entropy Article Automatic Analysis of Archimedes’ Spiral for Characterization of Genetic Essential Tremor Based on Shannon’s Entropy and Fractal Dimension KarmeleLopez-de-Ipina1,* ID,JordiSolé-Casals2 ID,MarcosFaúndez-Zanuy3 ID, PilarM.Calvo1,EnricSesa3,JosepRoure3,UnaiMartinez-de-Lizarduy4,BlancaBeitia5, ElsaFernández1,JonIradi6,JosebaGarcia-Melero7andAlbertoBergareche8 1 SystemsEngineeringandAutomationDepartment,EleKinResearchGroup,UniversityoftheBasque CountryUPV/EHU,20018Donostia,Spain;[email protected](P.M.C.); [email protected](E.F.) 2 DataandSignalProcessingResearchGroup,UniversityofVic—CentralUniversityofCatalonia, 08500Barcelona,Spain;[email protected] 3 EscolaSuperiorPolitècnicaTecnocampus,UniversitatPompeuFabra,08002Barcelona,Spain; [email protected](M.F.-Z.);[email protected](E.S.);[email protected](J.R.) 4 DepartmentofElectronicTechnology,EleKinResearchGroup,UniversityoftheBasqueCountryUPV/EHU, 20018Donostia,Spain;[email protected] 5 DepartmentofMathematics,EleKinResearchGroup,UniversityoftheBasqueCountryUPV/EHU, 1006Gasteiz,Spain;[email protected] 6 DepartmentofEnterprisesOrganization,EleKinResearchGroup,UniversityoftheBasqueCountry UPV/EHU,20018Donostia,Spain;[email protected] 7 DepartmentofMechanicalEngineering,EleKinResearchGroup,UniversityoftheBasqueCountry UPV/EHU,1006Gasteiz,Spain;[email protected] 8 DepartmentofNeuroscience,BioDonostiaHealthInstitute,20014Donostia,Spain; [email protected] * Correspondence:[email protected];Tel.:+34-943018667;Fax:+34-943017230 (cid:1)(cid:2)(cid:3)(cid:1)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1) (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7) Received:20May2018;Accepted:11July2018;Published:16July2018 Abstract: Amongneuraldisordersrelatedtomovement,essentialtremorhasthehighestprevalence; infact,itistwentytimesmorecommonthanParkinson’sdisease. ThedrawingoftheArchimedes’ spiralisthegoldstandardtesttodistinguishbetweenbothpathologies. Theaimofthispaperisto selectnon-linearbiomarkersbasedontheanalysisofdigitaldrawings. Itbelongstoalargercross studyforearlydiagnosisofessentialtremorthatalsoincludesgeneticinformation. Theproposed automaticanalysissystemconsistsinahybridsolution: MachineLearningparadigmsandautomatic selectionoffeaturesbasedonstatisticaltestsusingmedicalcriteria.Moreover,theselectedbiomarkers comprisenotonlycommonlyusedlinearfeatures(staticanddynamic),butalsoothernon-linear ones: ShannonentropyandFractalDimension. Theresultsarehopeful,andthedevelopedtoolcan easilybeadaptedtousers;andtakingintoaccountsocialandeconomicpointsofview,itcouldbe veryhelpfulinrealcomplexenvironments. Keywords: essentialtremor;automaticanalysisofdrawing;spiralofArchimedes;entropy;fractal dimension;automaticselectionoffeatures 1. Introduction Essentialtremor(ET)isaworldwidediseasethatistwentytimesmorecommonthanParkinson’s disease(PD)[1]. Inthewesternworld,theprevalenceofthismotordisorderisabout0.3%to4.0%. Itsincidenceis23.7per100,000individualsperyear,andbothmenandwomenof40yearsoldare Entropy2018,20,531;doi:10.3390/e20070531 www.mdpi.com/journal/entropy Entropy2018,20,531 2of15 affectedmoreorlessinthesameproportion;therearenodifferenceswithregardtogender. Studies pointoutthat50–70%ofETcasesareestimatedtobegenetic[1]. ETisarhythmictremor(4to12Hz) whichonlymanifestswhenthemuscleisexerting,andthetremor’samplitudevariesasthepatient becomesolder. ETischaracterizedbybothkineticandposturaltremor,andmostofthetimesitaffects thehands. Almosteveryothertremoroccurscombinedwithhandtremor[2]. Theclinicaldiagnosisofthefirstmanifestationsofthediseaseiscrucialtopalliateandmanage thesymptoms,sincetheyproduceproblemsintheperformanceindailyactivities. Inrecentyears wehavewitnessednotoriousadvancesinearlydiagnosistechniquesandinthecreationofusable clinical markers. The differential diagnosis between PD and ET is clinical. When the diagnosis is doubtful, a SPECT-DAT scan can be performed, which consists in evaluating the striatal uptake of 123I-2b-carbomethoxy-3b-(4-iodophenyl)-N-(3-fluoropropyl)-nortropane using a SPECT (Single PhotonEmissionComputedTomography),atestthatisextremelyexpensive,andthereforenoteasily accessibleinhabitualclinicalpractice[3]. Inspiteoftheusefulnessofthesebiomarkers,theexpenses andrequirementsoftheneededtechnologymakeitunfeasibletousethesetestswitheverypatientthat hasmotorconditions. Inthiscontext,intelligenttechniquesthatarenon-invasiveforthediagnosiscan beveryusefultoolsforearlydiagnosisofneurodegenerativedisorders. Staffwhoarenon-technical intheenvironmentofthepatientcouldbeabletousethesetechniquesandotherbiosignalswithout interferingwiththepatient’sdailylife,asanalysisofspeech,analysisofhandwritingoranalysisof drawingsarenotconsideredstressfulbymostpatients. Furthermore,thecostofthesetechniquesis lowandtheirrequirementsregardinginfrastructureormedicalequipmentarenotextensive,whereas theyprovideinformationinaneasy,quickly,andinexpensiveway[4–8]. Medicalstaffcommonlyuse handwritingtasksfordiagnosingessentialtremor. Nowadays,Archimedes’spiralisthegoldstandard testinclinicaldiagnosis[9]. The analysis of handwriting has traditionally been performed off-line, as only the strokes on the paper were available for the analysis. Nowadays, new devices, such as digitizing tablets and pens (with ink or without), can collect both written data and their temporal information (on-line handwritinganalysis). Newdigitizingtabletsgathernotonlythexandycoordinatesofthemovement of the device of writing, but they also gather other data, for example: the pressure on the surface of the writing device, the azimuth (angle between the pen and the horizontal plane), the altitude (anglebetweenthepenandtheverticalaxis),oreventhein-airmovementwhenthereisnocontact betweenthepenandthewritingsurface. Allthesecapabilitiesmakeit. possibletoanalyzebothstatic anddynamicfeatures[10,11]. Inthefieldofmedicine,theanalysisofhandwritinghasfoundtobeusefultomakeadiagnosisand trackneurodegenerativediseases. Forexample,thedegradationofhandwritingskillsandAlzheimer’s Disease(AD)seemtobecloselyrelated[12,13],andseveralhandwriting-relatedaspectscanbeusedas indicatorsfordiagnosingit[12],orcoulddifferentiatemildADfrommildcognitiveimpairment[12,13]. Handwritinganalysishasalsobeenshowntobehelpfultostudyhowsubstancessuchasalcohol[14], marijuana[15],orcaffeine[16]affectpatients. Psychologyhasbenefited,aswell,fromtheanalysis ofhandwritingthankstonewdevicesofacquisition. Forexample,Rosenblumetal.[17]reportthe relationbetweentheabilityofthewritersandthedurationofthein-airtrajectoriesofthehandwriting. Theprogressivedegreeofimpairmentcanbenoticedbyvisuallyinspectingthepen-downimage, whenthedrawingismoredisorganized,andtheeffectinthreedimensionsisonlynoticedformild casesofETandPDpatients. TherearecurrentlyreferencesthatanalyzethedetectionofETbymeans ofdigitalplatforms. In[18],newnumericaltechniquesareproposedinordertoevaluatethemwith regardtotheagreementwithvisualratingandreproducibility. Forexample,[19]analyzestheincrease inthevariabilityofthedrawingofthespiralinpatientswithfunctional(psychogenic)tremor. Louise presentsaninterestingstudyoverfourcohortsofET[20]. Ontheotherhand,recentreferences[21–24] presentadvancesinPDanalysisbymeansofdigitaldevicesaswell. Thevisualinformationfromthe in-airtrajectoriesofthedrawingsofunhealthypeoplealsoshowsdisorganizationandaprogressive impairmentwhenthosepeopletrytoplantheirdrawings.Itisworthnoticingthatifpen-upimagesare Entropy2018,20,531 3of15 comparedbetweenunhealthyindividualsandsubjectsinthecontrolgroup,italsorevealsimportant differenEcnetrsop.yI 2n01a8d, 2d0,i xt io ntoalongerin-airtime,therearemorehand-movementsbeforethey3p ouf 1t5t he penbackonthesurface. Therefore,thesemeasuresofgraph-motortypecanbeusefulwhenapplied unhealthy individuals and subjects in the control group, it also reveals important differences. In totheanalysisofwritinganddrawinginordertoanalyzetheprecisenatureandprogressionofthe addition to a longer in-air time, there are more hand-movements before they put the pen back on the writingsaunrfdacder. aTwheirnegfodreis, othredseer msreealsautreeds otof gnreauphro-mdeogtoern teyrpaet icvaen dbeis uesaesfeusl. when applied to the analysis of Thwerwitinogrk anpdr edseranwteindg inin tohridserp atop earnailsyzpea trhteo pfraecilsaer gneartucrreo asnsds tpurdogyreasismioend ofa tthdei awgrnitoinsgin agnda nd charactderraizwiningg edsissoerndteiarsl rtreelamteodr .toT nheiusrroedseegaerncehraitsivceo dnidseuacsteesd. attheBiodonostiaHealthInstitute,and it analyzesTfahme iwlioersk wpirtehseindteedn tiinfi etdhisg epnapeteirc ilso pciar[t2 5o,f2 6a ].larNgeerv cerrothsse lsetussd,yt haeimsetdu dayt dpiraegsneonsitnegd ainndt he characterizing essential tremor. This research is conducted at the Biodonostia Health Institute, and it manuscript is aimed at the characterization of essential tremor, and its objectives are focused on analyzes families with identified genetic loci [25,26]. Nevertheless, the study presented in the establishing the limits related to the natural variability of ET and the ability to discriminate ET manuscript is aimed at the characterization of essential tremor, and its objectives are focused on from physiological tremor. Therefore, the subject of study is not the ability to make a differential establishing the limits related to the natural variability of ET and the ability to discriminate ET from diagnosiswithPD.Regardingthegatheringofnon-linearbiomarkersfromhandwritinganddrawings, physiological tremor. Therefore, the subject of study is not the ability to make a differential Archimedes’spiralisthetesttobeexplored. Thistestisalsoabletodetectotherdiseasessuchasstress, diagnosis with PD. Regarding the gathering of non-linear biomarkers from handwriting and whichcdoruawldinagpsp, Aearrchinimceodnetsr’o slpiinradl iivs itdhue atelss.t Iton bthe eexnpelxotreSde.c Ttihoisn ste,swt ies walsilol aabnlae ltyoz deenteoctt oonthleyr cdoimseamseosn ly used dysuncahm aisc satrnesds,s wtahtiicchl icnoeualdr afepapteuarre isn, cbounttrootl hinedrinviodnu-allisn. eIanr thoen nesexat sSewcteiolln,si,. ew.,e fwrailclt aanladlyimzee nnosti on and entornolpy yc.omPmotoennltyi aulsebdio dmynaarmkeicr sanwdi slltabtiec liidneeanrt iffieaetdurbesy, bmuet aonthsero nfosne-vlienreaalr aounetos masa wtiecll,c il.aes.,s firfiaccatatli on techniqduiems.enLsaisotnly ,atnhde eqnutaroliptyy.o Pfothteenstiealle cbtieodmfaerakteurrse swwilli lbleb eidaeunttoifmieadt ibcayl lmyeaannasl yozfe dsebveyrauls ianugtoMmaactihci ne classification techniques. Lastly, the quality of the selected features will be automatically analyzed Learningparadigms. by using Machine Learning paradigms. 2. Materials 2. Materials 2.1. SystemofAcquisition 2.1. System of Acquisition The acquisition is performed with an Intuos Wacom 4 digitizing tablet. Figure 1 shows the The acquisition is performed with an Intuos Wacom 4 digitizing tablet. Figure 1 shows the information, which is captured by the USB pen. The tablet gathers 200 samples/second: spatial information, which is captured by the USB pen. The tablet gathers 200 samples/second: spatial coordinates (x, y), the applied pressure of the pen, and a pair of angles, which are the azimuth coordinates (x, y), the applied pressure of the pen, and a pair of angles, which are the azimuth (angle (anglebbeettwweeeenn ththee ppene nanadn tdhet hheorhizoorniztaoln ptlaalnpe)l,a anned) ,tahne dalttihtuedael t(iatnugdlee b(eatnwgeleenb tehtew peeenn anthde thpee nveartnicdalt he verticalaaxixsi)s [)1[01,101,1].1 C].olCleocltliencgt ionnglyo tnhleyset hdyesneamdiycn faeamtuicrefse, aatfuterrewsa,radfst,e mrwoarer dinsf,omrmoarteioinn fcoarnm bae tiinofnercreadn. be inferredF.oFr oerxaemxapmlep, lvee,lvoeciltoyc, itayc,caelcecrealteiorant, iotann,gteanntgiaeln atciacellaecracetiloenr,a tcieonnt,ricpeentatrl ipacectealleraactcioenle, raantigolne ,oafn tghlee of theinstiannsttaannteaoneuosutsr tarjaejcetcotorryy,,i innssttaannttaanneeoouus sddisipslpaclaemceemnte, nrat,driuasd oiuf scuorfvactuurrve,a etutcr. e[2,6e]t.c I.n[ F2i6g]u.rIen 1F, tihgeurree 1, is a modern digitizing tablet, and the graphic representation of the angles of the azimuth and thereisamoderndigitizingtablet,andthegraphicrepresentationoftheanglesoftheazimuthand altitude. It also represents the capture of a handwritten text (the Spanish word altitude. Italsorepresentsthecaptureofahandwrittentext(theSpanishword“biodegradable”— “biodegradable”—i.e., biodegradable in English but with a different pronunciation), and two captures i.e.,biodegradableinEnglishbutwithadifferentpronunciation),andtwocapturesofthedrawingof of the drawing of a house. Please note that every capture comprises both on-surface and in-air ahouse. Pleasenotethateverycapturecomprisesbothon-surfaceandin-airtrajectories. trajectories. Figure 1. Information from the digitizing tablet. Figure1.Informationfromthedigitizingtablet. Entropy2018,20,531 4of15 2.2. DatabaseofIndividuals ThispaperisbasedonadatabasenamedBIODARW,whichcontainssamplesfrombothhands of50people: 21controlindividuals(CR)and29ETpatientswithidentifiedgeneticloci. Thedataset belongstoalargerresearchstudyonthediagnosisofETconductedbytheBiodonostiaHealthInstitute aimedatcharacterizingthisimpairment. Theparticipantsarefamilieswithidentifiedgeneticsloci. SubjectswereselectedamongpatientsofapreviousstudythatcomprisesfamiliarandsporadicET cases,whereascontrolindividualswererecruitedintheMovementDisordersUnitoftheDonostia UniversityHospital(inSanSebastian,Spain).Anewinformedconsentwassignedbyeveryparticipant inourstudy,whichwasapprovedbytheethicscommitteeoftheDonostiaUniversityHospital[25,26]. Foreachindividual,anelectrophysiologicaltest(EPT)andfMRIareavailable. Inthetest,individuals draw a line, the Archimedes’ spiral, and do some handwriting both with the dominant and the non-dominanthand[25,26].Inthisstudy,weuseonlythespiralofArchimedes(Figure2).Thedatabase presentsvariabilityregardingtheamplitude,frequency,andpatternofthetremor;scaleofdiagnosis; anddemographicdata(genderandage). Asfarastheindividualsselectedforthestudyareconcerned, asubsetofsamplesofthespiralofArchimedesispickedupfromtheoriginaldatabaseanalyzing geneticandclinicalassessment. Withregardtothesamplesofthespiralsstudied,thesesamples(or spiraldrawings)comefromsubjectsdiagnosedwithETaccordingtoclinicalandelectrophysiological criteria[27]. Tremorratingscalesprovidecrude,non-linearandsubjectiveassessmentsoftheseverity oftremor. Amongthemostcommonlyusedscales,theFahn-Tolosa-Maríntremorratingscale(FTM) usesagradingof0–4toevaluatetremorinArchimedes’spiraldrawings. ThescaleofBainandFindley uses a grading of 0–10. Both scales have a strong logarithmic relationship with the amplitude of thetremormeasuredbymeansofadigitaltablet(consistentwiththeWeber-Fechnerlaw). Actually, thedigitizingtabletsaremuchmoreaccuratethantheclinicalclassifications,butthisadvantageis mitigatedbythenaturalvariabilityofthetremor,andhence,thisiswhythestudythatwepresent makessense[28]. WithinthegroupofindividualswithET,onlythesamplesofthehandaffectedby essentialtremorareused,butfivepatientsarenottakenintoconsiderationduetothepoorqualityof theirsamples. Forthecontrolindividuals,thebestsampleisselected(mostlythedominanthand),and forsomeindividualsthenon-dominanthandisalsoused. Therefore,thissub-databaseBIODARWOis madeof51samplesselectedfromtheoriginaldataset:27samplesselectedfromthecontrolindividuals, and24samples(thehandaffectedbytremor)fromtheETindividuals. Ahigherdatabasewouldbe very useful. Unfortunately, as often happens, the study has a limited number of samples. This is aclassicalBiomedicalEngineeringproblem. Inthebibliography,wecanfindmanyrealexamplesin thesamescenario. Oneofthetypicalcasesisthemodellingandexploringofthegenesthatcaused aspecifictypeofcancer. Inthiscase,therearemorethan4000genes-features,soafewofthemare selectedtobeusedasfeaturestofeedaclassifier. Additionally,underHospitals’EthicalCommittees’ advisors,thisisastandardstudy,andworkingwithothersubsetsofthedatainapreviousphase, wealreadyobtainedresultsusingasmallersamplesize,butalsoasmallersetoffeatures,reaching similarnumericalaccuracies. ThecomplexityoftheproblemliesinthedifficultyinobtainingfMRI samplesinthesepatientsthatcanonlybegatheredaspartofaresearchstudy,notinhabitualclinical practice. Forthisreason,thesampleislimited. Thestudycanbeconsideredafeasibilitystudy,whichis notthesameasapilotstudy[29]. Entropy 2018, 20, x 5 of 15 EEnnttrrooppyy 22001188,, 2200,, 5x3 1 55 ooff 1155 Figure 2. Example of (X, Y) and pressure (P) o f the Archimedes’ spiral carried out by two individuals wFFiiiggthuu rrEeeT 22:.. lEEoxxwaa mmleppvlleeel oo(flfe ((fXXt ,, iYYm)) aaagnnedd); p parrneedssss uuhrirgeeh ((P Pl)e) vooeff ltt hh(reei gAAhrrtcc hhimiimmaeegddeee).ss ’’T ssipmpiireraa lel vccaoarrlruriiteeidodn oo uuistt sbbhyyo ttwwwnoo iiinnn dddiivivfifiddeurueaanllsts wcwoiilttohhu ErEsTT (::b lloeogwwin llneeivvneegll (i(nllee bfftltu iimem aaanggdee) )e;;n aadnniddn ghh iiigngh hr elledevv).ee ll ((rriigghhtt iimmaaggee)).. TTiimmee eevvoolluuttiioonn iiss sshhoowwnn iinn ddiiffffeerreenntt ccoolloouurrss ((bbeeggiinnnniinngg iinn bblluuee aanndd eennddiinngg iinn rreedd)).. 3. Methodology 33.. MMeetthhooddoollooggyy 3.1. Derived Measures 33..11.. DDeerriivveedd MMeeaassuurreess Thanks to on-line drawing it is possible to evaluate the time spent and the pressure exerted whileTT hhwaarnniktkisns tgto,o o fronon-ml-ilni neae dq rduarawanwitniitgnagtiti viitse pipso opsisnoitbs sloeibft love ieetvowa e.l uvFaaotleru atihntees tttaihnmec eet,is mpitee ni sst pwaenendltl tahknendop wrtehnses tuphrraeetse AsxueDrrte e pdeaxtweierhntietldes wswprheintiliden gmw,rofirrtoeinm tgim,a ferq oaumnadn a tpi trqaoutdiavunectiept asotiiminvtep olpefrov ianinetwd o .sfo Fvfotieerrw isn.t srFtoaoknre csien [,s1it2ta]ni,s cwewh, eiiltlel ipks newoopewlllen kwtnhiotahwt EAnT Dt hapnatda t AiPeaDnrt kspinastpsioeennn’tdss Dmspiosereneadst iemm poerraoend tdiumcpeer o aldensudsc epsrtsaoimbdlupec leper rseaimsnsdpulrseeor f ptaenartdtse tsrroonfkste e[sr1 [s21t,r23o]0,k,w3e1sh] [.i1 leI2n]p, e[w2o9hp]il,le etw hpeiet ohapbElieTl iwtayni tdohf PE atThr kea innidnsod Pniav’rsikdDiuniassleosan si’ess prDerlioasdeteaudsce e tploer sospdsrutecasebs ulleeresps r easstnasdub lreei npp-araeitrts esrutnrrasej e[p1ca2tot,3tre0ier,3ns1 s] (.[t1hI2ne,i3r[0 2,93l]e1,n].tg hItehn) a [b2[i93l]i2t, –yt3ho5ef] . tahbAeilniitnoydt hiovefri d thuneae lwsin idsiinvretiedlarueteasdtlsin tigos pcrherelaasrtsaeucdrt eeratiosnt dicpi nroe-fas sirmutroread jeeacrnntod ro ienisn-l(-itanhiere irhtlareannjdegcwtthor)irt[ii3en2sg – 3(a5thn].eaAilry nsoilset hniegsr thtnh)e awt[ 3ii2nt –te3cra5en]s. tibnAegancro htihanre armc teninreidwst i icniofnoftremmreoasdtteiionrnng cocohnla-lleriacntceetdehr awinsthdiciwl eor tifth ienm gpoeadnne adrlnoy esosis nni-oslitnt heexa ethriatt npcdarnewsbrsiuetairnerg ion nam nthainely dssuiisnr ffaioscr emt. hFaaitgti ouinrt ecc o3al nrlee cpbtreeedasrew nihntsi ltmehteihn dedr paiewnnfiondrgom eoasf tntihooent secpxoeilrlreatclp toerfed s Awsurhcrhielieom nthetedh epese sncua rdrfroaieceesd .n Fooiutg teu xbreeyr 3tt hprrerepeesr esiusnerdnei tvosindth utehaelds sr: auawr fciaoncnget.o rFofilgt hiunerdesi p3vi irrdaeulpaorlef,s Aaennrc thEs iTtmh ieen ddderisvawicdaiurnragile odaft o tahunet ebsapyritlryha rls eoteafg inAed,r ciavhniiddmu aeandls eE:sTa c cainornrditiervdoid loiunuadtl ibavyti d atuhnar ael,dea vinnaEndTciveiidnd dustiaavlgisd:e u a[a 2cl6oa]n.t tParenonle -iadnrodlywivsnitd a(ugoaenl,-,as aunnrd fEaacTne EitnrTdaijienvcditdiovuriaidelsu a)a tal araent saehnaoralwyd nvs taianng cbee,ld uaens,t daw gahenil [eE2 6Tp] e.innP-deuinpv- id(dionuw-aalni ra (tto rananj-e sacutdrovfraaicenesc)et ardar ejse tscahtgooerw i[e2ns6 )i]na. r Preeesdnh.- odFwoigwnunrien ( o4b nlru-esepu,rrewfsahecnielt estr ptahejeenc -wtuoprriite(iisnn)g -aa orierf ttshrhaeoj ewwctnoo rriidne sbB)lIauOree,D swEhhGoiwRlenA piDennAr-eBudLp.E F( iinfgr-uoarmier t4rtharejeep crCteoRsre inegstrs)o atuhrpee wsthoro itbwienntgt eionrf ritelhldue.sw Ftroiagrtuder Beth I4Oe rDepEpeGrne-RsueApnD tasA ntBdhLe pEwefnrrio-tdimnogwt honef iCtnhRfeo grwrmooaurtpdio tnBo.I bOTehDtteeE riGnilR-lauAisrDt rinaAtfBeoLrtmhEe afptrieoonmn-,u opthr aepn eCdnRp-u egpnr,- odisuo pwsh ntooiw nbfnoe ritntme rra etiidlol,nu a.sntTrdha tetehi net -hoaein r-pisneufnrof-raumcpea atiinnofdno ,rpomreanpt-iedonon-wu opnr, piisnesfnoh-rodmwoawntinion nisr. esTdhh,oeaw ninnd -iatnhi reb ilonunfeo-.s rumrafaticoeni,n oforr pmeant-iuonp,o irs psehno-wdnow inn riesds,h aonwdn tihneb olune-s.urface information or pen-down is shown in blue. Figure 3. Sketching of the spiral of Archimedes carried out by (from left to right): a CR individual, an Figure3. SketchingofthespiralofArchimedescarriedoutby(fromlefttoright): aCRindividual, EFTig uinrde i3v.i dSkueatlc ahti nagn o ef atrhley ssptairgael ,o af nAdr cahni mEeTd einsd ciavrirdieuda lo autt bany (afdrovman lceefdt t ost raiggeh.t )B: oat hC Rp einnd-divoiwduna (lt, oapn an ET individual at an early stage, and an ET individual at an advanced stage. Both pen-down iEmTa gined iinv ibdluuea)l aant da np eena-rulyp (sbtaogtteo, man idm aagne E inT riendd)i vairde usahlo awt na sni maduvltaannceeodu ssltya.g e. Both pen-down (top (topimageinblue)andpen-up(bottomimageinred)areshownsimultaneously. image in blue) and pen-up (bottom image in red) are shown simultaneously. Entropy2018,20,531 6of15 Entropy 2018, 20, x 6 of 15 FFiigguurree 44.. PPeenn--ddoowwnn ((ttoopp iimmaaggeess)) aanndd ppeenn--uupp ((bboottttoomm iimmaaggeess)) ttrraajjeeccttoorriieess ooff ddiiffffeerreenntt ccaappttuurreess ooff bbiiooddeeggrraaddaabbllee ffoorr tthhee ccoonnttrrooll ggrroouupp.. Therefore, the movements of the hand when drawing or writing some text can be divided into Therefore,themovementsofthehandwhendrawingorwritingsometextcanbedividedinto two categories: on-surface trajectories (pen-down), and in-air trajectories (pen-up) [35]. The twocategories: on-surfacetrajectories(pen-down),andin-airtrajectories(pen-up)[35]. Thepen-down pen-down corresponds to the movements of the writing device when it touches the writing surface. corresponds to the movements of the writing device when it touches the writing surface. Each of Each of these on-surface trajectories is a visible stroke. The pen-up corresponds to the movements of theseon-surfacetrajectoriesisavisiblestroke. Thepen-upcorrespondstothemovementsofthehand the hand when changing from one stroke to the next, that is, there is no pressure at all on the writing whenchangingfromonestroketothenext,thatis,thereisnopressureatallonthewritingsurface. surface. Our past studies on recognition of people based on biometry showed that these two kinds of Ourpaststudiesonrecognitionofpeoplebasedonbiometryshowedthatthesetwokindsofdata data are complementary [10], and their discrimination capability is similar, even when using a arecomplementary[10],andtheirdiscriminationcapabilityissimilar,evenwhenusingadatabaseof database of 370 individuals [11,34]. 370individuals[11,34]. 33..22.. EExxttrraaccttiioonn ooff FFeeaattuurreess IInn tthhiiss ssttuuddyy,, wwee pprreesseenntt aa pprreelliimmiinnaarryy eexxppeerriimmeenntt wwhhoossee mmaaiinn ggooaall iiss ttoo eessttaabblliisshh tthhrreesshhoollddss ffoorr ssoommee hhaannddwwrriittiinngg rreellaatteedd bbiioommaarrkkeerrss.. AAss mmeennttiioonneedd aabboovvee,, iitt bbeelloonnggss ttoo aa bbrrooaaddeerr rreesseeaarrcchh ssttuuddyy aaiimmeedda taet aeralyrldyi adginaogsniossoisf EoTf. TEhTe. reTfhoerere,ftohrees, etahrec hsoefarfecaht uorfe sfeinatouurerss tuind youairm sstuadtpy reacilminsi caatl pevreacluliantiicoanl oervieanlutaatteiodnt oworaierdntsattheed dteofiwnaitridons otfheef fedcetfiivneittieosnts ofof redfifaegcntiovsei ntgesEtTs [f9o,1r2 ,d20ia,2g5n,o26si]n.g ET [9,12,20,25,26]. 3.2.1. LinearFeatures 3.2.1. Linear Features TheaimistoautomaticallydifferentiatebetweenETindividualsandhealthypeoplefromthe analyTshise oafimse vise rtaol aliunteoamrafetiactaulrlyes d(iLffFe)r,ewnthiaicthe abreetwateternib uEtTe sinodrivvaidriuaabllse sanlidn ehaeralylthdye ppeenodpelen tfroonmd athtae, aannadlyvsairsi aonf tsse(vmerinal, mlinaexa,rm feeadtiuarne,sa (nLdF)m, weahni)c,he xatrrea cattetrdibfurotems tohre virarhiaanbdlews rliintienagrsl:y dependent on data, and variants (min, max, median, and mean), extracted from their handwritings: 1. Time-relatedmeasures: timeon-surface,timein-air,andtotaltime(on-surfaceplusin-airtimes). 12.. TSipmatei-arleclaotmedp omneenatssuarnesd: vtaimriaen tosn:-csouorrfdaicnea, tetismXea nind-aYi,ra, ltaintudd etoatnagl letimO,ea z(oimn-ustuhrfaancgel epAlu,sa nignl-eaiZr taimndesm). odulusRpolarcomponentsandtheirprojectionsforbothin-airandon-surfacesignals 2. S(Fpiagtuiarle c5omshpoownse,nints tahnedd vraawriainngtso: fcothoerdinindaitveisd Xua alnwdi tYh, EalTt,itaundiel launsgtrlaet Oio,n aozfimthuethd iasntogrltei oAn, oafntghlee Zp oalnadr cmomodpuolnuesn Rts p).olar components and their projections for both in-air and on-surface signals (Figure 5 shows, in the drawing of the individual with ET, an illustration of the distortion of the 3. Pressureandvariants(min,max,mean,median,std). polar components). 4. Features of dynamic nature and variants: speed and acceleration for both in-air and 3. Pressure and variants (min, max, mean, median, std). on-surfacesignals. 4. Features of dynamic nature and variants: speed and acceleration for both in-air and on-surface 5. Zerocrossingrate. signals. 6. Frequency related features: spectral components for both in-air and on-surface signals 5. Zero crossing rate. (Figure6)[26]. 6. Frequency related features: spectral components for both in-air and on-surface signals (Figure 6) [26]. Entropy2018,20,531 7of15 Entropy 2018, 20, x 7 of 15 Entropy 2018, 20, x 7 of 15 Figure 5. Projections of the polar components of the drawing of the spiral of Archimedes carried out Figure5.ProjectionsofthepolarcomponentsofthedrawingofthespiralofArchimedescarriedout Figure 5. Projections of the polar components of the drawing of the spiral of Archimedes carried out by a control person (left image) and a patient with ET (right image). byacontrolperson(leftimage)andapatientwithET(rightimage). by a control person (left image) and a patient with ET (right image). Figure 6. Spectrum of the coordinate X of the pen-down signal for a individual of the control group Figure 6. Spectrum of the coordinate X of the pen-down signal for a individual of the control group Fig(ulerfet 6im.Sapgeec)t raunmd aonf itnhdeicvoidorudailn watiethX EoTf (trhigehpte inm-daogwe)n. signalforaindividualofthecontrolgroup (left image) and an individual with ET (right image). (leftimage)andanindividualwithET(rightimage). 3.2.2. Non-Linear Features 3.2.2. Non-Linear Features 3.2.2. Non-LinearFeatures The objective is to use the analysis of non-linear features (NLF) extracted from the subjects’ The objective is to use the analysis of non-linear features (NLF) extracted from the subjects’ samThpeleos bojfe chtaivnedwisrtiotinugs eint hoerdaenra tloy sdisisotifnnguonis-hli naeuatormfeaattiucarellsy (bNeLtwF)eeenx thraecatlethdyf rinodmivtihdeuaslusb ajencdts E’T samples of handwriting in order to distinguish automatically between healthy individuals and ET samsupffleesreorfs.h Tahned wNrLiFtisn gunindeorr dcoenrstiodedriasttiionng uairseh: faruatcotaml adtiimcaellnysiboentsw (eFeDn) haenadlt hthyeiinr dviavriidanutasl s(manind, EmTax, sufferers. The NLFs under consideration are: fractal dimensions (FD) and their variants (min, max, sufmfeeraenrs a.nTdh emNedLiFasnu) nfodre rthceo npsriedsesurartei oanndar teh:ef rsapcatatiladl icmomenpsoionnesn(tFs Dx,) ya,n cdotmhepiurtveadr iuasnintsg( mthien ,Hmigauxc,hi mean and median) for the pressure and the spatial components x, y, computed using the Higuchi meaalgnoarnitdhmm e(HdiFaDn)) faonrdt htheep Creassstuigrelioannid atlhgeorsipthamtia (lCcFoDm)p; othnee netnstrxo,pyy, c(oEm) apnudte idtsu vsainriganthtse (Hmiignu, cmhiax, algorithm (HFD) and the Castiglioni algorithm (CFD); the entropy (E) and its variants (min, max, algmoreiatnh man(dH FmDe)daiannd) thfoer Cthaset ipgrlieosnsuiraelg aonridth tmhe( CspFDat)ia;lt hceomenptroonpeynt(sE )x,a nyd; aitnsdv asrigiannatls d(mistinor,tmioanx ,by mean and median) for the pressure and the spatial components x, y; and signal distortion by meaannaalynzdinmge ldoicaanl )pfeoarktsh.e pressureandthespatialcomponentsx,y;andsignaldistortionbyanalyzing analyzing local peaks. localpeTahkes .fractal dimension of an object is a measure of its self-similarity. A self-similar object is The fractal dimension of an object is a measure of its self-similarity. A self-similar object is apTphroexfimraacttaellyd siimmeinlasri oton ao pfaarnt oofb ijtescetlfi,s thaamt iesa, sifu wree oefxaitmsisneelf i-ts icmloislearr,i twy.e Awislle flifn-sdim smilaarlleorb vjeecrtsiiosns approximately similar to a part of itself, that is, if we examine it closer, we will find smaller versions apopfr oitx. iImn aotredlyers tiom qiluaarntotifayp saerlft-osifmitislealrfi,tyth, awteis c,aifnw meeeaxsuamrei ntheei tnculmosbeer,r wofe bwuiillldfiinngd-bslmocakllse rthvaetr asrioen psart of it. In order to quantify self-similarity, we can measure the number of building-blocks that are part ofoitf. aIn poartdteerrnt;o thqius amnteiafysusreel fi-ss itmheil asroi-tcya,lwleed cfarnacmtael adsiumreenthsieonnu (mFDb)e.r Foofrb au gilidvienng -wblaovcekfsorthma,t tahreerep aisr tno of a pattern; this measure is the so-called fractal dimension (FD). For a given waveform, there is no ofraepfeartetnercne ;otfh tihsem FeDas vuarleuies itth sehsoou-lcda lhleadvefr. aAcdtadlidtiiomnaenllysi,o ans (sFigDn)a.lFso arrae gniovte unswuaalvlye fsotarmtio,nthareyre, iins nmoost reference of the FD value it should have. Additionally, as signals are not usually stationary, in most refteercehnnciequofesth, eshFoDrtv sailguneailt ssehcotuiolndsh aarvee .uAseddd iitnio onradlleyr, atos sciaglncaullsataer efenaottuuressu afrlloyms tathtieo nwaarvy,eifnormmo. sFtor techniques, short signal sections are used in order to calculate features from the waveform. For tecehxnaimqupelse,, sthhoer tssigignnaall csaecnt iboen ssaprleitu isnetdo inshoorrdte priteocceasl ciunl aotredfeera ttuor ecsaflrcoumlatteh ethwea vfeeafoturmre.s Foofr eeaxcahm ppilee,ce, example, the signal can be split into short pieces in order to calculate the features of each piece, thewshigicnha lisc atnheb etescphlnitiqinuteo tshhaot rwtpei ehcaevsei ntaokredne rintotoc aaclccuoluantet. tIhne ofetahteurr ewsoorfdes,a cwhep cieocme,pwuthei cthhei sftrhacetal which is the technique that we have taken into account. In other words, we compute the fractal tecdhinmiqeunseiothna otfw sheohrat vseegtmakeenntsi notfo thaec csoigunnat.l, Ianndo tohbesrewrvoer dthse, weveocluotmiopnu otef tthhee rfersauctltasl adliomngen tshieo nwohfole dimension of short segments of the signal, and observe the evolution of the results along the whole shosirgtnsaelg, mtryenintsg otof tdhiescsoigvnera lf,raanctdalo cbhsaerravcetetrhiesteicvso tlhuatito cnouolfdt hbee rheesluplftusla tloo nidgetnhteifwy hdoiflfeerseignnt aell,emtryeinntgs of signal, trying to discover fractal characteristics that could be helpful to identify different elements of totdhiastc osivgenraflr. aTchtaelrec haarera scetvereirsatli caslgthoraitthcmousl dthbaet chaenlp bfeu ultsoedid teon mtifeyasduirfefe trheen tFeDle. mIne onutsr ostfutdhya,t tshieg nfoacl.us that signal. There are several algorithms that can be used to measure the FD. In our study, the focus Thiesr eona rtehes eovpetriaolnasl gesoprietchimallsyt shuaittacbalne bfoeru tsheed atnoamlyseiass uofr etitmhee sFeDri.eIsn wohuirchst uddo yn,otht enefoedcu asniys opnretvhioeus is on the options especially suitable for the analysis of time series which do not need any previous opmtioondseleisnpge coifa ltlyhes usiytastbelmef.o Srothmeea noaf lythseisseo fmtiemtheosdesri eisncwluhdiceh tdhoe nHoitgnuecehdi aanlgyoprirtehvmio u[s36m],o tdheel inKgatz modeling of the system. Some of these methods include the Higuchi algorithm [36], the Katz ofathlgeosryitshtmem [.3S7]o manedo fththee Cseasmtiegtlhioondis ainlgcoluridthemth e[3H8]i.g uHcihgiuaclhgio arinthdm C[a3s6ti]g,ltihoeniK watezrea lgsoelreitchtemd, [3a7s] in algorithm [37] and the Castiglioni algorithm [38]. Higuchi and Castiglioni were selected, as in previous works in under-resourced conditions, this approach has proven to be more accurate previous works in under-resourced conditions, this approach has proven to be more accurate [7,8,39–41]. [7,8,39–41]. Entropy2018,20,531 8of15 andtheCastiglionialgorithm[38]. HiguchiandCastiglioniwereselected,asinpreviousworksin under-resourcedconditions,thisapproachhasproventobemoreaccurate[7,8,39–41]. Ontheotherhand,physicalsystemsentropyisadisordermeasure. Forasinglediscreterandom variableX,theentropyHmeasuresitsaverageuncertainty. Shannonentropy[42]iscomputedas: ∑ H(X) = − p(x )logp(x ) (1) i i xi 3.3. SetsofFeatures Next,wepresentthesetsoffeaturesusedintheexperimentation: 1. SetofLinearFeatures(LF),thissetwillbedescribedin3.3.1. 2. SetofNon-LinearFeatures(NLF),whichconsistsofLFandthesetthatwillbedescribedin3.3.2. (Higuchi(HFD),Castiglioni(CFD),entropy(E),andfractaldimension(F)):LFHFD,LFCFD,LFE, LFHFDEandLFCFDE. 3. TheresultingsetoffeaturesafterselectionbymeansofANOVA:SLF(SelectionofLinearFeatures), SLFHFD,SLFCFD,SLFE,SLFHFDEandSLFCFDE. Table1showsthecompletelistofacronymsandtheirmeanings. Table1.Completelistofacronymsandtheirmeanings. Acronym Description ET EssentialTremor PD Parkinson’sdisease FD FractalDimension HFD HiguchiFractalDimension CFD CastiglioniFractalDimension E Shannon’sEntropy LF LinearFeatures NLF Non-linearFeatures LFXFD LinearFeaturesandFractalDimension LinearFeaturesandFractalDimensionand LFXFDE Shannon’sEntropy SLF SelectionofLinearFeatures SLFXFD SelectionofLinearFeaturesandFractalDimension SelectionofLinearFeaturesandFractalDimension SLFXFDE andShannon’sEntropy SVM SupportVectorMachine MLP MultiLayerPerceptron NNHL Neuronnumberinthehiddenlayer k-NN K-NearestNeighbor NF NumberofFeatures TS TrainingStep CER ClassificationErrorRate Acc Accuracy ACCER AccumulativeClassificationErrorRate 3.4. AutomaticSelectionofFeatures The automatic feature selection is based on statistical tests (ANOVA). The MatLab one-way balancedanalysisofvariancefunction(anova1)isused[43]. Two or more columns of data that represent independent samples that contain mutually independentobservationsarecomparedwithregardtotheirmeans. Thetestcomputestheprobability (p-value)underthenullhypothesisthateverysampleisdrawnfrompopulationswiththesamemean. Ifpisneartozero,wecansaythatthenullhypothesisisveryunlikely,sowecanassumethatthe Entropy2018,20,531 9of15 mean of at least one of the samples is significantly different from the means of the other samples. Commonlyusedsignificancelevelsare0.05or0.01;inourcaseitis0.05. ThestandardANOVAtest differentiatesthevariabilityofthedataintotwotypes: ontheonehand,thevariabilityduetothe discrepancybetweenthesamplemeans,orvariabilitybetweengroups;andontheotherhand,the variabilityduetothedifferencesbetweenthedataineachsampleandthesamplemean,orvariability inside the groups. The size of the F-statistic and the p-value are derived from the box plot of the samples. SignificantdifferencesbetweenthecenterlinesoftheboxesleadtolargevaluesofF,and thereforesmallvaluesofp. 3.5. AutomaticClassification Theprincipalaimofthisstudyistheselectionofpotentiallyusefulhandwritingfeaturesuseful forpreclinicalevaluationtodefinetestsinordertodiagnoseET.Suchfeaturesshoulddifferentiate betweenthegroupofcontrolindividuals(CR)andthegroupofindividualswithessentialtremor (ET).Asthesetechniqueshavetobeusefulinrealenvironmentsforreal-timeapplications,asecond objectiveistooptimizethecomputationalcost. Therefore,themodelofautomaticclassificationwill beperformedbearinginmindthesetwogoals. ThreedifferentclassifiersimplementedintheWEKA softwaresuite[44]wereused. Allconfigurationswereorientedtorealtime: 1. ASupportVectorMachine(SVM)usingpolynomialkernel(e=1.0). 2. AMultiLayerPerceptron(MLP)usingneuronnumberinthe(unique)hiddenlayer(NNHL)= max(Attribute/Number+Classes/Number)andtrainingstep(TS)=NNHL×10. 3. K-NearestNeighborAlgorithmork-NN(usingk=1). Inordertoevaluatetheresults,weusedAccuracy(Accin%),ClassificationErrorRate(CERin%) and Accumulative Classification Error Rate (ACCER in %) for every model [5,24]. In the steps of trainingandvalidationweusedk-foldcross-validation(k=10). Cross-validationisarobusttechnique fortheselectionofvariables[45]. Thankstorepeatedcross-validation(ascomputedbyWEKA),robust statisticaltestsarepossible. Weusedthemeasurementprovidedbythefunction“Coverageofcases” ofWEKA(confidenceintervalat95%level)aswell. 4. ResultsandDiscussion The experiments were performed using the balanced subset BIODARWO. The aim of the experimentwastoassessthepotentialofselectedfeaturestoautomaticallymeasurethedegradation oftheArchimedes’spiraldrawnbysubjectsthatsufferET.Therefore,thepreviouslydefinedsetsof featureshavebeenassessedtodifferentiateproperlybetweencontrolandessentialtremorgroups. ExtractionofLinearandNon-LinearFeatures Inafirstphase,linearandnon-linearfeatureswereextractedusingthetechniquesofSection3.4 (seeTable2),andanautomaticclassificationwasperformedbymeansoftheclassifiersenumeratedin Section3.5. Figure7summarizestheresultingCERs(%)forthedifferentclassifiers,withlinearand non-linearfeaturesets. Intheseresultswecanobservethat: • Thesetsofnon-linearfeaturesprovidebetterresultsforeveryclassifier. • MLPandSVMhavethebestresultsinmostofthecases. • BothFDfeaturesareabletoprovidebetterresultsforeveryclassifier. • ShannonentropyhasthebestperformanceforMLPwheniscombinedwithCFD. • ThebestoptionisLFCFDEwithCastiglioniFD,entropyandMLP,wheretheCERis13.73%with 225features. EntEronptyro2p0y1 280,1280,, 2503,1 x 10o1f01 o5f 15 Figure7.CER(%)forthethreeclassifiersconsideredwithlinearandnon-linearsetsoffeatures. Figure 7. CER (%) for the three classifiers considered with linear and non-linear sets of features. Table 2. Sets of features and number of features (NF). Table2.Setsoffeaturesandnumberoffeatures(NF). LF LFHFD LFCFD LFE LFHFDE LFCFDE SLF SLFHFD SLFCFD SLFE SLFHFDE SLFCFDE NF 186 2L1F3 LFH2F13D LFCF19D8 LFE 225L FHFDELF22C5F DESL7F0 SLFH7F3D SLFCFDS7L7F E SLF7H6 FDE SL7F9C FDE 86 NF 186 213 213 198 225 225 70 73 77 76 79 86 In a second stage, an automatic selection of features is performed by applying the ANOVA test described in Section 3.5. Figure 8 shows the details for the selected feature time-up (in-air), which Inasecondstage,anautomaticselectionoffeaturesisperformedbyapplyingtheANOVAtest represents the time that the pen is not on the writing surface. In Table 2, we can see that the number describedinSection3.5. Figure8showsthedetailsfortheselectedfeaturetime-up(in-air),which of features (NF) for linear and non-linear cases is reduced by about 65% after the selection process. representsthetimethatthepenisnotonthewritingsurface. InTable2,wecanseethatthenumber Finally, an automatic classification process was applied to the entire database using the previously offeatures(NF)forlinearandnon-linearcasesisreducedbyabout65%aftertheselectionprocess. described classification algorithms. Figures 9 and 10 graphically summarize the CERs (%) yielded by Finally,anautomaticclassificationprocesswasappliedtotheentiredatabaseusingthepreviously each algorithm when considering linear and non-linear sets of features. In these results, we can describedclassificationalgorithms. Figures9and10graphicallysummarizetheCERs(%)yielded observe that: byeachalgorithmwhenconsideringlinearandnon-linearsetsoffeatures. Intheseresults,wecan observethat: • The automatic selection of features improves the results for most of the algorithms when linearfeaturesareconsidered(Figure9). Withtheselectedfeatures,MLPisthealgorithmthat performsworst. • Whenthenon-linearselectedfeaturesareincluded,theresultsarebetterinmostcaseswithregard totheLFs. • Both non-linear features (FD and entropy) improve the results regardless of the applied classificationalgorithm. • ShannonentropyperformsbetterwithSVM,andwhencombinedwithFD. • ThebestoptionwithregardtofeaturesetsisSLFCFDwithCastiglioniFDandMLPwithNF=77 features(CER=5.89%). ThesamefeaturesetprovidesthebestresultsforSVMwhenapplying theHiguchialgorithmforFD. • Goodresultsarealsoobtainedwithk-NN,requiringalowercomputationalcostwiththesetsthat includefeatureselectionbyANOVAandFD. Figure 8. Example of ANOVA analysis: time-up for the CR group in blue and the ET group in red.
Description: