• Tidak ada hasil yang ditemukan

Epidemics - ResearchOnline@JCU

N/A
N/A
Protected

Academic year: 2025

Membagikan "Epidemics - ResearchOnline@JCU"

Copied!
9
0
0

Teks penuh

(1)

ContentslistsavailableatScienceDirect

Epidemics

jo u r n al h om ep ag e :w w w . e l s e v i e r . c o m / l o c a t e / e p i d e m i c s

Optimally capturing latency dynamics in models of tuberculosis transmission

Romain Ragonnet

a,b,∗

, James M. Trauer

a,c,d

, Nick Scott

b,c

, Michael T. Meehan

e

, Justin T. Denholm

a,d,f

, Emma S. McBryde

a,e

aFacultyofMedicine,DentistryandHealthSciences,UniversityofMelbourne,Australia

bBurnetInstitute,Australia

cSchoolofPopulationHealthandPreventiveMedicine,MonashUniversity,Australia

dVictorianTuberculosisProgram,Melbourne,Australia

eAustralianInstituteofTropicalHealthandMedicine,JamesCookUniversity,Australia

fRoyalMelbourneHospital,Melbourne,Australia

a r t i c l e i n f o

Articlehistory:

Received13February2017

Receivedinrevisedform30May2017 Accepted14June2017

Availableonline16June2017

Keywords:

Tuberculosislatency Mathematicalmodelling Riskofdiseaseactivation Parameterestimation

a b s t r a c t

Althoughdifferentstructuresareusedinmoderntuberculosis(TB)modelstosimulateTBlatency,it remainsunclearwhethertheyareallcapableofreproducingtheparticularactivationdynamicsempiri- callyobserved.Weaimedtodeterminewhichofthesestructuresreplicatethedynamicsofprogression accurately.Wereviewed88TB-modellingarticlesandclassifiedthemaccordingtothelatencystruc- tureemployed.Wethenfittedthesedifferentmodelstotheactivationdynamicsobservedfrom1352 infectedcontactsdiagnosedinVictoria(Australia)andAmsterdam(Netherlands)toobtainparameter estimates.Sixdifferentmodelstructureswereidentified,ofwhichonlythoseincorporatingtwolatency compartmentswerecapableofreproducingtheactivationdynamicsempiricallyobserved.Wefound importantdifferencesinparameterestimatesbyage.Wealsoobservedmarkeddifferencesbetween ourestimatesandtheparametervaluesusedinmanypreviousmodels.Inparticular,whentwosuc- cessivelatencyphasesareconsidered,thefirstperiodshouldhaveadurationthatismuchshorterthan thatusedinpreviousstudies.Inconclusion,structuresincorporatingtwolatencycompartmentsand age-stratificationshouldbeemployedtoaccuratelyreplicatethedynamicsofTBlatency.Weprovidea catalogueofparametervaluesandanapproachtoparameterestimationfromempiricdataforcalibration offutureTB-models.

©2017TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction

Tuberculosis (TB) is a major health issue with10.4 million active cases and 1.8 million deaths worldwide in 2015 (WHO, 2016).Furthermore,aroundonequarteroftheworld’spopulation isestimatedtobeinfectedwithTB(HoubenandDodd,2016),rep- resentingahugereservoirofpotentialdisease.Accordingly,fully understandinglatentTBinfectioniscrucialforassessingthefuture epidemictrajectory and designing effectiveTB control policies.

Despitethis, muchreinfection occursinhighincidencecohorts, hampering accurate estimation of latency dynamics. Therefore insightsintotheactivationdynamicsfollowingasingleinfection episodeofMycobacteriumtuberculosisprovidedbyrecentstudies

Correspondingauthorat:85CommercialRoad,BurnetInstitute,Melbourne,VIC 3003,Australia.

E-mailaddress:[email protected](R.Ragonnet).

inverylowtransmissionsettingsareparticularlyvaluable(Trauer etal.,2016a;Slootetal.,2014).Theseworksprovidedetailedinfor- mationonpatternsofactivation,highlightingthatmostactivecases occurwithinthefirstfewmonthsofinfection.

MathematicalmodellinghasinformedTBcontrolprogramsby simulatinginterventions,orbyexplainingthemechanismsunder- lyingobservedepidemiologicaltrends(VynnyckyandFine,1997;

Gomesetal.,2004;Castillo-ChavezandFeng,1997;Abu-Raddad etal.,2009;Cohenetal.,2008;Dye,2012;Menziesetal.,2012;

Traueretal.,2016b),yetlittleisknownaboutwhethersuchmod- ellinghasbeen abletocapturelatencydynamics accurately.In thepast, TB modelshave beenconstructed tocapturethelife- longprobability ofdisease and,although somemodelsallowed for markeddifferencesbetweenthe earlyand late dynamicsof infection(Dowdyetal.,2013;Linetal.,2011;AparicioandCastillo- Chavez,2009),estimatesfortheassociatedparametershavenot beenfitcloselytolongitudinaldata.Despitethis,ithasbeenshown

http://dx.doi.org/10.1016/j.epidem.2017.06.002

1755-4365/©2017TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.

0/).

(2)

thatwhenmodellinginfectious diseases,itis criticaltoemploy appropriatedistributionsoflatentperiods(Wearingetal.,2005).

Focusingonemerging infectiousdiseases,WallingaandLipsitch furtherdemonstratedthatcapturingthemeanofthegeneration timesisnotsufficienttocharacterisetransmissionaccurately,as theshapeofthedistributionofthegenerationintervalsalsoplays acriticalroleininfectiondynamics(WallingaandLipsitch,2007).

AlthoughTBisanancientdisease,itsepidemiologyiscontinuously evolving.Inparticular,changesinTBepidemiologyinresponseto emergingphenomena,suchasintroductionofdrug-resistantforms ofTBorstrongercontrolprograms,arelikelytoaffecttheshape ofthegenerationtimedistribution.Therefore,therecentdetailed characterisationof TB activationdynamicsrepresent a valuable opportunitytoreview and improvemodelling practices forthe simulationofTBlatency.

Compartmentaldynamictransmissionmodels−themostcom- montypeofTBmathematicalmodel−simulateTBlatencywith various levels of complexity. While some modellers employ a singlelatencycompartmentthatprecedestheactivediseasecom- partment (Colijn et al., 2008; Blower and Chou, 2004), others incorporateasecondlatencycompartmentinordertocapturetwo differentratesofprogressionfromlatentinfectiontoactivedis- ease(Hilletal.,2012;Cohenetal.,2006;Traueretal.,2014).When twolatencycompartmentsareincorporated,theycaneitherbe positionedinseriesorinparallel,involvingdifferentunderlying assumptionsregardingtheprogressionpathwaystoactivedisease.

First,theserialstructureimpliesthatnewlyinfectedindividuals remainathighriskofdiseaseduringtheinitialphaseandthen, ifTBactivationhasnotoccurred,theytransitiontoanothercom- partmentwheretheirriskofdevelopingactiveTBisreduced.By contrast,withaparallelcompartmentalstructure,theunderlying assumptionisthataproportionofinfectedindividualsbelongto ahighriskcategory,whiletheremainderareatlowerriskofTB disease.WhileTBmodellinghasbeenusedextensivelyforover40 years,itremainsunclearwhichofthesestructuresarebestadapted tothenaturalhistoryofTB.

Inthisstudy,weaimtodeterminethemostappropriatemodel structurestosimulateTB latencyandprovideestimates forthe parametersassociatedwiththesestructuresacrossdifferentage categories.Weusethedistributionoftheestimatedtimesfrom infectiontoTBactivationin1352infectedcontactsofindividuals withactivepulmonaryTBfromVictoria(Australia)andAmster- dam(Netherlands)tocalibratethelatencystructuresofdifferent candidatemodelstothedynamicsobservedinthedata.

2. Methods

2.1. Literaturereview

Oursearchwasbasedontheliteraturereviewofmathematical and economic TB modelling articles provided by the TB Mod- ellingandAnalysisConsortium,availableonlineathttp://tb-mac.

org/Resources/Resource/4(seeAppendixinSupplementaryfilefor moredetails).Fromthisdatabaseweidentifiedall88publications reportingtheuseofadeterministiccompartmentaltransmission dynamicmodel.Allselectedpaperswerereviewedindependently bytwoauthors(RR,JMT)whoclassifiedthemanuscriptsaccording tothestructureusedtomodelTBlatency.Thesetwoindependent investigationsledtothesameclassificationwhichispresentedin theAppendixinSupplementaryfile(TableS1).

2.2. Analyticalsolution

Foreachlatencystructurefoundintheliterature,weassoci- atedabasicdynamicmodelcomprisedofthelatencystructurein

combination with compartments representing susceptibility to infectionandactivedisease.We thenfoundanalyticalsolutions for the TB activation dynamics corresponding to each model.

Namely,consideringthat individualswereinfected attimet=0, wedeterminedtheproportionI(t) ofinfectedindividualsthathad developedactiveTBaftereachtimet(t≥0).Analyticalexpressions arealsopresentedforthetotal proportionofinfectedindividu- alsprogressingtoactivedisease,obtainedbycalculatingthelimit ofI(t) astapproachesplusinfinity.Thedetailedmethodusedto obtaintheanalyticsolutionsisdescribedintheAppendixinSup- plementaryfile.

2.3. Datausedtocalibratethemodels

Themodelsdescribedabovewerecalibratedtoindividualdata onclosecontactsofindividualswithactivepulmonaryTBnotified intheAustralianstateofVictoriafromJanuary2005toDecember 2013.Thesedataarederivedfromaverylowendemicsettingand weredescribedindetailbyTraueretal.Theyconsistof613infected contactsofwhom67(10.9%)developedactiveTBduringthestudy period(Traueretal.,2016a).Toenhanceourdataset,wealsoused thepublisheddata onclose contactsof pulmonary TB patients fromAmsterdam(Netherlands)notifiedbetween2002and2011, asreportedbySlootetal.(Slootetal.,2014).Thesedatainclude 739infectedindividuals,ofwhom71(9.6%)developedactiveTB.

Thedetailedapproachesusedtodeterminebothdatesofinfection andactivationinindividualsinthetwostudiesarepresentedin therespectivemanuscripts.Theactivationtimesmeasuredinthese datawereusedtocalibratethedifferentmodels.Inordertovali- dateourapproachinvolvingmergingoftwodatasets,wepresent acomparisonoftheestimatesobtainedfromtheseparatefittings tothetwodatasets(AppendixinSupplementaryfileSection8.2).

TheapproachusedtoextractdatafromSlootandcolleagues’arti- cleisdescribedindetailintheAppendixinSupplementaryfile, alongwithavalidationanalysisoftheextractionmethodwhile thedistributionofthetimestoactivationmeasuredinthetwo datasets(VictoriaandAmsterdam)ispresentedinFig.S7(Appendix inSupplementaryfile).

Traueretal.alsoproposedanimputationmethodwhichtakes intoaccountthecensorshipformigration,death,andpreventive treatment(Traueretal.,2016a).Weusedthisapproach,whichis associatedwithhigherestimatesconcerningtheriskofTBactiva- tion,inasupplementaryanalysis.

2.4. Modelfitting

Modelfittingtodatawasmadebybuildingthesurvivallike- lihood defined asfollows. For a given model associated witha givensetofparameters,,weobtainananalyticalsurvivalfunc- tionS(t) whichrepresentstheprobabilitythatactivationhasnot occurredyetattimetgiventhatinfectionoccurredatt=0.This function is associated with a hazard function (t) definedby (t)=−S(t)/S(t),characterisingthechancethatprogression toactiveTBoccursatpreciselytimet,given survivaluptothat time.Then,foreachinfectedcaseiofourdataset,forwhomtidesig- natesthetimeofeitherTBactivationorendoffollow-up,wedefine anindividuallikelihoodcomponentby ifthecasewas notknowntodevelopactiveTB;and ifthecase effectivelyactivatedTBattimeti.Finally,weaimtomaximisethe multi-dimensionallikelihoodobtainedbymultiplyingalltheindi- viduallikelihoodcomponentstogether: .Thisproblem isequivalenttomaximisingthefollowinglog-likelihoodthatwe defineasthefittingscore: .

Anotherfittingmethodwasusedforvalidationandwhenthe datadidnotallowforthesurvivallikelihoodtobeutilised.Specif- ically,aleastsquaresoptimisationwasperformedtominimisethe

(3)

Fig.1.Representationofthedifferentmodelstructures.Solidlinesrepresentpro- gressivetransitionsbetweencompartmentsthatareparameterisedwithrates, whiledashedlinesrepresentinstantaneoustransitionsbetweencompartmentsthat areassociatedwithproportions.Theflowsrelatedtonaturaldeatharenotrepre- sentedandareassociatedwiththenaturalmortality.

distancebetweenthesurvivalcurvesgeneratedbythedataandby themodel.Thetime-pointsconsideredwhenperformingthisopti- misationwereequallyspacedbyonedayandrangedbetween0 andthemaximaltimetimeasuredinthedata.

Thetwomethodsintroducedaboveareconceptuallydifferent.

Ineffect,thesurvivallikelihoodapproachsearchesforthesetof modelparametervaluesthatmaximisestheprobabilityofobserv- ingthedata,givenaparticularmodelandaparticularparameter set.Incontrast,theleastsquaresoptimisationisbasedonameasure ofadistancebetweentwocurves:theactivationcurveobserved fromthedataandtheoneproducedbyamodel.Althoughthese twoapproachesarefundamentallydistinct,theyarebothformsof optimisationappliedtothemodel’sparametervalues.Thisoptimi- sationisbasedonaniterativemethodwhichallowstheparameters tovaryinamulti-dimensionalspace,fromwhichweretainonlythe parametersetthatyieldstheoptimalmeasure,i.e.thegreatestlike- lihood(fortheprobabilisticapproach)orthesmallestdistance(for theleastsquaresmethod).ThisprocesswasobtainedusingRv.3.3 anditsincorporatedfunction“constrOptim”.

Modelfittingwasperformedseparatelyusingthreeagecate- gories(“<5years old”,“5to14 years old” and“≥15years old”) inadditiontoapooledanalysisincludingthewholepopulation.

Inordertofullyexploretheparameterspacesand toreportall acceptableparametersets, weusedaMetropolis-Hastingsalgo- rithmpresentedindetailintheAppendixinSupplementaryfile.

3. Results

3.1. Literaturereview

Fromourreviewof 88publications related toTBmodelling, wefoundthat sixdifferentcompartmentalstructureshadbeen employedtomodelTBlatencyandtheseformthebasisofthefol- lowinganalysis.Thesixmodelsarerepresentedin Fig.1,along withthelabelsusedfor theassociated parameters.Thelevelof complexityrangesfroma modelincorporatingnolatencycom- partment(Model1)tomodelsbasedontwolatencycompartments positionedeitherinseries(Models4and5)orinparallel(Model 6). Some structures incorporate a bypass from the susceptible compartmenttotheactivediseasecompartment,allowingforcon- siderationofinstantaneousactivationofTBdiseaseafterexposure

(Models 1,3and 5).TableS1(AppendixinSupplementaryfile) reportsourclassificationofthe88reviewedstudiesaccordingto thelatencystructureemployed.

3.2. Modelcalibration

Fig.2presentsthedynamicsofTBactivationobtainedfromeach modelwhenoptimallycalibratedtoourdata(seeMethodspara- graph for fittingapproach). Onlythemodels incorporatingtwo latencycompartments(Models4,5and6)suitablyreplicatethe dynamicsofTBactivation.Model1,whichisnotshown,doesnot containanyfreeparametersandassumesthatallinfectedpatients progresstoactivediseaseimmediatelyafterexposure.Thisisnot compatiblewithourdatanorourknowledgeofthelifelongriskof TBdiseaseininfectedindividuals(Traueretal.,2016a;Slootetal., 2014).BothModels2and3produceunreasonablypoorfitstothe data.IntheAppendixinSupplementaryfileweshowthattheequa- tionsdescribingModels2and3onlyinvolveasingleexponential function,whichisnotsufficienttoreplicatethetwodistinctpat- ternsobservedinthedynamicsofactivation—ahighriskofdisease activationoverthefirstfewmonths,followedbyadramatically lowerrisk inasecondphase. Incontrast,in Models4, 5and6, whichincludetwolatencycompartments,theactivationdynam- icsaredrivenbytwoexponentialcomponentsthatareassociated withtwoindependentgrowthrates,leadingtoaccuratereplication ofthepatternsempiricallyobservedforeachagecategory.Thefit- tingscores(seeFig.2)forModels4,5and6indicatethatnone providesasignificantlybetterfitthantheothers;however,Model 5presentsahigherlevelofcomplexitythanModels4and6,as itinvolvesestimatingoneadditionalparameter.Sincetheinclu- sionofthisparameterdidnotimprovemodelfitting,weconsider thatitisunnecessarytoemploythisstructure.Thispropositionis stronglysupportedbyamoredetailedanalysisofModel5where differentapproachesallsupportthattheadditionalparameterdoes notimprovefitting(seeAppendixinSupplementaryfile).Accord- ingly,theremainderofouranalysisrelatesonlytoModels4and 6.

Table1presentstheparametervalues(withunitsofdays−1) obtainedfromthecalibrationofModels4and6tothedata.We observethattherateofre-activation()ismuchlowerintheage category“<5yearsold”thanintheotheragecategoriesforboth Model4andModel6.Incontrast,therateofrapidprogressionis higherinthe“<5yearsold”categorythanintheotheragecategories whenconsideringModel4;andwhen consideringModel 6,the proportionofinfected individualsexperiencingahighriskofTB disease(1−g)ishigherforchildrenthanfortheothercategories (35%for“<5y.o.”,19%for“5to14y.o.”,5%for“15y.o.andmore”, and9%forallagestogether).

ThelifelongriskofTBactivationininfectedindividualscanbe calculatedanalyticallyforthedifferentmodels(seeAppendixin Supplementaryfile,p.4).TheparametervaluesreportedinTable1 forModels4and6correspondtototalproportionsofactivation rangingbetween10%(“15y.o.andmore”)and32%(“<5y.o.”).

Given theverylow valuesobserved fortherateof reactiva- tion()inthe“<5yearsold”category,weundertookanadditional analysisinordertoexplorethepossibilityoftheabsenceoflate reactivationinthiscategory.Inthisanalysis,Models4and6were fittedtothedataundertheconstraintthat=0,withbothmod- elsfoundtoperformequallywellinabsenceofreactivation.Fig.3 presentsthecorrespondingsimplifiedmodelsaswellasthebestfits obtained.Afurtherexplorationofthecontributionofendogenous reactivationversusprimaryactivationispresentedintheAppendix inSupplementaryfile(Section3.)forModel4andforthedifferent agecategories.Thisanalysisdemonstratesthatendogenousreacti- vationcontributesverylittletotheburdenofactiveTB,especiallyin youngindividuals(“<5yearsold”and“5to14y.o.”).Althoughthis

(4)

Fig.2. CalibrationsobtainedwiththedifferentmodelsforthepercentageofactiveTBamonginfectedindividualsovertimesinceinfection.Theblacklinesandgreyshade representtheestimates(centraland95%CI)obtainedfromtheKaplan-Meieranalysisofourdata.ThebluelinerepresentsthepercentageofactiveTBovertimeobtained fromthedifferentmodels,whenoptimallycalibratedwiththesurvivallikelihoodmethod.Thex-scaleswerechoseninordertoallowforadecentvisualisationoftheearly stagesofinfectionanddonotcovertheentiretimewindowscorrespondingtothedataset.Fittingwasrealisedbymaximisingthefittingscore(FS).(Forinterpretationofthe referencestocolourinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.)

contributionismoresubstantialinthe“≥15yearsold”category,we estimatethatonly1%ofinfectedindividualsofthiscategorywould haveprogressedtoactivediseasetransitioningfromcompartment LBafterfiveorlessyearsfrominfection(usingEq.(23)inAppendix inSupplementaryfile).

OuranalysisoftheanalyticalsolutionsassociatedwithModel 4andModel 6demonstratedthatthetwo expressionscoincide iftheparametervalues ofthetwomodelssatisfythefollowing relationships:

ε64+46=4g6= 4 444

wherethesubscriptsindicatetowhichofModels4or6the parametersapply.Thisfindingindicatesthatthedynamicsofacti- vationsimulatedbythetwomodelsareidentical,differingonlyin thevalueoftheparametervaluesthatshouldbeapplied.

3.3. Probabilitydistributionofparameters

We useda Bayesianframework toinfertheposterior distri- butions for the parameters in themodel. Uniform priors were usedand toolsof Bayesianinference appliedincludingMarkov Chain Monte-Carlo exploration using the Metropolis-Hastings acceptancealgorithm.Theresultingposteriordistributionsforthe

(5)

Table1

Parameterestimates.Calibrationissuedfromthesurvivallikelihoodmaximisationappliedtothemergeddataset(VictoriaandAmsterdamdata).Pointestimatescorrespond totheparametersmaximisingthelikelihoodwhilevaluesintobracketsindicatethenarrowestintervalcontaining95%oftheacceptedvaluesduringtheMetropolis-Hastings simulation.Ratesarepresentedasdailyvalues.

ε g

Model4

Allages 1.1e−3 (8.4e−4–1.5e−3)

1.0e−2 (8.5e−31.4e−2)

5.5e−6 (2.5e−61.1e−5)

<5y.o. 6.6e−3 (4.4e−39.5e−3)

1.2e−2 (8.5e−31.8e−2)

1.9e−11 (5.0e−91.6e−5)

5–14y.o 2.7e−3

(1.7e−33.9e−3)

1.2e−2 (7.9e−31.6e−2)

6.4e−6 (6.7e−71.9e−5)

≥15y.o 2.7e−4 (1.6e−45.1e−4)

5.4e−3 (3.5e−31.1e−2)

3.3e−6 (1.9e−61.0e−5)

Model6

Allages 1.1e−2 (9.2e−31.5e−2)

5.5e−6

(3.4e−61.0e−5)

0.91 (0.890.93)

<5y.o. 1.9e−2 (1.2e−22.5e−2)

3.4e−11

(2.7e−92.0e−5)

0.65 (0.550.73) 5–14y.o 1.4e−2

(9.4e−31.8e−2)

6.4e−6

(8.0e−72.2e−5)

0.81 (0.750.86)

≥15y.o 5.6e−3 (3.8e−39.6e−3)

3.3e−6

(7.3e−79.3e−6)

0.95 (0.940.97)

Parameterestimates.Calibrationissuedfromthesurvivallikelihoodmaximisationappliedtothemergeddataset(VictoriaandAmsterdamdata).Pointestimatescorrespond totheparametersmaximisingthelikelihoodwhilevaluesintobracketsindicatethenarrowestintervalcontaining95%oftheacceptedvaluesduringtheMetropolis-Hastings simulation.Ratesarepresentedasdailyvalues.

Fig.3. SimplifiedmodelstructuresadaptedtosimulateTBlatencyinyoungchildren (<5yearsold)Here,itisassumedthatprogressionfromthepreviouscompartments LBofModel4andModel6toactivediseaseIcannotoccur.ThecompartmentLB

thereforebecomesaprotectedstatewhichislabelledRinthisillustration.Fitting wasrealisedbymaximisingthefittingscore(FS).

differentparametersofModels4and6arepresented(Table1).

Proposedstatisticaldistributionsthatcouldbeusedtogenerate similarsetsofparametersareavailableintheAppendixinSupple- mentaryfile,alongwiththeposteriordistributionsobtainedforall parameters.Theacceptablerangesofvaluesfortheparameterare wideforthecategory“<5y.o.”,duetothesmallnumberofreacti- vationcasesinourdataset.However,ouranalysisshowedthatthe rateofreactivationislimitedbyanupperboundofaround2.3e5 inallcategories,whichrepresentsarelativelysmallannualriskof re-activationof0.8%.

Byanalysingthedistributionsofeachoftheretainedparameter setsinpairs,weobservedacollinearitybetweentheparameters andεforModel4.Thiscorrelation(representedinFig.4)sug- geststhatwhenindividualsareassumedtostabiliseinfectionmore rapidly(increases),therateofrapidprogressiontoTBdisease(ε)

tendstoincreasetocompensate.Thisaffinerelationshipbetween parameters␬and␧isalsodescribedthroughaformalmathematical analysis(AppendixinSupplementaryfile,Section10.1.2.).

Anotherresult provided bythe Metropolis-Hastings simula- tionwasthedistributionoftheaveragedurationspentinthefirst latencycompartmentforModel4.Thisquantitywasobtainedvia theformula+ε1+whererepresentsthenaturaldeathrate.Fig.5 presentsthedistributionofthesedurationsforthedifferentage categories.Theaveragedurationspentinthefirstlatencycompart- mentisestimatedat82daysforallagecategoriescombined,this durationbeingshorterforchildren(52daysfor“<5y.o.”category) thanforolderindividuals(70daysfor“5to14y.o.”and146days for“≥15y.o.”).

3.4. Validationandsensitivityanalyses

Ourfindingswereconsistentwhenweemployedanalternate fittingmethod(least squaresminimisation),performedseparate fitsforeachdataset(VictoriaandAmsterdam)insteadofthesingle mergeddataset,orusedalternateextractionmethodsforthedataof Slootetal.WhenfittingModel3usingleastsquaresminimisation, weobtainedamodelcalibrationthatdifferedfromthatobtained withthesurvivallikelihoodmethod(SeeAppendixinSupplemen- taryfile,Fig.S14).However,whilethissecondcalibrationallowed forabettersimulationofthelatestagesoflatencycomparedtothe baselinefittingmethod(Fig.2),itcompletelyfailedtocapturethe earlydynamicsofactivation.Moredetailsaboutthedifferentsen- sitivityanalysesareavailableintheAppendixinSupplementary file.

Finally, by using imputed dataas described in Traueret al.

accountingforthecensorshipformigration,death,andpreventive treatmentandthereforeassociatedwithhigherestimatesforthe riskofTBdisease(Traueretal.,2016a),ourconclusionsregarding optimalmodelselectionremainedunchanged,whileweobtained slightlydifferentoptimisedparametervalues(seeAppendixinSup- plementaryfile).Inparticular,inModel4,imputationledtoaslight increaseintherateofrapidprogression(ε)combinedwithaslight reductionintherateoftransitiontowardsthelatelatencycom- partment().ConcerningModel6,thebestcalibrationtoimputed datawasobtainedwithahigherproportionofinfectedindividuals transitioningtothehighriskcompartment(1−g)whiletherateof rapidprogression(ε)wasslightlyreduced.

(6)

Fig.4. RepresentationofthecollinearityobservedbetweentheparametersandεforModel4.Thereddotsrepresentthe10,000acceptedparametersetsobtainedfrom theMetropolis-Hastingssimulation.Theblacklinerepresentstheaffinemodelapproximatingthedatawithaleastsquareminimisation.(Forinterpretationofthereferences tocolourinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.)

Fig.5. DistributionofthetimesspentinthefirstlatencycompartmentLAformodel4.Distributionassociatedwith10,000acceptedparameterssetsfromtheMetropolis- Hastingssimulation.

3.5. Comparisonofourresultswithpreviousworks

Wereviewedtheparametervaluesthathadbeenemployedin thepreviousstudiesincorporatingthelatencystructuresofModel 4andModel6andcomparedthesevaluestoourestimates(see AppendixinSupplementaryfile,Section9).ConcerningModel4, wefoundthat therateofprogression fromcompartment LAto compartmentLB()wasconsiderablylowerthanourestimatein allpreviousstudies,whichtypicallyassumelongperiodsspentin earlylatency(2–5years).Similarly,therateoffastprogressionto activeTB(ε)usedintheexistingliteraturewasmuchlowerthanour estimate.ForModel6,theproportionofslowprogressors(g)found inpreviousworkswasclosetoourestimate,whereastherateoffast progression(ε)hadbeenmarkedlyunderestimated.Finally,while therateofslowprogressiontoactiveTB()wasgenerallyunder- estimatedinstudiesincorporatingeitherstructure(Models4or6), twostudiesusedpointestimatesthatfallinsideofthe95%CIwe report,andnineotherstudiesusedparameterrangesoverlapping our95%CI.

3.6. Rapidestimationoftheparametersforfutureworks

Atheoreticalanalysisoftheequationsgoverningthedynamic modelsallowedustodevelopamethodtoestimaterapidlythe differentparameters of Models4 and 6 from anydataset.This approachinvolvessimplegraphicalmeasurementsperformedon the curve representing the proportion of active cases among infectedindividuals over timefrominfection (denoted),such asthecurvesrepresentedinFig.2.Thedetaileddescriptionand demonstrationofthemethodisavailableintheAppendixinSup- plementaryfileandthemainresultsarepresentedhere.Theprofile ofcanbedecomposedintotwophases(seeFig.2anddetailed

descriptioninAppendixinSupplementaryfile).Themethodcon- sistsoffirstmeasuringtheinitialslopeof(denoteds0)aswell asthecharacteristicsofatangenttoonapointsituatedatthe beginningofthesecondrisephase(slopedenoteds,andy-intercept denotedy0).Then,estimatesforthedifferentparameters(ˆε,,ˆ ˆ and ˆg)canbeobtainedbyusingthethreemeasures(s0,sandy0) asfollows:

ForModel4: ˆε=s0ˆ= εˆ

y0 −εˆ−ˆ=s

ˆ

+εˆ+

ˆ

ForModel6: ˆε= s0

y0 −ˆv= s0s

s0−y0(s0+)gˆ= εˆ−s0 ˆ ε−ˆv wheredesignatesthenaturalmortalityrate.

4. Discussion

4.1. Maincontributionsofthestudy

Inthisstudy,wedeterminethemostappropriatemodelstruc- turesforaccuratelysimulatingTBlatency.Wedemonstratethatof thestructuresemployedinthepast,onlythosethatincorporatetwo compartmentsforlatentinfectionareabletoreproducethespecific dynamicsofTBactivation.Suchapproachesthereforeinvolvetwo differentlevelsfortherateofactivation,allowingformoreTBcases tooccurafterrecentinfectionthanthroughlatereactivation.This workalsoprovidesadetailedandflexible catalogueofparame- tervaluesassociatedwiththeretainedmodelstructures,thatwas validatedbytheuseoftwoindependentfittingmethodsleadingto similarestimatesandthathighlightedmarkedgapswithparameter valuesemployedinpreviousworks.Thisstudymaybecomearefer-

(7)

enceforcalibrationoffutureTBmodelsandourapproachinvolving estimationswithandwithoutagestratificationwillallowourfind- ingstobedirectlyappliedwhethermodelsareage-structuredor not.

4.2. Serialversusparallelstructure

Wedemonstratethatthetwocompartmentsrequiredtomodel TBlatencycouldbeplacedeitherinseriesorinparallel,andwould leadtoidenticalactivationdynamics.Thedifficultyinmakingthis distinctionbetweenthetwomodelsisunfortunatebecausethey reflecttwoalternatebiologicalmechanismsthatcouldexplainthe higherburdenofdiseaseobservedafterrecentinfection,eachlead- ing to differentrecommendations for TB prevention strategies.

First,theserialstructuremodelsadecreasingriskofactivationover time ineveryindividual,indicating thatpreventivecareshould betargeted at themostrecent infections as a prioritythrough interventionssuchascontacttracing.Ontheotherhand,thepar- allelstructuresuggeststhatindividualpredispositionsmaymake someinfectedindividualsmoreorlesslikelytodevelopdisease, regardlessofthetimefrominfection.Ifthissecondscenariowas verifiedwithepidemiological data, theseresultswould suggest that identifyingand findingpriority populationscould dramat- ically enhanceefficiency ofinterventions. Somefactorssuchas HIV-infection,diabetesorsmokingarealreadyknowntoenhance theriskofTBactivation(Bishwakarmaetal.,2015;Houbenetal., 2010;Horsburghetal.,2010;JeonandMurray,2008),althoughour parameterestimatesundertheparallelstructurescenariosuggest thataround9%ofinfectedindividualswouldbelongtoahigh-risk groupinwhichtheriskofTBactivationwouldbeashighas2000- foldthatofthelow-riskgroup.Alternatively,fortheseriesstructure ofthelatencycompartments,rapididentificationandearlytreat- mentofinfectedpeoplewouldbethepriority.

4.3. Agedependency

Ourestimationshighlightimportantdiscrepanciesbetweenage categories,withparameterestimatesforyoungchildren(<5years old)differingmarkedlyfromtheotheragecategories.Inparticular, ourstudysuggeststhatforyoungchildren,therateofdiseaseacti- vationinthehigh-riskpopulation(earlylatencyforserialstructure orhigh-riskgroupforparallelstructure)ismuchgreaterthanfor 15+yearolds.Incontrast,amongthelowerriskpopulation(late latencyforserialstructureorlowriskgroupforparallelstructure), youngchildrenseemtobeatlowerriskofactivationthantheother individuals.Thesefindingsindicatethattheyoungestinfectedpop- ulationpresentsamuch higher riskof TBactivationearlyafter infection,whiletheirriskofactivationreducesdramaticallyinthe laterstagesofinfection.Previousworkshavealreadyidentifiedage asanimportantfactorinfluencingthenaturalhistoryofTB(Trauer etal.,2016a;Slootetal.,2014;VynnyckyandFine,1997),andour analysisbringsadditionalinsightthatreinforcestheimportance ofprovidingyoungchildrenwithearlypreventivetreatmentfor infectionaslatetreatmentswouldbeuseless.Byrevealingsuch age-specificcharacteristics,ourstudyalsoimpliesthatfutureTB modellingworks shouldincorporateage-stratifiedstructures in ordertoreplicateTBactivationdynamicsaccurately.Ouranalysis alsodemonstratesthatsimplifiedmodelstructuresincorporating noreactivationareadaptedtosimulateTBlatencyinthe“<5years old”category,suggestingthatasignificantproportionofinfected childrenwillneverreactivate.Dependingonthemodelstructure employed,theseyoungindividualscouldeitherbecomeimmune aftersomedurationofinfection(serialstructure)orbeprotected immediatelyafterinfection(parallelstructure).

4.4. Comparisonwithpreviousworks

Whentwolatencycompartmentsarepositionedinseries,the rateof transitionfromearlylatencytolate latencyreflects the periodfor whichindividuals remainathighriskofdisease.We estimatethattheaverageearlylatencyperiodis52,70daysand 146days,forchild(<5),adolescents(5–14)andadults,respectively.

Mostpreviousworksemployingtheserialstructurehaveused5 yearstodefineearlylatency,basedonapreviousconventionto defineprimarydiseaseversusendogenousdisease(Vynnyckyand Fine,1997;Holm,1969),andthereforemayhavegreatlyoveres- timated theactualduration ofhigh risk.Our review of articles showedthatthemodelswiththeselongdurationsathighriskwere alsoassociatedwithlowerratesoffastprogressiontoactivedis- easethantheratesestimatedfromourstudywhichcompensates forthelongearlylatencyandleadstosimilaroverallrisksofacti- vationoveralifetime.However,wehaveshownthatourapproach, whichhasdramaticallyshorterdurationsforthetimespentinearly latencyandathighrisk,isrequiredinordertoreproducetheprofile ofactivationdynamicsaccurately.

Ourestimatefortherateoflatereactivationcorrespondstoa riskof0.20casesper100person-yearsforallages,butismuch loweramongthe‘<5years old’category.However,althoughour analysisclearlyhighlightsdifferentpatternsofreactivationbyage, accurateestimationofthereactivationparameterforthe‘<5years old’ category was limited by thesmall number of reactivation casesobservedinthissub-group,whichexplainsthewidevari- ationintheassociatedconfidenceintervals.Whilepreviousworks reportedsomewhatlowerestimatesforthelatereactivationrate forallages,rangingfrom0.04to0.16casesper100person-years (VynnyckyandFine,1997;Horsburghetal.,2010;Comstocketal., 1967;Ferebee,1970),ourfindingssuggestthattheseresultscan onlybeinterpretedinthecontextofagemixinthestudies,and wecanspeculatethatsomeofthisvariabilitymaybeaccounted forbydifferentproportionsof“<5yearolds”inthestudypopula- tions.Detaileddatabyagearenotavailablefrompreviousworks soformalanalysisbyagehasnotbeenpossible.Anotherpossible explanationforthedifferencebetweenourestimateandprevious valuesforthereactivationrateisthatthedefinitionofinfectionin ourdatasetsmaybemorespecificthanwhatwasusedinprevious studies.Inparticular,looserdefinitionsforinfectionmayinclude falsepositiveswhichwouldtendtoincreasethenumberofpersons

“atrisk”andthereforereducetheinferredreactivationrate.

WhileourstudyshowsthatpreviousTBmodelshavenotalways incorporatedtheoptimalstructureorparameterisationtoaccount forthespecificpatternsobservedinTBactivationdynamics,the potentialconsequencesofusingsuboptimalapproachesremains undetermined.However,ananalysisbyDowdyandcolleaguesof themostinfluentialparameterstoTBdynamicsdemonstratedthat theparametersdescribingearlylatencyareamongthosewiththe greatestimpactonmodelpredictionsofsteady-stateTBincidence (Dowdyetal.,2013).Thissuggeststhatitiscriticaltoreproduce earlydynamicsofTB infectioncloselyinordertoprovideaccu- rateinsightintotheepidemicstrajectory.Therefore,futureworks investigatingtheconsequencesofemployinginappropriatemodel structuresorparameterisationareneeded.

4.5. Potentialforotherfutureworks

OuranalysiswasbasedondatafromverylowTBendemicset- tingswherere-infectionisexpectednottoplayanimportantrole (Wangetal.,2007).Thisallowedourestimationtofocusonthe potentialprogressiontoactivediseasefollowingasingleinfection eventandtoavoidtheconfusionthatwouldemergefromrepeated infectionswhenestimatingthetimefrominfectiontoactivation.In theeventthatsimilardatasetsbecameavailableinhigherendemic

(8)

settings, a follow-up study could be conducted by integrating our“re-infectionfree”parameterisationintoamodelthatwould includeanadditionalpathwayforre-infectionduringlatency.By fittingsuchamodeltothenewdata,there-infectionparameter couldthenbeestimated,thusprovidingvaluableinsightsintothe relativecontributionofre-infectioncomparedtotheriskoffirst infection.

Itisimportanttorememberthatthecalibrationspresentedin thisstudyareassociatedwiththespecificmodelsthatweselected.

Accordingly,suchestimatescouldnotdirectlybeusedinmodels thatincorporateadifferentstructureorthatdonotbelongtothe category ofdeterministic compartmentaltransmission dynamic models.However,providedthatboththestructureandthenature ofthemodelcorrespondtothoseusedinthisstudy,ourworkcould alsobeusedtore-estimateparametersassociatedwithmanydiffer- entsettings.Tothisend,weprovideastep-by-stepmethodwhich allowsforrapidestimationofthedifferentmodelparametersby performingsimplemeasuresonthereactivationfailurecurve.Con- sequently,ourstudycouldalsobeusedtoinformmodelsinsettings thatpresentdifferentcharacteristics,suchashighHIV-endemic settingswheretheactivationdynamicsareknowntobedifferent (Houbenetal.,2010;Horsburghetal.,2010).

One natural limitation of this work is that it is linked to theepidemiological challengesoftiminginfectionandreactiva- tionaccurately.However,theestimateddurationsthatweusein this studyare derived fromtwo recently publishedworks that employedrigorousdefinitionsforbothdatesofinfectionandacti- vation(Traueretal.,2016a;Slootetal.,2014).

5. Conclusions

Onlymodelsemployingtwolatencycompartmentsareableto reproduceTBlatencydynamicsaccurately.We provideparame- tervaluestooptimallysimulateepidemiologicalobservationsin suchmodelsalongwithanapproachtoobtainingsuchvaluesfrom future epidemiological studies. Our analysis alsoreinforces the importanceofage-stratificationforcapturingthedramaticdiffer- encesbetweenagegroupsinpatternsofreactivation,whichimply fundamentalbiologicaldifferencesbetweenagegroups.However, wealsodemonstratethatdataofthetypethisanalysisisbased uponcannotbeusedtodeterminetheidealconfigurationforthe twolatencycompartments.

Competinginterests

Wehavenocompetinginterests.

Author’scontributions

RR,JMTandESMdesignedthestudy.RRperformedtheanalysis.

RR,JMT,NS,MTM,JTDandESMinterpretedtheresults.RR,JMT,NS, MTM,JTDandESMwrotethemanuscript.

Funding

RRwassupportedbyanAustralianGovernmentResearchTrain- ingProgramScholarship.

Acknowledgments

WethankthenursesandotherstaffoftheVictorianTBProgram forcollectingthedataInparticular,wewouldliketothankNompilo MoyoandEe-LaineTaywhocompiledthedataset.

AppendixA. Supplementarydata

Supplementarydataassociatedwiththisarticlecanbefound,in theonlineversion,athttp://dx.doi.org/10.1016/j.epidem.2017.06.

002.

References

Abu-Raddad,L.J.,Sabatelli,L.,Achterberg,J.T.,Sugimoto,J.D.,LonginiJr.,I.M.,Dye, C.,Halloran,M.E.,2009.Epidemiologicalbenefitsofmore-effective tuberculosisvaccines,drugs,anddiagnostics.Proc.Natl.Acad.Sci.U.S.A.106 (33),13980–13985.

Aparicio,J.P.,Castillo-Chavez,C.,2009.Mathematicalmodellingoftuberculosis epidemics.Math.Biosci.Eng.6,209–237.

Bishwakarma,R.,Kinney,W.H.,Honda,J.R.,Mya,J.,Strand,M.J.,Gangavelli,A.,Bai, X.,Ordway,D.J.,Iseman,M.D.,Chan,E.D.,2015.Epidemiologiclinkbetween tuberculosisandcigarette/biomasssmokeexposure:limitationsdespitethe vastliterature.Respirology20(4),556–568.

Blower,S.M.,Chou,T.,2004.Modelingtheemergenceofthe‘hotzones’:

tuberculosisandtheamplificationdynamicsofdrugresistance.Nat.Med.10, 1111–1116.

Castillo-Chavez,C.,Feng,Z.,1997.Totreatornottotreat:thecaseoftuberculosis.

J.Math.Biol.35(6),629–656.

Cohen,T.,Lipsitch,M.,Walensky,R.P.,Murray,M.,2006.Beneficialandperverse effectsofisoniazidpreventivetherapyforlatenttuberculosisinfectionin HIV-tuberculosiscoinfectedpopulations.Proc.Natl.Acad.Sci.U.S.A.103, 7042–7047.

Cohen,T.,Colijn,C.,Finklea,B.,Wright,A.,Zignol,M.,Pym,A.,Murray,M.,2008.

Aresurvey-basedestimatesoftheburdenofdrugresistantTBtoolow?Insight fromasimulationstudy.PLoSOne3(6),e2363.

Colijn,C.,Cohen,T.,Murray,M.,2008.Latentcoinfectionandthemaintenanceof straindiversity.Bull.Math.Biol.71,247–263.

Comstock,G.W.,Ferebee,S.H.,Hammes,L.M.,1967.Acontrolledtrialof community-wideisoniazidprophylaxisinAlaska.Am.Rev.Respir.Dis.95(6), 935–943.

Dowdy,D.W.,Dye,C.,Cohen,T.,2013.Dataneedsforevidence-baseddecisions:a tuberculosismodeler’s‘wishlist’.Int.J.Tuberc.LungDis.17(7),866–877.

Dye,C.,2012.Thepotentialimpactofnewdiagnostictestsontuberculosis epidemics.IndianJ.Med.Res.135,737–744.

Ferebee,S.H.,1970.Controlledchemoprophylaxistrialsintuberculosis:ageneral review.Bibl.Tuberc.26,28–106.

Gomes,M.G.,Franco,A.O.,Gomes,M.C.,Medley,G.F.,2004.Thereinfection thresholdpromotesvariabilityintuberculosisepidemiologyandvaccine efficacy.Proc.Biol.Sci.271,617–623.

Hill,A.N.,Becerra,J.,Castro,K.G.,2012.ModellingtuberculosistrendsintheUSA.

Epidemiol.Infect.140,1862–1872.

Holm,J.,1969.DevelopmentfromTuberculousInfectiontoTuberculousDisease.

KNVC,TheHague,TheNetherlands.

HorsburghJr.,C.R.,O’Donnell,M.,Chamblee,S.,Moreland,J.L.,Johnson,J.,Marsh, B.J.,Narita,M.,Johnson,L.S.,vonReyn,C.F.,2010.Revisitingratesof reactivationtuberculosis:apopulation-basedapproach.Am.J.Respir.Crit.

CareMed.182(3),420–425.

Houben,R.M.,Dodd,P.J.,2016.Theglobalburdenoflatenttuberculosisinfection:a Re-estimationusingmathematicalmodelling.PLoSMed.13(10),e1002152.

Houben,R.M.,Glynn,J.R.,Mallard,K.,Sichali,L.,Malema,S.,Fine,P.E.,French,N., Crampin,A.C.,2010.Humanimmunodeficiencyvirusincreasestheriskof tuberculosisduetorecentre-infectioninindividualswithlatentinfection.Int.

J.Tuberc.LungDis.14(7),909–915.

Jeon,C.Y.,Murray,M.B.,2008.Diabetesmellitusincreasestheriskofactive tuberculosis:asystematicreviewof13observationalstudies.PLoSMed.5(7), e152.

Lin,H.H.,Langley,I.,Mwenda,R.,Doulla,B.,Egwaga,S.,Millington,K.A.,Mann, G.H.,Murray,M.,Squire,S.B.,Cohen,T.,2011.Amodellingframeworkto supporttheselectionandimplementationofnewtuberculosisdiagnostic tools.Int.J.Tuberc.LungDis.15,996–1004.

Menzies,N.A.,Cohen,T.,Lin,H.H.,Murray,M.,Salomon,J.A.,2012.Population healthimpactandcost-effectivenessoftuberculosisdiagnosiswithXpert MTB/RIF:adynamicsimulationandeconomicevaluation.PLoSMed.9, e1001347.

Sloot,R.,SchimvanderLoeff,M.F.,Kouw,P.M.,Borgdorff,M.W.,2014.Riskof tuberculosisafterrecentexposure:a10-yearfollow-upstudyofcontactsin Amsterdam.Am.J.Respir.Crit.CareMed.190(9),1044–1052.

Trauer,J.M.,Denholm,J.T.,McBryde,E.S.,2014.Constructionofamathematical modelfortuberculosistransmissioninhighlyendemicregionsofthe Asia-Pacific.J.Theor.Biol.358,74–84.

Trauer,J.M.,Moyo,N.,Tay,E.L.,Dale,K.,Ragonnet,R.,McBryde,E.S.,Denholm,J.T., 2016a.Riskofactivetuberculosisinthefiveyearsfollowinginfection.15%?

Chest149(2),516–525.

Trauer,J.M.,Denholm,J.T.,Waseem,S.,Ragonnet,R.,McBryde,E.S.,2016b.Scenario analysisforprogrammatictuberculosiscontrolinwesternprovince,Papua NewGuinea.Am.J.Epidemiol.183(12),1138–1148.

Vynnycky,E.,Fine,P.E.,1997.Thenaturalhistoryoftuberculosis:theimplications ofage-dependentrisksofdiseaseandtheroleofreinfection.Epidemiol.Infect.

119(2),183–201.

(9)

WHO,2016.GlobalTuberculosisReport.WorldHealthOrganization.

Wallinga,J.,Lipsitch,M.,2007.Howgenerationintervalsshapetherelationship betweengrowthratesandreproductivenumbers.Proc.Biol.Sci.274(1609), 599–604.

Wang,J.Y.,Lee,L.N.,Lai,H.C.,Hsu,H.L.,Liaw,Y.S.,Hsueh,P.R.,Yang,P.C.,2007.

Predictionofthetuberculosisreinfectionproportionfromthelocalincidence.

J.Infect.Dis.196(2),281–288.

Wearing,H.J.,Rohani,P.,Keeling,M.J.,2005.Appropriatemodelsforthe managementofinfectiousdiseases.PLoSMed.2(7),e174.

Referensi

Dokumen terkait

CHAPTER 1 INTRODUCTION 1.1 Epidemiology of HIV infection In 2005, UNAIDS/WHO reported that in 2005 approximately 40.3 million people worldwide were infected with HIV/AIDS and 3

tuberculosis,[4] and new infections occur in about 1% of the population each year.[5] In 2007, an estimated 13.7 million chronic cases were active globally,[6] while in 2013, an

doi:10.15171/jcs.2019.005 Introduction About one fourth of all cancers in women worldwide are breast cancer BC.1 GLOBOCAN reported, an estimated 1.7 million cases of BC were found in

Introduction Substance use is a major public health problem globally.1Available evi- dence suggests that the number of people who used drugs worldwide in- creased by 30% between 2009