HAL Id: hal-00294838
https://hal.archives-ouvertes.fr/hal-00294838
Submitted on 29 Nov 2011
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Interactive evolution for cochlear implants fitting
Pierrick Legrand, Claire Bourgeois-Republique, Vincent Pean, Esther Harboun-Cohen, Jacques Lévy Véhel, Bruno Frachet, Evelyne Lutton, Pierre
Collet
To cite this version:
Pierrick Legrand, Claire Bourgeois-Republique, Vincent Pean, Esther Harboun-Cohen, Jacques Lévy
Véhel, et al.. Interactive evolution for cochlear implants fitting. Genetic Programming and Evolvable
Machines, Springer Verlag, 2007, 8 (4), pp.319-354. �hal-00294838�
PierrikLegrand
1,2
,ClaireBourgeois-Republique
4
,VinentPéan
5
,
EstherHarbounCohen
6
,JaquesLevy-Vehel
2
,BrunoFrahet
6
,
EvelyneLutton
2
,PierreCollet
3
1
LaboratoiredeMathématiquesappliquéesdebordeaux,UMR5466 CNRS,U.F.R
SienesetModélisation, UniversitédeBordeaux2,146rueLéoSaignat,33076
Bordeauxedex,Frane
Legrandsm.u-bordeaux2.fr
2
COMPLEXTeam-INRIARoquenourt,B.P.105, 78153LeChesnayedex,
Frane
Pierrik.Legrandinria.fr,Evelyne.Luttoninria.fr,Jaques.levy-vehelinria.fr
http://fratales.inria.fr
3
Laboratoired'InformatiqueduLittoral,ULCOBP719, 62100Calaisedex,Frane
Pierre.Colletuniv-littoral.fr
http://lil.univ-littoral.fr
4
LE2I,UMR5158CNRS,9avenueA.Savary,B.P.47870,21078Dijonedex,Frane
Claire.bourgeois-Republiqueu-bourgogne.fr
5
CRTInnoteh,1PromenadeJeanRostand,93005Bobignyedex,Frane
vinent.peaninnoteh.fr
6
HpitalAvienne,ServieORL,125rtedeStalingrad,93000Bobigny,Frane
bruno.frahetav.ap-hop-paris.fr
Abstrat. Cohlear implantsare deviesthatbeome moreandmore
sophistiatedandadaptedtotheneedofpatients,butinthesametime
theybeomemoreand morediult to tune.After adeafpatienthas
beensurgiallyimplanted,aspeialisedmedialpratitionerhastospend
hoursduringmonthstopreiselyttheimplanttothepatient.Thispro-
essisaomplexoneimplyingtwointriatedtasks:thepratitionerhas
totunethe parametersofthe devie(optimisation)while thepatient's
brainneedstoadapttothenewdatahereeives(learning).Thispaper
presentsastudythatintendstomaketheimplantmoreadaptabletoen-
vironment(auditiveeology)andtosimplifytheproessoftting.Real
experimentson volunteer implanted patients are presented, that show
theeienyofinterativeevolutionforthispurpose.
1
1
Thisworkhas partiallybeenfundedbytheFrenhANR-RNTSHEVEAprojet
Cohlear Implants (CI) [Nih89℄ allow totally deaf people to hear again pro-
videdtheirauditorynerveandohleararestillfuntional:aomputerproesses
sounds piked up from a mirophone,to stimulate diretly the auditory nerve
throughseveraleletrodesinsertedinsidetheohlea(f.g.1).
As one an imagine, there are hundreds of parametersthat an be tuned,
andin thesametimethepatienthasto learntohear usingnewinformations
provided to his auditory nerve. Thetuning of suh a devie is thus extremely
omplex, and highly dependent on thepatient. This proess is urrentlydone
byhand bymedialpratitioners,andlookslikeanoptimisationproessbased
ontrialand error. This proessis so deliatethat sometimes,nosatisfatory
ttinganbefoundforsomepatients.
Hene,itseemsinterestingtouseaninterativeevolutionaryalgorithm(IEA)
to help ndingthe best valuesfor implantparameters. This is themain topi
of the HEVEA projet, whih is aollaboration between omputer sientists,
signalproessingexpertsandmedialresearhers.Theaimisatuallytwofold:to
failitatetheinitialttingofohlearimplants,andtoautomatisetheadaptation
ofohlearimplantstovarioussoundenvironments.AsimpleIEAwasdeveloped
with this in mind, and tested on a verybasi feature, the rangeof intensities
thataspeieletrodeantakewhenstimulatingtheauditorynerve.TheIEA
hasbeenimplementedonaPDAandtestshavebeenperformedonvolunteering
patientswithsatisfyingresults.
Thepaperisorganisedas follows:setion2presentsohlearimplants,and
setion3desribeshowtheyareurrentlytuned bymedialpratitioners.The
approahoftheHEVEAprojetisdevelopedinsetion4,andarstimplemen-
tation of an IEA is detailed in setion 5. Experimentson several patients are
reported insetion6,yielding goodresultsas wellas importantonlusionson
manual tting proedures.This rst validation step is important: an analysis
ofthesuessandfailuresraisesnewquestions thataredevelopedinsetion7,
related tothe well-knownuserfatigue problemof IEAs,and to thefat that
dierent sound environments havean important inuene on implants tting.
Automati adaptationof the devie to sound has been investigated, based on
asoundsignallassiationsheme,whihisdetailed in setion7.Conlusions
andperspetivesaredesribedin setion8.
2 Cohlear Implants
A ohlearimplantis asurgially implantable devie[GFM
+
98℄ that provides
hearingsensationstoindividualswithseveretoprofoundhearingloss,andan-
not benet from hearing aids. In a normal ear, sound energy is onverted to
mehanialenergybythe middleear, whih is thenonverted to eletrialim-
pulses by the inner ear (see gure 1). In order to perform this last stage, the
ohlea(part oftheinner ear) ontainsauidwhihis set into motionbythe
oval window whih is onneted to the middle ear. Within the ohlea, sen-
auditorynerve.Cohlearimplantsaredesignedtobeasubstituteforthemiddle
ear,ohlearmehanialmotion,andsensoryells,transformingdiretly sound
energy into eletrial energy that will initiate impulses in the auditory nerve
[B.C95℄,[Coh89℄thanksto adigitalsignalproessor.
Fig.1.Allimplantdevieshavethefollowingfeaturesinommon:soundisolletedby
amirophone(1)andsenttoeletroniomponentswithinaspeehproessor(2).The
speehproessor analyzes theinput signal (sound)and onverts it intoaneletroni
signal (eletrial).Thisodetravelsalongaable(3)tothetransmittingoil(4)and
is sent aross the skin viafrequeny modulated (FM) eletro-magnetiwavesto the
implant pakage(5).Basedonharateristisofthe odetransmittedtotheinternal
devie,eletrodeontats withinthe ohlea (6)provideeletrial stimulationto the
spiral ganglion ells and dendrites extending into the modiolus. Eletrial impulses
thentravelalongtheauditorynerve(7),asendingauditorypathwaystothebrain.
Cohlearimplants havebeen verysuessful in restoring partial hearing to
profoundly deaf people [ALM95℄, [Osb97℄. In2006, around 70 000deaf people
are implanted withsuh deviesaround theworld.Eienyis quite variable,
ranging from totallydeaf patientsthat havefully reoveredtheiraudition and
are apable to follow telephone onversations and enjoy musi, to others who
hear strange sounds they an't benet from, to a point where they prefer to
swithotheimplant[COM94℄,[GTBVC01℄,[BTE04℄,[Rom98℄.
For many people, it is still diult to fully take advantage of the devie
beause it is not easy to tune the parameters of digital signal proessor and
adjustthemfortheharateristisforeahpatient,sineallpatientsaredierent
(auseofdeafness,numberofyearsbetweentotaldeafnessandimplantation,age,
depthof eletrodeinsertion,...).
Researhhasbeengoingonsinenearly 50yearsagoonhowto eletrially
with one eletrode only, somelukypatientsmanaged to hear again and even
understandspeeh.Nowadays,itistehnologiallypossibletousemorethanone
eletrode,in order to stimulate moreofthethousands ofneuronsthe auditory
nerveismadeof[PCMF79℄[CFML83℄.However,themoreeletrodes,themore
parameterstotune.
The ohlea is used to interfae eletrodes and the auditory nerve. The
ohleaisabiologialdeviethatmainlyallowstomapdierentsoundfrequen-
iesontodierentneurons.Itisshapedlikeasnailshell.Onlylongwavelengths
(lowfrequeny sounds)an reah thefar endof theohlea,whileshort wave-
lengths(highfrequenysounds)arestoppedattheentraneoftheohlea.The
ideaisthenforsurgeonsto usethisfrequenydisriminatorandinsertintothe
ohleaathinsilionwire,bearingseveraleletrodes.
Stimulatinganeletrodeon thefar endof thewirewill therefore makethe
patienthearalowpithsound,whilestimulatinganeletrodeneartheentrane
oftheohleawillresultinthepatienthearingahighpith sound.
3 Cohlear Implant tting
3.1 Complexity ofthe problem
Being able to use morethan one eletrode to stimulate dierent neuron areas
isindeed agreatimprovement,butthenumberofparameterstotuneinreases
drastially.Conerningeletrodesonly,manyquestionsarise,amongwhih:
Whihfrequeniesshould bemappedtowhiheletrodes?
Whihrangeofintensitiesshouldbeappliedtowhiheletrodes?
Howmanyeletrodesshouldbestimulatedsimultaneously?
Shouldtheproessorprohibit neighboureletrodesto bestimulatedsimul-
taneouslyinordertoavoiddiaphony(rosstalkbetweennearbyeletrodes)?
Findinggood answersto thesequestionsis adiultoptimisationproblem.
Thisnotonlydueto theextremelylargesize ofthesearhspaebutto several
other reasons.First of all, thequalityof attingis atwostage proess where
subjetivityplaysalargerole:thepratitionerhastointerpretthequalityofthe
tting (rstsubjetiveproess)from the answersgivenby thepatient (seond
subjetiveproess).Thedisparityofpatientbehaviourwithrespettolanguage
andsensitivitytovariousthresholds,aswellastheharaterofthepratitioner
deeply inuenes the results. For example the well known psyhologialPyg-
malion eet biases answers of the patient, who often unonsiously tries to
satisfythepratitioner'sexpetations.
Thesoundenvironmentisanotherauseofvariabilityofresults,asthetting
session usually takes plae in a small room at hospital with the pratitioner.
Howevertheohlearimplantmustalsobeusedinreallife,andaorrettting
athospitalmayrevealveryunomfortableor unusefulwheninthestreet,orin
to test many possibleparameter sets during a singlesession, so the proess is
verylongandneedssometimes weeksto obtainasatisfyingresult.Inthesame
time,attingthatmaynotappearimmediatelyassatisfying,mayimprovewhen
testingitonalongerperiod (brainhasaplastiitythat annotbenegleted).
There are many fatorsthat make this problem highly irregular.However,
it has been proved that an aeptable or even good tting is reahable by a
manual searh onduted by an experiened pratitioner. We desribe below
thismanualttingtehnique,whih ismainlyahuman-guidedtrialanderror
proess,resemblingaloal searh.
3.2 Manualtting
Nowadays,dependingonthemanufaturer,thenumberofeletrodesvaries be-
tween8and22.Cohlearimplantttingisperformedbyanexpertpratitioner,
whoproeedsinthefollowingway:
Right after the surgial intervention, the pratitioner tries to determine
whiheletrodesarefuntional(aneletrodeisfuntionalifthepatienthears
asoundwhenurrentisapplied totheeletrode).
For eah funtional eletrode, thepratitionertries to determinetherange
ofintensitiesthatanbeused.Thelowestintensityabovewhihthepatient
pereives a sound is alled
T
(for Threshold). The maximum onfortable intensity (loudest sound the patient an bear for a reasonable amount oftime)isalled
C
(forComfortthreshold).Determining the
T
andC
values for eah eletrode takes time (ommuni-ation with a deaf patient, a young hild, or with an old patient an be
diult), and due to the inreasing number of eletrodes, some manufa-
turersnowadvise to determine
T
andC
valuesforone everythree or foureletrodes,andextrapolatethevaluesfortheothereletrodes.See[Rou01℄,
[Hes02℄ formoreinformationsonthistopi.
Othermanufaturersevenset averagevaluesfor
T
andC
, basedonneuralresponseorevenstatistis.
Then,one the
C − T
rangeis maximised for allthe eletrodes, the realttingbegins. The pratitioneruses his expertise to map frequeny bands
logarithmiallyontothedierentfuntionaleletrodes,andstartstotunethe
gainandsensitivitydependingonsoundfrequenies,thentunesthenumber
of simultaneously ative eletrodes,... while at the same time asking the
patientwhethertheyunderstandbetterorworse,whetherthesoundquality
is omfortable or not, a.s.o..In ertain ases, the pratitionerwill slightly
reduethe
C − T
rangeforsomeeletrodes,whenhehasthefeelingthattheneurologibandwidthislimited,andthattheneuronsfaingtheeletrode
aregettingsaturatedatonlymoderateauditorylevels.
Resultsarevariable,butoftengood.Usually,attingsessionstartswiththe
pratitioneraskingwhethertheurrentttingisbetterorworsethantheprevi-
ousone.Thebest ofthereentttingsistakenas abasisthatthepratitioner
tries to modify some parameters to help solving the problems. Two or three
parametersanbehangedduringa30to90minutesttingsession.Then,the
patientleaveswiththenewsettingsthathekeepsforaoupleofmonths,before
he omes bakfor another tting session. Thewhole proess is therefore very
long(severalyearsforproblematipatients).
4 Desription of the Problem
Asseenabove,ttingohlearimplantsisdonethroughaset oforrelatedpa-
rameters[LPD00℄,andpereptionandomfortthresholdsarelinkedtohistopatho-
logialfatorsspeitothepatient[KSC
+
98℄.Inmostases,thettingstrategy
simplyonsistsinmaximisingthenumberofeletrodesandmaximisingtheirdy-
namirange[BPG
+
92℄.Thisoftengivesgoodresults,butforsomepatientsthis
approah does not work.Moreover, the following observations have also been
reported:
Betterresultsmightbeobtainedbydereasingthedynamirange[FXP03℄.
Onlyusingasubsetofeletrodesmightimprovespeehreognition[ZCW97℄.
Holesinspetralrepresentationan existintonotopirepresentation(map-
ping of the sound frequenies on the eletrodes) and spetral information
redistributionaroundtheholesdoesnotinreasesresults[SGD02℄.
Moreover:
Mostofthepatientsdonotusealltheinformation givenbytheeletrodes
[Fis96℄.
All the eletrodes are not neessary to obtain maximal speeh pereption
performaneinsilent[DDML89,LWZF96,Fis96,KVR
+
00℄andnoisyenviron-
ments[FSBW01℄(partofthisouldbeduetoeletrialinterationbetween
hannels[SLM
+
06℄).
Thesepublishedobservationsshowthathoosingagoodsubsetofeletrodes
an havean inuene on speeh understanding, as well as the dynami range
on theeletrodes.Finally, taking intoaounta realsound environmentould
inreasespeehunderstandingforsomepatients.
Theworkpresentedinthispaperwill trytoaddressbothproblems.
5 Desription of the Interative Evolutionary Algorithm
Itseemsthatmanypatientswhoarenotsatisedwiththeirohlearimplantare
stuk in aloaloptimum: nomodiationproposedby theexpert would bring
anyimprovement.
Thistriggered the ideato use evolutionaryalgorithms, that are bothquite
goodatoptimising parametersandnoteasilytrappedin loal optima.Thege-
netiloopisthefollowing:theEAsuggestsasetofparametersthatarediretly
evolutionary algorithms, [Dur02,Tak01,Tak02℄, but they onern only onven-
tional hearing aids, with a relatively small number of parametersthat an be
tuned. Toourknowledge,nobodyhastriedtoapplyevolutionaryalgorithmsto
CohlearImplantstting.
5.1 Managing the runs
In an interativeevolutionary algorithm, a humanuser evaluates the dierent
individualsproposedbythealgorithm.
ThomasBäk'sresults([Bae05℄),suggestthatanevolutionaryalgorithmmay
do as well (if not better) than ahuman expert on a number of evaluations of
thesameorderthanthenumberofrealparameterstooptimise.Therefore,ifthe
problem hasaround 100 parameters to tune, performing only 100evaluations
should already allow to obtain interesting results. If it is possible to nd an
evaluationproedurethat takesaround5mn,arun wouldlast around8hours.
However, it is also important to take psyhology and human fatigue into
aount: awell tuned onvergene speed over100 evaluations ould seemdis-
ouragingfor a human patient, who may think that improvement is too slow.
Besides, sine it is not possible to have an 8 hour run in one go, an elegant
solution onsists in frationing the experimentation into several partial fast-
onvergingruns, with arestartat the end of eah run [Jan02℄. Dividing the8
hourruninto 5makesfor51h30runs,thatarequitemanageable.
Ratherthannding waysto avoidprematureonvergene,it isonthe on-
traryaveryfastonvergenethatissoughtontheseshortrunsofapproximately
20runs.Thisisquitenie,sineevolutionaryalgorithmsareknowntoonverge
quitefast,ifnoounter-measuresaretaken.
Thispoliy allowstouse averyfastonvergingalgorithm tryingtoexploit
loal minima,ratherthanaslowonvergingalgorithmtryingtowidelyexplore
thesearhspae,lookingfortheglobalminimum.Theonsequenesofpremature
onvergenearedealtwiththankstotheperiodialrestarts.Duringthelastrun,
oneanrestartthealgorithmwiththebestindividualsfoundinthe4rstruns,
soas tobenetfromtheresultspreviouslyfound.
Populationsizeandnumberofhildrenpergeneration. Foranidential
number of evaluations, twopossibilities exist:either many hildren per gener-
ation and a small number of generations, or a small number of hildren per
generationandmanygenerations.
Outofthesetwopossibilities,itisthealgorithmthatmaximisesthenumber
ofgenerationsthatwillfavourmostonvergene.ThissuggestsaSteadyStatere-
plaementpoliy,ora
(µ+λ)
withaveryreduedλ
(numberofhildren)[Bae95℄.Thenin orderto notspend toomanyevaluationsin theinitialpopulation,one
analso redueitasisdoneinmiro-GAs[Kri89℄.
Extremely lowvaluesan beused, suhas 3to 6individuals forthe initial
4previousruns.
Thealgorithmhosenforthisspeiinterativeoptimisationwilltherefore
beamodernevolutionaryalgorithm,inthesensethatitdoesnottakeafterany
of thefour usual paradigms(EvolutionStrategies, GenetiAlgorithm, Geneti
Programming,EvolutionaryProgramming)[DJ05℄.
Aording to Bäk [Bae05℄, using an Evolution Strategie paradigm for100
evaluations should allowto optimiseup to 100real variables. In CohlearIm-
plantstting,however,oneanstartwithtryingtondthebest
T
andC
valuesforeaheletrode.WiththeMXM15eletrodesCIusedforthisexperiment,the
genomeisthereforeanarrayofonly30realvalues,meaningthatthehanesto
ndagoodttingaremuh higher.
5.2 Initialisation
Onehardonstraintneedstoberespeted:thealgorithmshouldnotgobeyond
themaximumintensityforeahoftheeletrodesforfearofdestroyingsomeof
the patient's auditory neurons. Therefore,for eah newpatient, arst session
withapratitionerisrealisedtodeterminethemaximumadmissibleintensityfor
eaheletrode,thatisalledapsyhophysial test.Inordertoreduethesearh
spae, a minimal intensity below whih the patient does not hear anything is
alsodetermined.
The initialisation of eah individual therefore simply onsists, for eah of
the15eletrodes,to pikuptworandom valueswithin the
[min, max]
intervaldetermined during thepsyhophysialtest, and to takethelowervalueas a
T
threshold,andthehighervalueasa
C
thresholdfortheeahofthe15eletrodes.5.3 Seletion of the parents
Parents seletion is dierent from the replaement stage, in that it an selet
an individual several times. Whenever a hild must be reated, two dierent
individuals are seleted among the parent's population, that an be seleted
againtoreateanotherhild.
Sinetheseletionpressureofproportionalseletiondepends onthetness
landsape of the problem to be solved(whih is unknown), astohasti tour-
nament is seleted [BT97℄, with a 90% probability, that onsists in randomly
seleting2individualsandto takethebest ofthetwowitha90%probability.
5.4 Crossover
Thegenesarerealvalues,whihouldhavesuggestedsomekindofbaryentri
rossover(suhasusedin EvolutionStrategies),whereeahgeneofthehildis
an average between thetwogenes of his parents.But sineit is intervals that
mustbeevolved,thistypeofrossoverwouldhaveledtoreduingtheintervals