j ou rn a l h om epa ge :w w w . i n t l . e l s e v i e r h e a l th . c o m / j o u r n a l s / c m p b
Check
Your
Biosignals
Here:
A
new
dataset
for
off-the-person
ECG
biometrics
Hugo
Plácido
da
Silva
a,∗,
André
Lourenc¸o
a,b,
Ana
Fred
a,
Nuno
Raposo
c,
Marta
Aires-de-Sousa
caInstitutodeTelecomunicac¸ões,InstitutoSuperiorTécnico,1049-001Lisboa,Portugal bInstitutoSuperiordeEngenhariadeLisboa,1959-007Lisboa,Portugal
cEscolaSuperiordeSaúde,CruzVermelhaPortuguesa,1300-125Lisboa,Portugal
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:Received12March2013 Receivedinrevisedform 26November2013 Accepted28November2013 Keywords: Biometrics Electrocardiography Off-the-person Dataset
a
b
s
t
r
a
c
t
TheCheckYourBiosignalsHereinitiative(CYBHi)wasdevelopedasawayofcreatinga datasetandconsistentlyrepeatableacquisitionframework,tofurtherextendresearchin electrocardiographic(ECG)biometrics.Inparticular,ourworktargetsthenoveltrendtowards off-the-persondataacquisition,whichopensabroadnewsetofchallengesand opportuni-tiesbothforresearchandindustry.WhiledatasetswithECGsignalscollectedusingmedical gradeequipmentatthechestcanbeeasilyfound,foroff-the-personECGdatathesolution isgenerallyforeachteamtocollecttheirowncorpusatconsiderableexpenseofresources. Inthispaperwedescribethecontext,experimentalconsiderations,methods,and prelim-inaryfindingsoftwopublicdatasetscreatedbyourteam,oneforshort-termandanother forlong-termassessment,withECGdatacollectedatthehandpalmsandfingers.
©2013ElsevierIrelandLtd.Allrightsreserved.
1.
Introduction
Aresearchfieldthathasseenarecenttrendtowardstheuseof biosignalsisbiometrics,whichfocusesonidentityrecognition basedon physiological orbehavioral properties ofan indi-vidual[1].Electrocardiographic(ECG)dataisoneofthenovel biometrictraitswhereagrowinginterestisevidentwithinthe referenceliterature[2–6].Aparticularlyimportantaspect,that istransversaltoalltheworkdonetodate,istheaccesstolarge datasetstoevaluatetherobustnessofthedevisedmethods acrossresearchteams.Overtheyears,severalinitiativeshave greatlycontributedtomitigatethisproblem,mostlyby pro-vidingrepositoriesfordatacollectedinclinicalandlaboratory settings[7].
∗ Correspondingauthor.Tel.:+351919315567.
E-mailaddresses:hsilva@lx.it.pt(H.P.daSilva),alourenco@deetc.isel.ipl.pt(A.Lourenc¸o),afred@lx.it.pt(A.Fred),nraposo@esscvp.eu
(N.Raposo),msousa@esscvp.eu(M.Aires-de-Sousa).
Our researchgroup recentlypivotedtowards adifferent approachentitled“off-the-person”[8–11].Arecurrentproblem facednotonlybyus,butalsobyotherresearchers through-out the world,isthe accesstolargedatasets, inparticular collected in arepeatable manner,and that canbe publicly accessible.Thisconcernisespeciallyimportantwhendealing withECGacquired inanon-traditionalfashion, with mini-mallyconstrainedorunconstrainedscenarios,andthatcan accountforchangesintheenvironmentalconditions,aging, andotherfactorssurroundingthesubjects.
The Check Your Biosignals Here initiative (CYBHi) was motivatedbythelackofdatasetsofECGsignalscollectedat thehandpalmsandfingers(i.e.off-the-person).Whatis gen-erallyfoundinthestate-of-the-artisthateachteamcollects
0169-2607/$–seefrontmatter©2013ElsevierIrelandLtd.Allrightsreserved.
Table1–SummaryofdatasetsfoundintheECGbiometricsliterature.
Name MIT-BIH[7] AHA[7] PTB[7] DRIVEDB[7] CYBHi Chan[3] Zhao[6] Odinaka[4]
Numberofsubjects 48 155 290 17 125+ 50 22 265
Samplingrate 360Hz 250Hz 10kHz 496Hz 1kHz 1kHz 200Hz 1kHz
Resolution 11-bit 12-bit 16-bit 14-bit 12-bit 12-bit n.a. n.a.
Channels 2 2 14 1 2 1 1 1
Contactpoints n.a. n.a. n.a. 3 2 2 3 n.a.
Electrodes Gel Gel Gel Gel DryAg/AgCl+electrolycras DryAg/AgCl Gel Gel
Placement Chest Chest Chest Chest Handpalms+fingers Fingers Fingers Chest
Publiclyavailable Yes Yes Yes Yes Yes No No No
itsowncorpora,consistingofonlyafewdozensof individ-ualsorless,atgreatfinancialandtimeexpense,whichisthen onlyusedinthecontextoftheirownresearch,andoftenlacks severalrelevantdetailsregardingacquisitionandsubject con-ditions.Assuch,wedevisedadataacquisitionframeworkand experimentalsetup,forlargescaledatacollectionfromalarge groupofsubjectsthroughaneasilyrepeatableand efficient procedure.
Therest ofthe documentis organized as follows: Sec-tion2outlinesthecurrentlandscapeintermsofECGdatasets; Section3describestheexperimentalsetup,highlightingthe maintechnicalchallengesandoptionstoovercomethem; Sec-tions4and5describetheadoptedmethodsandprocedures fordataacquisition;Section6presentsthefindingsresulting fromresearchdonetodatewithCYBHi;Section7describesthe resourcesmadeavailableasaresultofourwork;andfinally, Sections8and 9provide ashort discussionand insighton futureworkresultingfromtheCYBHi.
2.
Background
Recognizingtheneedandusefulnessofcentralizeddatasets thatcanbeusedasacommonreferenceforresearchers world-wide,severalinitiativeshavebeencontributingwithresources tomitigatethis.Physionet [7] iscurrentlyoneofthe main forumsfordisseminationandexchangeofbiomedicalsignals ingeneral.Mostpubliclyavailabledatasetsarecurrently cen-tralizedonthePhysioBankrepository,andcontainmultiple parameters from healthy and pathological conditions (car-diorespiratory,neuralandothers).However,inwhatconcerns ECG datasets, public resources found to dateonly contain signalscollectedatthechestwithclinicalgradeequipment.
Table1 highlightsthe specifications ofdatasets commonly foundinECGbiometricsliterature.
Electrocardiographic signal acquisition methods can be classifiedaccording totheirintrusivenessasfollows[8]:(a)
In-the-person:referringtoimplantabledevicessuchas artifi-cialcardiacpacemakersandimplantableloop recorders;(b)
On-the-person:referringtodevicesthatneedtobeattachedto thebodyofthesubject,generallyrequiringconductivepaste orgel;and(c)Off-the-person:referringtodevicesthatare inte-gratedinobjectsorsurfaceswithwhichthesubjectsinteract with(e.g.acomputerkeyboard),anddonotrequireanyspecial preparationofthesubject.
Researchsofarhasshownbroadevidenceregardingthe applicability of ECG signals collected in an on-the-person approachforbiometricpurposes(see[2,4,5] andreferences therein).Nonetheless,thisapproachishighlyintrusiveforthe
subjects,thuslimitingthepotentialindustrialapplicationsof ECG-basedbiometrics.Morerecently,researchershavebeen pivoting towards the off-the-person approaches [3,6,10–13], inanefforttoovercometheintrusivenessproblemswidely criticizedinECGbiometrics.Recentworkfromourteamhas provided experimental evidence that the signals obtained through the off-the-person approach using dry electrodes, presentvirtuallythesamemorphologyasthoseobtainedat theLeadIderivationofastandard12-leadmedical-gradeECG system[8,14].
Still, tofurther progresson the study ofthis particular research topic,no public datasets currentlyexist (Table1). Furthermore,inthe non-publicoff-the-persondatasets, the amountofinformationprovidedregardingthematerialsand methods oftenlacksseveral relevantdetails.In this paper, weseektocontributetofillthisgap,byprovidingadetailed description ofour experimentalsetupfordataacquisition, withthepurposeofcreatingalargeandextendablecorpora ofdata,targetingthespecificclassofoff-the-personand min-imallyconstrainedsignalacquisitionscenarios[8].Ourgoal is to set the ground for a scalable family of datasets that enables researchers to benchmark their algorithms taking into accountfactorssuchasthe permanenceofthe biosig-nalsthroughouttime,emotionalarousal,amongothers.Initial acquisitionsessionstargetedthecollectionofECGsignalsin highly non-intrusive scenarios, and our dataset is publicly availableforthecommunityatlarge.
3.
Design
considerations
3.1. Overview
Giventhemotivationofourstudytowardstheoff-the-person approach,weproposedasetupforECGdataacquisitionatthe hand palmswithdryAg/AgCl electrodesandatthefingers withElectrolycras,toassessthebiometricpotentialofsignals collectedattheseanatomicregionsandcomparethe perfor-manceofbothmaterials.Moreover,ourprotocolincludesboth neutraltasksandemotionalelicitationtasks[15,16].Thelater wereintroducedasawayofinducingintra-subject variabil-ity,whichisanimportantaspecttotakeintoaccountwhen benchmarkingECG-basedbiometricalgorithms.Forthis pur-pose, electrodermalactivity (EDA)datawas simultaneously acquiredasawayofprovidinggroundtruthinformationabout the arousalstate ofthe subject, andwhich canbeusedin correlationwiththeECGdata.
Theinclusionofemotionelicitingstimuliwasmotivated by the factthat authors in the field ofECG-biometrics are
Fig.1–Overallworkbenchsettingfordataacquisition.
Fig.2–Electrodermalactivitysensorandelectrodes.
startingtoexplore the relationbetween emotionalarousal andchangestotheECGheartbeatwaveform[17,18].In[17], visual stimuli were used to induce positive and negative emotionsforvalenceandpassivearousalperformance eval-uation, and a video game was used to attempt to induce active mental stress.In [18] the ECG was acquired during human–computer interaction on different cognitive activi-ties.Bothstudiesreporttheexistenceofslightmorphological changestotheECGheartbeatwaveformassociatedwiththe presenceofdifferentemotional/affectivestates.
Fig.1,showstheoverallworkbenchsetting,comprisedof adevicetodisplaymultimediacontenttothesubject,theset ofsensorsusedintheexperiments,andthebiosignal acquisi-tionequipment,togetherwiththebasestationsforreal-time datarecording.Thearrangementwassuchthatthesubject satononesideofthetable,andtheexpertsontheotherside. Allparticipantsweresubmittedtoaninitialbriefingwiththe supportofaninformedconsentdocument(Fig.9).Thisform wasreadtothesubjectbythecoordinatoroftheexperiments, andclearlystatedthematerialsinvolvedintheexperiments,
Fig.3–Electrocardiographysensorsandelectrodes arrangement;onthetopwecanseetheelectrolycrastrips, andonthebottomthedryAg/AgClelectrodes.
thepurposeofourresearch,andthedifferentstepsthatwould beperformed.
To acknowledge the participation as a volunteer and grant the anonymous use of the collected data, subjects were requiredtosign thedocument.Participantswerealso inquiredaboutthepossibilityofparticipatinginfuturedata acquisition sessions, and primarycontact informationwas collectedforthispurpose.1Demographicinformationwasalso
requestedforstatisticsandsamplecharacterizationpurposes,
1 Itisimportanttofurtherreinforcethatcontactinformationis
storedseparatelyfromthedataandthatthedatabaseis anonymized.
Table2–Videosequencescharacterization.
Seq. Type Timeframe Content
A1 Low 0–55s SonyBraviapaintad
Arousal 55–60s Monkeyfallsoftree
A2 High 0–55s SonyBraviabouncyballsad
Arousal 55–60s FinalpartofRECmovietrailer
andaspaceforadditionalnotesorcommentsregardingthe
experiments wasalsoadded.Itisimportanttohighlightthat
allthedatapublicly availableintheCYBHidataset is
com-pletelyanonymizedtopreservetheprivacyoftheusers.
3.2. Computertask
Inthedesignoftheexperiment,weintroducedasetoffilm
clipsthatwouldserveasanactivityfortheusertoperform
whilethesignalacquisitionprocesswastakingplace,butthat
wouldalsoelicitnon-specificaffectiveresponses.Inorderto
potentiate noticeablereactions inthesubjects we
manipu-lated the movies to introducea stimuli.Two MPEG4 video
sequenceswereprepared,witha1280×720pixelresolution
and total duration of 1min; a background soundtrackwas
alsomaintainedforgreaterimmersionandstimulatory
poten-tial.Inthesesequences,thelast5swereadjustedinorderto
presentatriggeringstimuli.
Toallow synchronization with the biosignal acquisition
system,andalsotomarkthebeginningofthevideoandthe
stimulipresentation,awhitesquarewith120×120pixelswas
insertedonthetopleftcornerofthescreen,whichappeared
for2sinthebeginningofthevideoandinthemomentwhere
the triggering stimuli is presented. An Apple iPad 2tablet
devicewasusedtodisplayandreproducethemultimedia
con-tent; the device wasplaced ona verticalstand infront of
theperson(Fig.12(a)),andasetofheadphoneswereused,to
providegreaterfocusonthereproducedcontentand maxi-mizetheimpactofthepreparedstimuli(Fig.12(b)).
Table2describesthecontentusedtoproduceeachstimuli. SequenceA1wasdesignedasamoreamusingvideo,to stim-ulatelowarousal;it started withanexcerptofanad with afairlyintenseaudiosoundtrackandsurprisingmoments,2
whichswitchedtoashortfunnyvideoneartheend.3Sequence
A2wasdesignedasamoreintensevideotostimulatehigh arousal;itstartedwithapeacefulsequenceextractedfroman ad,4andswitchedtoanexcerptofahorrormovietrailer.5Fig.7
showssampleframesfrombothaffectiveelicitationvideos.
3.3. Biosignalsensors
Electrocardiographic (ECG) data acquisition was performed withacustom,twoleaddifferentialsensordesignwith vir-tual ground, proposed in [19]; Table 3 outlines the main specifications.TwoECGsensorswereusedinourexperimental setup,oneforsignalacquisitionatthehandpalmswithdry
2 http://www.youtube.com/watch?v=wwO-wo892pI. 3 http://www.youtube.com/watch?v=zea6SCzjtfw. 4 http://www.youtube.com/watch?v=0bx8bnCoiU. 5 http://www.youtube.com/watch?v=YGJjPKOj1c.
Table3–SpecificationsoftheECGandEDAsensors.
ECG EDA
Filtering 1–30Hzbandpass 3Hzlowpass
InputImpedance >1M >1T
CMRR 110dB 110dB
Gain 1000 –
Ag/AgCl electrodes,and anotherforsignalacquisitionwith
Electrolycrasattheindexandmiddle fingers.Forimproved
comfortandgreaterefficiency,theECGsensorswerefittedto
aleatherbase,withtheintendedhandplacementsignaledin
anunequivocalway.
Fig.3depictsthedevisedsensorandelectrodes arrange-ment. One of the ECG sensors was connected to the dry Ag/AgClelectrodesthatwereplacedatthebaseofthehand palms,nearthethenareminence;theotherECGsensorwas connectedtotheElectrolycrastripsplacedalongtheindexand middlefingers.Thetransferfunctionforthissensorisgiven byEq.(1);VO istheoutputvoltageofthesensorandVECGis
theECGsignalvoltage,bothinVolt.Eq.(2)providesan equiv-alentrepresentation,whichusesthedigitaloutputcodesDO
oftheADCdirectly.Inoursetupweusedasystemwith12-bit resolution(n=12),Vcc=5VandVss=0V.
VECG= 1000VO (1)
VECG= DO∗Vcc
2n×103 (2)
Aspreviouslymentioned,simultaneousacquisitionofEDA datawasalsoperformedtoassessthearousallevelofeach par-ticipant.WerecurredtothecommerciallyavailableedaPLUX,6
which has two independentsensor leads, allowing a high level offlexibility in the sensorplacement. Pre-gelled and self-adhesiveAg/AgClelectrodes7wereusedasinterfacewith
theskinforimprovedconductivity.Table3outlinesthemain specificationofthesensor,andFig.2depictsthesensorand electrodes.
ThetransferfunctionforthissensorisgivenbyEq.3;VO
istheoutputvoltageofthesensorinVolt,andGSistheskin
conductanceinS(microSiemens).Eq.4providesan equiv-alentrepresentation,whichusesthedigitaloutputcodesDO
oftheADCdirectly.Inoursetupweusedasystemwith12-bit resolution(n=12),Vcc=5VandVss=0V.
GS=0V.2O (3)
GS=D2On××0V.2cc (4)
3.4. Dataacquisition
RawbiosignalswereacquiredwithabioPLUXresearch,8
Blue-tooth wireless biosignal acquisition unit; this device was
6 http://www.biosignalsplux.com. 7 http://www.spesmedica.it. 8 http://www.biosignalsplux.com.
Table4–SpecificationsofthebioPLUXresearch.
Connectivity ClassIIBluetoothconnectivity
Range upto10m
Resolution upto12bit
Samplingrate upto1000Hz
Weight 74g
Size 84mm×53mm×18mm
Battery Li-On;7.4V;800mAh
usedina12-bitresolution,1kHzsamplingfrequency
config-uration.ToguaranteeelectricalisolationbetweenbothECG
sensorsusedintheexperiments,twoindependentbiosignal
acquisitionunitswereused.Table4describesthemain
spec-ificationofthebiosignalacquisitionunit.
Fig.8depictsthehardwareconfigurationusedinour exper-imentsforbiosignalacquisition.Aspreviouslydescribed,an iPad was used for to reproduce the multimedia contents presentedduring the computer task/acquisition; two Light DependentResistor(LDR)sensorsweretapedwithblacktape (fordiscreetnesspurposes),tothecorneroftheiPadscreen abovetheareawherethevideo-embeddedwhitesquarewould appear.TheECGsensorarrangementwasplacedinfrontofthe iPad,insuchwaythatthepersoncouldcomfortablyrelaxthe handswhileperformingthecomputertask,thatis,watchthe videoclips.Electrodermalactivitysensorswereplacednear theECGsensorarrangementtoenablefastapplicationtothe person.Allthesensorswereconnectedtoeachofthe biosig-nalacquisitionunitsthatwereplacedclosetotheremaining apparatus.Table5presentsthechannelmappingforeachunit.
3.5. Timesynchronization
Time synchronization between both biosignal acquisition unitswasperformedopticallytoensurefullelectrical decou-pling,usingelementsfromasyncPLUXsynchronizationkit.To oneoftheunits,weconnectedaswitch,whichsimultaneously activatedalightemittingdiode(LED),andtriggeredaTTL sig-naltothedigitalinputportofthedevice;totheotherunit, weconnectedaLDR,whichwasdirectlyattachedtotheLED. Withthissetup,whenevertheswitchwaspressed,acommon signalwasrecordedbybothdevices,enablingtimealignment inthepost-processing(Fig.4).
A second synchronization mechanism was also imple-mented using LDR sensors, to guarantee synchronization betweenthebiosignalacquisitionunitsandthemultimedia contentdisplayed. For this purpose,two LDR sensorswere appliedtotheiPadscreenandconnectedtoeachofthe biosig-nal acquisition units. In a pre-test stage, the videos were playedto capturea “videoprint”based onthe screen light intensityvariations,whichcanbeusedtorelatethebiosignals activitywiththestimulibeingpresented.
Furthermore, this mechanism allowed us to introduce redundancyinthesynchronizationprocess,asthe informa-tionextractedfromthevideoprintsalonecanbeusedfortime alignmentofthesignalscollectedbyeachindependentunit. SincetheoperatingprincipleoftheLDRsensorsisalso opti-cal,electricaldecouplingbetweenbothindependentbiosignal acquisitionunitsisonceagainguaranteed.
Fig.4–Examplesofthesignalsextractedfromourinitial experiments.Allsignalsarenormalizedforimproved legibility.OnthetopchartweshowtheLED+TTLSwitch synchronizationsignals,whiletheECGsignalscollected withbothelectrodematerialsareshowninthebottom chart.
4.
Short-term
dataset
4.1. Participants
Theexperimentalsetupwasplacedonanunrestrained set-ting,inthelobbyofthebuildingwhereoneofourresearch teams is located: Torre Norte of the Alameda Campus of InstitutoSuperiorTécnico.Togenerateawarenessforthe ini-tiativeandincreasecompliance,asetofpromotionalbanners wascreatedandplacedclosetotheworkbench;avolunteer counter,wasregularlyupdatedtofurthermotivatepotential candidates. Fig.11 showsthedeviseddesigns,seenalsoin theirintendeduseinFig.1(a).
Sessionstookplaceovertwofulldays,andintheenddata hadbeencollectedforanoveralltotalof65participants,the majorityofwhichwereengineeringstudentsandresearchers. The demographics shown 49 males and 16 females, with anaverageageof31.1±9.46years.Noneoftheparticipants reportedanyhealthproblems,reasonforwhichweconsider thecollecteddatatoberepresentativeofthehealthy popula-tion.
Asarewardforparticipatingintheexperiments,agiftcard createdfromthecollectedbiosignalswasprepared(Fig.10). Withtheindividualheartbeatwaveforms,anorangecolormap imagewascreated,whereeachlinecorrespondstoan indi-vidualheartbeatwaveform,andeachcolumncorrespondsto the amplitude ofeach sample ofthe heartbeat wave data. WiththeindividualEDAevents,abluecolormapimagewas createdusingthesameprinciple,butinthiscaserotatedof −90◦.
Table5–Channelmappingsandsensorreferences.
Ch. Sensor Placement
00:07:80:42:0F:8BbioPLUXresearch
1 ECG-3-82 Handpalms(withAg/Clelectrodeandnogel)
2 EDA-1-25 Lefthand
3 LDR-1-4 iPadscreen
4 LDR-1-7 LED
00:07:80:42:0F:85bioPLUXresearch
1 ECG-3-83 Indexandmiddlefingers(withelectrolycra)
5 LDR-1-1 iPadscreen
i/o SYNC Syncadapter+LED(indirectcontactwithLDR-1-7)
4.2. Experimentalprocedure
Thewholeexperimentwasdesignedtohaveanaveragetotal
completiontimeofapproximately5min,scattered
through-out threestages:(a) Informedconsent(CI);(b)Lowarousal
video(A1);and(c)Higharousalvideo(A2).Onceparticipants
willinglyshowedinterestinbecomingpartoftheexperiments,
thefirststagewouldtakeplace,andconsistedonhavingthe
coordinator oftheexperiment going throughthe informed
consent(Fig.9),explainingtheprocedureindetail,goalsofthe
study,andrelatedbackgroundwork.Duringthisphase,only ECGsignalsatthehandpalmsandfingerswererecorded,and noemotionelicitingstimuliwerepresented.
Forthe secondphase,subjectswere fittedwiththeEDA sensorontheleft hand,inwhichoneterminalofthe sen-sorwas placedon themiddle phalanx oftheindex finger, andtheotheronthemiddlephalanxoftheringfinger.The headphoneswereplacedovertheears,andthelowarousal videowasstartedontheiPadbythecoordinatorofthe exper-iment.Forthethirdphase,theheadphoneswerekeptonthe subject,andthevideoontheiPadswitchedtothehigharousal videosequence.Duringboththesephases,ECGsignalsatthe palms,andEDAsignalswererecorded.
Ateachstage,recordingsfromeachindividualdevicewere storedinaseparatefileformoreefficientpost-processing,and labeledwiththedate,identificationofthesystem,andacode assignedtothesubject.Inthebeginningandintheendofeach recordingtheswitchwouldbetriggeredinordertoproducethe synchronizationsignal,necessarytofindthecommontime baseaspreviouslydescribed(Fig.4).
5.
Long-term
dataset
5.1. Participants
Thelong-termdatasetwascollectedoveraperiodofseveral days and different settings, reason for which a slimmed-downversionoftheproposedacquisitionsetupwasadopted. Weusedasinglebiosignalacquisitionsystemwithjustone sensortomeasureECGfittedwiththedryAg/AgClelectrodes. Theexperimentalsetupwaspreparedatthe cardiopneumol-ogylaboratorylocatedattheEscolaSuperiordeSaúdedaCruz VermelhaPortuguesa,thatwaspreparedtoreceivethe partic-ipants.
Therecruitmentofvolunteerswasperformedby advertis-ing thedaysinwhich dataacquisition sessionswere held, andbyprovidinganoverviewoftheactionduringclasses.A
totalof63subjects(nursingandhealthtechnologiesstudents), wereenrolledintheexperimentandparticipatedinthetwo acquisitionsessions.Thedemographicsshowed14malesand 49 females,withanaverageageof20.68±2.83years.None oftheparticipantsreportedanyhealthproblems,reasonfor whichweconsiderthecollecteddatatoberepresentativeof thenormalpopulation.
5.2. Experimentalprocedure
Twodataacquisitionmomentsseparatedbya3-month time-frametookplace,andenableddatarecollectionfromtheset ofvolunteersenrolledintheexperiment,withthepurposeof studyingthechangesintheECGmorphologyovertime.Inthe firstmoment,thecoordinatoroftheexperimentpresentedthe informedconsenttoprospectiveparticipants(Fig.9), explain-ing the procedure indetail,goalsofthe study,and related backgroundwork.Onceparticipantswillinglyshowed inter-estinbecomingpartoftheexperiments,theywererequired tosignthedocumentandwereenrolledinthedatacollection program.
In both moments only ECG signals at the fingers were recorded,andineachofthesessionsthesubjectswereasked tositfor2mininarestingposition,withtwofingers,onefrom theleftandanotherfromtherighthand,placedineachofthe dryAg/AgClelectrodes.Therecordingswerestoredin individ-ualfilesformoreefficientpost-processing,andlabeledwith the date,identification ofthesystem, and acode assigned to thesubject. The factthat it hastwo sessions separated byseveralmonthsapartmakesthisdatasetparticularly use-fulforbenchmarkingtheperformanceofidentityrecognition algorithms taking into account potential variations in the heartbeatwaveformovertime.
6.
Status
report
6.1. Outline
Ourexperimentstodatehavealreadyenabledseveral interest-ingfindingsusingthecollecteddata.First,asshowninFig.4
(bottom), the optical synchronizationmethod enablestime alignment withinthemillisecond,allowingthe signals col-lectedbyeachindependentunittobecomparedinthesame timebase.Also,aswecanseefromtheexamplepresented inFig.5,boththeECGsignalscollectedwithdryAg/AgCland Electrolycrasatthehandpalmsandatthefingersretainahigh levelofdetailwithrespecttotheheartbeatwaveform.
Fig.5–ECGsignalsatthehandpalms(dryAg/AgCl)and fingers(Electrolycras).
Fig.6–ExampleofafilteredandsegmentedECGsignal collectedatthehandpalmswithdryAg/AgCl.Theblack linedepictstheaverageofallwaveforms,whilethedashed linesdepictthestandarddeviation;thegraylinesshowan overlayofthesegmentedindividualheartbeatwaveforms foroneoftheacquisitionsessions.
Inparticular,not onlythe so-importantQRScomplex is clearlynoticeable,butalsothePandTwavescanbefound. Moreover,wecanseethatthereisahighmorphological sim-ilaritybetweenbothsignals,leadingustobelievethatboth theelectrolycrasandthedryAg/AgClelectrodescanbeused asinterfacewiththeskin,givinghigherflexibilityand con-venienceto the biometric system designer. Thisis further reinforcedbytheanalysisperformedin[8,14].
Fig.7–Sampleframesfromthevideoandstimuli sequences.
Fig.8–Fullexperimentalsetup.
Anadditionalexampleofthecollectedsignalsispresented inFig.6,wherewecanseeanexampleofafilteredand seg-mentedrecordingcollectedatthehandpalmswiththedry Ag/AgClelectrodes.
Still,signalscollectedatthehandsandfingers,especially using dry materials as electrodes, have a lower signal-to-noiseratio,whichraisesseveralchallengesnamelyinterms offiltering,segmentation,andoutlierdetection.Afirstsetof guidelinesregardingthepre-processingofthesesignalscan befoundintheworkbyLourenc¸oetal.[9,10];morerecently, toenhancetheECGsignalqualityandincreasethe signal-to-noiseratiointheECGbiometricscontext,westartedtoadopt a300orderbandpassFiniteImpulseResponse(FIR)filter,with aHammingwindow,andcutofffrequenciesof5Hzto20Hz. Thesespecificationstake intoaccounttheECGinformation bandwidth and empiricalconsiderations obtainedfrom the dataitself.
For researchers working on fiducial or partially fiducial approaches,segmentation isakey aspectinterms ofdata analysis.WereferthereadertothepapersbyBachleretal.[20]
Fig.9–Informedconsentdocumentthatwasreadtoandsignedbythevolunteers.
andbyCanentoetal.[21],inwhichadescriptionand compar-isonofseveralECGsegmentationalgorithmsispresented.
6.2. StudieswithCYBHi
Inpartorinfull,thisdatasethasalreadybeenusedin sev-eralpeerreviewedandacceptedstate-of-the-artpublications. In[13]astudy ofdifferentsegmentationalgorithmsis per-formed, which addresses the problem of online vs. offline segmentationofhand/fingersECGdata;63recordsfromthe
CYBHi short-termdatasetwereused,and theauthorshave concludedthatonlineapproachespresentcompetitiveresults bothintermsofthenumberofoveralldetectedsegments,and onthepercentageofthosethatcanbeconsideredtobevalid. In[12], astudyon thebiometric performanceoffullvs. partialwaveformispresented.Atotalof32recordsfromthe CYHBi short-term dataset were used, to evaluatehow the useofthewaveformcomprehendingtheP-QRS-Tsegments compares tousing onlythe R-STportionofthe waveform. Experimentalevaluationwasperformedbothondatafromthe
Fig.10–Participantgiftcardofferedtovolunteersenrolledintheexperiments,showinghighlightsoftheirownsignals.
dryAg/AgClelectrodesandElectrolycras,andresultsrevealed thattheP-QRS-Tprovideshigherrecognitionratesduetothe additionalinformation;inthebestcasescenarioa1.66%±2.28 identificationerror(EID)wasachievedforgroupsof2users. Nonetheless,withamarginaldecrease inperformance,the RS-Tsegmentbasedapproachisabletoprovideanadequate compromisesolution,giventhatithasalowertemporaland spacialcomplexity;forgroupsof2users,thisapproachhas revealed2.40%±3.96EID.
Targetingthereal-timeclassificationscenario,in[9]atotal of32recordsfromtheshort-termCYBHiwereused,to com-paretheperformanceofaNearestNeighborclassifierwithan
approachbasedonSupportVectorMachines(SVM). Experi-mentalresultshaveshownthat,dependingontheapplication, each approachhas its own advantages,and that the SVM basedapproachenableshighrecognitionrates,forexample,in anauthenticationtask,foranoperatingpointwheretheFalse AcceptanceRate(FAR)is0%,theFalseRejectionRate(FRR)can beaslowas13.91%±4.55.
In[22],twoPCA-basedmethodsarepresented,whichuse the fullCYBHi short-term dataset to explorethe concepts ofIndividualEigen-HeartbeatandOverallPopulation Eigen-Heartbeat, asamethod offeatureextractionand template creation.Resultshaveshownthata0%EIDcanbeachieved.
Fig.11–Promotionalbannerscreatedtosignalthe initiativetoprospectiveparticipants.
Fig.12–Multimediasetupforemotionelicitation.
In [23,24], the authors tried to cope with outliers and time-variance introduced by acquisitions over time using wavelets and clustering-based methods. Using 63 records, andthetwotemporalseparatedsessionsfromthelong-term CYBHi dataset,Equal Error Rates(EER) of11.5% and 12.4% wereachieved,revealingthatthisisachallengingtask.More recentlyin[11],wefurtherevaluatedthetemporalstabilityof ECGsignalscollectedatthefingers,reachinganEERof9.10% usingSVMsalsofortheCYBHilong-termdataset.
7.
Resources
7.1. Materials
Allmaterialsandoutputsofthedataacquisitionsessionswere groupedintoafolderstructure,whichcanbeusedforfuture
Table7–Acquireddatafilestructure.
Col. Content
1 7-bitsamplesequencenumber(fromthedevice) 2 Digitalinput(fromthedevice)
3 Digitaloutputtoggle(fromthesoftware) ≥4 Analogchannels≥1(fromthedevice)
reference. Theseincludegraphicmaterials,videos, the
col-lecteddata,andsupportdocumentation;Table6outlineswhat
canbefoundineachfolder.
7.2. Dataformat
Eachfilecontainsthedatacollectedfromthebiosignal acqui-sitionandsynchronizationsensors.Includeddatacomprises themodalitiesdescribedinSection3;thedifferentmoments oftheexperimentalprocedureacquiredwiththetwo biosig-nal acquisition units are stored in separate files. Files are identifiedwithindicationofthedate,acodeassociatedwith thesubject,andthemomenttowhichtheycorrespond(e.g. YYYYMMDD-AAA-BB-CC.txt),thegeneralnotationbeinggiven by:
<filename> ::=<date>“−”<code>“−” <moment>“−”<unit>“.txt” <date> ::=YYYYMMDD
<code> ::=up to three characters <moment> ::=“CI”|“A0”|“A1”|“A2” <unit> ::=“8B”|“85”
Intheabovenotationusedtodistinguishthemoment,“CI” denotestheinformedconsentpartofthetest,while“A1”and “A2”denoterespectivelythelowandhigharousalvideosinthe short-termdataset;“A0”denotestheacquisitionsofthe long-termdataset.Inthenotationusedtodistinguishthebiosignal acquisitionunit,“8B”and“85”refertothe“00:07:80:42:0F:8B” and “00:07:80:42:0F:85” units,respectively. Data isstoredin standardASCIItextwithabioPLUXfileformat,whereeachline correspondstoonesample;thecolumnsmappingisdescribed inTable7.Theheaderhaseightlines,precededbythe# sym-bol,andthestoredinformationisthefollowing:
# bioPlux Text File Format # Version: 1 # StartDateTime: YYYY-MM-DD HH:MM:SS # SamplingFrequency: 1000 # SampledChannels: 1 2 3 4 5 6 7 8 # SamplingResolution: 12 # AcquiringDevice: XX:XX:XX:XX:XX:XX # EndOfHeader
Table6–Descriptionoftheprojectresources.
Folder Content
data Rawdatacollectedduringtheexperimentsstoredinindividual.txtfiles
documents Supportdocumentationusedintheexperiments:informedconsent,participantlist,andacquisitionsetupspecification
graphics Graphicmaterialspreparedfortheexperiments:bannersandgiftcards
photos Picturesofdifferentaspectsoftheexperimentalsetup
results Graphicsanddatafileswithresultsfrominitialdataprocessing
8.
Lessons
learned
WhiledatasetswithECGcollectedatthechestusing clinical-gradeequipmentare readily available,the samecannot be saidregardingsignalscollectedatthehandpalmsorfingers inanoff-thepersonapproach.Amongotherpotential applica-tions,theseunconventionallocationsarefundamentalforthe growingresearchfieldofECGbiometrics,asthechest place-mentisfartoointrusiveforfeasiblereal-worlddeployment. ThesenoveltrendsinECGbiometricsrequiremoreconvenient signalacquisitionmethodshowever, inthisparticulararea, researchersstilllackpubliclyavailabledatasetsthatcan pro-motebenchmarkingandcomparisonofresultsacrossteams. TheCheckYourBiosignalsHereinitiative(CYBHi)dataset, provideshighresolutionrawdatafromsignalscollectedatthe handpalmsandfingersinanoff-the-personapproach,using dryAg/AgClelectrodesandElectrolycrasasinterfacewiththe skin.Wehavedescribed abiosignalacquisitionframework, devisedtoovercometheneedforarepeatableand easy-to-applyprocedurefordatacollectioninlargepopulationgroups. Thissetupbuildsuponpriorworkfromourgroupinthearea ofbiosignalresearch,asatooltoidentifypotentiallytime vari-antpropertiesinthesignals,andtoevaluatetherobustness ofalgorithmsandmethodstochangesintheenvironmental conditionsandotherdataacquisitioninfluencingfactors.
Westartedtocompileashort-termdatasetwhereasingle acquisitionsessionwasperformedtocollectsignalsthatcan comparedifferentelectrodematerials,aswellastheeffectof arousal-elicitingstimuliintheheartbeatwaveformtemplates. Along-termdatasetwasalsostarted,inwhichtwodata acqui-sitionsessionswereperformed,enablingtheevaluationand benchmarkingofthepersistenceoftheheartbeatwaveform templatesovertime.Several studieshave alreadyusedthe CYBHidatasetandprovidedvaluableinsightonthepotential ofthecollecteddata,aswellasguidelinesonhowtoperform pre-processingofthedata.Ourdatasetisalreadymade avail-ableforresearchersworldwide.9Furthermore,wealsocreated
asetoftoolsthatresearcherscanusetofurthercontributeto theCYBHidataset,andreplicatetheexperiments,in particu-lar,byprovidingaccesstothebiosignalacquisitionhardware itself.10
9.
Future
plans
Ongoingworkisfocusingonfurtherextendingthenumberof subjectsbothintheshort-termandlong-termdatasets,and alsoenrichingthepotentialofourdatasetbyrecallingthe sub-jectsthatwerepreviouslyassessedinordertohavearecord ofthecardiacactivityofeachsubjectindifferentmoments intimeseparatedbyseveralmonthsoryearsapart.Weare currentlyrecallingthepopulationenrolledintheCYBHi long-termexperiments,inordertorecord1-yearapartdata.
Inparallel, we are also workingto provide morphologi-calground-truthdatabymanuallabeling offiducialpoints inthe recordings,whichwillextendtheuse ofourdataset
9 http://camoes.lx.it.pt/CYBHi.rar. 10 http://www.bitalino.com.
beyondthefieldofECGbiometrics,enablingforexamplethe assessmentandbenchmarkingoffilteringandsegmentation algorithms[25–27]. Anotheraspectthat willbeanalyzedin futureresearchworkisthecorrelationbetweentheECG sig-nalsandthevariationsofthearousallevelofeachsubjectas measuredthroughthecollectedEDAdata.
Regardingthedatasetitself,wewillseekthepossibilityof makingitavailablethroughPhysionet[7]inaneartime,and makingitcompliantwiththeISOstandardsforbiometricdata storageandexchange.Finally,thedatasetisnowprovidedas standardASCIItextfiles;futureworkwillalsofocuson migrat-ingthedatasettoastructureddatabaseformat[28–30].
Acknowledgements
ThisworkwaspartiallyfundedbytheFundac¸ãoparaa Ciên-ciaeTecnologia(FCT)undergrantsPTDC/EEI-SII/2312/2012, SFRH/BD/65248/2009andSFRH/PROTEC/49512/2009,bytheIT –InstitutodeTelecomunicac¸õesunderthegrant“Android Bio-metricSystem”,andbytheDepartamentodeEngenhariade ElectrónicaeTelecomunicac¸õesedeComputadores,Instituto SuperiordeEngenhariadeLisboa(ISEL), whosesupportthe authorsgratefullyacknowledge.Theauthorswouldalsolike toacknowledgeJoanaSantosandAnaTabordafromtheEscola Superior de Saúde da Cruz Vermelha Portuguesa for their invaluablesupportinpartoftheexperimentaldataacquisition forthelong-termdataset.
r
e
f
e
r
e
n
c
e
s
[1] A.K.Jain,A.Ross,K.Nandakumar,Introductionto Biometrics,SpringerUS,NewYork,USA,2011.
[2] Y.Wang,F.Agrafioti,D.Hatzinakos,K.N.Plataniotis, Analysisofhumanelectrocardiogramforbiometric recognition,EURASIPJournalonAdvancesinSignal Processing2008(2008)19.
[3] A.Chan,M.M.Hamdy,A.Badre,V.Badee,Waveletdistance measureforpersonidentificationusingelectrocardiograms, IEEETransactionsonInstrumentationandMeasurement57 (2)(2008)248–253.
[4] I.Odinaka,P.-H.Lai,A.D.Kaplan,J.A.O’Sullivan,E.J. Sirevaag,J.W.Rohrbaugh,ECGbiometricrecognition:a comparativeanalysis,IEEETransactionsonInformation ForensicsandSecurity7(6)(2012)1812–1824.
[5] S.A.Israel,J.M.Irvine,B.K.Wiederhold,M.D.Wiederhold, TheHeartbeat:TheLivingBiometric,Wiley-IEEEPress, Hoboken,NewJersey,2009,pp.429–459(Chapter17).
[6] C.Zhao,T.Wysocki,F.Agrafioti,D.Hatzinakos,Securing handhelddevicesandfingerprintreaderswithECG biometrics,in:Proc.oftheIEEEInt’lConf.onBiometrics: Theory,ApplicationsandSystems(BTAS),2012,pp.150–155.
[7] A.Goldberger,L.Amaral,L.Glass,J.Hausdorff,P.Ivanov,R. Mark,J.Mietus,G.Moody,C.Peng,H.Stanley,PhysioBank, PhysioToolkit,andPhysioNet:Componentsofanew researchresourceforcomplexphysiologicsignals, Circulation101(23)(2000)215–220.
[8] H.Silva,C.Carreiras,A.Lourenc¸o,A.Fred,Off-the-person electrocardiography,in:Proc.oftheInt’lCongresson CardiovascularTechnologies(CARDIOTECHNIX),2013,pp. 99–106.
[9] A.Lourenc¸o,H.Silva,A.Fred,ECG-basedbiometrics:areal timeclassificationapproach,in:Proc.oftheIEEEInt’l
WorkshoponMachineLearningforSignalProcessing (MLSP),2012,pp.1–6.
[10] A.Lourenc¸o,H.Silva,A.Fred,Unveilingthebiometric potentialofFinger-BasedECGsignals,Computational IntelligenceandNeuroscience2011(2011)5.
[11] H.Silva,A.Lourenc¸o,A.Fred,A.Jain,FingerECGsignalfor userauthentication:usabilityandperformance,in:Proc.of theIEEEInt’lConf.onBiometrics:Theory,Applicationsand Systems(BTAS),2013.
[12] H.Silva,A.Lourenc¸o,A.Fred,In-vehicledriverrecognition basedonhandECGsignals,in:Proc.oftheACMInt’lConf. onIntelligentUserInterfaces(IUI),2012,pp.25–28.
[13] A.Lourenc¸o,H.Silva,P.Leite,R.Lourenc¸o,A.Fred,Realtime electrocardiogramsegmentationforfingerbasedECG biometric,in:Proc.oftheInt’lConf.onBio-InspiredSystems andSignalProcessing(BIOSIGNALS),2012,pp.49–54.
[14] C.Carreiras,A.Lourenc¸o,H.Silva,A.Fred,Comparative studyofmedical-gradeandoff-the-personECGsystems,in: Proc.oftheInt’lCongressonCardiovascularTechnologies (CARDIOTECHNIX),2013,pp.115–120.
[15] C.Stickel,M.Ebner,S.Steinbach-Nordmann,G.Searle,A. Holzinger,Emotiondetection:applicationofthevalence arousalspaceforrapidbiologicalusabilitytestingto enhanceuniversalaccess,in:C.Stephanidis(Ed.),Universal AccessinHuman–ComputerInteraction.Addressing Diversity,Vol.5614ofLectureNotesinComputerScience, Springer,Berlin,Heidelberg,2009,pp.615–624.
[16] R.W.Picard,AffectiveComputing,TheMITPress, Cambridge,USA,2000.
[17] F.Agrafioti,D.Hatzinakos,A.K.Anderson,ECGpattern analysisforemotiondetection,IEEETransactionson AffectiveComputing3(1)(2012)102–115.
[18] C.Carreiras,H.Aidos,H.Silva,A.Fred,ExploratoryEEG analysisusingclusteringandPhase-Lockingfactor,in:Proc. oftheInt’lConf.onBio-inspiredSystemsandSignal Processing(BIOSIGNALS),2013,pp.79–88.
[19] H.Silva,A.Lourenc¸o,R.Lourenc¸o,P.Leite,D.Coutinho,A. Fred,Studyandevaluationofasingledifferentialsensor designbasedonelectro-textileelectrodesforECG biometricsapplications,in:Proc.oftheIEEESensorsConf., 2011,pp.1764–1767.
[20] M.Bachler,C.Mayer,B.Hametner,S.Wassertheurer,A. Holzinger,OnlineandofflinedeterminationofQTandPR intervalandQRSdurationinelectrocardiography,in:Q.Zu,
B.Hu,A.Eli(Eds.),PervasiveComputingandtheNetworked World,Vol.7719ofLectureNotesinComputerScience, Springer,Berlin,Heidelberg,2013,pp.1–15.
[21] F.Canento,A.Lourenc¸o,H.Silva,A.Fred,OnrealtimeECG segmentationalgorithmsforbiometricapplications,in: Proc.oftheInt’lConf.onBio-inspiredSystemsandSignal Processing(BIOSIGNALS),2013,pp.228–235.
[22] M.Santos,A.Fred,H.Silva,A.Lourenc¸o,Eigenheartbeats foruseridentification,in:Proc.oftheInt’lConf.on Bio-inspiredSystemsandSignalProcessing(BIOSIGNALS), 2013,pp.351–355.
[23] C.Carreiras,A.Lourenc¸o,H.Silva,A.L.N.Fred,Aunifying approachtoECGbiometricrecognitionusingthewavelet transform,in:Proc.oftheInt’lConf.onImageAnalysisand Recognition(ICIAR),2013,pp.53–62.
[24] A.Lourenc¸o,H.Silva,C.Carreiras,A.L.N.Fred,Outlier detectioninnon-intrusiveECGbiometricsystem,in:Proc.of theInt’lConf.onImageAnalysisandRecognition(ICIAR), 2013,pp.43–52.
[25] J.Abawajy,A.Kelarev,M.Chowdhury,Multistageapproach forclusteringandclassificationofECGdata,Computer MethodsandProgramsinBiomedicine112(3)(2013) 720–730.
[26] S.Das,K.Maharatna,Fractionaldynamicalmodelforthe generationofECGlikesignalsfromfilteredcoupledVan-der Poloscillators,ComputerMethodsandProgramsin Biomedicine112(3)(2013)490–507.
[27] T.Ji,Q.Wu,Broadbandnoisesuppressionandfeature identificationofECGwaveformsusingmathematical morphologyandembeddingtheorem,ComputerMethods andProgramsinBiomedicine112(3)(2013)
466–480.
[28] C.Carreiras,H.Silva,A.Lourenc¸o,A.Fred,StorageBIT:a metadata-aware,extensible,semantic,andhierarchical databaseforbiosignals,in:Proc.oftheInt’lConf.onHealth Informatics(HEALTHINF),2013,pp.65–74.
[29] K.K.-Y.Lee,W.-C.Tang,K.-S.Choi,Alternativestorelational database:comparisonofNoSQLandXMLapproachesfor clinicaldatastorage,ComputerMethodsandProgramsin Biomedicine110(1)(2013)99–109.
[30] J.Meyer,S.Ostrzinski,D.Fredrich,C.Havemann,J.Krafczyk, W.Hoffmann,Efficientdatamanagementinalarge-scale epidemiologyresearchproject,ComputerMethodsand ProgramsinBiomedicine107(3)(2012)425–435.