-tionoftheirneurologistcolleagueswhoreferredthepatients,and

areespeciallygratefultoElifKocasoyOrhanandBaharErbasfor

theirhelpduringdatacollection.

References

[1]J.W.Maghdery,D.B.McDougal,Electrophysiologicalstudiesofnerveand reflexinnormalman.I.Identificationofcertainreflexesinthe

electromyogramandtheconductionvelocityofperipheralnerveﬁbers,Bull. JohnsHopkinsHosp.86(1950)265–290.

[2]F.Mesrati,M.F.Vecchierini,F-waves:neurophysiologyandclinicalvalue, Neurophysiol.Clin.34(2004)217–243.

[3]M.A.Fisher,F-waves–physiologyandclinicaluses,Sci.WorldJ.7(2007) 144–160.

[4]G.D.Dawson,P.A.Merton,Recurrentdischargesformotoneurons,Proceedings oftheSecondInternationalCongressofPhysiologicalScience(1965)221. [5]J.Thorne,Centralresponsestoelectricalactivationoftheperipheralnerves

supplyingtheintrinsichandmuscles,J.Neurol.Neurosurg.Psychiatry28 (1965)482–495.

[6]C.H.Wulff,R.W.Gilliatt,F-wavesinpatientswithhandwastingcausedbya cervicalribandband,MuscleNerve2(1979)452–457.

[7]M.A.Fisher,Fresponselatencydetermination,MuscleNerve5(1982) 730–734.

[8]J.H.Petajan,F-wavesinneurogenicatrophy,MuscleNerve8(1985)690–696. [9]S.Peioglou-Harmoussi,F.R.Fawcett,D.Howel,D.D.Barwick,F-response

frequencyinmotorneurondiseaseandcervicalspondylosis,J.Neurol. Neurosurg.Psychiatry50(1987)593–599.

[10]W.N.Macleod,RepeaterFwaves:acomparisonofsensitivitywithsensory antidromicwrist-to-palmlatencyanddistalmotorlatencyinthediagnosisof carpaltunnelsyndrome,Neurology37(1987)773–778.

[11]C.J.Argyropoulos,C.P.Panayiotopoulos,S.F.Scarpalezos,FandM-wave conductioninamyotrophiclateralsclerosis,MuscleNerve1(1978)479–485. [12]R.F.Mayer,R.G.Feldman,ObservationsofthenatureoftheFwaveinman,

Neurology17(1967)147–156.

[13]M.Zappia,P.Valentino,L.P.Marchello,M.Panniccia,P.Montagna,F-wave normativestudiesindifferentnervesofhealthysubjects,Electromyogr.Clin. Neurophysiol.89(1993)67–72.

[14]E.Chroni,I.S.Tendero,A.R.Punga,E.Stålberg,Usefulnessofassessingrepeater F-wavesinroutinestudies,MuscleNerve45(2012)477–485.

[15]D.W.Stashuk,T.J.Doherty,A.Kassam,W.F.Brown,Motorunitnumber estimatesbasedontheautomatedanalysisofF-responses,MuscleNerve17 (1994)881–890.

latenciesandconductionvelocitiesinsinglemotorﬁbres,MuscleNerve15 (1992)1204.

[24]T.E.Feasby,W.E.Brown,Variationofmotorunitsizeinthehumanextensor digitorumbrevisandthenarmuscles,J.Neurol.Neurosurg.Psychiatry37 (1974)916–926.

[25]S.K.Yates,W.E.Brown,CharacteristicsoftheFresponse:asinglemotorunit study,J.Neurol.Neurosurg.Psychiatry42(1979)161–170.

[26]L.Puksa,E.Stålberg,B.Falck,ReferencevaluesofFwaveparametersin healthysubjects,Clin.Neurophysiol.114(2003)1079–1090.

[27]M.A.Fisher,ThecontemporaryroleofF-wavestudies.F-wavestudies:clinical utility,MuscleNerve21(1998)1098–1101.

[28]A.Hachisuka,T.Komori,T.Abe,K.Hachisuka,RepeaterF-wavesaresignsof motorunitpathologyinpoliosurvivors,MuscleNerve51(2015)

680–685.

[29]E.Chroni,D.Veltsista,C.Papapaulou,E.Trachani,GenerationofrepeaterF wavesinhealthysubjects,J.Clin.Neurophysiol.34(2017)236–242. [30]J.Kamel,R.Knight-Sadler,M.Cook,L.Roberts,Single-ﬁberFwavescompared

withconventionalsurfaceFwaves,andtheirutilityindetectingearlydiabetic neuropathy,MuscleNerve58(2018)665–670.

[31]M.deCarvalho,R.Dengler,A.Eisen,J.D.England,R.Kaji,J.Kimura,K.Mills,H. Mitsumoto,H.Nodera,J.Shefner,M.Swash,Electrodiagnosticcriteriafor diagnosisofALS,Clin.Neurophysiol.119(2008)497–503.

[32]ALSCNTFTreatmentStudy(ACTS)PhaseI-IIStudyGroup,Theamyotrophic lateralsclerosisfunctionalratingscale.Assessmentofactivitiesofdailyliving inpatientswithamyotrophiclateralsclerosis,Arch.Neurol.53(1996) 141–147.

[33]S.J.Oh,ClinicalElectromyography:NerveConductionStudies,thirded., LippincottWilliams&Wilkins,Philadelphia,2003.

[34]K.Kira,L.A.Rendell,Apracticalapproachtofeatureselection,Proceedingsof InternationalConferenceonMachineLearning(1992)249–256.

[35]M.Robnik-Sikonja,I.Kononenko,TheoreticalandempiricalanalysisofReliefF andRReliefF,Mach.Learn.J.53(2003)23–69.

[36]E.Chroni,N.Taub,C.P.Panayiotopoulos,Theimportanceofsamplesizefor theestimationofFwavelatencyparametersintheulnarnerve,MuscleNerve 17(1994)1480–1483.

[37]M.A.Fisher,ComparisonofautomatedandmanualF-wavelatency measurements,Clin.Neurophysiol.116(2005)264–269.

[38]X.Kong,E.A.Lesser,J.T.Megerian,S.N.Gozani,Repeatabilityofnerve conductionmeasurementsusingautomation,J.Clin.Monit.Comput.20 (2006)405–410.

[39]X.Kong,P.Bansal,J.T.Megerian,S.N.Gozani,PeronealF-wavecharacteristics undersubmaximalstimulation,Neurol.Neurophysiol.Neurosci.1(2006) 1–13.

[40]R.J.Guiloff,H.Modarres-Sadeghi,Preferentialgenerationofrecurrent responsesbygroupsofmotorneuronsinman,Brain114(1991)1771–1801. [41]T.J.Doherty,W.F.Brown,Theestimatednumbersandrelativesizesofthenar

motorunitsasselectedbymultiplepointstimulationinyoungandolder adults,MuscleNerve16(1993)355–3661.

[42]M.deCarvalho,P.E.Barkhaus,S.D.Nandedkar,M.Swash,Motorunitnumber estimation(MUNE):Wherearewenow?Clin.Neurophysiol.129(2018) 1507–1516.

[43]T.J.Doherty,D.W.Stashuk,W.F.Brown,Determinantsofmeanmotorunit size:impactonestimatesofmotorunitnumber,MuscleNerve16(1993) 1326–1331.

[44]M.P.Slawynch,C.A.Laszlo,C.Hershler,Areviewoftechniquesemployedto estimatethenumberofmotorunitsinamuscle,MuscleNerve13(1990) 1050–1064.

[45]M.B.Bromberg,Updatingmotorunitnumberestimation(MUNE),Clin. Neurophysiol.118(2007)1–8.

[46]C.L.Gooch,T.J.Doherty,K.M.Chan,M.B.Bromberg,R.A.Lewis,D.W.Stashuk, M.J.Berger,M.T.Andary,J.R.Daube,Motorunitnumberestimation:a technologyandliteraturereview,MuscleNerve50(2014)884–893. [47]H.Bostock,EstimatingmotorunitnumbersfromaCMAPscan,MuscleNerve

53(2016)889–896.

[48]S.Grimaldi,L.Duprat,A.M.Grapperon,A.Verschueren,E.Delmont,S. Attarian,Globalmotorunitnumberindexsumscoreforassessingthelossof

(13)

lowermotorneuronsinamyotrophiclateralsclerosis,MuscleNerve56 (2017)202–206.

[49]A.B.Jacobsen,R.S.Kristensen,A.Witt,A.G.Kristensen,L.Duez,S.Beniczky,A. Fuglsang-Frederiksen,H.Tankisi,Theutilityofmotorunitnumberestimation methodsversusquantitativemotorunitpotentialanalysisindiagnosisofALS, Clin.Neurophysiol.129(2018)646–653.

[50]J.Furtula,B.Johnsen,P.B.Christensen,K.Pugdahl,C.Bisgaard,M.K. Christensen,J.Arentsen,M.Frydenberg,A.Fuglsang-Frederiksen,MUNIXand incrementalstimulationMUNEinALSpatientsandcontrolsubjects,Clin. Neurophysiol.124(2013)610–618.

[51]J.M.Shefner,M.L.Watson,L.Simionescu,J.B.Caress,T.M.Burns,N.J. Maragakis,M.Benatar,W.S.David,K.R.Sharma,S.B.Rutkove,Multipoint incrementalmotorunitnumberestimationasanoutcomemeasureinALS, Neurology.77(2011)235–241.

Tu˘grulArtu˘gis anassistantprofessorin the Depart-mentofElectricalandElectronicsEngineeringatIstanbul ArelUniversity.Hehaseightyearsofacademic expe-rience at different universities. Artu˘gholds a BSc in ElectronicsandCommunicationEngineeringfromKocaeli University.HereceivedhisMScandPhDdegreesin Elec-tronicsfromYildizTechnicalUniversityin2010and2015 respectively.Hisresearchinterestsarebiomedicalsignal processing,imageprocessingandneuralnetworks.Heis alsoexperiencedinmicrocontrollers.Hehasparticipated ininternationalsymposiumsandsubmittedconference paperstoseveralnationalcongresses.

NerminGorkemSirinisworkingasafellowinclinical neurophysiologydepartmentinMarmaraUniversitysince 2017.Shewasgraduatedfrommedicalschoolin2006and startedNeurologyresidencyinthesameyear.After ﬁn-ishingherresidency,sheworkedinCizreStateHospital asaneurologistforherobligatoryservicebetween2011 and2013.Afterﬁnishingherservice,sheworkedin sev-eralhospitalsinIstanbulasaneurologist.Shehadher MasterofSciencedegreeinclinicalneurophysiologyin IstanbulUniversityinJune2018.Heracademicinterests areclinicalneurophysiology,neuromusculardiseasesand electromyography.

EmelOguzAkarsuwasborninSilifke/MersinonApril 1982. She graduatedfrom IstanbulUniversity Cerrah-pasaSchoolofMedicine(EnglishProgram)in2007.She completedherresidencyprograminHasekiTrainingand ResearchHospitalin2011andshecompletedIstanbul UniversityIstanbulMedicalFacultyClinical Neurophys-iologyfellowshipprogramin2017.Shehasbeenworking asaclinicalneurophysiologistinSakaryaUniversity Train-ingandResearchHospitalsince2017.Sheisinterestedin epilepsy,neuromusculardisordersandelectrophysiology.

M.BarisBaslowasborninOctober1970inIstanbul. HegraduatedfromIstanbulUniversityMedicalFacultyin 1993.Hecompletedhisresidencyprograminthe Neurol-ogyDepartmentofthesamefacultyin1998.Hestudied neuroscienceatMayoClinicbetweenAugust1996and July1997.Heworkedasaconsultantneurologistuntil 2003whenhebecameanassociateprofessor.In2009, hewaspromotedtoProfessorofNeurologyatthesame departmentandtwoyearslaterhegothisdegreein “Clin-icalNeurophysiology”ashissubspecialty.Hisresearch mainlyfocusesonelectrophysiologyofneuromuscular disorders.

A.EmreÖgeisaprofessorinthedepartmentsof Neu-rologyandClinicalNeurophysiology,IstanbulUniversity, IstanbulFacultyofMedicine.Hismainareasofinterestand researchareperipheralnervediseases,facialnerve phys-iology, quantitativeelectromyographyincludingmotor unitnumberestimationstudiesandmagneticstimulation methodsintendedforproximalperipheralnerve disor-ders,corticalexcitabilityandsensory-motorintegration.