Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
3
Pleasemindthegap
4
5
Disclaimer
ThismanuscriptisfullywritteninEnglish.
SUMMARYINFRENCH
Cemanuscritétant rédigéenanglais,un résuméen françaisestplacéendébutde
chaquepartie,encadré,souscettemiseenforme.
REFERENCE
Thereferencetopapersandcommunicationsareincludedintogreenboxes.
6
7
Acknowledgements
AMonsieurleProfesseurPierre‐AntoineGourraud,membredujury
Je vous suis extrêmement reconnaissante d’avoir accepté de siéger à ce jury.Trouvez ici le témoignagedemon respect etdemaprofonde considération.Votreparcoursestinspirant.Adeprochainescollaborations,jel’espère.
AMadameleProfesseurJoëlleMicallef,membredujury
Je suis très touchéepar l’enthousiasmedont vous avez fait preuve à l’égardde cetravail et très honorée que vous ayez accepté d’assister à la soutenance de cettethèse. J’ai sincèrementappréciévotreprésencestimulanteetvotreavisexpertsurles projets lors de nos diverses séances de travail ces dernières années. Lesoccasionsdevousvoirsonttoujourstroprares.
AMonsieurleProfesseurJean‐LouisMontastruc,membredujury
Vousmefaitesungrandhonneurenayantacceptédesiégerdansce jury. Jevousremercie infiniment pour votre soutien, votre rigueur et votre bienveillance,maisaussi pour votre confiance renouvelée. J’espère continuer à contribuer, à monhumbleniveau,àlabonneréalisationetàl’aboutissementdetouscesbeauxprojetsmenésauseinduservicePharmacologieMédicalequivousestsicher.Quecetravailsoitl’expressiondemareconnaissanceetdemaprofondeconsidération.
AMonsieurleProfesseurEmmanuelOger,rapporteur
Je souhaite vous exprimer toutema gratitude et mes remerciements respectueuxpouravoiracceptéd’évaluermathèseentantquerapporteur.Jevousremerciepourlesremarquesconstructivesdontvousm’avezfaitpart.
AMadameleProfesseurCatherineQuantin,rapporteur
Vousme faites l’honneur de siéger dans ce jury et d’avoir accepté d’être l’un desrapporteursdema thèse. Soyezassuréede toutemagratitudeetdemonprofondrespect.
8
AMadame leDocteurMaryseLapeyre‐Mestre,directricedethèseetmembredujury
Je vous remercie de m’avoir accueillie il y a déjà plus de 6 ans et d’avoir sum’encourageravecbienveillanceetcompréhensiondans laréalisationd’unmasterpuisdecetravaildethèse.Votreexpertiseetvotrehumanitéontcontribuéplusquetoute autre chose à la bonne réalisation et à l’aboutissement de ce travail, quej’espèredignedevosattentes.
AMonsieurleDocteurFabienDespas,directeurdethèseetmembredujury
Merci pour ta confiance accordée dans la réalisation de cette thèse et pour tonécoute. Je te remercie de m’avoir toujours accueillie dans la bonne humeur. Soisassurédemaprofondesympathie.
J’adresse tout particulièrementmes remerciements àMonsieur le Professeur GuyLaurentpourm’avoirpermisd’intégrer leprojetCAPTORetd’avoirainsi rendu laréalisation de cette thèse possible. Je remercie également tous les membres duprojet CAPTOR avec lesquels j’ai eu l’honneur de travailler dans le cadre de cettethèse.
Merci à Monsieur le Docteur Bourrel, de la DRSM Midi‐Pyrénées, pour avoiraccompagné la transmission des données utilisées dans le cadre de cette thèse.Merci égalementàMadame leDocteurCholley, àqui je tiens àexprimer toutemareconnaissancepourm’avoiraccordésaconfiance.
Merci aux poissons (Manuela, Emilie P, Bérangère, Cécile, Marie‐Céline, Ha...) etapparentés poissons (Emilie J) pour avoir suivi avec attention les avancées de cetravail. Merci à Camille, Petite Etoile, pour sa bonne humeur et son sens del’étiquette lors de nos échanges («OGrandEspadon»).Merci aussi au ProfesseurAnneRoussinpoursesencouragements.
A Edmonde, Nathalie et Jean‐Michel qui ont accueilli avec bienveillance mespassages désespérés et leurs lots de soupirs bruyants, qu’ils soient causés par undigicodedebureaurécalcitrantoutoutsimplementparunpeudelassitude.
Merci François pour avoir suivi de près les avancées de cette thèse et tesencouragements.
9
A Mathilde pour ta bonne humeur et ton enthousiasme. J’espère avoir su tecommuniquer l’importance de la ponctualité en matière de collations. Nosdivergences initiales sur les pandas roux n’ont jamais su gâcher notre belleambiancedetravail.
AGaëllepoursonefficacitéredoutableetsabonnehumeur.
MerciàJulie,monDupont,mondoubleenblond,undesraresspécimensquitientduhobbitetquipeutainsilecomprendre.Asacapacitéàsoulagerinstantanémentunepetitephaseàvideàl’aided’unchallengedeprogrammationSASetdedosettesdeChococino.
Merci Guillaume, pour ton exigence et ton humanité. Ton avis et ton soutien sontprécieux.
JesaluetousmescollèguesduservicedePharmacologieMédicaleetClinique.Jenepeuxpascitertout lemonde,maisvousaveztous,d’unefaçonoud’autre,suivi lesavancéesdecettethèse.
AuDocteurLaurenceCadieux,pourm’avoirsoutenuependanttoutescesannées.Jevoussuisinfinimentreconnaissante.
ATOUSMESPROCHES…
Amesparentsdecœur,sansquiriendetoutcelan’auraitétépossible.AChepoursonsoutienetsonamourindéfectibles,etsondévouementdepuistoutescesannées.APierre,poursonécoute,sesconseilsetsapatience,pourlesplatscuisinés«pourlecerveau», et pour toutes les petites collations et boissons apportéescérémonieusement,brastendusetairgrave, lorsde laphasefinalede larédactionde cette thèse. A leurs amis et proches qui ont suivi mon parcours et m’ontencouragée.
AMamieMarie‐Louise,poursonsoutiendiscret,maisnéanmoinstrèsprécieuxpourmoi.Quellebellevivacitéintellectuelleà97ans,tuesadmirable.
AGaia,qui a eu lemérited’accompagnerde saprésenceencourageanteetparfoisronflantelarédactiondemonmémoiredemasteretdumanuscritdethèse.
AMarc, pour cette belle rencontre, ton soutien durant la dernière phase de cettethèse, ton authenticité, ton esprit brillant, et ta présence à la fois apaisante etstimulante.
10
ÀMamie, à qui je dois le goûtde la lecture et dumot juste. J’ai fait, paraît‐il,mespremierscalculsdanstacuisineavecdesboutonsàcoudre.Mercipourcespaisiblesmomentsdemonenfance.
APapi, j’aurais tellementaiméque tusoisencoreparminous, tonsouvenirnemequittepas.
A Lise, mon petit moustique, pour m’avoir obligée à prendre quelques pausessalutaires,aveclacomplicitéd’unepelucheàl’airréprobateursavammentposéesurmonordinateurlemomentvenu.Formidablementingénieux.
AValérie,pourtabonnehumeur,taprécieuseécouteettesencouragementsrépétés.Mercipourl’amieprochequetuesdevenue.
Àtousceuxquej’aipeut‐êtreeul’indélicatessed’oublier…
11
12
13
Tableofcontents
Disclaimer5
Acknowledgements7
Tableofcontents13
Listoftables15
Listoffigures&illustrations16
Listedesrésumés17
Listofabbreviations19
I. Introduction23
A. Importanceofthesubject25
B. Handlingdrugexposure:backgroundandcommonstrategies32
C. Reflecting the diversity and complexity of real‐life patterns: from groups to
trajectories38
D. Potentialcontributionsfromotherfields41
E. FrenchHealthinsurancedatabases42
F. Summary47
II. Researchhypothesesandthesisstatement49
III. Researchquestions51
IV. Objectives53
V. Contextoftheprojects57
VI. Fieldofthethesis59
VII. Resultsoftheprojectsimplemented61
A. Chapter 1: Anticipating gaps in longitudinal data availability: “Overview of
drug data within French health insurance databases and implications for
pharmacoepidemiology”67
14
B. Chapter2:Illustratingtheimpactofmethodologicalchoiceonriskestimates
andtheinterestoftime‐dependentexposure:“Benzodiazepinesandriskofdeath:
resultsfromtwolargecohortsstudiesinFranceandUK”83
C. Chapter 3: Estimating the impact of immeasurable exposure periods due to
hospitalizationsonriskestimatesinmedico‐administrativedatabases109
D. Chapter 4: Dealing with longitudinal data and multiple concomitant
exposures in specific contexts: “Identifying cancer treatment regimens inFrench
healthinsurancedatabases:anapplicationinmultiplemyelomapatients”143
E. Chapter5:Dealingwithlongitudinaldataandmultipleconcomitantexposures
in specific contexts: “Analysing longitudinal exposure to produce automated
indicatorsonpotentialdrug‐druginteractions”175
F. Chapter 6: Improving the exploration of longitudinal drug data: “Data
visualizationfordrugexposureinpharmacoepidemiology”213
G. Complementary Chapter: “Identifyingmultiplemyeloma patients using data
fromtheSNIIRAMandPMSI:validationusingtheTarncancerregistry“231
VIII. Generaldiscussion263
IX. Perspectives277
A. Knowledgeofsecondarydatasourcesanddataexploration:perspectivesand
futureresearch283
B. Impactofmethodschosenandbias:developmentsplanned285
C. Etiologically‐compatible modelling: further perspectives on integrating
concomitantdrugs287
X. Recommendations295
XI. Conclusion303
XII. Bibliographie305
XIII. Appendices323
15
Listoftables
Table1.Typicalstructureofdatasources36
Table2.Contextoftheprojects(listofgrants)58
Table 3. Methods for drug exposure measurement in studies investigating the
associationbetweenbenzodiazepinesuseandmortality103
Table4.Factorsconsideredforimplementingthemultinationalstudy107
Table 5. Proposed steps for implementing a strategy for accounting for
immeasurabletimebias.286
Table6.Stepsforidentifyingtreatmentlinesinclaimsdatabase288
Table7.Plannedparametersforasystematicassessmentofchemotherapybuilding
algorithmsonthebasisofalistofrecommendedregimens290
Table8.Stepsforadaptingthecompendiumofinteractionforautomateddetection
inclaimsdatabases291
16
Listoffigures&illustrations
Figure1.Patientsprofile.Individualdrugsequences219
Figure 2. Sankey diagram. Trajectories according to age class and chemotherapy
regimensinthefirst6monthsoftherapy(n=236).220
Figure 3. Sankey diagram. Trajectories of doses in patients receiving incident
lenalidomide(first6monthsoffollow‐up,n=200)221
Figure4.Streamgraphrepresenting trends inprevalenceof selecteddrugclasses.
Incidentmultiplemyelomapatients (12monthsbeforeand12monthsafter index
date)223
Figure 5. Aggregated longitudinal drug exposure patterns (patients receiving
incidentlenalidomide,n=200)224
Figure 6. Heatmap of frequencies of potential drug‐drug interactions according to
the main (level I) ATC class (multiple myeloma cohort for pDDI identification,
n=506)225
Figure7.Longitudinaldataavailability226
Figure8.Dstributionofthedelaybetweenindexdateandthefirstdateofexposure
227
Figure9.Exposureprofile228
17
Listedesrésumés
Résumé1.Introduction23
Résumé2.Hypothèsesderecherche49
Résumé3.Questionsderecherche51
Résumé4.Objectifs53
Résumé5.Contextedesprojets57
Résumé6.Champdelaréponse59
Résumé 7. Mieux connaître les bases de données de l’assurance maladie pour
réduirelesbiaispotentielsdanslecadredelamesuredel'expositionaumédicament
69
Résumé 8. Illustrer l’impact de la méthode de mesure de l’exposition sur les
estimateursderisque:applicationà l’étudede l’associationentrebenzodiazépines
etmortalité85
Résumé 9. Evaluer l’impact des périodes inobservables lors des
hospitalisationspourlesétudespharmacoépidémiologiques111
Résumé 10. Mieux appréhender des situations impliquant des données
longitudinalesetcomplexes:«Reconstituer les lignesde traitementreçuesenonco‐
hématologieàpartirdesdonnéesduDCIRetduPMSI:applicationàl’étudedescycles
dechimiothérapiedanslemyélomemultiple»145
Résumé 11. Mieux appréhender des situations impliquant des données
longitudinalesetcomplexes:«Analysededonnéeslongitudinalespourlaproduction
d’indicateurs automatisés sur les Interactions médicamenteuses potentielles:
applicationauxbasesdedonnéesdel’assurancemaladie»177
18
Résumé 12. Visualisation de données pour l’exposition médicamenteuse en
pharmacoépidémiologie:uneétudedecasdanslemyélomemultiple215
Résumé 13. S’assurer de la validité de l’identification des cas pour lesmodèles à
l’étude:validationdel’algorithmed’identificationdumyélomemultipleàpartirdu
registredescancersduTarn233
Résumé14.Résumédesprincipauxrésultatsobtenus263
Résumé15.Perspectives277
Résumé16.Propositionderecommandations295
Résumé17.Conclusiongénérale303
19
Listofabbreviations
ANSM («Agencenationaledesécuritédumédicament»):Frenchmedicinesagency
ALD(«affectionde longuedurée»): long‐termandseriousconditionsthatare fully
coveredbythenationalhealthcarescheme
ATIH(«Agencetechniqued’informationsurl’hospitalisation»):TechnicalAgencyfor
HospitalisationInformation
ATU(«AutorisationTemporaired’Utilisation»):TemporaryAuthorizationforUse
CIP(«ClubInter‐Pharmaceutique»):nationalcodingschemeforidentifyingasingle
drugpackage
CPRD,ClinicalPracticeDatalink
CNAMTS («Caisse nationale de l’assurance maladie des travailleurs salariés»):
NationalHealthInsuranceFundforSalariedWorkers
CNIL («Commission nationale de l’informatique et des libertés»): National Data
ProtectionCommission
DCIR(«Donnéesdeconsommationinterrégimes»):inter‐schemehealthcaredata
EGB(«échantillongénéralistedesbénéficiaires»):Permanentbeneficiariessample
EHPAD («établissement d'hébergement, pour personnes âgées dépendantes»):
nursinghomes
EMA,EuropeanMedicinesAgency
FDA,USFoodandDrugAdministration
FOIN («Fonction d'Occultation des Identifiants Nominatifs»): algorithm used to
anonymisepatientsID
GHM(«Groupeshomogènesdemalades»):DiagnosisRelatedGroups(DRGs)
20
GHS(«groupehomogènedeséjour») :homogeneousstaygroup intheDRG‐related
system
HAD(«hospitalisationàdomicile»):homehospitalization
IDS(«Institutdesdonnéesdesanté»):HealthDataInstitute
INCA,InstitutNationalDuCancer
INSEE («InstitutNational de la Statistique et des Études Économiques»): French
NationalInstituteforStatisticsandEconomicStudies
ISAC,IndependentScientificAdvisoryCommitteeforMHRADatabaseResearch
MHRA,MedicinesandHealthcareproductsRegulatoryAgency
MSA («Mutualité Sociale Agricole”): national health insurance scheme for
agriculturalworkersandfarmers
NIR («numérod'inscription auRépertoireNational d'IdentificationdesPersonnes
Physiques»):nationalhealthinsurancenumberofabeneficiary
OMOP,ObservationalMedicalOutcomesPartnership
PMSI(«programmedemédicalisationdessystèmesd’information»):Programforthe
MedicalizationofInformationSystems
RAMQ,Régiedel'assurancemaladieduQuébec
RECORD, REporting of studies Conducted using Observational Routinely collected
Data
RG («Régime Général»): main health insurance scheme, see National Health
InsuranceFundforSalariedWorkers
RSI (“Régime socialdes indépendants»): nationalhealth insurance scheme for self‐
employed
SEER,Surveillance,Epidemiology,andEndResultsProgram
21
SNDS,SystèmeNationaldesDonnéesdeSanté
SNIIRAM(«Systèmenationald’informationinter‐régimedel’assurancemaladie»)
SSR(«soinsdesuiteetderéadaptation»):Postoperativeandrehabilitation
STROBE,Strengtheningthereportingofobservationalstudiesinepidemiology
T2A (“tarification à l’activité”,): the activity‐based diagnosis Related Groups
paymentsystemofpublicandprivatehospitals
UCD (“unités communes de dispensation”): national coding scheme used for
identifyinghospitaldrugs.
22
23
I. Introduction
Résumé1.Introduction
SUMMARYINFRENCH
Importancedusujet
Lorsde lamisesur lemarché, les connaissancessur leseffets indésirablesdes
médicamentssontlimitéesetjustifientunsuivipost‐autorisationrapproché.Les
étudespharmacoépidémiologiquesrépondentàcetobjectif.
Cependant, le poids donné à ces études observationnelles a parfois été sous‐
estiméencomparaisonau«gold‐standard»queconstituentlesessaiscliniques.
Laplaceaccordéeauxétudesobservationnellesreprésenteunenjeumajeur,qui
dépendétroitementdelarobustessedesrésultatsobtenusetdelaconfiancequi
leur est accordée par les différentes parties prenantes (agences de régulation,
etc.).
Le contexte multinational de pharmacoépidémiologie, et l'augmentation du
nombre d'études, ont suscité des inquiétudes en lien avec la production de
résultats contradictoires ou faussement significatifs, et avec le constat de
l’impactpotentiellementmajeurdeschoixméthodologiquessur lesconclusions
produites.
Danscedomaine, laméthodedemesurede l’expositionauxmédicamentset la
fenêtrederisqueconsidéréepourraientêtredesfacteursmajeursdevariabilité
desrésultatsobtenuspourunemêmequestionderecherche.
Mesuredel’exposition:méthodesusuellesdepriseencompte
Les comparaisons inter‐groupes traditionnellement utilisées présentent
l’inconvénient majeur de négliger le caractère changeant de l’exposition
24
médicamenteuse,etomettentdeprendreencompteleséventuelschangements
dedoses,interruptions,etc.1.
Desapprochespermettantdeprendreencompte lesvariablesdépendantesdu
tempsontétédéveloppées,permettantunemodélisationplusflexible2,3.
Expositionmédicamenteuseenvieréelle:del’approchegroupeàlanotionde
trajectoiresd’exposition
Les expositions médicamenteuses sont multiples et discontinues, les
comparaisonseffectuéesàpartirdegroupesexclusifsreflètentmallaréalité.
Avec notamment le passage successif par différentes lignes de traitement
(exemple des chimiothérapies), le problème de catégorisation de l’exposition
peut se ramener à une reconstitution de «trajectoire d’exposition». Il est
cependant difficile de rendre compte de schémas de traitement complexes et
discontinusenutilisantdesstatistiquesdescriptivesconventionnelles.
Lesbasesdedonnéesdel'assurancemaladie(SNIIRAM)offrentunpotentieltrès
importantpourlarecherchepharmacoépidémiologique,enlienaveclataillede
la population couverte et le champ de ces données, à la fois ambulatoires et
hospitalières.Cepotentielestcependantloind'êtreexploitéàsajustevaleur,du
faitdecontraintesd’accès,maisaussienraisondelacomplexitédesdonnées.
Desdéveloppementsméthodologiquessontnécessairespouraméliorer laprise
en compte de l’exposition, en particulier dans des contextes spécifiques
impliquantdestrajectoiresd’expositioncomplexes.
25
Introduction
This section outlines the importance of exposure measurement in
pharmacoepidemiology, in particular in the context of concerns regarding the
impact ofmethodological choices on study results. It also provides a background
withmethodsappliedforconstructingtreatmentepisodesonthebasisofpatients‐
basedclaims.A thirdpartoutlines the lackof researchonhow tohandlecomplex
treatmentepisodesinspecificareas,andwhatmethodologicaldevelopmentsonthe
French health insurance databases might have to offer for longitudinal
pharmacoepidemiologicalstudies.
A. Importanceofthesubject
1. Increasingnumberofstudiesusingsecondarydatasources
forstudyingdrugutilization,drugsafetyoreffectiveness
When placed on themarket, the knowledge of a drugmight be quite limited and
wouldrequireaclosepost‐approvalmonitoring.Pharmacoepidemiologicalstudies,
inthesamewayaspharmacovigilance,areinlinewiththisobjective.However,the
weightgiventotheseobservationalstudieshassometimesbeenunderestimatedin
comparison to the "gold standard" represented by randomized controlled clinical
trials. In addition, there is a strong need to provide a robust conceptual data
managementandmethodologicalframework,tostrengthentheircredibilityfacedto
datafromclinicaltrialsandtoenhanceconfidenceintheconclusionsderived.
An essential prerequisite: transparency of dataa)
sourceandmethodologicalsupport
Thegrowingmultinationalcontextofpharmacoepidemiologyhasgeneratedaneed
to develop common protocols and data models 4 and the need of documenting
database content. Before going further in the implementation of multi‐source
studies,thecontentofthedatabaseshouldhavebeensufficientlydocumented,and
the crucial content should have been validated 5. Pharmacoepidemiological
databaseswhichhavebeenparticularlyusedareoftenaccompaniedbya rangeof
26
methodological work ensuring sufficient robustness of the data ormethods used.
ThisisthecaseforinstancefortheClinicalPracticeResearchDatalink(CPRD)6,7.
Increasing the confidence of stakeholders in the conclusions derived from
observationalstudiesiscrucial,andiscloselyrelatedtoanappropriatestructuring
of the discipline. Examples towards a robustmethodological framework comprise
thedevelopmentofgoodpracticesandnetworks.
The European Network of Centres for Pharmacoepidemiology and
Pharmacovigilance (ENCePP) was created by the European Medicines Agency in
2006inorder“tostrengthenthemonitoringofthebenefit‐riskbalanceofmedicinal
products in Europe by […] facilitating the conduct of high quality, multi‐centre,
independent post‐authorization studies (PAS) with a focus on observational
research;bringing togetherexpertiseandresources inpharmacoepidemiologyand
pharmacovigilanceacrossEuropeandprovidingaplatformforcollaborations;[and]
developing and maintaining methodological standards and governance principles
forresearchinpharmacovigilanceandpharmacoepidemiology”.
The methodological approaches for multi‐source pharmacoepidemiology studies
have been addressed by the ENCePP Work Plan 2013‐2014 8. The report of the
Working Group on data sources andmulti‐source studies prepared a synthesis of
currentpracticeandlessonslearnedfromthesestudies,onthebasisofasurveyof
researcherscoordinatingmulti‐sourceprojectsfundedbytheEuropeanCommission
9.Inaddition,an“InventoryofEUdatasourcesandmethodologicalapproachesfor
multi‐source studies” had been planned as a mandate of an ongoing ENCePP
WorkingGroup10.
Generalguidancea)
In addition to the creation of networks, methodological standards and good
practices inPharmacoepidemiologyrepresentan importantstep for increasingthe
quality and robustness of pharmacoepidemiological studies. The ENCePP has
publishedaGuideonMethodologicalStandardsinPharmacoepidemiology11.
27
In parallel, the International Society For PharmacoEpidemiology (ISPE) has
developed Guidelines for Good Database Selection and use in
PharmacoepidemiologyResearch(PharmacoepidemiolDrugSaf2012;21:1‐10).The
International Society for Pharmacoeconomics and Outcome Research (ISPOR) has
designed “Good research practices for designing and analysing retrospective
databases for comparative effectiveness research”12–14. The German Society for
Epidemiology (DGEpi) has also formulated Good Practice in Secondary Data
Analysis15.TheFoodandDrugAdministration(FDA)“BestPracticesforConducting
andReportingPharmacoepidemiologicSafetyStudiesUsingElectronicHealthCare
DataSets”arealsorelevantinthisarea16.
Otherqualityassessmenttoolsfortheevaluationofpharmacoepidemiologicalsafety
studies have been reviewed 17 and other methodological papers 18–20 are also
contributive.
Outside general guidances which could provide some recommendations on
reporting,dedicatedrecommendationshavebeendesigned.Foryears,theSTROBE
(Strengthening the Reporting of Observational Studies in Epidemiology) applied
directlyforpharmacoepidemiologicalstudies21.However,followingseveralcallsfor
reporting guidelines22, a checklist for REporting of studies Conducted using
ObservationalRoutinely‐collectedData (RECORD)23wasdevelopedon thebasisof
STROBE (The REporting of studies Conducted using Observational Routinely‐
collectedhealthData(RECORD)Statement).
2. Conflicting results, spurious associations: a need to
increaseconfidenceandlevelofevidence
The multinational context of pharmacoepidemiology, and the resulting increased
number of multi‐source studies have also generated concerns in relation with
conflicting results, spurious associations, and the question of the impact of
methodological choices on study results. In the section “interpreting
pharmacoepidemiologyresults",Strometal.discussedtheissueoferroneoussafety
issue 24, stating that “misinterpretation of epidemiologic studies perpetuates the
impressionthatthedisciplineisweakbygeneratingcontroversyoverstudyresults”.
28
Theyconcludedthat“thedisciplineofpharmacoepidemiologymaybeimprovedby
focusingsupport,assessingstudyqualityandadvancingagreaterunderstandingof
thefield”.
To overcome this issue, several international programs have been launched. The
paragraph above highlights the lessons learned from methodological projects on
observationalstudies.
Lessons learned from PROTECT on commona)
protocolsformulti‐databasestudies
The outputs of the PROTECT projects (Pharmacoepidemiological Research on
OutcomesofTherapeuticsbyaEuropeanConsorTium)areofparticularinterest in
thisarea.TheoverallobjectiveofPROTECTwasto“strengthenthemonitoringofthe
benefit‐riskofmedicinesinEurope”,andoneofitsspecificobjectivewas“toidentify
andhelpresolveoperationaldifficultieslinkedtomulti‐siteinvestigations”.Onecase
studyhasparticularlyillustratedtheimpactofthestudydesignandthechoiceofthe
risk window 25. This study aimed to assess risk estimates of hip/femur fractures
associatedwithbenzodiazepinesuse,using2designsandtwodatasources(Basede
datosparalaInvestigaciónFarmacoepidemiológicaenAtenciónPrimaria,BIFAPand
CPRD). For sensitivity analysis purposes, exclusion of the 30‐day pre‐exposure
periodfromthereferenceperiodresulted inamajor impactonriskestimates: the
incidence rate ratio (IRR) was 0.73 (0.63 ‐ 0.84), but 6.47 (5.91‐ 7.09) after
excludingthispre‐exposuretime.Thisexampledoesnotstrictlyreflecttheimpactof
designchanges,neverthelessanadaptationtomeettheconditionforuse.Thesame
modelofhip/femur fracturesassociatedwiththeconsumptionofbenzodiazepines
wasalsousedtoassessinconsistenciesacrossdatabasesfromthePROTECTproject
26.
One of the final recommendations of this project was to test the robustness of
findings by conducting multiple sensitivity analyses, “using multiple designs (e.g.
cohort/case ‐control vs case‐only)” exposure and outcome definition, and
confounding adjustment 27. The harmonisation of methods with a single study
designisnotsufficienttoavoid“consistentlymeasuringincorrectestimates”27.
29
LessonslearnedfromtheUSExperiences:OMOPb)
The Observational Medical Outcomes Partnership (OMOP) was a public‐private
partnershipdedicatedtothe investigationof“theappropriateuseofobservational
healthcare databases for studying the effects ofmedical products”. A specific aim
wasto“conductmethodologicalresearchtoempiricallyevaluatetheperformanceof
various analyticalmethods on their ability to identify true associations and avoid
falsefindings”(http://omop.org/).Themajorityoftheresultsfromthissystematic
assessment were expected, nevertheless “a disturbingly large number did not
replicate the anticipated ground truths”28. The findings of OMOP highlighted a
certain“fragilityofstandardapproaches”appliedforstudyingsafetyissuesthrough
healthcaredatabases.
The main points related to inconsistencies or unexpected variability across data
sources,studydesignandanalyticapproaches.Somewell‐knownassociationscould
not be detected (benzodiazepines and hip fractures in a self‐controlled design),
while some expected harmful associationswere unexpectedly protective (tricyclic
antidepressants and acute myocardial infarction). Conversely, some drug‐events
pairs considered as negative were found as strongly associated, like typical
antipsychoticsanduppergastrointestinalbleeding.
Madiganetal. 29havesystematicallyexamined the impactof thestudydesignand
analyticalchoicesonriskestimatesfor53drug/eventpairs,usingthepercentileof
the distribution of the relative impact on the estimated risk as an indicator. For
cohort studies, the parameters included the risk window, covariate eligibility
window, analysis strategy and covariates included in the propensity score.
Modifying the risk window has the greatest impact, with 50 % of the analyses
(correspondingtothe50%percentile)givingariskestimatemodifiedby1.36factor,
andevenby2.24in10%ofthecases(90%percentileofthedistribution).
Inaddition, facedwithanheterogeneity in the results at theendof the study, the
roleofthedatabaseshouldbequestioned30.Thisissueisahighlycurrentconcern,
as illustrated by its citation in the IMEDS research program. The Innovation in
Medical Evidence Development and Surveillance (IMEDS) program is intended to
30
extentthefindingsoftheOMOPprogram,andincludessomeobjectivesinrelation
withthedatasourcesandimplicationsofheterogeneity28.
Otherdemonstrationsoftheimpactofmethodologicalchoicesonstudyresultshave
been published. Revera et al. for instance examined the association between
psychotropicdrugexposureandmotorvehicleaccidentsusingtwodifferentdesigns
(case–crossover and case–time–control compared to case–control study) 31.
However, the PROTECT andOMOP experiences are very illustrative in relation to
theirsystematicapproachfortestingcombinationsofdesignchoices.
3. Variabilityofriskestimates:drugexposuremeasurement
andriskwindowmightbecrucial
Aspresentedabove,factorsaffectingriskestimatescomprisestudydesign,database
and population covered, exposure and outcome definitionmeasurement, but also
methodsfordealingwithconfounding27.Amongthesefactors, themethodologyof
drugexposuremeasurementandtheriskwindowmightplayamajorrole.
Theinfluenceofdrugexposuremeasurementhasbeenlessexploredthantheimpact
ofthestudydesign.Thefollowingstudiesprovidedexamplesofsuchinfluence.Ina
study on antidepressants, Gardarsdottir et al. have highlighted the impact of the
methodsformeasuringdepressionrelapse/recurrence,usingconventionalmethods
orafterderivingtreatmentepisodes32.Theriskratiowas1.58(95%CI,1.02‐2.45)
usingthefixedexposure,and0.77(95%CI,0.49‐1.21)usingthesecondmethod.In
aneffectivenessstudycomparingtwomethodsforanalysingexposuretostatins33,
the time‐dependent exposure definition was found to be “more accurate”, in the
sensethatestimatesweremoreconsistentwiththosefromrandomizedcontrolled
trials.
There is no straightforwardmethod for choosing one strategy formodelling drug
exposureandcontrollingthisproblem.Performingsensitivityanalysesappearsasa
minimal requirement, but further methodological investigations are needed to
assesstheimpactofdifferentchoicesintheresultsproduced.
31
4. Modellingdruguse:datamanagementalsomatters
Highimpactaffairsinthefieldofbiomedicalresearchhavealsoraisedawarenesson
theneedformakingbiomedicalresearchmorereproducible.ThejournalNaturehas
publishedaspecialissueon“Challengesinirreproducibleresearch”34.Thisconcern
of reproducibility has been recently discussed by Afonso et al. 35 through a
comparisonoffindingsoninhaledlong‐actingbeta‐2‐agonistsandtheriskofacute
myocardialinfarctioninEuropeanprimarycaredatabasesandareplicationinaUS
claimsdatabase.
The different international initiatives in pharmacoepidemiology have developed
standard procedures for data collection, data management and analysis. In the
OMOPproject,thistooktheformofacommondatamodel.TheMini‐Sentinelproject
wasanotherexampleofalargeconsortiumaimingtofacilitatetheuseofroutinely
collectedelectronichealthcaredataforsafetysurveillance.Intheframeworkofthe
elaborationofacommondatamodel,theyhavealsoreleasedaseriesofcomputing
codesfordataformattingandanalysis.Arecentpaperhascomparedthestrategies
fordatamanagement in four consortiumsusing secondarydata sources, including
OMOPandMini‐Sentinel36.Inthesameway,thelibraryofstatisticalcodeshavebeen
planned27asafurtherdevelopmentofPROTECTprojects.
Theproblemofdrugexposuremodellingshouldnotbe limited tothechoiceofan
appropriate statistical model. Even though it should not be a data management
problem (i.e. technical), data handling could be subjected to numerous sources of
errorsordeviations.Evenadetailedprotocolmaybeinsufficienttoaccountforall
casesencountered.Inadditionthemethodologicalandstatisticaldevelopments,the
importanceofhelping researchers to implement theproposed strategies and then
providing standard methodology and detailed principles of computing should be
highlighted.Thisalsoinvolvesbetterreportingofthemethodologyused.
The problem of data handling is closely related to the reproducibility issue. In a
report published in 2016, the Academy of Medical Sciences has proposed 7
strategies to overcome the reproducibility problem 37. These strategies include in
particular automation and open methods. Automation is defined as “finding
32
technologicalwaysofstandardisingpractices,therebyreducingtheopportunityfor
humanerror”.
5. Place of scientific rationale inmethodological choices: a
needforetiologically‐compatiblemodelling
Facedtounexpectedfindingsattheendoftheproject,OMOPinvestigatorscarefully
investigatederroneousfindingsconsideredas“falsepositive”(benzodiazepinesand
acuterenalfailureanduppergastrointestinalbleeding).Someofthesefindingswere
generated by the permutation of design choices, which was the basis of the
approach. They finally concluded that “one cause of reproducible “error” may be
repeated failure to tie design choices closely enough to the research question at
hand“ andstated that it “is likely that all surveillanceprogramswill need tailored
designsthatreflectpharmacologicandclinicalknowledge”38.
These projects have then highlighted the need for an etiologically‐compatible
modelling, and the place of pharmacological rationale in the choice of essential
parameters (risk window). As developed by Lee et al. 39, drug exposure
measurement should not be considered apart from the outcome of interest, a
“conceptualframework”ofthelinkbetweendrugexposureandtheeventofinterest
shouldguidethechoiceofexposuremodelling.Furtherdevelopmentsareneededto
integrate more systematically and closely the pharmacological rationale into the
methodologicaldesign.
B. Handlingdrugexposure:backgroundandcommonstrategies
Pharmacoepidemiology is concerned with the detection or confirmation of
relationshipsbetweendrugexposureandthehealthbenefitsand/orharmsofthese
medications. Central to these studies is the measure of drug exposure. However,
drug exposure across time can be measured in many different ways, and the
33
methodsusedcangreatlyaffecttheobservedassociationbetweenthedrugandthe
outcomeofinterest.
1. Definitionofdrugexposure
Drugexposureisintendedtorefertorealpatient’sintakeofagivendrug,whichis
oftencomplexanddiscontinuous.Definitionsarevaryingaccordingtothediscipline.
Fromapharmacologicalpointofview,drugexposureisoftenestimatedusingarea
under the curve (AUC) methods. When plasmatic levels are not available, drug
exposure could be approached using pharmacokinetic models. These models
representthelevelofdrugexposureovertime.Medicationassaysprovideestimates
thatcouldbeconsideredas themostcloselyrelatedtotheactualdrugexposure39.
However,drugmonitoringisnotfeasibleatalargescale,andpronetobeimpacted
bydesirabilitybias.
In pharmacoepidemiology, possible sources for assessing drug exposure include
prospective surveys, prescribing and dispensing/claim data. When working on
electronic healthcaredatabases, assumptions have to bemadeon actual exposure
status. Even if dispensingdata are givenon adaily basis, treatmentdurations are
often not recorded and have to be derived from quantity dispensed using
pharmacoepidemiological methods for building treatment episodes 40–42.
Ascertainmentofperiodsofexposure isprone tobesubject tomeasurementbias.
Moreover, it should be remembered that claims data refer only to the amount
dispensed and reimbursed, and that real patient intake always remains unknown
(which is also the case with all other databases, based on prescription or
reimbursementdata).
Thesourceofuncertaintyconcerningdrugexposureareinrelationwiththestartof
actual treatment, levelanddurationofexposure,andendofactual treatment(last
drugintake).Uncertainty isalsoinrelationwiththepharmacologicalpropertiesof
thedrugs:uptowhatdurationafterthelastpatient’sintakeshouldweconsiderthat
thereisnoresidualeffect(endoftheriskperiod).
34
2. Choiceofanindexdate
Identifyingnewusersa)
The new‐user design has been developed in order to control for the confounding
thatmayariseasaresultofpatientspastexposures,inparticularsurvivorshipfrom
past exposures43. Thenew‐userdesign avoids the adjustment onvariable that are
intermediateinthecausalchain(colliderbias).Theinterestofnewusersdesignhas
been described for risk‐based or comparative effectiveness research , but also for
adherencestudies44.
One of the main difficulties of applying the new‐user design using healthcare
databasesistheriskofmisclassifyingpatientsbyadoptingatooshorttimewindow
toexcludeallprevalentusers.There isnogeneral rule for thisobservationperiod
and this choice is highly dependent of the type of drugs and context of
administration,butdurationof6or12monthsaregenerallyencountered.However,
asdemonstratedbyRiisetal. forasthmamedications,severemisclassificationcan
be even encountered in periods as long as 2 years 45. Blanch et al. have recently
tested 10 different observation periods for selecting new‐users of antipsychotics
and opioid analgesics, together with the corresponding relative misclassification,
andalsofoundanon‐negligibleimpact46.
In order to adapt this choice to the dataset studied,Hallas et al. have proposed a
graphicalmethod,thewaitingtimedistribution,formakingamorerelevantchoice
oftheperiodofobservationbasedonobservedfirstdateofprescriptions47.
3. Quantifyingdosereceived
TheAnatomicalTherapeuticChemical(ATC)classificationsystemisroutinelyused
for drug utilization studies or other pharmacoepidemiology studies requiring an
assessment of overall doses received.ADefinedDailyDose (DDD) is the assumed
averagemaintenancedoseperdayforadrugusedforitsmainindicationinadults
48.
35
4. Buildingtreatmentepisodeswithinclaimsdatabases
Severalarticleshavedetailedthemethodologyforbuildingtreatmentepisodes(also
called “cycles”) 40–42 and the methods are varying across studies. In general, the
minimal requirement for building treatment episodes should include date of
prescription or dispensing, and quantity dispensed or end of treatment. To build
treatmentepisodes,interruptionshavealsotobedefined.
Treatment discontinuation: how to define gap ina)
drugexposure
Defining treatment discontinuation (“gap”) is a crucial issue, as it has been
demonstratedtohaveanimpactonriskestimates49.Thequestionis:whenshould
we consider that the patient has no more drug available and is likely to have
interrupted his treatment. Strict definitions (30 days for a 30‐day treatment
dispensed) are obviously not adapted. The conventional approach is to allow a
maximum duration (called grace period) for refill between two consecutive
dispensing or after the estimated end of treatment before considering that the
patienthasactuallystoppedhistreatment.Withthisapproach,anyadditionalrefill
afterthisperiodwouldthenbeconsideredasthestartofanewtreatmentepisode.
Definingthisgraceperiodshouldreflecttherealityofpractice.
Determination of accumulated dose andb)
accumulatedduration
As previously stated, the real duration of a treatment estimated through claims
databasesisingeneralunknown.Themaximaldurationforasingleprescriptionin
France is 30 days. Even in the case of renewable prescription, the quantity
dispensedisforonemonth.Tocontinuetheirtreatment,patientshavetoreturntoa
pharmacytobedispendedthequantityforthenext30days.Thedateofdispensing,
identificationofthedrugpackageandquantitydispensed(numberofpacks)isthen
automatically recordedand transmitted tohealth insurance informationsystemto
enablereimbursement.Asdiscussedinthepreviousparagraph,ifthenextrecordfor
the same substance exceeds a fixed number of days, the treatment is considered
36
interrupted.Sensitivityanalysescouldberealizedwithdifferentfixedduration.The
doses are derived from the package codes indicating package dose and size, and
from the variable indicating the number of individual product packs dispended.
TypicalstructureofdatasourcesisprovidedTable1.
Table1.Typicalstructureofdatasources
PatientId Eventdate DrugPackageCode
Itemperpack
Dosageoftheitem
QuantityDispensed
1 01AUG2016 XYZ1 28 25 11 02SEP2016 XYZ1 28 25 1
Theformulaforcomputingthedosereceivedonthebasisofinformationprovidedis
givenbelow.
Dosereceived DDDstrengh mg ∗ numberofunits ∗ numberofpacksdispended
DDDfortheactivesubtance mg
5. Biasesaffectinglongitudinaldrugexposuremeasurement
In addition to sources of uncertainty concerning start and end of treatment, drug
exposuremeasurementislikelytobeaffectedbyawiderangeofbiases.Inmostof
thecases,exposedpatientsmaybemisclassifiedasunexposed(orconversely)orthe
level of exposure could be underestimated due to incomplete data capture
(measurementbias).Misclassificationbiasmaybebidirectional,anddifferentialor
not among thosewith orwithout the event of interest 50. Epidemiological studies
have investigated the impactofexposuremisclassificationonriskestimates,using
cohortorcase‐controldesign51,andinstudiesusingmultiplelevelsofexposure52,53.
AmoregeneralframeworkonquantitativebiasassessmentwasofferedbyLashet
al54.
Immeasurable timebias isoneexampleofmisclassificationbias, inwhichexposed
patients are misclassified as unexposed due to unavailability of exposure data
duringspecificperiods(hospitalisationsingeneral). Indeed,drugsadministeredto
37
hospitalizedpatientsarenotavailablewithinalmostallhealthinsurancedatabases,
such as inMedicaid/Medicare in the US or inmedical claims database of Quebec
(RAMQ). This issue also affects clinical databases such as the CPRD in UK 55–57.
During hospitalization, the patients’ status for drug exposure could not be
ascertained, leading to an unobservable or immeasurable exposure time bias in
whichpatientsaremisclassifiedasunexposed56.Thepotentialimpactofthisissue
hasbeen illustrated forsafety 56,effectivenessstudies55,and forstudiesassessing
drugcomplianceandpersistence58.However,theseimmeasurableperiodsarevery
rarelytakenintoaccountinpharmacoepidemiologystudies.Inasystematicreview
investigatingtheimpactofcompliancetoosteoporosispharmacotherapyonfracture
risk59,themajorityofstudiesdidnotacknowledgethisbiasorexplicitlyignoredit
whenstudyingcompliancetomedicationregimens.
6. Modellingdrugexposureinassociationstudies
Standard strategies for modelling drug exposure include (i) fixed exposure (ii)
currentuse(timedependentbinaryvariable),(iii)accumulateduse(timeordose),
(iv) past, current and no use, and (v) more complex models derived from these
approaches.
Considering fixedgroupsofexposurehasbeencommonpractice foryears.Groups
were allocated on the basis of the first exposure encountered. The patients were
analysedasexposedwhateverthedurationofuse(intentiontotreatapproach).This
approachpresentedthemajorinconvenientofomittingthechangingnatureofdrug
exposure, and could lead to serious time misclassification bias 1. Then, several
studies have then integrated exposure as a time‐dependent variable 6050. One
important requirement for this approach is to derive drug episodes (or cycles),
definedbyperiodsofuninterrupteduse,asdescribedinthepreviouschapters(page
35).
To study a drug/outcome association of interest, one could also choose tomodel
accumulateddoses(accumulatedquantityreceivedsinceindexdateorothercustom
period). In the models using cumulative doses, the quantity dispensed from the
indexdate isadded,andentered inthemodelasacontinuousorclassvariable. In
38
caseofarational forbothcurrentandaccumulatedeffect,drugexposurecouldbe
modelledas“past,currentandnouse”61.
More flexible models have been proposed for modelling cumulative dose and
exposure duration 62–64. Themainmotivationwas that all themethods commonly
usedarerelyingonstrongassumptions:allpastdoseshavethesameimpactforthe
cumulativedosemodel,eveniftheyhavebeenadministeredsomeweeksormonths
ago. In the same way, the current dose model involves that doses previously
received do not have any impact. The authors introduced the concept of a more
generalmodel,theWeightedCumulativeExposure(WCE), inwhichthecumulative
effectismodelledasaweightedsumofallpastdoses”62–64.Thismodelwasapplied
to study the association between flurazepam use and fall‐related injuries in the
elderly64.
In addition to thesemodels, other approaches could be of potential interest, like
multistate/Markov models. Multistate models are not widely used in
pharmacoepidemiology, but previous experiences have shown their interest in
describingdruguse65,66andmodellingpersistence67orregimenchanges68.Markov
modelsprovideaninterestingalternativeforstudyingdrug‐eventassociationsand
the impact of medical conditions, and add flexibility in drug exposure modelling
whilereflectingreal‐lifeanddynamictrajectories.
C. Reflectingthediversityandcomplexityofreal‐lifepatterns:fromgroupstotrajectories
1. Lack of research in specific contexts of complex drug
exposure
Asthecomplexityofexposurepatternsincreases,groupsofeverusersarenomore
relevant for studying drug effects, in particular in cases of multiple, potentially
interacting,andconcomitantexposures.
39
Oneof themostspeakingexamples is theareaofoncology.Oncology isoneof the
areas of research thatwould benefit from additionalmethods taking into account
thecomplexnatureofdrugexposure.Indeed,thereisaneedtodevelopmodelsto
study drug exposure in such populations of highly treated patients, prone to be
exposed to multiple treatment sequences, with cumulative or delayed effect and
differentprognosisvalues.
Given the focus on health insurance databases through this thesis, data sources
collectingexplicitcombinationsofregimens(clinicaldatasource,registries)should
bediscriminatedfromthosecollectingdrugsonapatients‐daybasis,withoutlinking
aspecificdrugtoaparticularregimen.
Studies on cancer drugs in pharmacoepidemiology are still scarce. Most of the
studiesonSEER(Surveillance,Epidemiology,andEndResultsProgram)databases
wereconductedonfirst‐linepatientsonly,andweremainlyfocusedonasingledrug
or on the description of patients' trajectories through sequences of treatment
(surgery, chemotherapy, radiotherapy, etc.), without determining the nature and
historyoftreatmentlinesreceived69,70.
Indeed,cancerpatientsaregenerallyexposedtoseveraltreatmentlines,composed
of one ormore drugs. It is therefore essential to take into account the particular
characteristics of drug exposure in oncology, and to move from a ‘single drug’
approach toward a ‘multidrug, multiline’ perspective when modelling drug
exposure. The complexity of treatment patterns for cancer is growing and the
numberofpossibleregimensincreasesaccordingly.Inthecontextofobservational
studies, it has become more and more difficult to consider past lines and the
durationofpreviouslineswhencomparingmultidrugregimens71”.Studiestreating
thisparticularaspectareveryrare72.
In case of treatment changes over time (chemotherapy), the problem drug
classification or quantification may become a problem of finding patients
trajectories. Modelling such trajectories has already been implemented as
aggregatedmodels.Anexampleisprovidedbythestudiesonbreastcancerpatients
trajectoriesFrance73,74.However,thedescriptionneedstobemoredetailedinorder
40
tobe relevant forpharmacoepidemiology.Theproblemofmultiline therapy isnot
limitedtothecancerarea:thesameproblemarisesforHIVtherapyforinstance.
2. Patientsdrughistory:thequestionofdelayedeffects
More complexways of handling exposure are of particular interest in the area of
long‐term effects. For instance, the relation between proton pump inhibitors and
fractureshouldbeinvestigatedatlongterm,from5to7years75,76.Thisrequiresan
exhaustive assessment of patient exposure. In the same way, the question of
secondary malignancies after lenalidomide exposure in multiple myeloma 77,78
highlights the complexity of such questions (i) lenalidomide is administered in a
condition which is a risk factor for developing some additional haematological
malignancies,(ii)itcouldbeco‐administeredwithotherdrugsincreasingtheriskof
malignancies,and(iii)compositionofpastlinesanddurationofpreviouslinesmust
betakenintoaccount.
3. Commonindicatorsofexposureovertimeandlimitations
Indicatorsofexposureovertimecouldfallintotwomaincategories,corresponding
todifferentobjectives.Intheareaofpatient’sadherence,theobjectiveistoenhance
the lack of exposure in comparison to a reference (expressed in dose received,
number of days). Several indicators have been defined, like persistence, or the
proportionof days covered (PDC). These indicatorswere also adapted for several
concomitantdrugs.
Inothercontexts,conventionalindicatorsareused,likedurationofuse,numberof
treatmentepisodesormeandose. In thecontextofanalysisof largedatasets, it is,
however, difficult to account for complex treatment schemes or discontinuous
exposure using these conventional indicators, and further methods or reporting
methodswouldbeofinterest.
41
4. Takingaccountofconcomitantdrugs
Twodifferentsituationscouldbeencounteredwhenattemptingtotakeintoaccount
ofconcomitantdrugs.Insomespecificcases,drugstakenconcomitantlybelongtoa
recommended scheme, and are included in a broad therapy strategy. In these
particularcases,andat leastfordescribingdrugexposure, itmaybeinterestingto
describeconcomitantdrugsinaformthatreflectscurrentpractice.
Thesecondsituationoccurswhendrugstakenconcomitantlyhaveanimpactonthe
outcome.Bothofthesesituationscouldoverlapinsomecases.Ingeneral,strategies
for managing drug taken concomitantly include (i) no particular strategy (ii)
baseline use taken as a covariate, (iii) concurrent use taken as a time‐dependent
variable(iv)thosewithconcurrentuseareexcluded(riskofselectionbias).
The impactof timingofconcomitantdruguse inpharmacoepidemiological studies
hasbeenrecentlyhighlighted79,usingtheexampleofbenzodiazepines(concomitant
exposure)andantidepressants(exposureofinterest).Thisstudyrevealedthevery
highprevalenceofconcomitantbenzodiazepinesusers,butalso,moreinterestingly,
theimpactthetimingofstartanddurationonriskestimates.
Thesituationmaybemorecomplicatedwhenbuildingacontrolgrouponthebasis
onthesamedrugsof interest.Whensuchanactivecontrol isused, it is frequently
selected among exclusive users (non‐users for the whole period of observation).
However, it is important to avoid selection bias and to have the opportunity to
observeconcomitantpatternsforthesamepatients,asinreal‐life.
D. Potentialcontributionsfromotherfields
When the complexity is too high, methods from areas outside
pharmacoepidemiologymight be useful to explore longitudinal drug records, and
helpfindingpatternsinheterogeneousdata.ThediffusionofgraphictoolsafterMini
Sentinel or OMOP projects are illustrative is this area. An interesting example is
42
provided by theMini Sentinel group, discussing briefly how “pictorialmodels can
help elucidate statistical models”80, proposing a set of visual types that could be
producedbeforeorjustafterthestudytofacilitatethechoiceofstudydesignwhile
verifying the underlying assumptions. In the same way, some visualizations
developedasapartof theOMOPprojecthavebeenreported inabookchapter 81.
Noneof thesepublicationswasbasedonapre‐plannedapproach,butbothshared
the visuals developed for supporting the discussions around the various projects.
Thestudypresentedwiththediscussiononconcomitantdrugs79providedanother
example in thisarea. Indeed, theauthorshaveusedvisualizations toascertain the
distributionofthedurationsofconcomitantuseofbenzodiazepinesaccordingtothe
durationsofantidepressantstreatmentepisodes.
The examples are not so numerous in the case of conventional
pharmacoepidemiology studies, but illustrate how data visualization tools might
help to gain insight into patterns of exposure and modelling in
pharmacoepidemiology. Therefore, data visualization approaches are worth
exploringintheareaofdrugexposuremeasurement,buttheplaceandthescientific
framework forusing thesemethodsmust absolutelybe specified in the contextof
hypothesis‐basedstudies.
E. FrenchHealthinsurancedatabases
1. PresentationofFrenchhealthinsurancedatabases
SNIIRAMa)
In France, most of the population is covered by 3 health insurance schemes: the
main health insurance scheme (RG) for salaried workers (including also retired,
unemployed and low‐beneficiarieswith universal healthcare coverage), the health
insurance scheme for agricultural workers and farmers (MSA) and the health
insurance scheme for self‐employed (RSI). These three healthcare schemes and
43
other supplementary health insurance schemes (government employees, public
education…)accountformorethan97.5%oftheFrenchpopulation82.Thedataare
collectedseparatelyforeachschemeandgatheredinanationaldatabase,theinter‐
scheme consumption data on ambulatory health expenditure (Données de
consommationInterRégimes,DCIR).Datafromhospitalizationstays(“Programmede
MédicalisationdesSystèmesd’Information”,PMSI)aremanagedwithinasinglecase‐
mixdatabaseoftheactivity‐basedpaymentsystem,(“tarificationà l’activité”,T2A).
The systemwas initiated formedical, surgical and obstetrics care (PMSIMCO) in
1991. Separated systems were implemented for postoperative and rehabilitation
(PMSISSR for “SoinsdeSuiteetderéadaptation”),homehospitalizations (HAD for
“HospitalisationàDomicile”), andpsychiatricwards.ThePMSI isheldby theATIH
(“Agence technique d’information sur l’hospitalisation”), and provides data on all
claims paid by the national health insurance system (whatever the specific health
insurancescheme) topublicandprivatehospitals.Thedataarekept for10years,
plusthecurrentyear.ThePMSIissecondarilylinkedtotheDCIR.
TheSNIIRAM(“Systèmenationald’informationinter‐régimedel’assurancemaladie”)
compriseslinkedambulatoryandhospitaldata,correspondingtodatafromnational
healthinsuranceschemes(DCIR)andhospitaldata(PMSI)
2. The permanent beneficiaries sample (“échantillon
généralistedesbénéficiaires”,EGB)
Thepermanentbeneficiariessample(“échantillongénéralistedesbénéficiaires”,EGB)
isa1/97threpresentativesampleof theSNIIRAM,comprisingmore than660,000
Frenchbeneficiaries,plannedfora20‐yearduration(from2003to2023)83.
The EGB includes longitudinal records of all reimbursed healthcare expenses,
includingconsultationsinprimaryandsecondarycaresettings,dispensingdatafor
all reimbursed medications (primary and secondary care) and diagnostic testing
performed.TheEGBdoesnotcontainmedicaldataorlaboratoryresults,butmajor
chronic diseases can be identified using International Classification of Diseases
(ICD)‐10codes.ThedateofdeathisprovidedindirectlybytheNationalInstituteof
44
StatisticsandEconomicResearch(INSEE).Thecauseofdeath isnotrecorded.The
EGBcontainsbasicdemographicdata(age,gender,areaofresidence)butdoesnot
record lifestyle data. This database has been linked with another large‐scale
information system containing data from hospital stays (PMSI), providing linked
dataafter2005andonwards, includingentryanddischargedates,proceduresand
diagnosesaccordingtoICD‐10.TheEGBhasbeenusedforpharmacoepidemiological
research23‐25.
3. Strengthening the potential of French Health insurance
databases
The SNIIRAM offers a great potential for pharmacoepidemiological research in
relation with its national coverage, linkage of ambulatory and hospital data, and
complementarydataondemographics,hospitaldiagnosisandlong‐termconditions,
andisincreasinglyusedforpharmacoepidemiologicalresearch57,83.
In France, there is no working interface offering structured tools for using these
data. However, some aspects are prone to change in relation to recent legislative
changes 84. With the creation of the « Système national des données de santé »
(SNDS), INSERMwillofferawiderangeofservices forresearchers.Theseservices
havetobespecified,butmayincludemethodologicalsupport.
Frenchdatabaseswerenotintegratedintothelargenetworksdiscussedinthefirst
partofthisintroduction,butanongoingproject(ALCAPONE)isplanningto“assess
the suitability of the French nationwide healthcare insurance system database
(SNIIRAMandEGB)fordrugsafetysignalgenerationbasedontheOMOPreference
setandmethodologies,andthecase‐populationapproach”85.
In addition, there is an initiative in relation with case findings algorithms
(REDSIAM). The validation of hospital data use for cancer research has been
addressed by the INCA (“Institut National Du Cancer”). The validation of the
Charlson’s score 86 in these databases is a good example of validation studies
45
participatingtoimprovetherobustnessofthemethods.However,tothebestofour
knowledge,nosuchprojectsareongoinginrelationwithdrugdata.
Thecomplexityofdatabasearchitecture,methodsofdatacollectionandreleasehas
been identified as one of the obstacles to the development of research on these
healthinsurancedatabasesintheReportonGovernanceandtheuseofhealthdata
87. The report from the French Courts of Auditors 88 has also pointed out several
limitationsoftheSNIIRAM,includingthelackofsupportforusers.
At the date of writing, there has been no detailed description of drug data or
methodological guidance concerning studies on drug use within French health
insurancesdatabases, inspiteofitswell‐recognizedcomplexity.Adeepknowledge
of database content, origin and release is, however, crucial in order to avoid bias
whendesigningpharmacoepidemiologicalstudiesusingtheFrenchhealthinsurance
databases.
46
47
F. Summary
Themultinationalcontextofpharmacoepidemiology,andtheresultingincreased
number ofmulti‐source studies have also generated concerns in relationwith
conflictingresults,spuriousassociations,andwiththeimpactofmethodological
choicesonstudyresults.
Increasing the confidence of stakeholders in the conclusions derived from
observational studies is crucial, and is closely related to the robustness of the
evidenceproduced.
Drugexposuremeasurementandriskwindowmightbecrucial
Theproblemofdrugexposuremodellingshouldnotbe limitedtothechoiceof
anappropriatestatisticalmodel.Technicalaspectsofthedatamanagementand
analysiscouldalsointroduceheterogeneity.
Some of the false associations retrievedmight be explained by failure to take
accountofthepharmacologicorclinicalrationale.
Asthecomplexityofexposurepatternsincreases,groupsbasedoneveruseare
nomorerelevantforstudyingdrugeffects.
In case of treatment changes over time (chemotherapy), the problem of drug
classification may become a problem of finding patients trajectories. It is
however difficult to account for complex treatment schemes or discontinuous
exposureusingconventionaldescriptivestatistics.
The French health insurance database (SNIIRAM) offers a great potential for
pharmacoepidemiologicalresearchinrelationwithitsnationalcoverage,linkage
of ambulatory and hospital data. This potential is far from being exploited, in
relationtotechnicalconstraintsduetothecomplexityofthedata.
Methodological development are needed to improve pharmacoepidemiological
studies in French medico‐administrative databases, in particular in specific
contextsinvolvingmultipleandconcomitantdrugexposures.
48
49
II. Researchhypothesesandthesisstatement
Résumé2.Hypothèsesderecherche
The way drug exposure is taken into account has potentially a great impact on
estimates and conclusionwhichwouldbedisseminated.Measuring its impact and
developing a methodological framework to overcome this issue and facilitate its
managementwithinpharmacoepidemiologicalstudiesispivotaltothisprocess.The
propositionsforaframeworkofdrugexposuremeasurementinFrenchinmedico‐
administrative databases, focused on complex exposures, could contribute to
improve measurement of drug exposure in the specific context of discontinuous,
concomitantexposureorinpresenceofimmeasurableperiods.
SUMMARYINFRENCH
Lafaçondontl'expositionaumédicamentestpriseencompteapotentiellementun
impact fort sur les résultats obtenus et donc sur les conclusions scientifiques
diffusées. En mesurant cet impact et en travaillant à l’élaboration d'un cadre
méthodologique permettant d’envisager ce problème de façon plus structurée, en
particulier dans le cas d’expositions complexes, ces travaux pourraient ainsi
contribueràaméliorer larobustessede lamesurede l’expositionmédicamenteuse
danslesbasesdedonnéesensanté.
50
51
III. Researchquestions
Résumé3.Questionsderecherche
SUMMARYINFRENCH
a)Commentunemeilleureconnaissanceaprioridessourcesdedonnéesetdeleur
origine contribue à réduire les biais potentiels dans le cadre de la mesure de
l'expositionaumédicament?
b) Comment positionner les comparaisons intergroupes classiques face à des
méthodes intégrant l’exposition dépendante du temps? Ces différentes méthodes
aboutissent‐elles à des résultats comparables? Comment interpréterd’éventuelles
différences?
c)Commentfairefaceauxrupturesdansladisponibilitédedonnéeslongitudinales?
Quel est l’impact du biais lié à ces périodes inobservables dans les études
pharmacoépidémiologiques?
d) Comment identifier les schémas de traitement de chimiothérapie ou des
combinaisonsmédicamenteusesd’intérêtàpartirdedonnéesdedélivrance?
e)Quelleestlacontributionpotentielledesoutilsdevisualisationdedonnéespour
améliorerl'explorationdesdonnéeslongitudinales?
f) Comment la proposition d’un cadre méthodologique adapté aux expositions
complexespeut‐ellecontribueraudéveloppementdesétudeslongitudinalesausein
desbasesdedonnéesmédico‐administrativesfrançaises?
52
Researchquestions
a) How a priori knowledge of data sources content and origin help to reduce
potentialbiasindrugexposuremeasurement?
b) How do traditional between‐group comparisons compare with methods
integrating the time dependent nature of drug exposure? Do different
methodsforhandlingdrugexposureproducedifferentriskestimates?
c) Howtodealwithgapsinlongitudinaldataavailability?Howtoimprovethe
integrationofimmeasurabletimebiasinfurtherstudies?
d) How to identify relevant drug regimens or drug‐drug combinations of
interestwithinlongitudinaldatawithmultipleconcomitantdrugs?
e) What is the potential contribution of data visualization tools for improving
theexplorationoflongitudinaldrugdata?
f) Howdoesaconceptualandmethodologicalframeworkhelpthedevelopment
oflongitudinalstudiesinFrenchmedico‐administrativedatabases?
53
IV. Objectives
Résumé4.Objectifs
SUMMARYINFRENCH
L'objectif de cette thèse est de développer desméthodes pour lamodélisation de
l’expositionauxmédicaments,de tenterdeproposerdessolutionspourassurer la
qualitédesdonnéeslongitudinalesetcontrôlerlerisquedebiaislorsdel'utilisation
desbasesdedonnéesmédico‐administratives.
a)Effectuerunerevuedesdonnéessurlemédicamentcontenuesdanslesbasesde
donnéesdel’assurancemaladie.
b) Evaluer les méthodes de comparaison intergroupes classiques face à des
méthodesintégrantl’expositiondépendantedutempsdanslecontexted’uneétude
cohorterétrospective.
c)Evaluerl’impactpotentieldubiaisliéauxpériodesinobservablesdanslesétudes
pharmacoépidémiologiques.
d) Proposer de nouvelles méthodes permettant d’identifier les schémas de
traitement de chimiothérapie ou des combinaisons médicamenteuses d’intérêt à
partirdedonnéeshétérogènesdedélivrance.
e) Evaluer la contribution potentielle des outils de visualisation de données pour
améliorerl'explorationdesdonnéeslongitudinales.
f) Proposer un cadre conceptuel et méthodologique pour contribuer au
développement des études longitudinales au sein des bases de données médico‐
administrativesfrançaises.
54
Objectifstransversaux
Aider à accroître la pertinence des descriptions ou des analyses de
l’expositionmédicamenteuse
o Par la productiond’outils permettant demieux rendre comptede la
réalitédel’expositionmédicamenteuse,
o Enprenantencomptelesaspectspharmacologiquesdansladéfinition
despériodesd’expositionàrisque.
Favoriser la cohérence et la reproductibilité des projets d’études
pharmacoépidémiologiques.
Faciliter l’exploration des données longitudinales dans des situations
complexesgrâceàlamiseenapplicationdeméthodesoriginales.
Promouvoir l’utilisation des bases de données française de l’assurance
maladie.
Faciliterletransfertd’expérience:formaliserlesélémentsclésdetellesorte
qu’ils puissent fournir une base utile pour les chercheurs dans le cadre de
futursprojets.
55
Mainobjective:
Theobjectiveofthisthesisistodevelopmethodsforincreasingtherobustnessand
the pharmacological relevance of longitudinal drug exposure modelling, and
guidance for researchers onhow to ensure longitudinal data availability,measure
drugexposureinspecificcontexts,andcontrolsometime‐relatedbiaseswhenusing
electronichealthcaredatabases.
Specificobjectives:
To review the sources of information on drug use in the French health
insurancedatabasesanddiscusstheriskoftimerelatedbias.
Toassesshowtraditionalbetweengroupcomparisoncomparewithmethods
integrating the time‐dependentnatureof drug exposure in the context of a
retrospectivecohortstudy
Toassesstheimpactofgapsinlongitudinaldataavailabilityinthecontextof
aretrospectivecohortstudy
To identify relevant drug regimens or drug‐drug combinations of interest
withinlongitudinalrecordswithmultipleconcomitantdrugs
Toassessthepotentialcontributionofdatavisualizationtoolsforimproving
theexplorationoflongitudinaldrugdata
Toproposeasetofrecommendationstosupportresearchersindealingwith
discontinuous,multipleconcomitantexposureorpresenceof immesureable
time
56
57
V. Contextoftheprojects
Résumé5.Contextedesprojets
SUMMARYINFRENCH
Lestravauxeffectuésutilisentdeuxmodèlesdistincts.Lepremiermodèleutiliséest
plutôtgénériqueetfacilementextrapolable,prototyped’uneexpositiondiscontinue,
c’estceluide lapopulationgénéraleexposéeàdesmédicamentstrèsrépandus, les
benzodiazépinesanxiolytiquesethypnotiques.
Dans une deuxième partie, les investigations sont menées dans le champ en
développement en pharmacoépidémiologie: celui de l’onco‐hématologie, avec les
médicamentsdechimiothérapiedumyélomemultiple commemodèled’exposition
complexe.
Lesmodèlesutilisésreflètentaussilesupportinstitutionnelobtenuaucoursdecette
période.L’étudesurlamortalitéliéeauxbenzodiazépinesaainsiétésupportéepar
l’agence européenne du médicament (EMA). Ainsi le troisième projet utilise les
données françaises issues de l’étude sur les benzodiazépines et mortalité pour
traiterlaquestiondespériodesinobservables(ANSM,appelsàprojetscibléssurles
produits de santé, “Comment considérer les « périodes à trous » dans un suivi
longitudinald’unepriseenchargemédicamenteuse”).
La deuxième partie de ces travaux concernait le traitement des expositions
multiples, avec une application dans le cas du myélome multiple, en lien avec la
thématique Pharmacologie sociale du projet CAPTOR (Cancer Pharmacology of
ToulouseandRegion).Ceprojetestl’undes2lauréatsnationauxdel'appelàprojets
du programme Investissements d'avenir « Pôles Hospitalo‐Universitaires en
Cancérologie»etabénéficiéd’unfinancementpour5ans.Levolet3estenlienavec
laPharmacologieSocialeetprévoitl’utilisationdeméthodesoriginalesintégrantla
télémédecinedans lesuivicliniqueprospectifdespatientset ledéveloppementde
bases de données à visée pharmacoépidémiologique utilisant en particulier les
donnéesduSNIIRAM.
58
Thethesisworksusedmodelsfromdifferentfields.Forthefirstworks,amodelof
frequent and discontinuous exposure in general population is used, with
benzodiazepinesuseisanexample.Then,toinvestigatelongitudinaldrugexposure
in more complex contexts, the model of chemotherapy regimens in an
haematological malignancy, multiple myeloma, was used as a model of complex
exposure.
Thesechoicesalsoreflecttheprojectssupportedduringthethesiscourse.Thework
onbenzodiazepinesandmortalitywasthefirstprojectperformed,andwasrealized
in the frameworkofarestricted tender fromtheEuropeanMedicinesAgency.The
articlewasadeliverableofthetender.Thelargestpartofthethesiswasfundedby
the Captor project. The scientific project of the Toulouse site CAPTOR (Cancer
PharmacologyofToulouseandRegion)isoneofthetwonationalawardwinnersof
the ‘Future Investments' call for projects program. With a funding of 10 million
Euros over 5 years, the project aims to promote the emergence of new cancer
treatment drugs. CAPTOR has 4 areas of investigation: fundamental research,
clinical research, social pharmacology, and education. Under the social
pharmacology workpackage, several types of study (prospective clinical studies,
retrospective studies of databases, surveys) are carried on several diseases
(multiple myeloma, chronic myelogenous leukemia, colon cancer, etc.). Their
common objective is to contribute to a better knowledge of cancer drugs in real
conditionsofuse.
Table2.Contextoftheprojects(listofgrants)
Grants
EuropeanMedicines Agency, restricted tender No. EMA/2012/20/PV, Association
betweenanxiolyticorhypnoticdrugsandtotalmortality.
National Research Agency (ANR: Agence Nationale de la Recherche) for the
“investissementd’avenir”(ANR‐11‐PHUC‐001,CAPTORproject)
LigueNationalecontre leCancer,Demandedesubventionsd'équipementet/oude
fonctionnementdelaboratoire
ANSM,targetedcallforresearchappliedtohealthcareproducts,2014
59
VI. Fieldofthethesis
Résumé6.Champdelaréponse
SUMMARYINFRENCH
Les éléments suivants ont joué le rôle de principes directeurs dans la façon de
répondreàlaquestionderechercheetàl’objectifgénéraldecontribueràaméliorer
la robustesse et la pertinence de la mesure de l’exposition au médicament. Ces
principesincluaient:
La formulation des principes méthodologiques et des leçons issues des
travauxde tellesortequ’ilspuissentaiderd’autreschercheursconfrontésà
desquestionssimilaires.
Un effort constant d’intégration du rationnel pharmacologique dans la
conduitedel’étude,desaconceptionàl’interprétationetàlacommunication
desrésultats.
Unaborddesproblématiquesàl’aided’étudesdecas.
Le développement d’outils transparents, reproductibles et transférables à
d’autreschamps.
Despropositionsdesolutionsprivilégiantlafacilitédemiseenœuvreplutôt
quelerecoursàdesdéveloppementsstatistiques.
Below are detailed the several guiding principles for answering the research
questionsandmeetingthegeneralobjectivetocontributetoincreasetherobustness
and the relevance of drug exposure measurement. These ways of attainment
included:
Theformulationoflessonslearnedandgeneralmethodologicalprinciplesin
suchaformthatitwillhelpotherresearchersinvestigatedrelatedarea
Asystematicattempttointegratepharmacologicalrationaleduringthestudy
course
60
Afocusoncasestudiesandmethodologicaldiscussion
Thedevelopmentoftransparent,reproducibleandhighlycustomisabletools
A proposal for easy to implement solutions rather than statistical
developments(avoiding“black‐box”solutions).
61
VII. Resultsoftheprojectsimplemented
1. Aurore Palmaro, Guillaume Moulis, Fabien Despas, Julie Dupouy, Maryse
Lapeyre‐Mestre. Overview of drug data within French health insurance
databases and implications for pharmacoepidemiological research.
Fundamentalandclinicalpharmacology,2016,DOI:10.1111/fcp.12214
2. AurorePalmaro,JulieDupouy,MaryseLapeyre‐Mestre.Benzodiazepinesand
riskofdeath:ResultsfromtwolargecohortstudiesinFranceandUK.
Europeanneuropsychopharmacology,2015;
DOI:10.1016/j.euroneuro.2015.07.006
3. Aurore Palmaro, Quentin Boucherie, Julie Dupouy, Joëlle Micallef, Maryse
Lapeyre‐Mestre. Unobservable drug exposure due to hospitalization in
medico‐administrative databases: which impact for Pharmacoepidemiology
studies?(PharmacoepidemiologyandDrugSafety,underreview)
4. Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐Mestre.
IdentifyingcancertreatmentregimensinFrenchhealthinsurancedatabases:
an application in multiple myeloma patients (Pharmacoepidemiology and
DrugSafety,underreview)
5. Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Analysing
longitudinal exposure to produce automated indicators on contraindicated
combinations and potential drug‐drug interactions: Application using the
French medico‐administrative database. (British Journal of Clinical
Pharmacology,submitted)
6. Aurore Palmaro, Maryse Lapeyre‐Mestre. Data visualization for drug
exposureinpharmacoepidemiology:acasestudyforcomplexdrugregimens
in multiple myeloma e‐Health Research 2016. How digital technologies
disrupt epidemiology and medical research. Paris, October 11‐12, 2016
(abstract)
7. Aurore Palmaro, Martin Gauthier, Cécile Conte, Fabien Despas, Pascale
Grosclaude, Maryse Lapeyre‐Mestre. Identifyingmultiplemyeloma patients
using data from the SNIIRAM and PMSI: validation using the Tarn cancer
registry(Medicine,underreview)
62
Listofcommunications
Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Association between
benzodiazepinesdrugsandtotalmortality:evidencefromastudyonCPRDdata.IX
annualCongressoftheFrenchSocietyofPharmacologyandTherapeutics,Poitiers,
France,April22‐24,2014(poster)
Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Benzodiazépines et
mortalité:étudedecohorteauRUetenFrance.VIIIcongressofYoungResearchers
in General Practice (Devenir Jeune Chercheur en Médecine Générale), Toulouse,
France,March14‐15,2014(oralcommunication)
AurorePalmaro,QuentinBoucherie, JulieDupouy, JoëlleMicallef,MaryseLapeyre‐
Mestre. Périodes d’exposition inobservables au cours des séjours hospitaliers en
PMSI MCO: quel impact pour les études pharmacoépidémiologiques? ADELF
(Association des Epidémiologistes de Langue Française)‐EMOIS (Evaluation,
Management,Organisation,Information,Santé)meeting,Dijon,France,March10‐11,
2016,(poster)
Quentin Boucherie, Julie Dupouy, Joëlle Micallef, Maryse Lapeyre‐Mestre, Aurore
Palmaro. Unobservable drug exposure due to hospitalization in medico‐
administrative databases: which impact for Pharmacoepidemiology studies? XI
annual Congress of the French Society of Pharmacology and Therapeutics, Nancy,
France,April19‐21,2016(oralcommunication)
Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐Mestre.
Identifying cancer treatment regimens in French health insurance databases: an
application in multiple myeloma patients. XI Congress of the French Society of
Pharmacology and Therapeutics, Nancy, France, April 19‐21, 2016 (oral
communication)
Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐Mestre.
Reconstituer les lignes de traitement reçues en onco‐hématologie à partir des
donnéesduSNIIRAMetduPMSI:applicationàl’étudedescyclesdechimiothérapie
dans le myélome multiple. ADELF (Association des Epidémiologistes de Langue
63
Française)‐EMOIS (Evaluation, Management, Organisation, Information, Santé)
meeting,Dijon,France,March10‐11,2016(oralcommunication)
AurorePalmaro,EmiliePatrasdeCampaigno,MathildeDupui,BerangèreBaricault,
Julie Dupouy, Fabien Despas, Maryse Lapeyre‐Mestre. Analyse de données
longitudinales pour la production d’indicateurs automatisés sur les interactions
médicamenteuses potentielles : application aux bases de données de l’assurance
maladie. ADELF (Association des Epidémiologistes de Langue Française)‐EMOIS
(Evaluation,Management,Organisation,Information,Santé)meeting,Nancy,France,
March23‐24,2017(oralcommunication)
Aurore Palmaro, Maryse Lapeyre‐Mestre. Data visualization for drug exposure in
pharmacoepidemiology: a case study for complex drug regimens in multiple
myeloma. e‐Health Research 2016. How digital technologies disrupt epidemiology
andmedicalresearch.Paris,October11‐12,2016(commentedposter)
AurorePalmaro,MartinGauthier,CécileConte,FabienDespas,PascaleGrosclaude,
MaryseLapeyre‐Mestre.Identifyingmultiplemyelomapatientsusingdatafromthe
SNIIRAM and PMSI: validation using the Tarn cancer registry. GRELL Meeting
(Group for Epidemiology and Cancer Registry in Latin Language Coutries, Nancy,
May4‐6,2016(poster)
64
65
66
67
A. Chapter1:Anticipatinggapsinlongitudinaldataavailability:“OverviewofdrugdatawithinFrenchhealthinsurancedatabasesandimplicationsforpharmacoepidemiology”
Aurore Palmaro, Guillaume Moulis, Fabien Despas, Julie Dupouy, , Maryse
Lapeyre‐Mestre. Overview of drug data within French health insurance
databasesandimplicationsforpharmacoepidemiologicalresearch.Fundamental
andClinicalPharmacology,2016,DOI:10.1111/fcp.12214
68
Consistencyoftheprojectinrelationwiththethesisobjectives
Aprioriknowledgeofthedatasource,includingfactorsaffecting
longitudinaldataavailability,shouldbeanessentialprerequisite
beforeimplementingfurtherinvestigations
Whatisalreadyknownandwhatthisstudyadds
Severalgeneralpapershavedescribed thecontentand interest
ofSNIIRAMdatabasesformedicalresearch
However,importantmethodologicalconsiderationsondrugdata
haveneverbeenpublished
This paper offers a comprehensive description of drug data
contained in the French Health insurance databases, with a
particularfocusongapsindataavailability
Keyresearchquestions
Howaprioriknowledgeofdatasourcescontentandoriginhelp
toreducepotentialbiasindrugexposuremeasurement?
Where are drug data located in the French health insurance
database?
What are themain points to considerwhen investigating drug
usethroughthesedatabases?
How to ensure longitudinal data availability when analysing
healthinsurancedataforresearchpurposes?
69
Résumé7.Mieuxconnaîtrelesbasesdedonnéesdel’assurancemaladiepour
réduire les biais potentiels dans le cadre de lamesure de l'exposition au
médicament
SUMMARYINFRENCH
Les bases de données de l’assurance maladie sont de plus en plus utilisées pour
répondre à des interrogations sur l’utilisation des médicaments ou leur sécurité
d’emploi en vie réelle. Mais que contiennent‐elles vraiment? D’où viennent les
données sur le médicament et comment sont‐elles restituées dans les bases de
données? Quels sont les principaux points de vigilance à respecter lors de leur
exploitation? C’est à ces questions que nous avons tenté de répondre dans cet
articlepubliédansFundamentalandClinicalPharmacology.
Danscetteétude,nousavonscherchéàfournirunaperçuactualisédesdonnéessur
les médicaments contenues dans les bases de données de l'assurancemaladie, le
datamart de consommation inter‐régimes (DCIR), et l’échantillon généraliste des
bénéficiaires (EGB). Cet article identifie les problèmes affectant la disponibilité et
l’exhaustivité des données: (i) les variations du niveau de prise en charge des
médicamentsd'intérêt(perteéligibilitéauremboursement,médicamentnonsoumis
à prescription médicale obligatoire), (ii) perte d’éligibilité des bénéficiaires
(changementderégime,etc.),et(iii)lescontraintestechniquesetrèglementaires.
L'impactdesrupturesdansladisponibilitédesdonnéesvadépendredelaquestion
de recherche, du médicament, du secteur de soin considéré et de la population
d'intérêt.
L’intégration d’une liste des éléments à vérifier et d’un panorama de la mise à
dispositiondesdonnées«médicaments»complètentcetaperçuquisevoulaitune
ressourcepréalableàl’exploitationdecesbasesdedonnées.
70
1. Presentationofthestudy
Several general papers have described the content and interest of SNIIRAM
databasesformedicalresearch57.However,tothebestofourknowledge,thereisno
detaileddescriptionofdrugdataormethodologicalguidanceconcerningstudieson
drug use within French health insurances databases. Indeed, SNIIRAM databases
presentparticularities(indatacollectedandindatabasearchitecture)thatarelikely
to introduce bias in pharmacoepidemiological studies. A deep knowledge of
database content, origin and release is then crucial in order to avoid bias when
designingpharmacoepidemiologicalstudiesusingFrenchdatabases.
2. Objectives
The objective of this paperwas to review sources of information on drug use for
pharmacoepidemiological purposes and particularities of the French health
insurance databases, using technical documentation provided by the database
holder,CNAMTS (“Caissenationalede l’assurancemaladiedes travailleurs salariés”,
FrenchNationalHealthInsuranceFundforEmployees).
3. Methods
Wemadeaninventoryofdrugdataaccordingtohealthcarescheme,period,sector
(public/private)andsettingconsidered,includingadescriptionofpotentialgapsin
dataavailability,andweprovideabriefchecklistforidentifyingtheseissues.
4. Publication
71
72
73
74
75
76
77
78
79
80
5. Discussion
Thispaperoffersastructuredandtransparentoverviewofdrugdatacontainedinthe
SNIIRAM and most common problems that could be encountered in
pharmacoepidemiological studiesusing thesedatabases. SNIIRAMpresents important
strengths:ahigh levelofcoverage,withpotentiallyallFrenchpeoplecovered,andan
universalhealthcareschemeinwhichdrugs,evencostlyandinnovative,areextensively
covered. There is a need to pursuit furthermethodological and validation studies to
promote accurate and transparent use of French health insurance databases for
pharmacoepidemiology.
Asstatedintheintroduction,gapsindrugdataavailabilityhavenotbeendescribedin
detailfortheFrenchdatabasesandthispaperiscontributiveinthisarea.
Some elements of discussion could be added. At the international level, some papers
havedealtwithdatatheissueofgapsindataavailability.InUK,thiswasdiscussedfor
The Health Improvement Network (THIN) primary care database
(http://csdmruk.cegedim.com/). Whereas in the French database gaps must be
consideredatthehealthcareplanlevel,inclinicaldatabasetheproblemisdefinedatthe
practicelevel.Asthegeneralpracticesadoptwereprogressivelycomputerizedandhave
adopted new software systems, there is a possibility for inconstant data quality,
improving over time. Then this study focused on finding the best indicator for
identifyingperiodsofacceptablecomputerusage,whichiscloselyrelatedtoourissueof
longitudinaldataavailability89.This issuehasalsobeeninvestigatedinanotherarticle
onTHINdatabase,butwithafocusondefiningperiodsofcompletemortalityreporting
6.SaskatchewanDrugPlandatabase,theissueofgapsindataavailabilityisfocusedon
oneprecisedataperiod(Validatingamethodthatdealswithmissingdruginformation
intheSaskatchewanDrugPlandatabase).
Afterthisoverview,theperspectivewouldbetointegrateproperlytheseconsiderations
inlongitudinalstudies.
81
82
83
B. Chapter2:Illustratingtheimpactofmethodologicalchoiceonriskestimatesandtheinterestoftime‐dependentexposure:“Benzodiazepinesandriskofdeath:resultsfromtwolargecohortsstudiesinFranceandUK”
Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Benzodiazepines and
riskofdeath:ResultsfromtwolargecohortstudiesinFranceandUK.European
neuropsychopharmacology.07/2015;DOI:10.1016/j.euroneuro.2015.07.006
Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Association between
benzodiazepinesdrugsandtotalmortality:evidencefromastudyonCPRDdata.
IX annual Congress of the French Society of Pharmacology and Therapeutics,
Poitiers,France,April22‐24,2014(poster)
Aurore Palmaro, Julie Dupouy, Maryse Lapeyre‐Mestre. Benzodiazépines et
mortalité : étude de cohorte au RU et en France. VIII congress of Young
Researchers in General Practice (Devenir Jeune Chercheur en Médecine
Générale),Toulouse,France,March14‐15,2014(oralcommunication)
84
Consistencyoftheprojectinrelationwiththethesisobjectives
Drugexposuremodellingcouldnotbeconsideredapartfromthe
outcome
Here the case study of benzodiazepine mortality enabled to
comparebetween‐groupcomparisonswithmethodsintegrating
thetime‐dependentnatureofexposure
Whatisalreadyknownandwhatthisstudyadds
Previous studies on benzodiazepines‐related mortality have
shownconflictingresults
Differentmethods(everuse,cumulativeuse,doseresponse)and
design(control,casecontrol)havebeenused
Keyresearchquestions
How do traditional between‐group comparisons compare with
methods integrating the time‐dependent nature of drug
exposure?
How to manage a common research protocol in case of
heterogeneousdatasources?
85
Résumé8. Illustrer l’impactde laméthodedemesurede l’expositionsur les
estimateurs de risque: application à l’étude de l’association entre
benzodiazépinesetmortalité
SUMMARYINFRENCH
Danslecadred’unappelàprojetrestreintdel’AgenceEuropéenneduMédicament
(EMA), nous avons mené une étude sur la question de la mortalité liée aux
benzodiazépines. Alors qu’une étude parue dans le BMJ Open en Janvier 2012 90
montraitunemortalité toutescauses3à5 foisplusélevéechez lesutilisateursde
benzodiazépines hypnotiques, l’état de la littérature ne permettait pas de fournir
desélémentsprobants,enraisonenparticulierdelaconfusionrésiduelle.
Méthodes
Nous avons donc mené une étude de cohorte rétrospective de type exposés/non
exposés à partir de 2 bases de données, en France avec l’EGB et auRoyaume‐Uni
avec le CPRD (Clinical Practice Research Datalink). Nous avons reconstitué une
cohortedenouveauxutilisateurs(«new‐userdesign»),etconsidérél’expositionaux
benzodiazépinesdefaçondépendantedutemps.
Lesutilisateursincidentsdebenzodiazépinesontétécomparésà2groupestémoins,
un groupe composé de nouveaux utilisateurs d’antidépresseurs et un groupe de
nouveauxconsommateursdesoins(consultationgénéraliste).Lespatientsexposés
ontétéappariésauxtémoinssurl’annéedenaissance(±5ans),legenre,ainsiquele
cabinet médical de rattachement). Les patients âgés de plus de 18 ans dont les
cabinetsderattachementavaientconsentiàparticiperauchaînageavecleregistre
de mortalité tenu par l’ONS (Office of National Statistics), étaient éligibles. La
relationentreexpositionauxbenzodiazépinesetmortalitéàunanaétéétudiéeà
l’aided’unmodèledeCoxstratifié,avecvariablesdépendantesdutemps.
Résultats
Al’issuedelasélection,lapopulationfinalecomprenait94123patientspargroupe
pour le CPRD, (57 287 pour l’EGB). La population comprenait une majorité de
femmes,avecunemoyenne(ET)d’âgede58ans(18.6).
86
Au sein du CPRD, la mortalité toutes causes à un an chez les exposés était
significativement plus élevée dans le groupe exposé [HR: 5.67, 95% CI 5,60‐6,69,
p<0,0001] et dans le groupe des utilisateurs d'antidépresseurs/anxiolytiques non
benzodiazépiniques(HRbrut,2,00;IC95%,1,86‐2,16),parrapportauxtémoins.En
présence d’une exposition aux benzodiazépines prise comme variable dépendante
dutemps,lamortalitétoutescausesàunanétaitsignificativementaugmentée[HR
brut: 4,77, 95% CI 3,93 – 5,80, p<0,0001]. Cette association persistait après
ajustementsurlesfacteurscliniques,liésaumodevieetsocioéconomique.
Dans l’EGB, les utilisateurs de benzodiazépine présentaient également un risque
plusélevédemortalitétoutescausesà12mois(HR1,99;1,74à2,29),demêmeque
les utilisateurs d'antidépresseurs/anxiolytiques non benzodiazépiniques (1,53; IC
95%, 1,32‐1,77) , en comparaison aux témoins. Après ajustement, leHazardRatio
étaitde1,26chezlesutilisateursdebenzodiazépines(IC95%,1,08‐1,48),etde1,07
(ICà95%,0,91‐1,27)chezlesutilisateursd’anxiolytiquesnonbenzodiazépiniques.
Conclusions
Grâce à un ajustement additionnel intégrant des variables traditionnellement non
prisesencompte(tabac,alcool,IMC)etàunemesureplusadéquatedel’exposition
médicamenteuse, nous avons ainsi pu montrer que la forte association observée
entre benzodiazépines et mortalité, déjà décrite dans la littérature, était ici
fortementatténuéeetinterrogeaitsurlapossibilitéd’uneconfusionrésiduelle.
Perspectives
Al’issuedecetravail,laquestiondespériodesd’expositioninobservablesetdeleur
poids sur les estimations obtenues s’est posée. En effet 12% des nouveaux
utilisateursdebenzodiazépinesdans l’EGBont effectuéun séjourhospitalierdans
l’année suivant l’initiation, et les approches traditionnelles ne prennent
malheureusement pas en compte ces périodes inobservables dans l’estimation du
risque. Cette étude a pu bénéficier d’un développement supplémentaire dans le
cadredeceprojetderechercheciblé.
87
1. Presentationofthestudy
Theworkonbenzodiazepinemortalitywasrealizedintheframeworkofarestricted
tender from the European Medicines Agency (No. EMA/2012/20/PV, Association
between anxiolytic or hypnotic drugs and total mortality). The article was a
deliverableofthetender.Thisstudyaimedtoinvestigatemortalityassociatedwith
anxiolyticorhypnoticdrugs in twomemberstates in theEuropeanUnion.To this
purpose, cohortstudieswereconductedon two largepopulation‐baseddatabases:
the Clinical Practice Research Datalink (CPRD) in the UK and the Echantillon
GénéralistedeBénéficiaires (EGB)database (arepresentativesampleof theFrench
beneficiariesofthenationalhealthinsurancescheme).
“A signal from the published literature has recently been reviewed by the Agency
relatingtoamatchedcohortstudythatfoundelevatedHazardsRatios(HRs)fordeath
in patients who received hypnotic prescriptions compared to those not prescribed
hypnotics.TheAgencyconsideredthattheresultsofthestudyshouldbetreatedwith
cautiondue tomethodological limitationsof theanalysisconducted, inparticular in
terms of controlling for known potential confounders. The Agency, however, also
considers that further focussedresearchon theassociationbetweenuseofhypnotics
andmortalityisrequiredduetoapotentialforsignificantpublichealthimpactgiven
thewidespreaduse of hypnotics.Technical specifications for restricted invitation to
tender
3. Subject of the tender: The Agency considers that it requires an in‐depth and
comprehensive study of mortality associated with hypnotics/anxiolytics. It is
anticipated that the research undertaken will further explore what risks can be
attributed to complex confounding by life‐style, psychological and socio‐economic
factorsandhistoryofpsychiatricdisordersandothercomorbidities.Theresultsofthe
research should inform on the need for regulatory action and risk management
planning.”
88
Studyapprovalandethicalaspectsa)
The draft and final protocol were approved by the study sponsor, the EMA. The
datasets were obtained from the CPRD after approval of ISAC, the Independent
ScientificAdvisoryCommitteeforMHRAdatabaseresearch.Thestudyalsoobtained
a seal from the European Network of Centres for Pharmacoepidemiology and
Pharmacovigilance(EnceppstudySeal).
2. Objectives
This study aimed to explore the impact of benzodiazepine use on short‐term (1
year) mortality. For this purpose, we conducted cohort studies using two large
population‐baseddatabasesfrom2countrieswithhighlevelofbenzodiazepineuse:
theClinicalPracticeResearchDatalink(CPRD)intheUKandtheGeneralSampleof
Beneficiaries (Echantillon Généraliste des Bénéficiaires, EGB) database (a
representative sample of French beneficiaries of the national health insurance
scheme).
3. Methods
Exposed‐unexposed cohorts were constructed with the Clinical Practice Research
Datalink (CPRD) in the UK and with the EGB in France. Benzodiazepine incident
users were matched to incident users of antidepressants/non‐benzodiazepine
sedatives and to controls (non‐users of antidepressants or anxiolytics/hypnotics)
according to age, gender and practice for the CPRD). Survival at one year was
studiedusingCoxregressionmodel.Thefirstanalysiswasbasedonanintention‐to‐
treatcomparisonbetweencohorts,withcontrolsasthereferencegroup.Treatment
episodes were derived to build time‐dependent covariate for benzodiazepine
exposure. The effect of benzodiazepine use as a time‐varying variable was then
examined separately offering the opportunity to compare between‐group
comparisonswithmethodsintegratingthetimedependentnatureofexposure.
4. Publication
89
90
91
92
93
94
95
96
97
98
99
100
101
5. Discussion
Theworkentitled“Benzodiazepinesandriskofdeath:resultsfromtwolargecohorts
studies inFranceandUK” provides an illustrationof the impact ofmethodological
choiceonriskestimatesanddemonstratestheinterestoftime‐dependentexposure
compared to traditional between‐group comparison. It illustrates that different
methods forhandlingdrugexposureare likely toproducedifferentriskestimates,
and that traditionalbetween‐groupcomparisonsshouldbecompletedbymethods
integratingthetime‐dependentnatureofdrugexposure.
Justificationofthechoiceofcontrolgroupsa)
Incontrasttomanyothercohortstudies,thestudyonbenzodiazepinesintegrateda
secondcontrolgroup inadditiontonon‐users.Withaviewtominimize indication
bias,non‐usersmightnotalwaysbe thebest controls.Thebest controlswouldbe
patients who have a similar baseline risk, and who would be likely to receive
benzodiazepines,butfinallydidnotreceiveit.Finally,usersofnon‐benzodiazepine
anxiolytics or antidepressants were selected. Antidepressants have different
indications,butinpractice,areoftenco‐prescribedwithbenzodiazepines,andeven
frequently initiated on the same day. In final, users of antidepressants were
hypothesizedtosharesimilarcharacteristicswithbenzodiazepineusers.
Drugexposuremeasurementinotherstudiesb)
Inadditiontothedetailsincludedinthepaper,methodologicalelementsfromother
studiesinvestigatingthesamequestioncouldprovideusefulinsightinrelationwith
the issue of drug exposuremeasurement. Indeed, for the same research question,
distinctmethodshavebeenimplemented,asillustratedTable3(page103).Inmost
of the studies, exposure to benzodiazepines was considered as a fixed variable.
Duration of use was rarely taken into account 91–93. A little number of studies
accounted for accumulated doses 90,94 or considered benzodiazepines as a time‐
varyingvariable95,96.Stratificationaccordingtohalf‐lifeeliminationwasretrievedin
twopreviousstudy97,98.
102
The example of benzodiazepine and related injuries have provided a series à
literature concerning drug modelling, which have been of great interest in the
contextofthestudyofbenzodiazepinesandmortality63,99,100.
Perspective on the multinational study onc)
benzodiazepinesandmortality
This study is an example of multinational study using a common protocol and
methodology. This project raised additional questions on the impact of
methodologicalchoiceonriskestimates.Moreprecisely,areaofdiscussionswerein
relationwith the followingobservations: (i) inspiteof theapplicationof thesame
inclusion criteria, thepopulations selectedareverydifferent (ii)despite the same
drugsofinterests,populationexposed,typeofdrugusedandpatternsofusevaries
extensively,and(iii)inspiteofthesameeventofinterest,withminorpotentialfor
misclassification,magnitudeofriskestimatesrangedfrom1to3.Considerationsfor
implementingmultinationalstudyandapplyingacommonprotocolontwodifferent
databasesaresummarizedinTable4(page107).Theseelementswerealsouseful
forexploringheterogeneityanttoexplainthevariationbetweencountryestimates
forthesamemethodology.
103
Table 3.Methods for drug exposuremeasurement in studies investigating the
associationbetweenbenzodiazepinesuseandmortality
Author(year) Population Exposureassessment Results
Huybrechts2011101
Exhaustivepopulationof>65yoBritishColumbiaresidentsinnursinghomes,initiatingapsychotropic(BZD,AD,conventionaloratypicalAP)Lengthoffollowup:6‐monthfollow‐upSamplesize:10900Numberandproportionexposed:4887incidentBZDusers/10900(44.8%)
Drugdatasource:PharmaNetdatabaseDrugsconsidered:AnxiolyticBZD(alprazolam,chlordiazepoxide,clonazepam,clorazepatedipotassium,lorazepam)andotherhypnoticagentsagents(diazepam,estazolam,flurazepam,oxazepam,temazepam,triazolam,zolpidem,zaleplon,diphenhydramine,lutethimide)Typeofanalysis:SurvivalanalysisDrugexposuremodelling:exposureconsideredastimedependant)
Mainoutcome:NoncancermortalityHighdimensionalPropensityscoreanalysis:aHR=1.20[0.96‐1.50]ascomparedtoatypicalAPnewusers
Kripke201290
PopulationservedbythePennsylvaniaGeisingerHealthSystemLengthoffollow‐up:Mean2.5yearfollow‐upSamplesize:34205Numberandproportionexposed:10531exposedtoanyhypnoticamong34205(44.5%)4338exposedtozolpidem,and2076totemazepam
Drugdatasource:ElectronichealthrecordDrugsconsidered:Hypnotics(BZDwerearound>90%)Typeofanalysis:Survivalanalysis,(exposureconsideredastimefixed)Drugexposuremodelling:Exposureexpressedasdosesperyear(0.4‐18pills,18‐132pills,>132pills)zolpidem:none/5‐130mgperyear/130‐800mgperyear/>800mgperyear
Mainoutcome:All‐causemortalityAnyhypnoticuse:aHR=4.56[3.95‐5.26]Accordingtothelevelofexposure:0.4‐18pills/yruse:aHR=3.60[2.92‐4.44]18‐132pills/yruse:aHR=4.43[3.67‐5.36]>132pills/yruse:aHR=5.32[4.50‐6.30]
104
Author(year) Population Exposureassessment Results
temazepam:none/1‐240mgperyear/240‐1640mgperyear/>1640mgperyear
Obiora201392
Community‐acquiredpneumoniapatientsTHIN(TheHealthImprovementNetwork)databaseSamplesize:4964Lengthoffollow‐up:2,8‐yearmeanfollow‐upNumberandproportionexposed:1269exposed/4961(25.6%)
Drugexposuremodelling:Baselineuse(i)Evervsneveruse(ii)dividedin:current(<30days),recent([31‐90days]),past(>90days)useaccordingtothelengthbetweenthelastprescriptionandpneumoniaindexdate(vsneverusers)Chronicuse(prescriptionsbothinthe30‐and90‐dayperiodsbeforethepneumoniaindexdate)
Mainoutcome:All‐causemortality2mortalityendpoints:30days(n=947),andduringthewholeperiod:(n=1547)Concerningoverallmortality:(i)HRa(ever/neverBZDuse):1.32[1.19‐1.47]Significantresultsalsoobservedforeachindividualagent(diazepam,chlordiazepoxide,lorazepam,temazepam)exceptzopiclone(ii)CurrentuseaHR=1.42[1.21‐1.67]RecentuseaHR=1.49[1.19‐1.85]PastuseaHR=1.24[1.09‐1.41]Chronicuse:aHR=1.37[1.20‐1.56]PropensityscoreadjustmentincreasedthemagnitudeoftheHRs(aHRever/neverBZD=1.49[1.30‐1.71])
Tiihonen201293
ExhaustiveincidentFinnishsubjectsdiagnosedforschizophreniaLengthoffollow‐up:4,2‐yearmeanfollow‐upSamplesize:2588Numberand
Drugdatasource:PrescriptiondatabaseoftheSocialInsuranceInstituteDrugsconsidered:BZDandBZDrelateddrugsTypeofanalysis:SurvivalanalysisDrugexposuremodelling:
Mainoutcome:all‐causemortalityAll‐causemortalityHRa(currentuse):1.91[1.13‐3.22]aHR(pastuse):0.99[0.97‐1.01]80%ofdeathsamongstBZDuserswereduringperiodswith>28DDD
105
Author(year) Population Exposureassessment Results
proportionexposed:904exposed/2588(34.9%)
exposureconsideredastimedependant)Currentandpastuse
(violationoftreatmentguidelines)
Vinkers200397
Leidenelderly(>85yo)residents(Netherlands)Samplesize:599Numberandproportionexposed:181exposedatbaseline(30%)Lengthoffollow‐up:3to5‐yearfollow‐up
Drugdatasource:ComputerizedPharmacyRegistrieswithatimeframeof3monthsBZD(diazepamequivalent)Typeofanalysis:SurvivalanalysisDrugexposuremodelling:exposureconsideredastimedependantBZDusedefinedbyaprescriptionduration>50%ofthe3‐monthtimeframeBZDuseaccordingtoshortorlong(diazepam,chlordiazepoxide,flunitrazepam,flurazepam,nitrazepam)halflife
Mainoutcome:All‐causemortalityaRR(anyBZD)=0.68[0.44‐1.04]
Winkelmayer200798
incidenthemodialysispatientsrandomsampleoftheUSRenalDataSystemSamplesize:3630Numberandproportionexposed:490exposed(13.5%)Lengthoffollow‐up:3to4‐yearfollow‐up
Drugdatasource:Medicalcharts(DialysisMorbidityandMortalitystudy)BZD(anxiolyticandhypnotic)andzolpidemTypeofanalysis:SurvivalanalysisDrugexposuremodelling:exposureconsideredastimefixedBaselineuseaccordingto(i)ever/neverusers(ii)thenumberofBZDused(iii)longacting(chlordiazepoxide,
Mainoutcome:All‐causemortalityaHR=1.15[1.02‐1.31]IncreasedriskwithshortactingBZD(aHR=1.17[1.02‐1.35])butnotlongacting(aHR=1.11[0.88‐1.39])vsnouse
106
Author(year) Population Exposureassessment Results
clonazepam,flurazepam,diazepam)vsshortacting(alprazolam,clorazepam,estazolam,lorazepam,oxazepam,temazepam,triazolam,zolpidem)
Baandrup201096
18‐53yoschizophrenic/othernon‐affectivepsychoticpatientsDanishNationalPatientRegisterSamplesize:2131Numberandproportionexposed:47.2%exposedtolonghalf‐lifeBZD(cases)vs479/1937i.e.24.7%amongcontrols
Drugdatasource:DanishRegisterofMedicinalProductdrugsDrugsconsidered:BZDderivativesandrelated(ATCclass)dividedbyeliminationhalf‐life:long(>24h),Intermediate(6‐24h)andshort(<6h)Typeofanalysis:MultivariateconditionallogisticregressionDrugexposuremodelling:Currentuse:>1prescriptionfilledwithin90daysbeforedeathorindexdate
Mainoutcome:AllnaturalmortalityCurrentBZDusewasassociatedwithincreasingmortality(datanotshown)BZDwithlongeliminationhalf‐life:aOR=1.78[1.25‐2.52]BZDwithintermediatehalf‐life:aOR=0.75[0.49‐1.15]BZDwithshorthalf‐life:aOR=1.16[0.77‐1.76]
Abbreviations : aOR :ajusted Odd Ratio, AP: antipsychotic, AD : antidepressant, AE : antiepileptic, BZD :
benzodiazepine, COPD : Chronic Obstructive Pulmonary Disease, cOR : crude Odd Ratio, ICD : international
Classification of Diseases, FGA : first generation antipsychotic, MTD : methadone, NSAIDs : Non‐steroidal Anti
InflammatoryDrugs,OTC:overthecounter,PY:personyear,SGA:secondgenerationantipsychotic,yo:yearold.
107
Table4.Factorsconsideredforimplementingthemultinationalstudy
Factorstoconsider
Whatarethevariablesavailable?
Restrictions?
Typeofbeneficiariesincluded
Population‐basedorsample
Objectiveofthedatacollection(administrative,research?)
Methodsfordatacollection?Automated,Mandatory?incitations?
Any significant evolution of the methods of data collection over the period of
observation?
Secondaryprocessingofdata
Conditionsforaccess
Drugsofinterest
Marketingauthorization
Availabilityinthemarket
Legalstatus,restrictionappliedtoprescribers
Coveredbythehealthinsurancesystem
Nationalrecommendations
Anyphenomenonofdiversiondescribed
Completenessofdatacollection
Levelofdetail
Classificationsystem
Quantification:ATC/DDDmethodology
108
109
C. Chapter3:Estimatingtheimpactofimmeasurableexposureperiodsduetohospitalizationsonriskestimatesinmedico‐administrativedatabases
Aurore Palmaro, Quentin Boucherie, Julie Dupouy, Joëlle Micallef, Maryse Lapeyre‐
Mestre. Unobservable drug exposure due to hospitalization in medico‐administrative
databases :whichimpact forPharmacoepidemiologystudies?(Pharmacoepidemiology
andDrugSafety,underreview)
Quentin Boucherie, Julie Dupouy, Joëlle Micallef, Maryse Lapeyre‐Mestre, Aurore
Palmaro.Unobservabledrugexposureduetohospitalization inmedico‐administrative
databases:whichimpactforPharmacoepidemiologystudies?XIannualCongressofthe
French Society of Pharmacology and Therapeutics, Nancy, France, April 19‐21, 2016
(oralcommunication)
Aurore Palmaro, Quentin Boucherie, Julie Dupouy, Joëlle Micallef, Maryse Lapeyre‐
Mestre.Périodesd’expositioninobservablesaucoursdesséjourshospitaliersenPMSI
MCO:quelimpactpourlesétudespharmacoépidémiologiques?ADELF(Associationdes
EpidémiologistesdeLangueFrançaise)‐EMOIS(Evaluation,Management,Organisation,
Information,Santé)meeting,Dijon,France,March10‐11,2016,(poster)
110
Consistencyoftheprojectinrelationwiththethesisobjectives
Facilitating the exploration of longitudinal data in complex
contexts
Whatisalreadyknownonthistopicandwhatthisstudyadds
In the study presented at chapter 2, 12% of benzodiazepines
users(EGB)hadatleastonehospitalizationintheyearfollowing
initiation.
Conventional approach does not account for the potential for
immeasurabletimebiasduringthesehospitalizations
Keyresearchquestions
Whatistheimpactofunobservabletimebiasonriskestimates?
Howtodealwithgapsinlongitudinaldataavailability?
How to improve the integration of immeasurable time bias in
furtherstudies?
111
Résumé 9. Evaluer l’impact des périodes inobservables lors des
hospitalisationspourlesétudespharmacoépidémiologiques
SUMMARYINFRENCH
Introduction:
L’essentiel des médicaments courants (non coûteux/innovants) est compris dans la
dotation globale des établissements de soins et n’est donc pas décompté
individuellement.Danslecadred’étudesmenéessurlesbasesSNIIRAM/PMSI,lestatut
d’unpatientparrapportà l’expositionnepeutdoncêtrepasêtreconnuaucoursdes
séjours hospitaliers. Or, ces périodes dites inobservables ne sont actuellement pas
prisesencomptedanslesétudespharmacoépidémiologiques.L’identificationetlaprise
en compte des périodes d’exposition inobservables sont nécessaires en
pharmacoépidémiologie autant pour des études portant sur l’estimation d’un risque
associéàunmédicamentquepourdesétudes sur l’observancemédicamenteuse.Une
réflexionméthodologiqueconcernantlapriseencomptedecespériodesinobservables
a donc été menée. Nous avons répondu à un appel d'offres de l'ANSM et obtenu un
financement de 18 mois pour mener des travaux méthodologiques pour permettre
d’améliorer la mesure de l’exposition médicamenteuse dans les études
pharmacoépidémiologiques issues des bases médico‐administratives. Le travail
présentéiciconstitueundeslivrablesduprojet.Ilutiliselesdonnéesdel’EGBissuesde
l’étudedel’associationentreexpositionauxbenzodiazépinesetmortalité(projet2).
L’objectif était de modéliser les périodes d’exposition inobservables et d’étudier
l’impactdeleurpriseencomptesurlesestimationsderisqueobtenues.
Méthodes:
Unecohortedetypeexposés/nonexposésaétémiseenplaceàpartirdesdonnéesde
l’EGBsur lapériode2006‐2012.Lesutilisateurs incidentsdebenzodiazépinesontété
comparés à 2 groupes témoins, un groupe composé de nouveaux utilisateurs
d’antidépresseursetungroupedenouveauxconsommateursdesoins.Larelationentre
112
expositionauxbenzodiazépinesetmortalitéaétéétudiéeà l’aided’unmodèledeCox,
avec variables dépendantes du temps (exposé/non exposé), et selon différentes
hypothèsesconcernantlestatutparrapportàl’expositionaucoursdeshospitalisations.
Lespériodesinobservablesétaientdéfiniesparlesdatesd’entréeetdesortieduPMSI
MCO(médecine,chirurgie,obstétriqueetodontologie).
Modèlesmulti‐états(oumodèlesMarkoviens)
Lesmodèlesmulti‐étatssontdeplusenplusutilisésenpharmaco‐économieouencore
en épidémiologie pour modéliser la survenue d’un évènement tout en prenant en
compte les différentes trajectoires en lien avec l’évènement. En pharmaco‐
épidémiologie,ilssontpourlemomentpeuutilisésalorsqu’ilspourraientpermettrede
prendreencompte lesdynamiquescomplexes liéesàuneexpositionmédicamenteuse
(par exemple) car constituée de nombreux états tels que l’arrêt d’un traitement, la
reprise,leswitch,lechangementdeposologie(Kildemoes2010,Leufkens2002).Dans
le cadre de cette étude, un modèle de Markov à 3 états
(observable/inobservable/décès)aétédéfini.
Résultats:
Au total, 171 861 patients ont été inclus (57 287 par groupe). En présence d’une
exposition aux benzodiazépines prise comme variable dépendante du temps, la
mortalité toutes causes à un an était significativement augmentée [HazardRatio brut
1,28(IC95%1,02‐1,60)].Enprenantencomptelespériodesinobservables,l’exposition
aux benzodiazépines n’était plus significativement associée [HR 0,99 (0,77‐1,18)]. La
modélisationmulti‐étataboutissaitàdesrésultatscohérents.
Conclusion:
Enmontrantl’impactdelapriseencomptedespériodesd’expositioninobservablessur
l’estimationd’unrisque,cetteétudesouligne lanécessitéd’identifieretdeprendreen
comptecespériodesinobservables,autantpourdesétudesportantsurl’estimationd’un
risqueassociéàunmédicamentquepourdécrirel’expositionmédicamenteuseaucours
dutemps.
113
1. Presentationofthestudy
In2014,theFrenchMedicinesAgencyhaslaunchedacall forspecificprojects(ANSM,
targeted call for research applied tohealthcareproducts, 2014).Oneof theproposed
areaconcernedunobservabletimebias.(“Commentconsidérerles«périodesàtrous»
dansunsuivilongitudinald’unepriseenchargemédicamenteuse”).Theprojectshould
investigate how to consider unobservable periods in the context of longitudinal drug
exposureassessment.
2. Objectives
Theaimofthisstudywastomodelunobservableperiodsduetohospitalizationandto
applyseveralmethodsforaddressingthisbiasandassessitsimpactonriskestimates.
Thisapproachwasappliedtothestudyoftheassociationbetweenbenzodiazepinesand
mortalityon thebasisof theusing theGeneralSampleofBeneficiaries(EGB)data for
presentedinchapter2102.
3. Methods
All‐cause mortality at one year (Cox regression model) was studied using time‐
dependent variables (exposed/unexposed or under two hypotheses, inpatients are
exposed or inpatients are unexposed), completed with a multistate model based on
observable/unobservable/deathstatus.
4. Publication
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
5. Discussion
Throughourcasestudy,weillustratehowassumptionsconcerninginpatientsexposure
toaccountforperiodsofimmeasurabletimecanimpactriskestimateinacohortstudy.
Asourstrategywasbasedoncasestudy,itraisesawarenessonthepossibleimpactof
unobservable time bias but does not provide a general answer to this issue, and the
impactinothercontextsispronetovary.
This contribution did not assess all possible combinations, but provided what could
assimilatedtoan“universeofpossibleestimates”asdefinedbyMadiganelal. 29.This
approachsupportsthat“account[ing]foruncertaintyduetoanalyticdesignchoiceneed
tobecomepartofstandardpractice”.
Morecomplexmodelswouldalsobeofinterest.Inourstudyanunobservablestatushas
beendefined.Withinthisstate,thepatientsiseitherexposedornot,buttherealstatus
isunknown.InthisparticularcontexthiddenMarkovmodelcouldbeofinterest.Inhis
study on «Estimation of Drug Effectiveness by Modeling Three Time‐dependent
Covariates: An Application to Data on cardioprotective medications in the chronic
dialysis population” Phasnis provide interesting insights concerning further
possibilities 103. InterestofPhadnis approach is twofold (i) thehiddenMarkovmodel
(Start drug ‐stop drug ) would account for changes during unobservable time (ii)
simulateHRvaluesagainstvariationinexposuredefinition.
142
143
D. Chapter4:Dealingwithlongitudinaldataandmultipleconcomitantexposuresinspecificcontexts:“IdentifyingcancertreatmentregimensinFrenchhealthinsurancedatabases:anapplicationinmultiplemyelomapatients”
Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐
Mestre. Identifying cancer treatment regimens in French health
insurance databases: an application in multiple myeloma patients
(PharmacoepidemiologyandDrugSafety,Underreview)
Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐
Mestre. Identifying cancer treatment regimens in French health
insurance databases: an application in multiple myeloma patients. XI
Congress of the French Society of Pharmacology and Therapeutics,
Nancy,France,April19‐21,2016(oralcommunication)
Aurore Palmaro, Martin Gauthier, Fabien Despas, Maryse Lapeyre‐
Mestre. Reconstituer les lignes de traitement reçues en onco‐
hématologieàpartirdesdonnéesduSNIIRAMetduPMSI:applicationà
l’étudedescyclesdechimiothérapiedanslemyélomemultiple.ADELF–
EMOIS meeting, Dijon, France, March 10‐11, 2016 (oral
communication)
144
Consistencyoftheprojectinrelationwiththethesisobjectives
Addingrelevancetodrugexposuremeasurement
Facilitating the exploration of longitudinal data in complex
contextsthroughthedevelopmentofnewtools
Whatisalreadyknownandwhatthisstudyadds
Studiesoncancerdrugsarescarceandweremainlyfocusedona
single drug or on aggregated patients' trajectories (surgery,
chemotherapy,radiotherapy,etc.)
Identificationof thedrugsentering in treatmentregimens isas
an essential prerequisite for further safety and effectiveness
studies
Previousattemptsforidentifyingchemotherapyregimensonthe
basisofclaimsdatabaseswerescarce.
Thisstudyprovidesanalgorithmforidentifyingthenatureand
sequence of drug regimens using data from the French health
insurancedatabases.
KeyResearchquestions
How to identify complex drugs regimens in oncology on the
basisofdispensingdata?
How to account for the potential of immeasurable timebias in
theidentificationprocess?
145
Résumé 10. Mieux appréhender des situations impliquant des données
longitudinales et complexes: «Reconstituer les lignes de traitement reçues en
onco‐hématologieàpartirdesdonnéesduDCIRetduPMSI :applicationà l’étude
descyclesdechimiothérapiedanslemyélomemultiple»
SUMMARYINFRENCH
Dans cette deuxième partie, les investigations sont menées dans un champ en
développement en pharmacoépidémiologie, celui de l’onco‐hématologie. Le premier
travail faitéchoàceluisurl’aperçudeladisponibilitédesdonnéesetsur lespériodes
inobservables,puisquel’étudedeschimiothérapiesalongtempsétéconsidéréecomme
non réalisable en raison du caractère inobservable de la plupart des médicaments
anticancéreux,àl’exceptiondesmédicamentsdits«innovants».
Reconstitutiondesparcoursdesoins:unprérequisessentiel
Un des obstacles au développement d’étude sur les bases de données de l’assurance
maladie est la complexité des données, tant dans leur structure que dans leurs
modalitésdecollecteetderestitution.Latranspositiondecesdonnéesadministratives
en des entités pertinentes cliniquement représente un enjeu important. A notre
connaissance,aucuneétudeneproposaitdeméthodologiedereconstitutiondescycles
de chimiothérapie sur les bases de données françaises de l’assurance maladie. Alors
qu’elles sont de plus en plus utilisées avec succès dans d’autres contextes, leur
utilisation en cancérologie est restée limitée. Une étude récente sur des données
comparablesauniveaurégionalaproposéuneméthodologiepermettantdedériverdes
cycles exclusifs de séquences de soins (chimiothérapie, radiothérapie, etc.) 73.
Cependant,unemesureplusfinedel’exposition(auniveaudumédicamentmême)est
requisepourl’évaluationdestratégiesmédicamenteuses.Ladescriptiondesstratégies
de traitement et l’évaluationde leur bénéfice‐risque imposent en effet de prendre en
comptelescombinaisonsreçuesainsiqueleurséquenced’administration.Al’imagedes
études de validation effectuées dans des bases médico‐administratives SEER
(Surveillance,Epidemiology,andEndResultsProgram)104,105,ons’intéresseradoncàla
146
reconstruction des lignes de traitements successifs dans le cas d’ association de
médicaments à visée anticancéreuse. On s’attachera à prendre en compte le fait que
certainsmédicaments vont être partiellement ou totalement inobservables. Ce travail
propose ainsi un algorithme de reconstitution des lignes de traitement successives à
partir des bases de données de l’assurancemaladie. L’objectif de cette étude était de
définirunalgorithmepermettantdereconstituerdeslignesdetraitementsuccessivesà
partir des données SNIIRAM et PMSI MCO. Lemodèle d’étude est celui du myélome
multiple.
Méthodes
Une cohorte de patients atteints demyélomemultiple et initiant un traitement a été
constituéeàpartirdesdonnéesSNIIRAMMidi‐Pyrénéespourlapériode2011‐2014.Les
patientsontétéidentifiésgrâceauxcodesmyélomemultiple(CIM10«C90»)desALD
ou diagnostics principaux des séjours PMSI MCO. Les médicaments considérés
comprenaient le bortezomib, les imids (thalidomide, lénalidomide), les agents
anticancéreux (cyclophosphamide,melphalan, bendamustine,doxorubicine, étoposide,
carmustine),ainsiquelescorticoïdes(prednisoneetdexaméthasone),identifiésàl’aide
desdonnéesduDCIR,desdonnéesde rétrocession et desmédicaments en sus (PMSI
MCO). Un algorithme a été appliqué afin de définir les combinaisons de traitement
reçuesaucours6premiersmoisde traitement (nombredecycleset changementsde
lignes). Les cycles faisant intervenir desmédicaments hors liste en sus (cisplatine et
vincristine) ont été identifiés à partir de la combinaison de médicaments traceurs
observables (ambulatoires, rétrocession ou hors GHS) selon une table de
correspondanceétablieaveclesthesaurusrégionauxdechimiothérapie.
Résultats
Parmi les 236 patients inclus, 48% ont reçu au cours de leur première ligne de
traitement l’association bortezomib‐melphalan‐prednisone (VMP) (n=112), 18%
bortezomib‐thalidomide‐dexaméthasone (VTD ou VTD‐PACE) (n=43), et 18%
melphalan‐prednisone‐thalidomide (MPT) (n= 43). Les autres lignes consistaient en
147
l’association melphalan‐prednisone (MP) (12%, n=29), lénalidomide‐dexaméthasone
(RD) (3%, n=8) et bortezomib‐bendamustine‐dexaméthasone (VBD) (0,4%, n=1). La
naturedescyclesetleurattributionparclassed’âge(+/‐65ans)étaientenaccordavec
lesrecommandationsdepriseencharge.
Conclusion
Cetteétudepermetdedémontrerlafaisabilitédereconstituerdescyclescomplexesde
traitementenhématologieàpartirdesdonnéesduSNIIRAM.
148
1. Presentationofthestudy
Data from the French health insurance system are a very valuable data source for
Pharmacoepidemiology and enables to describe real‐life treatment patterns at the
nationwide level. However, studies on cancer drugs are still scarce, andweremainly
focusedonasingledrugoronthedescriptionpatients'trajectoriesthroughsequences
of treatment (surgery, chemotherapy, radiotherapy, etc.), without determining the
natureandhistoryof treatment linesreceived73. Indeed,cancerpatientsaregenerally
exposedtoseveraltreatmentlines,composedofoneormoredrugs.Itisthenessential
to take account of the particular characteristics of drug exposure in oncology, and to
move from a ‘single drug’ approach toward a ‘multidrug,multiline’ perspectivewhen
modellingdrugexposure.Thecomplexityof treatmentpatterns for cancer isgrowing
and the number of possible regimens increases accordingly. In the context of
observationalstudy,ithasbecomemoreandmoredifficulttotakeaccountofpastlines
and duration of previous lineswhen comparingmultidrug regimens. Identification of
treatment lines should be considered as an essential prerequisite to enable further
safetyandeffectivenessstudiesbasedonclaimsdatabases.However,thisidentification
is not straightforward in thesedatabases and there is a need todevelopmethods for
identifyingmulti‐drugchemotherapyregimens.
1. Objectives
Thiswork aimed to develop an algorithm for identifying the nature and sequence of
drug regimens in multiple myeloma using regional data from the French health
insurancedatabase.
2. Methods
Throughthiscasestudy,itwasintendedtodevelopastandardapproachforidentifying
multidrugchemotherapyinFrenchhealthcaredatabases.
3. Publication
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
4. Discussion
To the best of our knowledge, this is the first study proposing an algorithm for
identifyingmultidrugregimensusingtheFrenchhealthinsurancedatabases.Identifying
multidrug lines is not simply a data management problem, but raises a lot of
methodological and clinical issues. Previous publications in this areawere extremely
scarce106,107. Thiswork integrates the findings of previous attempts of chemotherapy
identification, and further develops the approach by integrating the possibility of
unobservabledrugs.
Interestof furtherexternalvalidationagainstmedicala)
charts
Themainlimitationofthisstudyisthatidentifiedtreatmentregimensarenotvalidated
against medical charts. Two potential sources for regimen ascertainment have been
identified: multidisciplinary staff meetings (MSMs) and databases from pharmacy
hospital. MSMs are organized on a regional basis, which specific meetings for each
malignancy.For theMidi‐Pyrenéesarea, casesarenot computerized followingaMSM
formultiplemyeloma.Validationagainstdatafromhospitalsystemsisnotimpossible,
but would require identifying the centre in medico administrative data. No cohort
providinghistoryofdrugregimenscouldbeidentified.
Underdetection:patientsinclinicaltrialsb)
Experimentaldrugsadministeredtopatientsenrolledinclinicaltrialsarenotavailable
intheHealthinsurancesystems.UnpublisheddatafromacommunicationattheGRELL
Meeting could provide some useful estimates. This studywas performed in 3 French
area:Côted’Or,CalvadosandGironde.Inclusionratesinclinicaltrialsforthefirst line
regimenwererespectively34%,7%and5%.Thisstudyhighlightedthepotentialwide
variations in drug regimens received but also demonstrated important differences in
the 5 years Net survival (64%, 46%, and 42%) and Progression Free Survival (PFS).
173
There were also significant differences in autologous transplantation rates, which
shouldimpactthenatureofthefirstlinereceived.Inourstudypopulation,51patients
had code for hospital chemotherapy (Z51), but with not any recorded drugs. As
validatedchemotherapyregimenswithoutanyof therecentdrugs isquiterare, these
patientsmaybethosefromclinicaltrials.
174
175
E. Chapter5:Dealingwithlongitudinaldataandmultipleconcomitantexposuresinspecificcontexts:“Analysinglongitudinalexposuretoproduceautomatedindicatorsonpotentialdrug‐druginteractions”
AurorePalmaro,EmiliePatrasdeCampaigno,MathildeDupui,BerangèreBaricault,
JulieDupouy,FabienDespas,MaryseLapeyre‐Mestre.Analysinglongitudinal
exposuretoproduceautomatedindicatorsonpotentialdrug‐druginteractions:
applicationintheFrenchmedico‐administrativedatabase(BritishJournalof
ClinicalPharmacology,tobesubmitted)
AurorePalmaro,EmiliePatrasdeCampaigno,MathildeDupui,BerangèreBaricault,
JulieDupouy,FabienDespas,MaryseLapeyre‐Mestre.Analysededonnées
longitudinalespourlaproductiond’indicateursautomatiséssurlesinteractions
médicamenteusespotentielles:applicationauxbasesdedonnéesdel’assurance
maladie.ADELF(AssociationdesEpidémiologistesdeLangueFrançaise)‐EMOIS
(Evaluation,Management,Organisation,Information,Santé)meeting,Nancy,
France,March23‐24,2017(oralcommunication)
176
Consistencyoftheprojectinrelationwiththethesisobjectives
When exploring longitudinal drug records with multiple
concomitantexposures,occurrenceofspecificdrug‐drugpairsis
sometimesofinterest,inthesamewayasfordrugcombinations
forchemotherapyregimens(chapter4).
Whatisalreadyknownandwhatthisstudyadds
French health insurance databases represent a potentially
valuablesourceforstudyingpotentiallydrug‐druginteractions.
However,notoolsareavailabletoscreenthemassiveamountof
drugdataagainstexistingcompendiumofinteractions.
Thistooloffersageneralframeworkforimplementationofdrug‐
drug interaction studies in the French health insurance
databases.
Keyresearchquestions
How to identify relevant drug‐drug combinations of interest
withinmultipleconcomitantdrugs?
177
Résumé 11. Mieux appréhender des situations impliquant des données
longitudinales et complexes : «Analyse de données longitudinales pour la
production d’indicateurs automatisés sur les Interactions médicamenteuses
potentielles:applicationauxbasesdedonnéesdel’assurancemaladie»
SUMMARYINFRENCH
Introduction:
Les interactions médicamenteuses représentent une part non négligeable des cas
d’hospitalisationsetdedécèsliésauxmédicaments.Avecdesdonnéescouvrantplusde
65 millions de Français, les bases de l’assurance maladie constituent une source
potentiellementpertinentepourl'étudedesinteractionsmédicamenteusespotentielles
(IMP). Cependant, il n’existe aucun outil permettant d’évaluer la prévalence des
interactionsmédicamenteusesàpartirdecesdonnées.Nousavonsdoncmisaupoint
unoutilcompletpourcaractérisercesinteractionspotentiellesàpartirdesdonnéesde
remboursements, accompagné d'indicateurs quantitatifs. Cet outil est applicable
immédiatementauxbasesdedonnéesdel'assurancemaladiefrançaise,maisadaptable
àdessourcesdedonnéesvoireàdesthesaurusdifférents.Lespossibilitésdecetoutil
sont illustrées au travers d’une étude de cas menée sur une population de patients
prévalentsatteintsdemyélomemultiple.
Méthode:
Lethésaurusdesinteractionsmédicamenteusesélaboréparl’ANSMaétéutilisécomme
référentiel (dernière mise à jour Janvier 2016). Les interactions médicamenteuses
potentielles y sont classéesen4niveauxde contrainte: contre‐indication, association
déconseillée, précaution d'emploi, à prendre en compte. Les interactions retenues
devaientavoirunetraductioncliniquesignificative,pouvantprovoqueroumajorerdes
effets indésirables ou entraîner une moindre efficacité des traitements. La présence
d’une IMP a été définie par la présence concomitante de 2 médicaments ou classes
pendant au moins un jour. L’exposition longitudinale a été étudiée pour calculer le
178
nombre,letypeetladuréedeschémasconcomitants.L'outild'interactionaétéconçuà
l’aidedeSAS9.4,accompagnéduthesauruscodé.
Afind’illustrerlespossibilitésdecetoutil,uneétudedecohorterétrospectiveaétémise
enœuvredansunecohortedepatientsatteintsdemyélomemultiple,identifiésdansle
SNIIRAMMidi‐Pyrénéesàl'aidedesdiagnosticsprincipaux,reliésouassociésduPMSI
MCO(CIM‐10codesC90)etsuivispendant12mois. Ils’agitdepatientsnouvellement
diagnostiqués,traitésounonpourunepremièreligne.
Résultats:
Parmiles506nouveauxpatientsatteintsdemyélomemultiple(446avecaumoinsune
séquencedeprescriptionconcomitante),73.3%(n=327)ontétéexposésàaumoinsune
interactionmédicamenteusepotentielle,dont8,6%de“contre‐indications”(n=28)et
15,7%d’“associationsdéconseillées”(n=51).LesIMPimpliquaientessentiellementdes
médicamentsdestinésàtraiterlescomorbidités,etaucunecontre‐indicationimpliquant
desmédicamentsanticancéreuxn’aétéidentifiée.Lesmédicamentsimpliquésdansles
IMPprovenaientleplussouventdumêmeprescripteur(60%,n=10555).
Conclusions:
Cet outil offre un cadre général pour la mise enœuvre d’études sur les interactions
médicamenteuses à partir des bases de données de l'assurance maladie. A partir du
thesauruscomplet,desétudesdédiéespourrontêtreconduitessurd’autrespopulations
cibles, accompagnées éventuellement d’une étude de la survenue d’événements
spécifiques, permettant d’apporter des éléments qualitatifs contributifs. Les résultats
générés par cet outil pourraient ainsi permettre d’accroître les connaissances
concernantlesinteractionsmédicamenteuses.
179
1. Presentationofthestudy
Electronic healthcare databases are increasingly used in pharmacoepidemiology to
study drug safety in real life. Drug–drug interactions (DDI) represent an important
causeofhospital admissionanddeath,but tools for screening themassiveamountof
drug data in such databases against existing compendiumof interactions are lacking.
Among the existing initiatives, SFINX database has demonstrated the interest of an
integratedsystemtoreducetheprevalenceofDDI108.
RoutinelycollecteddatafromtheFrenchhealthinsurancedatabase,withmorethan65
million inhabitants covered, represent a potentially valuable source for studying
potentialdrug‐drug interactions(pDDIs).Wehavedevelopedacomprehensivetool to
characterizePotentialDDi(pDDI)fromclaimsdatabaseswithasetofquantitativeand
visualindicators.ThistoolisreadytoapplytolargedatabasesusingATCcodes,suchas
the French health insurance databases (SNIIRAM), but is adaptable to different
compendium.
Multiplemyelomaisaninterestingmodelforstudyingcoprescribing.Withamedianage
atdiagnosisaround70, thesepatients frequentlysuffer fromadditionalcomorbidities
requiring long‐term therapy. In addition, multiple myeloma therapy is based on
prolonged chemotherapy, in association with a wide range of supportive care
treatments. Indeed, those patients are particularly at risk for both drug–drug
interactions and related occurrence of serious adverse events and death. We have
already investigated the most appropriate methods for identifying cases of multiple
myeloma109andthecombinationofdrugsreceived110.Inaddition,completenessofthe
data source enables to access both ambulatory and hospital drugs.Wedemonstrated
thecapabilitiesofthistoolthroughacasestudyinmultiplemyelomapatientsidentified
intheSNIIRAMandfollowedfora6months,illustratingthecontributionofanticancer
drugs’,‘supportivecaredrugs’,and‘drugstotreatadditionaldiseases/comorbidities’in
thepDDIretrieved.
180
2. Objectives
Theobjectivewas to adapt the compendiumof interactions developedby the French
MedicinesAgencyforautomateddetectionofpotentialdrug‐druginteractions
3. Methods
Aretrospectivecohortstudywasimplementedamongpatientswithmultiplemyeloma
in regional healthcare database from 2011 to 2014. List of DDIs was based on the
compendiumelaboratedbytheFrenchMedicinesAgency(lastupdatedJanuary2016).
ApDDIwasdefinedasthepresenceofaminimumonedayoverlapfordrugslistedas
“interacting”. Longitudinal exposure was investigated to compute number, type, and
durationofoverlapforinteractingdrugs.Theinteractiontoolwasdesignedasasetof
SAS9.4computingcodes.
4. Publication
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
5. Discussion
Thistooloffersageneralframeworkforimplementingdrug‐druginteractionstudiesin
French health insurance database. In the sameway as coprescribed anticancer drugs
were used to identify relevant drug regimens in chapter 4, identification of relevant
concomitantsequenceswerebasedon ‘actualconcurrentuse’ 111.Dispensingdataare
available on a daily basis, and concurrent use of anticancer drugs dispensed in the
hospital pharmacy, in the sameway as other drugs (e.g., supportive care drugs’ and
‘drugstotreatadditionaldiseases/comorbidities’)couldbeinvestigated.
This work is likely to facilitate further research on DDIs through automated
computationandadaptabletools.OutputsofDDIsexplorationareintendedtoincrease
knowledge and raise awareness of different stakeholders on concomitant use of
contraindicated medication combinations, and may be applied for monitoring
prescribingquality.
211
212
213
F. Chapter6:Improvingtheexplorationoflongitudinaldrugdata:“Datavisualizationfordrugexposureinpharmacoepidemiology”
Aurore Palmaro,Maryse Lapeyre‐Mestre. Data visualization for drug exposure in
pharmacoepidemiology : a case study for complex drug regimens in multiple
myeloma.e‐HealthResearch2016.Howdigital technologiesdisruptepidemiology
andmedicalresearch.Paris,October11‐12,2016(commentedposter).
214
Consistencyofthearticleinrelationtothethesisobjectives
Addingrelevancetodrugexposuremeasurement
Facilitating the exploration of longitudinal data in complex
contextsthroughdevelopmentofnewmethods
Whatisalreadyknownandwhatthisstudyadds
In the context of analysis of large datasets, it is difficult to
account for complex treatment schemes or discontinuous
exposureusingconventionaldescriptivestatistics
Novelstrategiesforinformationintegrationarethenneeded.
Visualisationtoolsmightbeusefulinpharmacoepidemiologyfor
betterstudydesignandreporting.
Keyresearchquestions
Whatisthepotentialcontributionofdatavisualizationtoolsfor
improvingtheexplorationoflongitudinaldrugdata?
215
Résumé 12. Visualisation de données pour l’exposition médicamenteuse en
pharmacoépidémiologie:uneétudedecasdanslemyélomemultiple
SUMMARYINFRENCH
Introduction
Les bases de données de l'assurancemaladie (SNIIRAM) rassemblent unemasse très
conséquented’informationssurlesmédicamentsdélivrésparexemple.Danslecadrede
l'analyse de données massives sur les délivrances médicamenteuses, il est
particulièrementdifficilededécriredesschémasdetraitementcomplexes,marquéspar
des expositions multiples et discontinues, en utilisant des statistiques descriptives
conventionnelles. Le recours à des outils alternatifs se révèlerait donc dans ce cas
particulièrementpertinent.
Au cours de ce travail, différentes méthodes potentielles pour visualiser les cycles
d'expositionauxmédicamentsontétépasséesenrevue,ainsiqueleurapportpotentiel
pour améliorer la conception des études, la stratégie de modélisation, générer de
nouvelles hypothèses et mieux décrire l’exposition médicamenteuse en
pharmacoépidémiologie.
Méthodes
Différentes visualisations ont été générées à partir des données médicamenteuses
ambulatoiresethospitalières.Deuxprincipalestechniquesdevisualisationdedonnées
ontététestées:lesreprésentationstemporellesetlesreprésentationsenréseaux.
Conclusions
Cette étude illustre l'utilisation d'outils de visualisation de données pour décrire les
schémasd’expositionlongitudinauxauxmédicamentsetlessituationsdeconcomitance
enprésencede régimescomplexes.Cesoutilspourraient contribueràmieuxexplorer
les grandsensemblesdedonnées longitudinalesdesbasesdedonnéesde l'assurance
maladiefrançaiseetàgénérerdeshypothèsesconcernantlesmodesdeconsommation
envieréelle.
216
1. Presentationofthestudy
French health insurance databases (SNIIRAM) contain millions of patient records in
relation tomedicationsdispensedorhospitaldiagnoses for instance. In thecontextof
analysis of large datasets, it is difficult to account for complex treatment schemes or
discontinuous exposure using conventional descriptive statistics. Novel strategies for
information integration are then needed. Data visualization and visual analytics are
widelyused.However,toolsormethodsarenotwellknown,andtranspositiontodrug
treatmentdataisnotalwaysstraightforward.
Literature on exploration of electronic healthcare record is now abundant, but is
essentially for exploration and hypothesis generation purpose. A increasing range of
papers has been dedicated to the exploration of electronic healthcare databases for
knowledge discovery, as illustrated by the “Medication‐Wide Association Studies”112.
Otherdevelopments in thisarea include for instance interactivesystemsdesigned for
physicians in a personalized medicine perspective. In the context of a confirmatory
approach, such methods should be introduced with caution. Application of data
visualizationmethods within this framework has not been extensively discussed. An
interesting initiative is provided by Mini‐Sentinel group, discussing briefly how
“pictorialmodelscanhelpelucidatestatisticalmodels”80.Thisreportproposedasetof
visualtypesthatcouldbeproducedbeforeorjustafterthestudytofacilitatethechoice
of study design while verifying the underlying assumptions. In the same way, some
visualizationsdevelopedaspartoftheOMOPprojecthavebeendescribed81.
Oneofthemotivationsofthisthesiswastogaininsightintopatternsofexposureinreal
life. Several graphical tools were tested were developed throughout the different
projects, showing the potential of visual analytics to gain insights into complex drug
exposurepatterns.Thischapterillustratesthepotentialinterestofthesevisualizations
through thedifferentcasestudiesconductedanddiscuss their contribution forbetter
studydesign,hypothesisgeneration,andreportinginpharmacoepidemiology.
217
2. Methods
Visualizationstesteda)
Thetoolsdevelopedduringtheprojectsincludedapatientprofiletostudylongitudinal
exposure patterns, a Sankey diagram to investigate patients’ changes across
chemotherapy regimens, a stream graph to highlight prevalence of use according to
drugclass, aheatmap to investigatecoprescribingpatterns, andaclusteredgraph for
summarizingexposureprofiles.Inaddition,threeadditionalusefulrepresentationsare
discussedfortheirmethodologicalcontribution:acohortdiagramtoassesslongitudinal
dataavailability,adistributionapproachforoptimizingthechoiceofindexdate,anda
diagram todiscriminatepoint fromchronicexposure.Exampleof insightgained from
thesevisualizationsareprovided.Visualformatswerecategorizedintographsforsingle
drugpatterns(individualprofilesoraggregated),visualizingchangesacrosscategories,
andvisualizing changes inprevalenceofuse.VisualizationswereproducedusingSAS
9.4,Rv3.2.1,D3.jslibraryandGephiv0.8.2.
Datasetsforthestudyb)
Visualizations were developed in the context of the first four projects: the cohort of
benzodiazepineusers fromthestudypresentedatchapter2and3andthecohortsof
multiplemyelomapatientspresentedthroughchapter4(236incidentpatientsstarting
anewline)andchapter5(506incidentpatients).
Casestudyforlongitudinaldatacomprisedthenameofchemotherapyregimensforthe
first 6 months of follow‐up, derived from the work on drug regimen identification
(Chapter4,page143),andusedforproducingpatients’profilesillustratedFigure1and
Figure 2. Figure 3 is derived from a complementary investigation on 200 incident
lenalidomideusers,followedforupto6months.Thedatasetonallconcomitantdrugs
for the cohort of 236multiplemyeloma patientswas used for producing the stream
graph (Figure 4). To study relation between categorical data (Figure 6), a dataset
summarizingthemostfrequentinteractingdrugsclasseswasused(Chapter5:Dealing
with longitudinal data and multiple concomitant exposures in specific contexts:
218
“Analysing longitudinal exposure to produce automated indicators on potential drug‐
druginteractions”,page167).
For highlighting longitudinal data availability (Figure 7), the cohort of multiple
myeloma patients is used (Chapter 4). The final example of chronic versus point
exposure is based on the study on benzodiazepines and mortality (Chapter 2:
Illustrating the impact ofmethodological choice on risk estimates and the interest of
time‐dependentexposure: “Benzodiazepinesand riskofdeath: results from two large
cohortsstudiesinFranceandUK”,page75).
3. Results
Individualsequencesa)
Episodes of exposure could be considered as time spans,which could beplaced on a
horizontal timeline.When only exclusive, non‐overlapping sequences are considered,
one horizontal, interrupted bar chart could describe the whole treatment trajectory.
Whendrugs considered are overlapping, placing all drugs in a single line innomore
relevant,andGanttchartsmightberelevant.Hence,Ganttcharts (Figure1)wouldbe
moreadaptedtodisplaysinglepatient’sprofiles,withstartandenddateofanepisode
representedusinghorizontalbars.
Individual sequences graphs were used to support the validity of the regimens
identificationalgorithm(chapter4), inthecontextoftheabsenceofexternalstandard
for comparison. Indeed, once the limitation due to some unobservable drugs
acknowledged,individualdrugpatternsaresupposedtobeaccuratelyrecordedinthe
data source. The main source of misclassification is then linked to the grouping of
individualdrugsintochemotherapyregimensaccordingtotheproposedalgorithm.As
already proposed by Bikov 106,we have generated individual drugs sequences charts
(horizontal bar charts) for 100 randomly patients and visually examined the
consistency between these diagrams and the chemotherapy regimens attributed to
assess face validity of the algorithm.One example is provided Figure 1, inwhich the
algorithmaccuratelycapturedthefirstline(Melphalan‐Prednisone‐Thalidomide(MPT)
isattributed).Thesecondlinebeginswiththeadditionofanewdrugnotbelongingto
219
thefirstline,andcomprisedthedrugadministeredinthe28daysafterstartingthisnew
line (prednisone and melphalan). The regimen attributed is then VMP (Bortezomib‐
Melphalan‐Prednisone).
Figure1.Patientsprofile.Individualdrugsequences
Longitudinaldrugexposuretreatedassequentialdata:a)
Sankeydiagrams
In somecases, treatmentepisodescould fall ina finitenumberof states, andpatients
made transition between several possible sequences. A typical example is
chemotherapydatacomprisingsequentialtreatmentcycles.
Sankey diagrams were initially designed to represent flows in the energy industry.
Through these thesis projects, Sankey diagrams were used to represent patient’s
trajectories according to chemotherapy regimens, as a representation of movements
betweencategoriesover time.Flowsarerepresentedwithband, thebandwidthbeing
proportional to the size of the flow (number of patients). Vertical bar represent time
220
intervals,settoonemonth. Inthecasestudy,multiplemyelomapatientsarefollowed
upto6months.Thisdiagramenablestorepresentthetimingofregimenschanges,but
alsotheproportionofpatientsswitchingtoanotherregimen.Interruptedflowsarealso
takenintoaccount.
Figure2.Sankeydiagram.Trajectoriesaccordingtoageclassandchemotherapy
regimensinthefirst6monthsoftherapy(n=236).
Thisdiagramenables to assess the typeofdrug regimens and allocationbyage class
(+/‐ 65 years). Patients aged less than 65 received mainly VTD (Bortezomib‐
Thalidomide‐Dexamethasone) or VMP (Bortezomib‐Melphalan‐Prednisone), while
those age more than 65 were attributed VMP, MPT (Melphalan‐Prednisone‐
Thalidomide) or MP (Melphalan‐Prednisone), which is in accordance with current
recommendations113,114.Ascendingordescendingbandsrepresentthepartofpatients
moving from one regimen to another, or interrupting their treatment (interrupted
band).
221
Sankey diagram can also be of interest for studying doses concessions. As a
complementary investigationondrugregimens trajectories,a focuswasmadeonone
particularregimen:Lenalidomide‐Dexamethasone(RD).TheRDschemeconsistsin28‐
day cycles, with continuous lenalidomide (25 mg orally from Day 1 to 21) and
dexamethasone 20mg on days 1, 8, 15, 22. Three dose reduction levels are defined
(Dose1: 25mg, Dose 2: 15mg and Dose3: 5mg)115. The lenalidomide in available in
packagesdosedat5,15or25mg,enablingtofollowthechangesinthedosedelivered.
Older or comorbid patientsmay initiate at lower dose, and further dose concessions
couldbedecidedincertainpatientsinordertoreducetreatmentrelatedtoxicity116,117
whilemaintaining thepatientsunder treatment. For this case study, patients starting
this regimen (first dispensing of lenalidomide) were followed up to 6 months.
TrajectoriesofdosesareplottedFigure5.
Figure 3. Sankey diagram. Trajectories of doses in patients receiving incident
lenalidomide(first6monthsoffollow‐up,n=200)
A certain number of key points illustrate what could be derived from this Sankey
diagram:
Attheendofthestudyperiod,morethanhalfofthepatientshavestoppedtheir
treatmentorswitchedtoanothertherapy.
222
Doses concession (from 15‐25mg to 10‐15mg or 10‐15mg to 5mg) were
importantinthefirstmonthsafterinitiation.
Thereisnoobserveddirectconcessionfrom15‐25mg+to5mgfromonemonth
toanother.
Conversely,patientsaugmentingtheirdosedonotswitchdirectlytothe25mg
step(exceptionobservedbetweenmonths3and4)
Thestreamgraphisaspecialcaseofstackedareachart.Thestreamgraphwasusedto
display trends in prevalence of use of selection drug classes over time (Figure 4). In
contrasttoSankeydiagram,itdoesnotenabletoseetrajectoriesovertime(patientsin
the bandwidth are not necessarily the same). However, it could reveal interesting
trends, such as the pre‐index increase in druguse, in particular for analgesics (N02).
After index date, anticancer drugs are represented (L01, L02, L04), together with
specific supportive care drugs (antibiotics (J01) and vaccines (J05), bisphosphonates
(M05).
Interestingly, nonsteroidal anti‐inflammatory drugs (NSAIDs) are frequently used
before index date, but disappeared after start of multiple myeloma management. In
addition to other nephrotoxic drugs (contrast agents, etc.), these drugs should be
avoidedinmultiplemyelomapatients.
The graph reveals a very satisfying compliance to this recommendation. Another key
finding is the rapid decrease in the size of the bandwidth for anticancer drugs (L01,
L03).Attheendofthe12‐monthperiod,themajorityofthepatientsarenomoreunder
activechemotherapytreatment.
223
Figure4.Streamgraphrepresentingtrendsinprevalenceofselecteddrugclasses.
Incidentmultiplemyelomapatients(12monthsbeforeand12monthsafterindex
date)
Clusteringlongitudinaldatab)
Whensamplesizebecomestoohightovisualizeindividualdata,methodsforexamining
aggregatedtrajectoriescouldbeconsidered.Somemethodsareatthefrontierbetween
data visualization and analytics, and offer the opportunity to classify pattern of
exposure according to various algorithm (kmeans), like PROC TRAJ in SAS or Klm
packages47,118–123.
TheTraminer®packagewasusedtoproducethefollowingaggregatedgraphs.
224
Figure 5. Aggregated longitudinal drug exposure patterns (patients receiving
incidentlenalidomide,n=200)
Studyingconcomitantdrugpatternsc)
Inthecasestudy,heatmapswereusedtohighlightthemostfrequentinteractiondrugs
orprescribedclasswithinthedataset.Otherpotentialchartswerethoserepresenting
the relationbetween two categorical variables (area charts, node‐linkdiagrams, etc.).
Transversal diagram are less informative than those previously presented. However,
they highlight the main interacting drug classes observed. Cardiovascular drugs are
frequently retrieved, with potential interactions between other cardiovascular drugs,
drugs of the musculoskeletal system and interactions between two central nervous
systemsdrugs.
225
Figure6.Heatmapoffrequenciesofpotentialdrug‐druginteractionsaccordingto
themain (level I)ATC class (multiplemyeloma cohort for pDDI identification,
n=506)
Graphics for better study design and exposured)
modelling
(1) Ascertaining longitudinal availability
As highlighted in the first article (chapter 1), ensuring prior longitudinal data
availability iscrucial.However,additionalerrorsmayoccur through thedataanalysis
workflow,andmightresultsinincompletedatasetsforinstance.
Figure9shouldbeusefultoascertainthetimespanofthedataextracted.However,they
shouldbecompletedbyadditionalexplorationatthemonthlevel.
226
Figure7.Longitudinaldataavailability
(2) Optimisation of the choice of index date
When studying the prevalence of initiation of selected drugs after diagnosis of a
conditionorspecificevents,graphswouldbeofparticularinterest.Whentheindexdata
couldbe subjected to importantuncertainty (dateofdiagnosisof cancer inelectronic
healthcare databases), the event of interest in relation to the diagnosis could occur
somedaysbeforethedateretainedastheindexdate,resultinginanunderestimationof
true “newusers”. To overcome this issue, the delay between index date and the first
dateofexposurecouldbeplottedinordertostudyitsdistribution.Inthedistributionis
switched to the left, thiswould indicate that theprevalence of newusers is certainly
biased.
The waiting time distribution is a graphical method based on observed distribution
which isused for choosing themostappropriateobservationperiod fordefiningnew
users.47
227
Figure 8. Dstribution of the delay between index date and the first date of
exposure
(3) Assessing adequacy of time dependent models
Time‐dependentexposuremaylackofrobustnessincaseofveryimportantunbalance
between exposed and unexposed time. Graphics summarizing type of exposure have
then a potential interest for assessing adequacy of time dependent models 124.
Simulation showed more biased estimates of exposure–outcome associations if
proximitytofollow‐upstartwasnotconsidered125.
Throughtheprojectonbenzodiazepines,drugexposurepatternsofasampleofpatients
wererepresentedtoconfirmthenatureofdrugexposure.Theprincipleisverysimilar
tothe“StarAndStripes”diagramswereproposedbytheMini‐Sentinel’sMethodsCore
WorkgrouponCase‐BasedApproaches.Thisdiagram,closetohorizontaltimeline,was
intendedtodifferentiatepointexposuresprofilesfromchronicexposure(“stripes”)80.
228
Figure9.Exposureprofile
4. Discussion
Theapproachforintegratingvisualizationintopharmacoepidemiologicalstudyisinline
with the original definition provided by Tukey. Indeed, in the 70ies, Tukey have
developed the concept of exploratory data analysis 126, defined as: "procedures for
analyzing data, techniques for interpreting the results of such procedures, ways of
planning the gathering of data to make its analysis easier, more precise or more
accurate,andallthemachineryandresultsof(mathematical)statisticswhichapplyto
analyzingdata”.
This chapter was focused on drug exposure patterns only, which had been poorly
explored, except for interactive visualization purposes 127,128. A larger set of
visualizationshavebeendevelopedforstudyingtheassociationbetweenexposureand
anoutcome129orsignaldetection.Prescriptionsequencesymmetryanalysisareagood
229
example of such graphical method 122. Charts for signal detection have also been
discussedintheMini‐Sentinelreport80.
Whenintroducingvisualizationtoolsinthedecisionprocess,itshouldbeacknowledged
thatvisualrepresentationscouldbesubjectedtomisinterpretationordistortions.This
aspectisnotdiscussedhere,butdeservesfurtherconsiderations,asalreadyperformed
in the framework of the PROTECT project for benefit‐risk assessment130,131. In the
proposedfigures,numbersmightbeaddedtohelptheinterpretation.Inshouldalsobe
noted that graphical methods are often accompanied by numerical indicators to
overcome this issue (e.g. sequence ratios prescription sequence symmetry
analysis).Thereisnowarichliteratureonvisualizinghealthcaredataontimelines.The
fieldoflifehistorydatahasalsoprovidedrelevantcontributionsinthisarea.Themain
limitations of this short overview are in relation with the non‐systematic review of
visualizations.However,previouspublicationwithinthespecificareaofdrugexposure
patterns is relatively scarce, and this chapter might be useful to highlight potential
application of data visualization in the context of hypothesis based
pharmacoepidemiologicalstudies.
5. Conclusion
Thischapterillustratestheuseofvisualanalytictoolstocharacterizelongitudinaldrug
patterns in the presence of complex regimens. These tools could contribute to better
explorethelargelongitudinaldatasetsoftheFrenchhealthinsurancedatabasesandto
generatehypothesesconcerningpatternsofdruguse inreal life,orcouldbeuseful to
supportmethodologicaldecisionsduringstudydesignorearlydataexploration.
230
231
G. ComplementaryChapter:“IdentifyingmultiplemyelomapatientsusingdatafromtheSNIIRAMandPMSI:validationusingtheTarncancerregistry“
AurorePalmaro,MartinGauthier,CécileConte,FabienDespas,PascaleGrosclaude,
MaryseLapeyre‐Mestre.Identifyingmultiplemyelomapatientsusingdatafromthe
SNIIRAM and PMSI : validation using the Tarn cancer registry (Medicine, under
review)
AurorePalmaro,MartinGauthier,CécileConte,FabienDespas,PascaleGrosclaude,
MaryseLapeyre‐Mestre.Identifyingmultiplemyelomapatientsusingdatafromthe
SNIIRAM and PMSI: validation using the Tarn cancer registry. GRELL Meeting
(GroupforEpidemiologyandCancerRegistry inLatinLanguageCountries,Nancy,
May4‐6,2016(poster)
232
Consistencyoftheprojectinrelationtothethesisobjectives
Adding relevance to drug exposure measurement through
accurateselectionofthetargetpopulation
Facilitating the exploration of longitudinal data in complex
contextsthroughdevelopmentofnewtools
Whatisalreadyknownandwhatthisstudyadds
Misclassificationbiascouldimpactdrugexposure,asillustrated
inthethesis,butcouldalsoaffectcaseascertainment
Accuracyofidentificationandthenimplementationofvalidation
studiesisofprimaryimportance
This study provides an assessment of case identification
algorithmsformultiplemyeloma
Keyresearchquestions
What is the accuracy of case identification algorithms used for
identifying multiple myeloma through the French health
insurancedatabases?
Howdofirstdiagnosiscomparewiththedocumenteddateinthe
registry?
Do algorithms using longer periods of observation perform
better?
233
Résumé13.S’assurerdelavaliditédel’identificationdescaspourlesmodèlesà
l’étude:validationde l’algorithmed’identificationdumyélomemultipleàpartir
duregistredescancersduTarn
SUMMARYINFRENCH
Introduction
La validation des algorithmes d’identification doit accompagner lamise en place des
études sur les bases de données. Des initiatives internationales sur l’identification
d’affections ou d’évènements particuliers ont étémenées, à l’image des projetsMini‐
Sentinel et OMOP (Observational Medical Outcomes Partnership) aux États‐Unis ou
d’EU‐ADR en Europe 132. Une série importante de revues systématiques portant sur
l’identification d’un certain nombre d’affections, dont le lymphome 133 a été publiée
depuis2012(defaçonnonexhaustive:134–140.
Cependant, les paramètres de validité d’un algorithme ne se sont pas transposables
entre lesdifférentessourcesdedonnées.EnFrance, leréseauREDSIAMaentaméune
démarche généraliste avec une volonté structurante, qui s’accompagne de la mise à
dispositiond’algorithmes.Des travauxontégalementétémenéssur l’identificationde
cas de cancer 141. Cependant, aucune validation n’avait été menée pour le myélome
multipledanslesbasesdedonnéesdel’assurancemaladie.
Objectifs
Cette étude visait à évaluer les performances de plusieurs algorithmes basés sur les
diagnosticshospitaliersduPMSIMCOet lesaffectionsde longueduréepour identifier
lespatientsatteintsdemyélomemultiple.
Méthodes
Lescaspotentielsdemyélomeaucoursdelapériode2010‐2013ontétéidentifiéspar
laprésenced’aumoinsuncodedediagnosticprincipalpourlemyélomemultiple(CIM‐
10«C90»).Desalgorithmesalternatifsontégalementconsidérélesdiagnosticsreliéset
234
associés,encombinaisonounonaveclesaffectionsdelonguedurée.Lescasincidents
étaient ceux sans code «C90» au cours des 24 ou 12 derniersmois. La sensibilité, la
spécificité et les valeurs prédictives positives et négatives (VPP et VPN) ont été
estimées, les cas demyélomemultiple diagnostiqués en 2010‐2013 figurants dans le
registreducancerdeTarnétantpriscommeréférence.
Résultats
Les données sur les ALD concernaient 11 559 patients (22 244 pour les données du
PMSIMCO).Leregistrecontenait125casdemyélomemultiple.Lasensibilitéétaitde
70%enutilisantseulementlesprincipauxdiagnosticshospitaliers(spécificitéde100%,
VPP79%),76%en considérant également lesdiagnostics reliés (spécificitéde100%,
VPP74%),et90%aveclesdiagnosticsassociés(spécificitéde100%,64%PPV).
Conclusions
Les algorithmes intégrant les diagnostics hospitaliers présentaient des performances
relativementsatisfaisantes.L’algorithmeoptimalpouridentifierlespatientsatteintsde
myélomemultiple,etmaximisantàlafoisl’indicedeYoudenetlaspécificité,étaitcelui
exigeant«aumoins»undiagnosticprincipal,reliéouassocié«C90»,avecunepériode
d’observationde12mois(sensibilité:90%,spécificité:100%,VPP60%).
235
1. Presentationofthestudy
Relevance of drug exposure measurement and in particular external validity is
conditioned by a proper selection of the target population.Misclassification bias (i.e.
including falsemyelomapatients)could introduceconfusionand irrelevantresults. In
administrative databases, this selection is made on the basis of case definition
algorithms, and the validity of the coding is then of primary importance 142–144. As
algorithmsperformancecouldbeinmanywaysdatabase‐specific,therewasaneedto
implementthisvalidationstudyinFrenchhealthinsurancedatabases.Alotofprevious
validationsweremadeintheICD‐9databaseintheUSandvalidationstudiesarelacking
forEuropeanandNordicdatabase, inwhichICD10ismorefrequent144.Whileseveral
studieshavemeasuredthevalidityofcancercasesascertainmentinFrance145–147,none
focused on haematological diseases. Then, the validity of identification of multiple
myelomacasesthroughthesedatabaseshasnotbeenpreviouslyestablished.Thisstudy
aimed to assess the performance of several algorithms based on hospital diagnoses
(PMSI,“Programmedemédicalisationdessystèmesd’information”)anddiagnosesfrom
the long‐term diseases (LTD) scheme. Validation of case identification algorithms
representsanimportant issue,asdemonstratedbyrecentcalls144,butalsobyseveral
initiativesfromMini‐SentinelandOMOP(ObservationalMedicalOutcomesPartnership)
inUSorEU‐ADRinEurope132.Animportantseriesofsystematicreviewonmethodsfor
validating a wide range of disease, including lymphoma for instance 133, has been
publishedsince2012134–140.Lessonslearnedandproposalsforimprovementhavebeen
formulatedduringthesevalidationstudies148.However,literatureconcerningmultiple
myelomaisverypoor,andonlyoneresourcecouldbeidentified149.
2. Objectives
This study aimed to assess the performance of several algorithms based on hospital
diagnoses (PMSI, “Programme de médicalisation des systèmes d’ information”) and
diagnosesfromthelong‐termdiseases(LTD)scheme.
236
3. Methods
Potentialmultiplemyelomapatientsin2010‐2013wereidentifiedusingthepresenceof
hospital recordswithat leastonemaindiagnosis code formultiplemyeloma (ICD‐10
‘C90’). Alternative algorithms also considered related and associated diagnoses,
combinationwithlong‐termconditions,oratleast2diagnoses.Incidentpatientswere
those with no previous ‘C90’ codes in the past 24 or 12 months. The sensitivity,
specificityandpositiveandnegativepredictivevalues(PPVandNPV)werecomputed,
using a French cancer registry for the corresponding area and period as the gold
standard.
237
238
4. Publication
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
5. Discussion
Algorithms tested exhibited very different performances, ranging from poor
performance when using only main hospital diagnoses to very acceptable
parameters when all hospital diagnoses are used in combination with long‐term
conditions. The optimal algorithm to identify MM patients (maximizing both the
Youden’s index and specificity) was “at least 1 main, OR related, OR associated
hospitalMMcode”,witha12‐monthobservationperiod,whichhadasensitivityof
90%,aspecificityof100%,andaPPVof60%.Thesamealgorithmwitha24‐month
observationperioddemonstratedasimilarperformance;neverthelessthealgorithm
with the shortest period of observation should be preferred. Indeed, the study
design simulated theperformanceof algorithms thatwouldbebasedon the large
French health insurance databases (SNIIRAM) in further research. Using an
algorithmwith a restricted period of observation (12months as compared to 24
months)haspotentiallyagreatinterestforanincreasingsamplesizeandlengthof
possible follow‐up in the contextof limited longitudinaldata availability (data are
availablesince2006intheSNIIRAM).
This study could also be discussed with regards to other validation studies
conducted in France. Other validations realised were performed on the same
principle:gettingnonanonymizeddatafromregistriesandhospitaldata,linkingthe
data, and assessing performance of cases finding algorithms. Compared to the
strategyofQuantinetal. forcolorectal cancercases 147,wedonotapply thesame
strategyfordata linkage(softwarecalledANONYMATbasedonhashcoding150,vs
proceduresforinternaluseoftheregistry,strategybasedon24matchingattempts
combining 5 identifying variables). Health insurance data for medical procedures
werenotavailable,thusvalidationusingtheinitialproceduresofinterest(surgical
procedures or endoscopic investigation for instance), as proposed for colorectal
cancers 147 was not realizable. Maybe the best initial procedure for identifying
myelomawouldbeamyelogram,neverthelessitcouldalsoberepeatedovertime.In
addition, available data within the registry revealed that only 94% of confirmed
casesofmyelomahavebenefitedfromamyelogram.Thesecondalgorithmbasedon
diagnosiscodes 147 ismorecomparablewithourstrategy,but theperiodrequired
260
withoutdiagnosis codewas longer (5yearsvs2or1year inour study), together
with the type of diagnosis considered (related diagnoses not taken into account).
The impact of coding practices for main, related and associated diagnoses in the
PMSIhasnotbeenextensivelydiscussedinthemanuscript,neverthelessshouldbe
acknowledged.
261
262
263
VIII. Generaldiscussion
Résumé14.Résumédesprincipauxrésultatsobtenus
SUMMARYINFRENCH
1. Importancede la connaissancedes sourcesdedonnées secondairesetde
l'explorationdesdonnéespourréduirelerisquedebiais
a)L'intégrationdesconnaissancesaprioridansd'autresétudes
Dans le premier article présenté au chapitre 1, une description complète des
données sur le médicament contenues dans les bases de données de l'assurance
maladie est proposée, avec un accent particulier mis sur les ruptures dans la
disponibilitédesdonnées.
Aucoursdesdiversprojets,l’importancedelavaliditédesdonnéeslorsdelamise
enœuvredesétudespharmacoépidémiologiquesafait l’objetd’unevigilancetoute
particulière,enparticulierpourlesrupturesdansladisponibilitédesdonnéesoule
choix des codes. Comme les algorithmes d'identification de la maladie sont
égalementsujetsaumêmetypedeconsidérations,l’étudedevalidationutiliséepour
identifier l’affection d’intérêt dans 2 des études a été intégrée dans un chapitre
complémentaire de ce manuscrit de thèse (" Identifier les patients atteints de
myélomemultipleàpartirdesbasesdedonnéesduSNIIRAM:validationàpartirdu
registredescancersduTarn»).
b) Mise en évidence de l’impact du biais lié aux périodes d’expositions
inobservablesdanslecadred’uneétudedecohorte
Alors que le premier projet a permis de mettre l’accent sur les ruptures dans la
disponibilité des données lors des séjours hospitaliers, le travail présenté au
264
chapitre 3 permet demontrer l’impact possible de ces ruptures au travers d’une
étudedecassurunschémad’étudedetypecohorte.
c) Réhabiliter l’analyse exploratoire des données dans les études
confirmatoires : mieux connaître les données observées pour accroître la
pertinencedelamodélisationdel'expositionaumédicament
L’expositionaumédicamentesttraditionnellementdécriteentermesdenombrede
délivrances, de cycles ou sous forme de quantité cumulée. Cependant, cette
description ne permet pas de rendre compte de la réalité des trajectoires de
traitement, et de nouvelles stratégies pour l'intégration de l'information étaient
nécessaires.
Un projet dédié a alors cherché à examiner des méthodes potentielles pour
visualiser les épisodes d'expositions auxmédicaments ("Visualisation de données
pour l’exposition médicamenteuse en pharmacoépidémiologie: une étude de cas
dans le myélome multiple "). Les outils permettant de représenter des données
longitudinalesouenréseauontétéplusparticulièrementabordés.D’aprèsWilliam
S.Cleveland151 , «Lavisualisationestunaspectessentielde l’analysededonnées.
Elleoffreuneligned’attaquefrontale,révèlelastructurecomplexededonnéesqui
ne pourraient être comprises d’aucune autre façon. Elle permet de découvrir des
résultatsinattendusetderemettreenquestionlesconclusionsattendues.»152.
2.Mettreenévidencel’impactdeschoixméthodologiquesetlesbiaisaffectant
lamesuredel'expositionaumédicament
a)Confrontationdes comparaisons inter‐groupesauxméthodes intégrant la
naturedépendantedutempsdel'expositionaumédicament
Lestravauxsurlesbenzodiazépinesfournissentensuiteuneillustrationdel'impact
deschoixméthodologiquessurlesestimationsderisqueetdémontrentl'intérêtde
l'exposition dépendante du temps par rapport à une comparaison inter‐groupe
traditionnelle.
265
b) Confrontation desméthodes intégrant ou non le biais lié aux périodes
inobservables
Lepremierarticleapermisdemettreenévidencelesrupturesdansladisponibilité
des données et en particulier lors de séjours hospitaliers pour la plupart des
médicaments. Grâce au travail sur les périodes inobservables, la nécessité de
prendre en compte l'impact de ces périodes d'expositions inobservables sur les
estimations a été rappelé, et mis en évidence dans le cas des bases de données
françaises.
3. Rôle du rationnel pharmacologique et clinique pour une modélisation
étiologiquementcompatibleetunemeilleureinterprétationdesrésultats
La question du rôle du rationnel pharmacologique a été mobilisée à différents
moments lorsdesdifférents travaux, comme lorsde la sélectiondesmédicaments
d’intérêt par exemple (benzodiazépines et mortalité). Selon le point de vue de
l’agence de régulation, seules les benzodiazépines anxiolytiques et hypnotiques
relevant des classesATCN05BA, CD ouCF devaient être prises en compte. Sur le
planpharmacologique, le casdesautresbenzodiazépinesne figurantpasdans ces
classes ATC posait problème. Il s’agissait du tétrazépam et du clonazépam. Nous
avons fait le choix d’adopter une solution intermédiaire permettant demodéliser
l’expositionauxbenzodiazépinesclasséesailleursentantquevariablesdépendantes
detemps.Cependant,comptetenududéséquilibred’effectifimportant,nousn’avons
pasmisenévidencel’impactdecesmédicaments.
La question du choix d'un comparateur pour étudier la mortalité liée aux
benzodiazépinesaégalement fait l’objetd’uneattentionparticulière.Uneréflexion
utilepeutêtreretrouvéedanslecasdel’étudeentrebenzodiazépinesetfractures38.
Les groupes sont basés sur une similarité d’indications (anxiolytiques et
hypnotiques). Dans le cadre des études OMOP, un éventail bien plus large
266
d’indication a été considéré (ulcère duodénal, schizophrénie). Cependant, les
contraintes pour la constitution étaient similaires: les médicaments témoins
devaient partager les mêmes indications avec les benzodiazépines, mais pas le
même mécanisme d’action 38. En dépit de cette contrainte, le midazolam a été
intégré dans le groupé témoin, tout comme d’autres médicaments (nortriptyline,
doxépine, buspirone, chlorpromazine, chlormezanone, prochlorperazine,
méprobamate). On y retrouve des antidépresseurs et les anxiolytiques non
benzodiazépines qui constituaient notre groupe témoin, mais aussi des
antipsychotiques.Aucoursd’uneanalyseultérieure, les investigateursd’OMOPont
décidédeconserverlesmédicamentscomparateurslesplusutilisés.L’hydroxyzine
n'aainsifinalementpasétéretenue38.
Lechoixde la fenêtrederisqueetde laméthodedemodélisationaégalement fait
appel à une réflexion sur la plausibilité et sur les mécanismes pouvant mener à
l’évènementd’intérêt.Lechoixde la fenêtrederisquea fait l’objetd’uneattention
particulière,avecunemodélisationdetypedépendantedutemps.
D’autres réflexions ont également eu lieu lors de la conception de l’étude sur les
interactions médicamenteuses potentielles, avec le choix d'une durée minimale
d’expositionconcomitante,oulorsduchoixdesinteractionsàmettreenavantselon
lapertinenceclinique.
c)Etudierlamultiplicité:l'intégrationdesmédicamentsconcomitants
Les réflexions sur la prise en compte de la concomitance se retrouvent dans
plusieursdesprojetsconduits.
Dans le cadre du développement d’un algorithme pour identifier les lignes de
chimiothérapies, c’est l’approche de la pertinence clinique qui motive la mise en
placede l’étude.Enpratiqueclinique, le traitementestappréhendésous formede
protocolesdechimiothérapie,etnondemédicamentsindividuels.Or,lesétudessur
lesbasesdedonnéesnepermettaientqu’unedescriptiondecertainsmédicaments,
267
etletravailadoncconsistéàpasserd’uneapprochemédicamentversuneapproche
«multi‐médicaments» et «multi‐ligne». Cesdéveloppements ont ainsi permisde
prendre en compte les expositions concomitantes multiples et de proposer une
approche standard pour identifier les poly‐chimiothérapies dans les bases de
donnéesdel’assurancemaladie.
Dans le casdes interactionsmédicamenteuses, ladiscussions’est traduiteparune
restitutiondesrésultatssousformedecontributionsrespectivesdesanticancéreux,
desmédicamentsdesupport,etdesautresmédicaments.Laméthodologiemetainsi
l'accentsurlesépisodesd'expositionconcomitantemultiplelorsquel’occurrencede
certainesprescriptionsconcomitantesconstitueunévènementd’intérêt.
Uneréflexionsurlerationnelcliniqueaégalementeulieufaceàl’hétérogénéitédes
résultats suite à l’étudemulti‐source sur les benzodiazépines.Malgré l’application
desmêmescritèresd’inclusion,lespopulationsfinalementsélectionnéesdifféraient
de façon importante selon lespays.Despratiquesdeprescriptionsdifférentesont
ainsi conduit à différentespopulations exposées,mais aussi à différentsprofils de
médicaments, ce qui suggère que ces éléments doivent être discutés au vu du
contextedel'étudepharmacoépidémiologique.
1. Importance of the knowledge of secondary data sources
anddataexplorationtoreducepotentialforbias
Atthebeginningdateofthisthesis, therehasbeennodetaileddescriptionofdrug
dataormethodologicalguidanceconcerningstudiesondrugusewithintheFrench
health insurances databases. In addition, the complexity and the multiple
particularities of the SNIIRAM databases were likely to introduce bias in
pharmacoepidemiological studies. The firstwork of this thesis (“Overview of drug
data within French health insurance databases and implications for
pharmacoepidemiological research”), intended to fill this gap, by offering a
268
comprehensive description of drug data contained in the French health insurance
databases, with a particular focus on gaps in data availability. It provided an
illustrationofhowaprioriknowledgeofdatasourcescontentcanhelp to identify
andlimitsourcesofbiasindrugexposuremeasurement.
Through this thesis, the importance of data validity when implementing
epidemiological or pharmacoepidemiological studies has been particularly
highlighted, especially for gaps in data availability and choice of drug or disease
codes. The majority of algorithms for disease identification used diagnosis codes
from the PMSI, sometimes in combination with long‐term conditions (ALD).
However, the use of some drugs as a proxy for specific diseases is sometimes
applied. In particular, an adaptation of the Charlson’s score for the SNIIRAM
databases86usedtheconceptof“packsize”fororalantidiabeticmedications,which
list(whenexpressedasCIPcodes),ispronetochangeovertime.Finally,inthesame
way as for drug exposure, all methods employing specific drugs for comorbidity
ascertainmentorasaproxyfortheoccurrenceofspecificeventswouldbeaffected
bytheissuesindataavailabilitydescribedinchapter1.Inaddition,therelevanceof
drugexposuremeasurementandinparticularexternalvalidity isconditionedbya
properselectionof the targetpopulation.Misclassificationbias (i.e. including false
myeloma patients) could introduce confusion and irrelevant results. In
administrative databases, this selection is realized on the basis of case definition
algorithms. The complementary paper then consisted in a validation of the
algorithmusedinthreeprojects(“Identifyingmultiplemyelomapatientsusingdata
fromtheSNIIRAMandthePMSI:validationusingtheTarncancerregistry”).
Integratingaprioriknowledgeinfurtherstudiesa)
The findings of the review on drug data availability were used to support
methodological design of the remaining studies. The issue of immeasurable time
during hospitalization developed through the project presented at chapter 3 is
discussed inthe firstchapterondrugdataavailability.Thisconsiderationhasalso
been integrated during the design of the algorithm for identifying treatment lines
(chapter4).
269
Raising the researchers attention on the impact ofb)
immeasurabletimebias
Thefirstpaperemphasisedtheimportanceofgapsindataavailability,inparticular
duringhospitalstays.Throughtheworkonthe“Impactofunobservabletimebiason
riskestimates”, the thirdpaperhighlighted theneed to takeaccount the impactof
unobservable exposure periods on risk estimates. Furthermore, it underlines the
interest of modelling unobservable periods for a better description of the time
courseofdrugexposure.
Rehabilitating exploratory data analysis inc)
hypothesis‐testing studies: Importance of exploring
observeddata to increase therelevanceofdrugexposure
modelling.
This issue was retrieved in different projects. First, as a complementary
consideration of the project, a method for the identification of gaps in data
availabilityinextracteddataisproposedinchapter1(checklist).Duringtheproject
on benzodiazepines (chapter 2), the interest of testing the relevance of the time‐
dependentapproachusingagraphicalapproachwasdiscussed.
The project on data visualization (chapter 6) further developed this idea. Drug
exposureistraditionallydescribedintermsofnumbersofdrugepisodes,etc.Inthe
context of the analysis of large datasets, it is difficult to account for complex
treatment schemes or discontinuous exposure using conventional descriptive
statistics. Novel strategies for information integration are therefore needed.
Consequently, the chapter 6 was designed to review potential methods for
visualizingdrugexposureepisodesandtodiscusstheircontributionforimproving
study design, hypothesis generation or testing, and reporting in
pharmacoepidemiology (“Data visualization for drug exposure in
pharmacoepidemiology: a case study for complex drug regimens in multiple
myeloma.”).
270
2. Raisingtheresearchersattentionontheimpactofmethods
chosenandbiasaffectingdrugexposuremeasurement
Documentingmethodologicalchoicesa)
Oneillustrationonhowthelackofreportingmayimpacttheusefulnessofastudy
was encountered during the literature review for the project on benzodiazepines
and mortality (chapter 2, page 109). While some studies detailed the drugs
considered (through list of ATC codes or substance names list 90–94,98,101), a non‐
negligiblenumberhavereportedtheexposureofinterestunderthenon‐informative
terms such as “sleeping pills”, “tranquilizers” or sleep related drugs 95,153–155. This
lack of information does not enable to discriminate relevant drug groups:
benzodiazepinesandnon‐benzodiazepinesdrugsforinstance.
Between‐group comparisons as compared tob)
methods integrating the time‐dependent nature of drug
exposure
Thework“Benzodiazepinesandriskofdeath:resultsfromtwolargecohortsstudiesin
FranceandUK” providesan illustrationof the impactofmethodological choiceon
riskestimatesanddemonstratestheinterestoftime‐dependentexposurecompared
to traditional between group comparison. It illustrates that differentmethods for
handling drug exposure are likely to produce different risk estimates, and that
traditional between‐group comparisons should be completed by methods
integratingthetime‐dependentnatureofdrugexposure.
Methods integrating immeasurable exposurec)
periodsversusthoseignoringit
Theworkpresentedatchapter3,“Unobservabledrugexposureduetohospitalization
inmedico‐administrativedatabases:whichimpactforPharmacoepidemiologystudies”
illustrated the differences between methods integrating immeasurable exposure
periodsandthoseignoringit.
271
Impact of computation parameters, a neglectedd)
aspect
Through the multisource study (chapter 2), the impact of computation has been
particularlydiscussed.Indeed,adetailedprotocolisnotsufficienttosetallpossible
choices for the analysts. During this thesis, different tools supporting the idea of
reproducibleresearchwereimplemented.Theconceptofliterateprogrammingwas
developedbyKnuth."Themainideaistoregardaprogramasacommunicationto
humanbeingsratherthanasasetofinstructionstoacomputer."Providingrawdata
does not enable to reproduce the results as statistical codes, minor choices
contribute to variability and divergence of possible results. Thus, tools such as R
Markdown, Sweave 156,157 were used to join statistical outputs with their
correspondingcomputingcode.
3. Role of pharmacological and clinical rationale for
etiologically‐compatiblemodellingandproper interpretationof
theresults
Choosing the drugs of interest: therapeutic anda)
regulatory perspective versus pharmacological approach
(case of other benzodiazepines tetrazepam and
clonazepam)
Thechoiceof thedrugsof interest forthestudyonbenzodiazepinesandmortality
was subjected to adiscussion. Fromapharmacologicalpointof view, all potential
benzodiazepinesshouldbeincluded.However,fromaregulatorypointofview,the
conclusions shouldbemadeon thebasis of the list of approveddrugs for anxiety
and insomnia (ATC codes N05BA‐CD‐CF), thus excluding clonazepam and
tetrazepam which are classified elsewhere (with antiepileptics for clonazepam
(N03)andwithmyorelaxants fortetrazepam(M03B)).Ourproposalwastomodel
separately these drugs and to include them in the analysis. This discussion was
272
illustrativeofthepotentialpitfallsofinvestigationsfocusedonselectedATCclasses
withoutproperpharmacologicaldiscussion.
Choosing a comparator for studyinga)
benzodiazepines‐relatedmortality
Incontrasttomanyothercohortstudies,thestudyonbenzodiazepinesintegrateda
secondcontrolgroupinadditiontonon‐users.Inordertominimizeindicationbias,
non‐usersarenotthebestcontrols.The“bestcontrols”wouldbepatientswhohave
a similar baseline risk, and who would be likely to receive benzodiazepines,
neverthelessdidnot receive it. Finally,usersofnon‐benzodiazepineanxiolyticsor
antidepressants were selected. Antidepressants have distinct indications from
anxiolytics, however, in practice, they are often co‐prescribed, frequently on the
same day. In final, users of antidepressants were hypothesized to be sufficiently
similarwithbenzodiazepineusers.Apropensityscoreadjustmentwasplanned,but
notmaintainedintheanalysisasthecomparabilitybetweenthegroupsrevealedto
bevery satisfying,making thepropensity score adjustmentnotveryuseful in this
case.
A discussion on comparators for the relation between benzodiazepines and hip
fracture could provide useful insight in this area 38. In our study, we focused on
anxiety and hypnotic indications, whereas the “OMOP‐accepted indications for
benzodiazepines included alcohol withdrawal delirium, alcoholism, anxiety
disorders,bipolardisorder,depressivedisorder,duodenalulcer,muscle spasticity,
neuralgiapartialandabsenceepilepsy,panicdisorder,psychoticdisorders,restless
legs syndrome, schizophrenia, sleep disorders, status epilepticus, substance
withdrawal,ticdisorders,andvomiting”.
Theconstraintsforchoosingacontrolgroupwerequitesimilarwithourapproach
(“drugsconsideredascomparatorssharinganindicationwithbenzodiazepines,but
not a mechanism of action”)38. Gruber et al. included midazolam in their control
group, together with other drugs (“hydroxyzine, amobarbital, chlorazepate,
midazolam (a benzodiazepine typically given for single‐dose or very short‐term
use), bromodiphenhydramine, diphenhydramine, methotrimeprazine, Kava
273
preparation, nortriptyline, doxepin, buspirone, chlorpromazine, chlormezanone,
prochlorperazine,andmeprobamate”).
Choosing the risk window and risk function forb)
modelling
(1) Benzodiazepines: accumulated use or current use?
Riskfunctionwasbasedonpharmacologicalpropertiesofthebenzodiazepines.The
issue of all‐causemortality as hypothesized to be a short‐termeffect, in the same
way as the risk function for falls or injuries following benzodiazepines intake.
Conversely,theissueofdementiahasbeeninvestigatedonanaccumulatedduration
basis,inrelationwithdistinctmechanismsofactions.Sometimes,theriskfunctionis
notknown,andshouldbeinvestigatedindedicatedstudies158.
The high crude mortality hazards observed in our study are consistent with
previousfindingsintheliterature,includingthosefromarecentlypublishedcohort
studyusingthesamedatasource(CPRD)94.Althoughanincreasedriskofdeathwas
observedinthetwocohorts,theplausibilityofacausaleffectmustbeconsidered.In
our study, themortality risk was significantly increased earlier after exposure in
newusers inbothsources.Theseresultsaremore in linewitha short‐termeffect
rather than with a cumulative effect, by contrast with results of two recent
studies90,94.Thechoiceoftheoutcomeinourstudyisconsistentwiththeunderlying
pharmacological mechanism of a benzodiazepine‐related acute or sub‐acute
mortality, and could even be shortened in further studies. Actually, high risks
reported with longer use should be attributed to indication bias. Additionally,
decreaseofriskovertimecouldbeexplainedbytolerancetothesedativeeffectof
benzodiazepinesamongsurvivors6,7,94.
(2) Potential drug‐drug interactions and choice of a
relevant overlap duration
274
Inthisstudy,thechoiceofoverlapdurationandperiodofexposureshouldtakeinto
account the residualpharmacological activityafter the lastdayofpatient’s intake.
For instance, in relation with their long duration of action, interactions with
monoamineoxidaseinhibitorsshouldbeconsideredupto15daysafterlastintake.
Thesameproblemneedstobeconsideredforlonghalf‐lifebenzodiazepines(upto
21days).
(3) Type of potential drug‐drug interactions and
clinical relevance
The area of drug‐drug interactions (DDI) is a good model for discussing clinical
relevance of the parameters derived from healthcare databases. Because only
potential(pDDI)andnotactualdrug‐druginteractionswereinvestigated,thechoice
ofthepDDIincludedisevenmorecrucial:toproduceclinicallymeaningfulresults,
thestudyhastofocusonasubsetofDDIwithaknownclinicalimpact.Inthisstudy,
thisissuewasmanagedbystratifyingtheanalysisbypDDItype,andfocusingonthe
descriptionofcontraindicatedandinadvisablecoprescriptions.
Choiceoftheoutcomec)
This area falls beyond the scope of the thesis, but, in the same way as the risk
period/function should be plausible, the outcome considered might also be
discussedinthissense(causeofdeath,cancerandbenzodiazepines).
Studyingmultiplicity: integratingco‐prescribedandd)
concomitantdrugs
(1) Developing a better way to model and report
exposure in oncology: moving from a ‘single drug’
approach toward a ‘multidrug, multiline’ perspective
As stated in the introduction, the pharmacoepidemiology of cancer drugs is
emerging.InFrance,averylittlenumberofworkwasproduced.Thework“Dealing
withcomplextreatmentschemes: identifyingcancertreatmentpatterns inoncology”,
275
accountsformultipleconcomitantexposureandproposesastandardapproachfor
identifyingmultidrugchemotherapyinhealthcaredatabases.
(2) Case of drug interactions: role of cancer drugs,
supportive care and other drugs
Themethodologyalsofocusedonepisodesofmultipleconcomitantexposureswhen
occurrences of specific drug‐drug pairs are of interest (“Analysing longitudinal
exposure to produce automated indicators on contraindicated combinations and
potentialdrug‐druginteractions:ApplicationusingtheFrenchmedico‐administrative
database”). In this study, prevalence of contraindicated drug combinations is
estimatedinapopulationofmultiplemyelomapatients.
Interpretingheterogeneous results: insights gainede)
afterthemulti‐sourcestudyonbenzodiazepines
Inspiteofthesameinclusioncriteriaapplied,populationsincludedinthecohortsin
each country exhibited different demographic and medical characteristics. These
differencescouldbeattributedtonationalpractices.Inaddition,typesofdrugsused
were very different. All these differences were discussed in the light of external
elements from the literature. One of the underlying issues is the impact of the
context.Differentprescribingpracticesledtodifferentpopulationsexposed,butalso
todifferentdrugsprofiles,suggestingthattheseelementsneedtobediscussedusing
bothpharmacologicalrationaleandknowledgeofthestudycontext.
276
277
IX. Perspectives
Résumé15.Perspectives
SUMMARYINFRENCH
Promouvoir l’utilisation des données de l’assurance maladie:
perspectives
Lepremierprojetapermisd’offrirunaperçudesdonnéessurlesmédicamentsau
seindesbasesdedonnéesdel’assurancemaladie,ainsiquelesimplicationspourla
validitéetl’exhaustivitédesdonnées.Encesens,ilfournituneréférenceutilepour
les lecteurs internationaux. Cependant, d'autres considérations méthodologiques
n’ont pu être intégrées dans le cadre d'unepublication internationale (limitations
liées au nombre de mots, adaptation au lectorat international, etc.), telles que
l'identificationexactedunomdestablesoudesvariables.Danslecadredestravaux
de thèse, cet article a constitué un point de départ aux investigations
complémentairessurlespériodesinobservablesousurlareconstitutiondescycles
de chimiothérapie. Cependant, la connaissance préalable de la source de données
représenteseulementuneétapedeladémarched’analysedesdonnées,etbeaucoup
d’élémentssupplémentairesnécessitentd’êtreprisenconsidération.
Danslecadrededéveloppementsultérieurs,ilestprévuderédigerdesdocuments
pour conduire une analyse de données dans les bases de données de l’assurance
maladie.L'objectifestdefournirunesérieétenduedepointsdecontrôlessimples,
mais systématiques, qui pourraient prévenir la survenue d’erreurs ou de biais
(spécificationpourl'extractiondedonnées,vérificationdesdonnéesextraites,oula
définition d'un ensemble minimal d’indicateurs d'exposition à rapporter dans le
cadre d’une étude longitudinale). Selon K. Fairman159, la majorité des erreurs
survenant dans le cadre de l’analyse de données secondaires sont évidentes, et
pourraient être évitées en utilisant de simples tableaux de fréquences. De plus,
commediscutéaucoursduchapitre6(page210),desoutilsvisuelspourraientêtre
278
d'un intérêt particulier pour ces étapes de validation. Ainsi, ces divers outils
pourraientêtreréunisdansuntableaudebord,quirésumeraitlescaractéristiques
de l’exposition au niveau populationnel, et pourrait permettre d’identifier
d’éventuelsproblèmes liésauxdonnées.Les futurstravauxautourde laqualitéde
l'analysedesdonnéesdel'assurancemaladieprévoientainsid’intégreràlafoisdes
considérationsméthodologiquesetdesétapesd’explorationdesdonnées.Uneveille
seraeffectuéepourprendreencomptel’apportdesinitiativesexistantesetàvenir,
commeparexemplelesapportseffectifsduréseauREDSIAM160danslapromotion
d’uneutilisationrationnelledesdonnéespourl’identificationdepathologiesoudes
évènements, et ce, en fonction des objectifs poursuivis (enquête de prévalence,
risque et algorithmes de définition des incidents, etc.). Une attention toute
particulièreseraégalementportéeausupportproposéàl’issuedel'enquêteréalisée
par l'INSERM sur les attentes des chercheurs concernant les données en santé.
L’INSERMprévoiteneffetlacréationd’uneinfrastructuredeservice,accompagnée
d’un «support et un partage de documentation sur les aspects réglementaire,
juridique, éthique, technico‐scientifique, et en data management et système
d’information»161.
Impactdes choixméthodologiquesetdesbiaisaffectant lamesurede
l'expositionaumédicament:développementsprévus
Projetencourspourfaciliterlapriseencomptedespériodesinobservables
L’étude présentée au chapitre 3 a permis de mettre en évidence l’impact des
périodes inobservables. Cependant, elle ne fournissait pas de solution générique
pouraideràprendreencomptecebiaisauseindesbasesdedonnéesdel’assurance
maladie.
UnprojetencoursconsistedoncàmettreaupointunensembledeprogrammesSAS
pourcomblercedéfautetintégrerunensembled’étapesprédéfiniespoureffectuer
un diagnostic de l’ampleur des périodes inobservables au sein d’une base de
données et faciliter la mise en œuvre d’analyses de sensibilité pour contrôler ce
279
biais.Cetoutilfourniraitdesstatistiquesdescriptivessurlespériodesinobservables.
Sur la base des informations fournies, les utilisateurs pourraient alors évaluer
l'exposition au traitement et ses effets sous diverses hypothèses et d'examiner
l'impact potentiel sur les résultats de l'étude. La macro permettra également
d'examinervisuellementlestrajectoiresindividuellesdepatients.
a) Evaluer l’impact des périodes inobservables dans d’autres schémas
d’étude
Au cours de la première étude de cas sur les périodes inobservables, l’étude
d’impact a été effectuée sur un schéma d’étude de type suivi de cohorte. Or, la
problématiquepourraitserévélerlégèrementdifférentedanslecasdesétudescas‐
témoins.Nousprévoyonsdoncd’étudiercettequestionàpartirdesdonnéesd’une
autre étude de type cas‐témoins niché, en prenant comme modèle le cas de la
survenue d’infections sévères suite à l’exposition aux corticoïdes dans la
thrombopénie immunologique. En effet, à l’issue des premières investigations
menées158,laprésenced’unehospitalisationdansles7joursprécédantladateindex
était associée à la survenue d’une infection sévère, rendant ce contexte
particulièrementintéressantpourétudierl’impactdespériodesinobservables.
A. Pourunemodélisationpluspertinente:perspectivessurlaprise
encomptedesexpositionsconcomitantes
1. Développementd’uneapprochepourmodéliseretdécrire
l’expositionenoncologie
Dans un article publié en 2010, Turesson et al. 71 mettait déjà l’accent sur les
difficultéscroissantesd’établirdescomparaisonsdesurvie fiablescompte tenude
l’accroissement de la diversité des médicaments et des protocoles de
chimiothérapies proposés, et de la variabilité des séquences selon les patients.
Pouvoirdisposerdel’historiquecompletdeslignesdetraitementreçupourraitêtre
particulièrement contributif dans la perspective d’études d’efficacité comparative
280
(ComparativeEffectivenessResearch,CER).Dansledomaineducancer,onretrouve
lesinitiativesnord‐américainesissuesdelabaseSEER.Desétudescomparativesont
par exemple été menées dans le lymphome B 162. Pour ce qui est des bases de
donnéesfrançaises,lesexpériencesd’efficacitécomparativerestentencorelimitées
163,164.
Al’aidedestrajectoiresreconstituées,onpeutégalementenvisagerdedisposerd’un
outilprécieuxd’étudedespratiquesetd’étudiercertainseffetsàlongterme,comme
la survenue de seconds cancers après exposition au lénalidomide. La possibilité
d’étudesmédico‐économiquesselonlestrajectoirespeutégalementêtreenvisagée,
àl’imaged’initiativedéjàmenéepourlemyélome165.Lespossibilitésoffertesparce
typededonnéesetlagénérationde«fulldiseasemodels»sontégalementillustrées
parlapublicationdeCidRuzafaetal72.
a) Après l’identification:prendre en compte les trajectoiresd’exposition
dansl’analyse
Leprojetprésentéauchapitre4permetdedémontrerlafaisabilitédereconstituer
descyclescomplexesdetraitementenhématologieàpartirdesdonnéesduDCIRet
duPMSIMCO. Iloffreainsi lapossibilitédereconstitueravecunniveaude finesse
très important les lignes de traitement reçues. Une des questions qui se pose est
l’intégrationde ces trajectoirespourmodéliser l’expositiondans le cadred’études
étiologiques. Cette intégration est hors du champ des travaux de thèse, mais on
pourra cependant citer plusieurs travaux utiles pour résoudre cet aspect. Un
exemple de stratégies analytiques possibles pour prendre en compte les
changementsdelignesdanslecadredumyélomepeutainsiêtrecité166.
Danslecadredesétudesd’efficacitéenvieréelle,l’intérêtdesmodèlesd’équations
structurelles,quipermettentd’ajusterlesrelationsdynamiquesentrelesdifférentes
lignesdetraitement,estégalementrappelé167.
b) Identifier les lignes de chimiothérapies: quelle transférabilité aux
autrescancersethémopathiesmalignes?
281
La question de la transférabilité de cet algorithme est essentielle. C’est la
représentation des médicaments de rétrocession ou en sus dans les différents
protocoles de chimiothérapie qui va déterminer la capacité à discriminer ces
protocolesentreeux,latransférabilitén’estaprioripasassuréepourl’ensembledes
cancersetdeshémopathiesmalignes.
L’adéquationpeutêtrevérifiéeàpartirdelalisteetdelanaturedesprotocolesde
chimiothérapie indiquées dans l’affection d’intérêt. Un protocole de vérification
permettraitdeconfronterlalistedesmédicamentsàlalistedespécialitéensusou
de rétrocession, et de confirmer ou non leur caractère observable lors des
hospitalisations.Une listedescombinaisonspouvantêtredistinguéepourraitalors
être établie. En fonction de l’étendue des protocoles pouvant être discriminés ou
non, l’investigateurpourraitprendre ladécisiondechercherounonàreconstituer
les trajectoires de traitement de chimiothérapie reçues dans l’affection d’intérêt à
partirdesbasesdedonnéesdel’assurancemaladie.
2. Reconstituer les épisodes de concomitance: perspectives de
développementdel’outildedétectiondesinteractionsmédicamenteuses
L’outil d’identificationdes interactionsprésenté au chapitre5 (page175)offreun
cadregénéralpourlamiseenœuvred’étudessurlesinteractionsmédicamenteuses
àpartirdesbasesdedonnéesde l'assurancemaladie.Lemodèleutilisénepermet
sans doute pas demontrer tout le potentiel d’identification en lien avec la faible
diversité desmédicaments anticancéreuxutilisés.De futures études conduites sur
des échantillons plus larges (extractions nationales) et portant sur des classes
potentiellement plus pourvoyeuses d’interactions (inhibiteurs de tyrosine kinases
parexemple)pourraientêtrepertinentesetgénérerdesdonnéesdansuncontexte
peu exploré des pratiques de prescriptions chez les patients exposés aux
médicamentsanticancéreux.
Defaçongénérale,àpartirduthesauruscomplet,desétudesdédiéespourrontêtre
conduites sur des populations ciblées, accompagnées éventuellement d’une
282
recherche de la survenue d’événements spécifiques, permettant d’apporter des
éléments qualitatifs réellement contributifs. Les résultats générés par cet outil
pourraient permettre d’accroître les connaissances concernant les interactions
médicamenteuses. Il est prévu de mettre à jour cet outil chaque année, et de le
rendre disponible sous une forme permettant une traçabilité très fine des
modifications apportées (https://github.com/), de lamême façonque cequi a été
proposé par C. Le Cossec et A. Filipovic‐Pierucci pour les indicateurs de
polymédicationdansleSNIIRAM168,169.
283
A. Knowledgeofsecondarydatasourcesanddataexploration:perspectivesandfutureresearch
The firstworkof this thesisoffereda comprehensiveanalysisofdrugdatawithin
health insurance databases, together with implications for drug exposure
completenessandstudyvalidity,inthecontextofthegrowinguseofsecondarydata
sourcesforpharmacoepidemiologicalresearch.Inthissense,itshouldbeusefulfor
presentingthedatabasetointernationalresearchers.
However,furthermethodologicalconsiderationscouldnotbedescribedoranalysed
properly in conventional articles and in the context of an international peer‐
reviewedpublication(word limits,adaptation toan internationalreadership,etc.).
Furthertechnicalelementswouldhavebeennecessarytocompletethisoverview,in
particularfortheFrenchreadership,suchasidentificationofthetablename,exact
labelofthevariables,etc.Theseelementswereprovidedinthethesismanuscript.
In final, in the context of the thesis, it represents an important basis, but prior
knowledgeofthedatasourceisonlyoneofthestepsofthedataanalysisworkflow,
andalotofadditionalelementshavetobeconsidered.
Then, in line with the first work on data sources, further elements would be of
interest to reduce potential for errors and increase the transparency and
reproducibility of database studies. As a further development, writing of quality
controldocumentisplanned.Thefinalobjectiveistoprovideanextendedseriesof
simple but systematic checks that could prevent essential of errors. This element
wouldincludeforinstanceworkingondefiningcleardataextractionspecifications
(whatdoescriteriasuchas“alllong‐termconditionforthecodeXX.Xfrom2010to
2015” mean exactly?), checking incoming data, or defining a minimal set of
indicatorsofexposuretobereported.
284
Itcouldbearguedthatsucherrorsareimprobable,but,asdevelopedbyK.Fairman,
the majority of errors are obvious, and might be avoided by using simple cross‐
tabulation checks. In the “Guidelines for Good Database Selection and use in
Pharmacoepidemiology Research”, Hall et al. have integrated a checklist for
investigatorsindatabaseresearch.Thischecklistincludedasectionon“qualityand
validation procedures” andmentioned the need for quality checks. This principle
hasalsobeenstatedbyHennessySetal.170,who,afterconductingananalysisofthe
integrity of US Medicaid claims databases, concluded that “Whenever possible,
investigators using administrative data should perform macro‐level descriptive
analyses on the parent data set. In particular, researchers should examine the
numberofmedicalandpharmacyclaimsovertime,lookingforgaps”.
Asdemonstratedinthesectionon“Ascertaininglongitudinalavailability”ofchapter
6(page225),visualtoolscouldbeofparticularinterestforthesevalidationsteps,
and further work on the quality of data analysis workflow using French health
insurancedatabasewillintegrateasetofvisualtools.
Thus, these various tools could be combined in a dashboard that summarizes the
characteristicsofdrugexposureatthepopulationlevel,andcouldidentifypotential
problems.Futureworkaround thequalityof theanalysisofhealth insurancedata
and data will integrate both methodological considerations and steps of data
exploration.
Particularattentionwillbemadetotheintegrationofexistingandfutureinitiatives,
suchastheactualcontributionsofREDSIAMnetwork160inpromotingthevalidation
and reporting of cases findings algorithms, adapted to different objectives
(prevalencesurvey,definingincidents,etc.).Particularattentionwillalsobepaidto
the proposed support after the survey conducted by INSERM on researchers'
expectationsonhealthdata. Indeed, INSERM isplanning to establish a service for
sharing of documentation on regulatory aspects, legal, ethical, scientific, technical,
anddatamanagementandinformationsystems"161.
285
B. Impactofmethodschosenandbias:developmentsplanned
1. Future research on integrating immeasurable exposure
periods
Through our case study, we demonstrate how assumptions concerning inpatient
exposuretoaccountforperiodsofimmeasurabletimecanimpactriskestimateina
cohort study. The bias generated by the failure to take account of these
immeasurableperiodsmaybeproblematicinstudiesfocusedonlong‐termexposure
or on chronic diseases requiring hospitalization 56. As our strategywas based on
casestudy,itraisesawarenessonthepossibleimpactofunobservabletimebiasbut
doesnotprovideageneralanswertothisissue,andtheimpactinothercontextsis
pronetovary.
In further work, we will then try to develop a framework for identifying and
modelling these periods in the particular context of French health insurance
databases.
The objective is to provide a framework for accounting immeasurable time in
Frenchhealth insurancedatabasethroughthedevelopmentofaSASmacro.These
SAS programs are intended to fill this gap andwill integrate a set pre‐computed
stepsforprovidingadiagnosisofthemagnitudeofunmeasurabletimeinthedataset
and facilitating the implementation of sensitivity analyses to control this bias is
currently under development. The macros compute descriptive statistics
immeasurableperiods,aswellasthenumberofperiodsduringfollow‐up.Basedon
the macros’ output, researchers will assess treatment exposure and treatment
effectsundervariousassumptions56andexaminepotentialimpactonstudyresults.
Themacroalsopermitstovisuallyexamineindividualtrajectories.
To use themacro, the followingmacro‐variables should be set by the user: study
startandstudyend(dateformat), thenameof thedatasetcontainingdrugdataof
286
interest(withdrugsidentifiedusingATCorCIPcode)andthenameofthedataset
forhospitalepisode(PMSIMCOorappendedepisodes).Thedatasetofdrugsshould
containstartandendofthedrugepisode(mightbesetto30daysbydefault),and
specify thenameof thesevariables.Userswillalsoneedtospecify thenameofan
additionalSASdatasetcontainingthefollowingvariables:patientsID,indexdate,an
eventindicator(1=event,0=censoring),thenameofthegroupvariable(drugclass
forinstance).
Table5.Proposedstepsforimplementingastrategyforaccountingfor
immeasurabletimebias.
Proposedstepsforimplementation
Checkdrugofinterestagainstlistofcostlydrugstoconfirmitsstatus
(immeasurable)
Createanuniquedatasetwithdrugandhospitalepisodes
Dealwithduplicate,embeddedoroverlappinghospitalepisodes
Createmutuallyexclusivesequences
Displaydescriptivestatistics(personyears)
Differentapproachestoaccountforunobservableexposuretime.
generateflagsforexcludingpatientswithunobservabletimepriortoindex
date
adjustforimmeasurabletimeorcensoranalysesatthestartofimmeasurable
exposureperiod(hospitalentry)
computethenumberofobservabledaysandnumberunderprespecified
threshold
Usethenumberofobservabledaysasacovariate.
makeassumptiononexposureduringhospitalization
287
2. Investigatingimpactofimmeasurabletimebiasinanother
design
Inaddition to the first investigationusinga cohortdesign,wewilluseamodelof
nested case control study and investigate whether corticosteroid risk function of
severe infection in primary immune thrombocytopenia adults is impacted by
immeasurabletimebiasinhospitalization.Inthefirstanalysesimplemented158,the
occurrence of a hospitalization of at least 7 days between start of follow‐up and
index date was independently associated with severe infection occurrence, thus
makingthiscontextappropriatetoconductanothercasestudy.
C. Etiologically‐compatible
modelling:furtherperspectiveson
integratingconcomitantdrugs
1. Developingabetterway tomodelandreportexposure in
oncology
According to Turesson et al. “Given that the treatment strategies for MM are
currentlychangingandnewertherapiesarecommonlyusedatdiseaseprogression,
mostpatientswithMMwilleventuallyreceiveallavailablenoveldrugs;mainly,the
sequence of different regimens will vary. Consequently, it will become harder to
establish survival differences between defined induction, consolidation, and
maintenancetherapiesinthefuture71”.Theabilitytobuildcompletedrughistoryof
drugregimenswouldbethenverycontributiveregardingthisissue.
Thealgorithmdescribedatchapter4isintendedtobeusedinfurthercomparative
safetyoreffectivenessresearchortoinvestigateissuesontreatmentandsurvivalin
multiplemyeloma patients, taking account of the nature and number of previous
treatmentlinesandtreatmentduration.Usingthesetrajectorieswillenabletobuild
288
full‐diseasemodelincreasetheknowledgeofcurrentpractices,ortostudydelayed
outcomes(e.g.issueoflenalidomideandsecondarymalignancies).Medico‐economic
studies should also benefit from this work. The capabilities of such data are
illustratedbythestudyofdiseasetrajectoryinmultiplemyeloma72.
2. After the identification: taking account of complex drugs
regimensinoncology
Oneprojectwasdedicatedtoamethodforidentifyingchemotherapyregimensand
thenoffersthepotentialofbuildingwholepatientstrajectories.Theinterestofthis
knowledgeisreal.Anextstepwouldbetointegratethesetrajectoriesinstatistical
analysis.The thesisdoesnotdevelop thispoint,butmarginal structuralmodelling
would be of particular interest to adjust for the “dynamic relationship between
durationoftimeondrug(s),confounders,andoutcomes”167.Anexampleinmultiple
myeloma is provided byKalinjuma 166,who estimated the effect of chemotherapy
regimens in patientswithmultiplemyeloma, taking account of treatment changes
(cross‐overbias).
3. Transferabilitytoothermalignancies
The question of transferability to other malignancies is of particular interest.
According to therepresentationofcostlydrugswithinthe"horsGHS"schemeand
theabilitytodiscriminateregimens,thetransferabilityisnotaprioriensuredforall
malignancies,andmustbehandledonacase‐by‐casebasis.Thestepsforidentifying
treatmentlinesinclaimsdatabaseisinsertedTable6.
Table6.Stepsforidentifyingtreatmentlinesinclaimsdatabase
Stepsforidentifyingtreatmentlines
Reviewrelevantrecommendationstolistalldrugswithanindicationinthe
diseaseofinterest
Checklegalstatus(andconditionsfordispensing)andpresencewithinthe
database
289
Stepsforidentifyingtreatmentlines
Createapatient‐daydatasetcontainingonlyfullyobservabledrugs
Considerthepossibilityofgroupingdrugthatcouldbeadministeredindifferently
(e.g.corticosteroids)
Setaperiodofeligibility(28days*),aperiodtodefinegap(90days*)and
determinegraceperiodsforalldrugs
Applythealgorithmondataaggregatedbypatients,dayanddrugordruggroup
Startwiththefirstdrugdispensing
Alldrugswithintheperiodofeligibilitywouldbegintothecurrentline
Ifnodrugisrefilledafterthemaximumperiod(90days),declaretheendof
theline
Declareanewlineafteralinegaporifanewdrugisintroducedoutsidethe
graceperiodofdrugsenteringinthepreviousline
Considerthepossibilityofdeclaringanewlineonlyifdurationofoverlap
withnewlyintroduceddrug(s)issufficient(7days*)
Applythisalgorithmuntilallthedrugsinthedatasethavebeenprocessed
Examineunknowncombinationsandconsiderapossibleoverlapbetweenlines.
Investigateconsistencywithcurrentpracticeusingexpertadvice(oncologist)and
externalsourcesifavailable
*Sensitivityanalysesshouldbeconductedusingdifferentdurations
Totestapplicabilityinotherdiseasescouldbecomputerized.Minimalrequirements
wouldincludenatureandcompositionofrecommendedchemotherapyregimens.A
proposal of parameters for a systematic assessment of chemotherapy building
algorithmsonthebasisofalistofrecommendedregimensisproposedtable7.
290
Table 7. Planned parameters for a systematic assessment of chemotherapy
buildingalgorithmsonthebasisofalistofrecommendedregimens
Parameters
User‐defined
parameters
Listalldrugsinvolvedindrugregimens(ATCi)
ListofallrecommendedcombinationsATC1‐ATCi
Outputs Test ATC i against list of costly drugs to assign status for
exposureduringhospitalization(immeasurable)
Listalldistinctcombinationsthatcouldbeidentified
Summary
statistics
Process the database and output exposure indicators: number
oflines,cycles,numberofnon‐standardcycles,timetostop
individualandaggregatedprofiles(flowdiagrams)
4. Studyingmultiplicity: perspectives for implementing the
methodforidentifyingdrug‐druginteractions
Thetoolpresentedatchapter5offersageneralframeworkforimplementingdrug‐
druginteractionstudiesinFrenchhealthinsurancedatabase.Themodeluseddoes
notprovideahighprevalenceofinteractionduetotherelativelylowrangeofdrugs
used. Studies on cancerpatients andmaybeon classeswith ahigherpotential for
interaction(TyrosineKinaseInhibitors)mightberelevant.
Existing compendium could be refined in order to select only relevant (“high
priority”, “clinically significant”) or unlisted drug‐drug pairs of particular interest.
These tools could also be used as a first step when studying the occurrence of
specific outcomes (drug‐related hospitalizations, death, etc.).Thiswork is likely to
facilitate further researchonDDIs through automated computation and adaptable
tools. Outputs of DDIs exploration are intended to increase knowledge and raise
291
awareness of different stakeholders on concomitant use of contraindicated
medication combinations, andmaybe applied forprescribingquality surveillance.
We plan to update the compendium yearly, and make it available for other
researchers. The steps for adapting the compendiumof interaction for automated
detectioninclaimsdatabasesaresummarizedTable8.
Table 8. Steps for adapting the compendium of interaction for automated
detectioninclaimsdatabases
Mainpointstobeconsidered
Qualitycontrolchecksperformedforadaptingthecompendium
Coherenceofthesourcenumberofpairsandnumberinthedescription
AllpDDIhaveadescriptionattributed
AllpDDIcouldbeclassifiedinto“contraindicatedcombinations”,“inadvisable
combinations”,“Precautionsforuse”or“Combinationstoconsider”,someofthem
belongtoseveralcategories
ForATCcoding,assignalltheATCcodesforwhichtheactivesubstanceisincluded
Refertothethesaurusofclassestoaccessallindividualsubstances
IdentifyingpDDIs
Createapatient‐daydatasetcontainingonlyobservabledrugs
Considerthepossibilityofgroupingdrugsofinterest
Computethenumberofdrugdaysdispensed
Setagraceperiod
Setanadditionalperiodtoaccountforresidualpharmacologicalactivity
Applythealgorithmondataaggregatedbypatientsandoverlappingsequence
(applyonlyinsequenceswithatleasttwo2distinctdrugs)
Applythisalgorithmuntilalloverlappingsequencesinthedatasethavebeen
processed
InvestigatingparticularpDDI
ReviewrelevanceofthepDDI
292
Mainpointstobeconsidered
LimitthecompendiumtothepDDIofinterest
Considerthepossibilityoftestingonlysomerouteofadministration
Considerthepossibilityoftestingonlydrugsforsomedosageonly
Checklegalstatus(andconditionsfordispensing)andpresencewithinthedatabase
InvestigatetherelevanceofthepotentialDDIretrievedusingexpertadvice(clinical
pharmacologist)andexternalsourcesifavailable
*Sensitivityanalysesshouldbeconductedusingdifferentduration
293
294
295
X. Recommendations
Résumé16.Propositionderecommandations
SUMMARYINFRENCH
Une des lignes directrices de ce travail a consisté à énoncer les leçons tirées des
diversesanalysessousformedeprincipesméthodologiquesplusgénéraux,detelle
sortequ’ilspuissentêtreplusfacilementtransposésdansdescontextesproches.
Aucoursdupremiertravailderevuedesdonnéesdumédicament171,celaaprisla
formed’unelistedespointsàvérifierpours’assurerdeladisponibilitélongitudinale
des données. De la même façon, le projet de reconstitution des lignes de
chimiothérapieinclutunetabledétaillantlesétapesnécessairespourconduirecette
reconstitutiondansd’autresaffections.Ceprincipeestégalementretrouvéausein
duprojetsur les interactionsmédicamenteuses,par lebiaisd’untableauretraçant
les points essentiels à considérer pour utiliser l’outil proposé, l’adapter un autre
compendiumoueffectuerunciblagesurdesinteractionsd’intérêt.
Conformément à cet objectif, cette thèse se termine par une série de
recommandationsdanschacundes3axesidentifiésàtraverslesdifférentsprojets.
1. Développer une connaissance approfondie des sources de données et
réhabiliterunephased'explorationdesdonnéeslongitudinales
S’assurer une bonne connaissance de l'origine de la source de données, de
sonmoded’alimentation,etdesoncontenu.
Examiner attentivement les médicaments d'intérêt pour identifier tout
problèmepotentieldansladisponibilitédesdonnéeslongitudinales.
Effectuer une analyse exploratoire des données brutes (outils graphiques)
pouridentifierlesrupturesinattenduesdansladisponibilitédesdonnées:
– Si aumoins un desmédicaments d’intérêt ne figure pas sur la liste des
spécialités en sus (statut par rapport à l’exposition non disponible au
cours des hospitalisations), considérer la possibilité de planifier une
296
méthode afin de quantifier l’ampleur de ce biais lié aux périodes
inobservablesettenterdeleprendreencompte.
– Envisager la possibilité d'utiliser des méthodes de visualisation de
donnéespourmieuxcomprendrelesmodalitésd’utilisationenvieréelle.
Effectuer une analyse exploratoire sur un sous‐ensemble de la base de
donnéespour confirmer lapertinencede lamodélisationde l'expositionau
médicament(expositionchroniqueouponctuelle).
Envisager lapossibilitéd'utiliserdesméthodesdevisualisationdedonnées
pourexploreretdécrirelesmodalitésd'expositionaumédicament.
2. Prendre desmesures appropriées pour réduire l'impact
desméthodesutiliséesdanslecadred’étudeslongitudinales
Envisager la possibilité d'utiliser plusieurs méthodes pour catégoriser
l'exposition ou pour estimer les paramètres de risque au sein de lamême
étude.
Documenterl’ensembledescodesdemédicamentsutilisés,lestablessources
(exemple:PMSIMED,UCD), ainsique lesméthodesdétailléespourdériver
lesdosesetlesdurées.
S’appuyer surunoutil commeRECORDpour rapporterde façon structurée
lesétudessurlessourcesdedonnéessecondaires:
o Ne pas négliger l'impact des étapes de calcul et 'envisager la
possibilité d'utiliser des outils permettant de faciliter la
reproductibilitédesanalyses(Knitr,Sweave).
o Prévoir la réalisation d’analyses de sensibilité dès lors qu’un
paramètre est basé sur un choix de l'investigateur (fenêtre de
risque,périodedegrâce,etc.).
o Évaluerlarobustessedesestimationsselonlesméthodesutilisées
etexplorerladirectiondesrésultats.
3. Interroger le rationnel pharmacologique ou clinique pour chaque choix
297
méthodologique
Lorsquelasélectiondesmédicamentsd’intérêtestbaséesurlaclassification
ATC, accorder une attention particulière aux médicaments avec un
mécanismesimilaireclassésailleurs.
Dans le cas d’une étude ciblant une classe demédicaments, discuter sur la
base des connaissances pharmacologiques de la possibilité d’un potentiel
effet différentiel entre les substances individuelles, et planifier un moyen
d'explorer cette question (analyses complémentaires au niveau de la
substanceactiveparexemple).
Lorsde la conceptiondesgroupes,envisager lapossibilitéde constituerun
groupe de comparaison actif en plus du traditionnel groupe constitué de
patientsnonexposés(réductiondubiaisd’indication).
Lorsduchoixdelafenêtrederisque,examinerattentivementlajustification
pharmacologique:
Éviterd’adopterunefenêtrederisqueàlongtermelorsqu’uneduréeplus
courteseraitpertinente(sélectiondessurvivants).
Préférerl'utilisationd’uneexpositiondépendantedutempslorsquetoute
exposition (quelle que soit la dose reçue) reste compatible avec la
survenue de l’évènement d’intérêt (hypothèse retenue pour les
benzodiazépines).
Compléter l'analyse avec une évaluation dose‐effet afin d'obtenir des
argumentssupplémentairespourdiscuterl’aspectcausaldel’association.
Envisager lapossibilitéd'utiliserdesméthodesplus flexiblespourrelier
l'expositionaumédicamentaveclerésultat.
Envisagerlapossibilitéd'incluredesmédicamentsconcomitantsd'intérêt
sousformedépendantedutemps.
Catégoriseretdécrirel'expositiond'unemanièrecliniquementpertinente
(protocoles de chimiothérapie en oncologie, catégories soigneusement
choisiespourlesdoses).
Faceàdes résultatshétérogènesentredes sous‐groupesoudes sources
298
de données, rechercher des facteurs potentiellement explicatifs (statut
légal,recommandationsetpratiquesnationales,etc.).
Recommendations
A general objective of this thesis was to formulate lessons learned and general
methodological principles in such a form that it will help other researchers.
Whenever it was relevant, a table summarizing methodological principles was
includedinthearticle,insuchaformthatitwillhelpotherresearchers.Inthearticle
onoverviewofdrugdataavailability171,ittakestheformofastructuredchecklistto
identify problems with data availability in SNIIRAM databases. In the project on
chemotherapyregimens identification, the “steps for identifying treatment lines in
claimsdatabase”wereinsertedinthemanuscript.Inthesameway,themanuscript
ondruginteractions,themainpointstoconsiderwhenadaptingthecompendiumof
interactionforautomateddetectioninclaimsdatabaseswereprovided.
In linewiththisobjective, thisthesisconcludeswithasetofrecommendationsfor
researchers,fallinginto3mainareasidentifiedthroughthedifferentprojects.
1. Develop a deep knowledge of the data source and rehabilitate a
properdataexplorationphaseoflongitudinaldata
o Ensureaproperknowledgeofthedatasourceorigin,content
o Carefully consider the drugs of interest to identify any potential
issueinlongitudinaldataavailability
o Make an exploratory phase (graphical tools) on raw data to
identifyanyunexpectedgapsindataavailability
o Whenfocusingondrugsnotrecordedduringhospitalizations,plan
amethodforhandlingwithimmeasurabletimebiasandreportthe
magnitudeofimmeasurabletimeascomparedtopatients’follow‐
up.
299
o Considerthepossibilityofusingdatavisualizationmethodstogain
insightintoreallifepatternsofuse
Makeanexploratoryphaseonadatabasesubsettoconfirm
therelevanceofdrugexposuremodelling(chronicorpoint
exposure?)
Considerthepossibilityofusingdatavisualizationmethods
to explore and report longitudinal patterns of drug
exposure
2. Takeappropriatemeasuresforreducingthe impactofthemethods
usedforhandlinglongitudinaldrugexposure
o Consider the possibility of using a combination of different
methods for categorizing exposure or for estimating risk
parameters
o Document all drug codes used, including detailed methods for
derivingdosesandduration
Consider the possibility to use a tool like RECORD for
reportingstudyonsecondarydatasources
o Donotneglect the impactof computation stepsand consider the
possibilityofusingreproducibletools(Knitr,Sweave)
o Made sensitivity analyses whenever a parameter is based on
investigator’schoice(gapduration,etc.)
o Estimate changes/robustness of the estimates according to the
methodsusedandexploredirectionoftheresults
3. Carefullyconsiderpharmacologicalorclinicalrationale formaking
methodologicalchoices
o WhenselectinglistofdrugcodesbasedonATCclassification,pay
particular attention to drugswith a similarmechanism classified
elsewhere
o When focusing on a drug class, carefully discuss the potential
differential effect of individual drugs based on pharmacological
300
knowledge and plan away to explore this issue (complementary
analysesattheactivesubstancelevel)
o When designing study groups, consider the possibility to include
an active comparator group in addition to the traditional “non‐
user”group
o When setting the risk window, carefully consider the
pharmacologicalrationale
Avoidingalong‐termriskwindowwhenashorterendpoint
wouldberelevant(selectionofsurvivors)
Prefercurrentusewhenanyexposure(atanydose)might
be sufficient to cause the outcome (example of
benzodiazepines)
Completetheanalysiswithdose‐effectassessmentinorder
togainadditionalargumentsfordiscussingcausality
Consider the possibility to use more flexible methods for
linkingdrugexposurewiththeoutcome
o Considerthepossibilityofincludingconcomitantdrugsofinterest
inatime‐dependentway
o Categorize and report drug exposure in a clinically relevantway
(drugregimensinoncology,carefullydesigneddosecategories)
Instudies investigatingdruguse,drugexposureshouldbe
reported in a way that is consistent with the
recommendationorcurrentmodalitiesofuse.Forinstance,
indiscontinuousdrugexposure,numberofintakecouldbe
an appropriate indicator (migraine and triptans use for
instance), whereas computing mean dose per month has
probablylittlesense.
o Faced to resultsheterogeneityamongsubgroupsordatasources,
consider potential explanatory factors (legal status,
recommendationsandnationalpractices,etc.)
301
302
303
XI. Conclusion
Résumé17.Conclusiongénérale
SUMMARYINFRENCH
Lesconclusionsdesétudesobservationnelles semontrent très sensiblesauxchoix
méthodologiques,enparticulieràlafonctionetàlafenêtrederisqueretenuespour
modéliserl’exposition,maisaussiàcertainsbiaisrarementprisencompte(périodes
inobservables).Enmettantl’accentsurl’importancedelaconnaissancedesdonnées
et du rationnel pharmacologique dans la modélisation, et en développant des
approches alternatives pour la prise en compte des expositions multiples, ces
travaux contribuent à accroître la pertinence et la robustesse des études
longitudinales conduites à partir de bases de donnéesmédico‐administratives, en
particulierdanslecasd’expositionsmédicamenteusesmultiplesetdiscontinues.
Conclusions are highly sensitive to methodological choices. By promoting prior
knowledge of the data sources and the implementation of simple but robust
methods, but alsoby reminding the central role of pharmacological rationale, this
thesiswas intended to improve the validity and the robustness of drug exposure
measurementinmedico‐administrativedatabasesinthecontextoflongitudinaland
multipleconcomitantexposures.
304
305
XII. Bibliographie
1. van Walraven C, Davis D, Forster AJ, Wells GA. Time‐dependent bias was common in survival analyses published in leading clinical journals. J Clin Epidemiol 2004; 57: 672–682. doi:10.1016/j.jclinepi.2003.12.008.
2. Abrahamowicz M, Beauchamp M‐E, Sylvestre M‐P. Comparison of alternative models for linking drug exposure with adverse effects. Stat Med 2012; 31: 1014–1030. doi:10.1002/sim.4343.
3. Sylvestre M‐P, Abrahamowicz M. Flexible modeling of the cumulative effects of time‐dependent exposures on the hazard. Stat Med 2009; 28: 3437–3453. doi:10.1002/sim.3701.
4. Bazelier MT, Eriksson I, de Vries F, et al. Data management and data analysis techniques in pharmacoepidemiological studies using a pre‐planned multi‐database approach: a systematic literature review. Pharmacoepidemiol Drug Saf 2015; 24: 897–905. doi:10.1002/pds.3828.
5. Andrews EB, Margulis AV, Tennis P, West SL. Opportunities and Challenges in Using Epidemiologic Methods to Monitor Drug Safety in the Era of Large Automated Health Databases. Curr Epidemiol Rep 2014; 1: 194–205. doi:10.1007/s40471‐014‐0026‐0.
6. Maguire A, Blak BT, Thompson M. The importance of defining periods of complete mortality reporting for research using automated data from primary care. Pharmacoepidemiol Drug Saf 2009; 18: 76–83. doi:10.1002/pds.1688.
7. Horsfall L, Walters K, Petersen I. Identifying periods of acceptable computer usage in primary care research databases. Pharmacoepidemiol Drug Saf 2013; 22: 64–69. doi:10.1002/pds.3368.
8. ENCePP. ENCePP Work Plan 2015 ‐ 2016. Available at: http://www.encepp.eu/publications/documents/ENCePPWorkPlan2015‐2016_rev1.pdf. Accessed August 5, 2016.
9. The European Network of Centres for Pharmacoepidemiology and Pharmacovigilance (ENCePP) survey of methodologies for European Union publicly funded multi‐database safety studies. Current practice in European Union multi‐database pharmacoepidemiology research (EMA/651167/2014). Available at: http://www.encepp.eu/publications/documents/Survey_Multi‐source_studies.pdf. Accessed September 1, 2016.
10. ENCePP. Mandate of ENCePP Working Group 3. Inventory of EU data sources and methodological approaches for multi‐source studies (EMA/196493/2011 rev.1).,
306
2015. Available at: http://www.encepp.eu/structure/documents/Mandate_WG3.pdf. Accessed August 5, 2016.
11. The European Network of Centres for Pharmacoepidemiology and Pharmacovigilance (ENCePP). Guide on Methodological Standards in Pharmacoepidemiology (EMA/ 95098/2010 Rev.5 ). Available at: http://www.encepp.eu/standards_and_guidances/documents/ENCePPGuideofMethStandardsinPE_Rev5.pdf. Accessed August 20, 2016.
12. Johnson ML, Crown W, Martin BC, Dormuth CR, Siebert U. Good Research Practices for Comparative Effectiveness Research: Analytic Methods to Improve Causal Inference from Nonrandomized Studies of Treatment Effects Using Secondary Data Sources: The ISPOR Good Research Practices for Retrospective Database Analysis Task Force Report—Part III. Value Health 2009; 12: 1062–1073. doi:10.1111/j.1524‐4733.2009.00602.x.
13. Cox E, Martin BC, Van Staa T, Garbe E, Siebert U, Johnson ML. Good Research Practices for Comparative Effectiveness Research: Approaches to Mitigate Bias and Confounding in the Design of Nonrandomized Studies of Treatment Effects Using Secondary Data Sources: The International Society for Pharmacoeconomics and Outcomes Research Good Research Practices for Retrospective Database Analysis Task Force Report—Part II. Value Health 2009; 12: 1053–1061. doi:10.1111/j.1524‐4733.2009.00601.x.
14. Berger ML, Mamdani M, Atkins D, Johnson ML. Good Research Practices for Comparative Effectiveness Research: Defining, Reporting and Interpreting Nonrandomized Studies of Treatment Effects Using Secondary Data Sources: The ISPOR Good Research Practices for Retrospective Database Analysis Task Force Report—Part I. Value Health 2009; 12: 1044–1052. doi:10.1111/j.1524‐4733.2009.00600.x.
15. Swart E, Gothe H, Geyer S, et al. Good Practice of Secondary Data Analysis (GPS): guidelines and recommendations. Gesundheitswesen Bundesverb Ärzte Öffentl Gesundheitsdienstes Ger 2015; 77: 120.
16. FDA. Best Practices for Conducting and Reporting Pharmacoepidemiologic Safety Studies Using Electronic Healthcare Data Sets. 2011. Available at: http://www.easconsultinggroup.com/EASeDocs/files/EASeDocs‐D039‐11.pdf. Accessed September 1, 2016.
17. Neyarapally GA, Hammad TA, Pinheiro SP, Iyasu S. Review of quality assessment tools for the evaluation of pharmacoepidemiological safety studies. BMJ Open 2012; 2. doi:10.1136/bmjopen‐2012‐001362.
307
18. Schneeweiss S, Avorn J. A review of uses of health care utilization databases for epidemiologic research on therapeutics. J Clin Epidemiol 2005; 58: 323–337. doi:10.1016/j.jclinepi.2004.10.012.
19. Harpe SE. Using Secondary Data Sources for Pharmacoepidemiology and Outcomes Research. Pharmacotherapy 2009; 29: 138–153. doi:10.1592/phco.29.2.138.
20. Hall GC, Sauer B, Bourke A, et al. Guidelines for good database selection and use in pharmacoepidemiology research. Pharmacoepidemiol Drug Saf 2012; 21: 1–10. doi:10.1002/pds.2229.
21. Vandenbroucke JP, von Elm E, Altman DG, et al. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. PLoS Med 2007; 4: e297. doi:10.1371/journal.pmed.0040297.
22. Etminan M. Reporting guidelines for pharmacoepidemiological studies are urgently needed. BMJ 2014; 349: g5511. doi:10.1136/bmj.g5511.
23. Benchimol EI, Smeeth L, Guttmann A, et al. The REporting of studies Conducted using Observational Routinely‐collected health Data (RECORD) statement. PLoS Med 2015; 12: e1001885. doi:10.1371/journal.pmed.1001885.
24. Strom BL, Kimmel SE, Hennessy S. Pharmacoepidemiology. John Wiley & Sons, 2011.
25. Abbing‐Karahagopian V, Kurz X, de Vries F, et al. Bridging Differences in Outcomes of Pharmacoepidemiological Studies: Design and First Results of the PROTECT Project. Curr Clin Pharmacol 2014; 9: 130–138. doi:10.2174/1574884708666131111211802.
26. Requena G, Huerta C, Gardarsdottir H, et al. Hip/femur fractures associated with the use of benzodiazepines (anxiolytics, hypnotics and related drugs): a methodological approach to assess consistencies across databases from the PROTECT‐EU project. Pharmacoepidemiol Drug Saf 2016; 25 Suppl 1: 66–78. doi:10.1002/pds.3816.
27. Klungel O. Lessons learned from PROTECT on common protocols for multi‐database studies. Available at: http://www.encepp.eu/publications/documents/5.2_OKlungel_Multidatabase_PROTECT.pdf. Accessed August 6, 2016.
28. Reagan‐Udall Foundation. Innovation in Medical Evidence Development and Surveillance (IMEDS): IMEDS‐methods research agenda., 2014. Available at: http://www.reaganudall.org/wp‐content/uploads/2013/04/IMEDS‐Methods‐Research‐Agenda.pdf. Accessed August 29, 2016.
308
29. Madigan D, Ryan PB, Schuemie M. Does design matter? Systematic evaluation of the impact of analytical choices on effect estimates in observational studies. Ther Adv Drug Saf 2013; 4: 53–62. doi:10.1177/2042098613477445.
30. Madigan D, Ryan PB, Schuemie M, et al. Evaluating the Impact of Database Heterogeneity on Observational Study Results. Am J Epidemiol 2013; 178: 645–651. doi:10.1093/aje/kwt010.
31. Ravera S, Rein N van, Gier JJ de, Berg LTW de J den. A comparison of pharmacoepidemiological study designs in medication use and traffic safety research. Eur J Epidemiol 2012; 27: 473–481. doi:10.1007/s10654‐012‐9689‐3.
32. Gardarsdottir H, Egberts TC, Stolker JJ, Heerdink ER. Duration of Antidepressant Drug Treatment and Its Influence on Risk of Relapse/Recurrence: Immortal and Neglected Time Bias. Am J Epidemiol 2009; 170: 280–285. doi:10.1093/aje/kwp142.
33. Maitland‐van der Zee AH, Klungel OH, Stricker BHC, et al. Comparison of two methodologies to analyze exposure to statins in an observational study on effectiveness. J Clin Epidemiol 2004; 57: 237–242. doi:10.1016/j.jclinepi.2003.07.011.
34. Challenges in irreproducible research : Nature News & Comment. Available at: http://www.nature.com.gate2.inist.fr/news/reproducibility‐1.17552. Accessed August 5, 2016.
35. Afonso A, Schmiedl S, Becker C, et al. A methodological comparison of two European primary care databases and replication in a US claims database: inhaled long‐acting beta‐2‐agonists and the risk of acute myocardial infarction. Eur J Clin Pharmacol 2016; 72: 1105–16. doi:10.1007/s00228‐016‐2071‐8.
36. Gini R, Schuemie M, Brown J, et al. Data Extraction And Management In Networks Of Observational Health Care Databases For Scientific Research: A Comparison Among EU‐ADR, OMOP, Mini‐Sentinel And MATRICE Strategies. EGEMs Gener Evid Methods Improve Patient Outcomes 2016; 4. doi:10.13063/2327‐9214.1189.
37. Academy of Medical Sciences. Reproducibility and reliability of biomedical research. Available at: http://www.acmedsci.ac.uk/policy/policy‐projects/reproducibility‐and‐reliability‐of‐biomedical‐research/. Accessed June 28, 2016.
38. Gruber S, Chakravarty A, Heckbert SR, et al. Design and analysis choices for safety surveillance evaluations need to be tuned to the specifics of the hypothesized drug–outcome association. Pharmacoepidemiol Drug Saf 2016; 25: 973–981. doi:10.1002/pds.4065.
309
39. Lee TA, Pickard AS. Exposure Definition and Measurement. Agency for Healthcare Research and Quality (US), 2013. Available at: http://www.ncbi.nlm.nih.gov.gate2.inist.fr/books/NBK126191/. Accessed July 21, 2016.
40. Nielsen LH, Løkkegaard E, Andreasen AH, Keiding N. Using prescription registries to define continuous drug use: how to fill gaps between prescriptions. Pharmacoepidemiol Drug Saf 2008; 17: 384–388. doi:10.1002/pds.1549.
41. Gichangi A, Andersen M, Kragstrup J, Vach W. Analysing duration of episodes of pharmacological care: an example of antidepressant use in Danish general practice. Pharmacoepidemiol Drug Saf 2006; 15: 167–177. doi:10.1002/pds.1160.
42. Nielsen LH, Keiding N. Validation of methods for identifying discontinuation of treatment from prescription data. J R Stat Soc Ser C Appl Stat 2010; 59: 707–722. doi:10.1111/j.1467‐9876.2010.00712.x.
43. Ray WA. Evaluating medication effects outside of clinical trials: new‐user designs. Am J Epidemiol 2003; 158: 915–20.
44. Maciejewski ML, Bryson CL, Wang V, Perkins M, Liu C‐F. Potential Bias in Medication Adherence Studies of Prevalent Users. Health Serv Res 2013; 48: 1468–1486. doi:10.1111/1475‐6773.12043.
45. Riis AH, Johansen MB, Jacobsen JB, Brookhart MA, Stürmer T, Støvring H. Short look‐back periods in pharmacoepidemiologic studies of new users of antibiotics and asthma medications introduce severe misclassification. Pharmacoepidemiol Drug Saf 2015; 24: 478–485. doi:10.1002/pds.3738.
46. Blanch B, Daniels B, Litchfield M, Pearson S‐A. Looking forward and looking back: the balancing act in new drug user designs for pharmacoepidemiological research. Pharmacoepidemiol Drug Saf 2015; 24: 1117–1119. doi:10.1002/pds.3848.
47. Hallas J, Gaist D, Bjerrum L. The waiting time distribution as a graphical approach to epidemiologic measures of drug utilization. Epidemiology 1997; 8: 666–670.
48. WHO Collaborating Centre for Drug Statistics Methodology. Guidelines for ATC classification and DDD assignment, 2015. Available at: http://www.whocc.no/atc_ddd_publications/guidelines/. Accessed July 31, 2016.
49. Gardarsdottir H, Souverein PC, Egberts TCG, Heerdink ER. Construction of drug treatment episodes from drug‐dispensing histories is influenced by the gap length. J Clin Epidemiol 2010; 63: 422–427. doi:10.1016/j.jclinepi.2009.07.001.
50. Ray WA, Thapa PB, Gideon P. Misclassification of current benzodiazepine exposure by use of a single baseline measurement and its effects upon studies of injuries. Pharmacoepidemiol Drug Saf 2002; 11: 663–669. doi:10.1002/pds.728.
310
51. Chyou P‐H. Patterns of bias due to differential misclassification by case‐control status in a case‐control study. Eur J Epidemiol 2007; 22: 7–17. doi:10.1007/s10654‐006‐9078‐x.
52. Correa‐Villaseñor A, Stewart WF, Franco‐Marina F, Seacat H. Bias from nondifferential misclassification in case‐control studies with three exposure levels. Epidemiology 1995; 6: 276–281.
53. Birkett NJ. Effect of nondifferential misclassification on estimates of odds ratios with multiple levels of exposure. Am J Epidemiol 1992; 136: 356–362.
54. Lash TL, Fox MP, MacLehose RF, Maldonado G, McCandless LC, Greenland S. Good practices for quantitative bias analysis. Int J Epidemiol 2014; 43: 1969–1985. doi:10.1093/ije/dyu149.
55. Cook EA, Schneider KM, Chrischilles E, Brooks JM. Accounting for unobservable exposure time bias when using Medicare prescription drug data. Medicare Medicaid Res Rev 2013; 3. doi:10.5600/mmrr.003.04.a01.
56. Suissa S. Immeasurable time bias in observational studies of drug effects on mortality. Am J Epidemiol 2008; 168: 329–335. doi:10.1093/aje/kwn135.
57. Moulis G, Lapeyre‐Mestre M, Palmaro A, Pugnet G, Montastruc J‐L, Sailler L. French health insurance databases: What interest for medical research? Rev Med Interne 2014; 36: 411–417. doi:10.1016/j.revmed.2014.11.009.
58. Cadarette SM, Burden AM. Measuring and Improving Adherence to Osteoporosis Pharmacotherapy. Curr Opin Rheumatol 2010; 22: 397–403. doi:10.1097/BOR.0b013e32833ac7fe.
59. Nikitovic M, Solomon DH, Cadarette SM. Methods to examine the impact of compliance to osteoporosis pharmacotherapy on fracture risk: systematic review and recommendations. Ther Adv Chronic Dis 2010; 1: 149–162. doi:10.1177/2040622310376137.
60. Stricker BHC, Stijnen T. Analysis of individual drug use as a time‐varying determinant of exposure in prospective population‐based cohort studies. Eur J Epidemiol 2010; 25: 245–251. doi:10.1007/s10654‐010‐9451‐7.
61. Nielsen LH, Løkkegaard E, Andreasen AH, Hundrup YA, Keiding N. Estimating the effect of current, previous and never use of drugs in studies based on prescription registries. Pharmacoepidemiol Drug Saf 2009; 18: 147–153. doi:10.1002/pds.1693.
62. Abrahamowicz M, Schopflocher T, Leffondre K, du Berger R, Krewski D. Flexible modeling of exposure‐response relationship between long‐term average levels of particulate air pollution and mortality in the American Cancer Society study. J Toxicol Env Health A 2003; 66: 1625–54. doi:10.1080/15287390306426.
311
63. Abrahamowicz M, Bartlett G, Tamblyn R, du Berger R. Modeling cumulative dose and exposure duration provided insights regarding the associations between benzodiazepines and injuries. J Clin Epidemiol 2006; 59: 393–403. doi:10.1016/j.jclinepi.2005.01.021.
64. Abrahamowicz M, Beauchamp ME, Sylvestre MP. Comparison of alternative models for linking drug exposure with adverse effects. Stat Med 31: 1014–30. doi:10.1002/sim.4343.
65. Kildemoes HW, Andersen M, Støvring H. The impact of ageing and changing utilization patterns on future cardiovascular drug expenditure: a pharmacoepidemiological projection approach. Pharmacoepidemiol Drug Saf 2010; 19: 1276–1286. doi:10.1002/pds.2039.
66. Kildemoes HW, Støvring H, Andersen M. Driving forces behind increasing cardiovascular drug utilization: a dynamic pharmacoepidemiological model. Br J Clin Pharmacol 2008; 66: 885–895. doi:10.1111/j.1365‐2125.2008.03282.x.
67. Huiart L, Ferdynus C, Dell’Aniello S, Bakiri N, Giorgi R, Suissa S. Measuring persistence to hormonal therapy in patients with breast cancer: accounting for temporary treatment discontinuation. Pharmacoepidemiol Drug Saf 2014; 23: 882–889. doi:10.1002/pds.3631.
68. Boucherie Q, Pauly V, Frauger E, Thirion X, Pradel V, Micallef J. Use of a multi‐state model in a claims database: illustration with methadone. Pharmacoepidemiol Drug Saf 2015; 24: 991–998. doi:10.1002/pds.3835.
69. Timmers L, Beckeringh JJ, van Herk‐Sukel MPP, Boven E, Hugtenburg JG. Use and costs of oral anticancer agents in the Netherlands in the period 2000–2008. Pharmacoepidemiol Drug Saf 2012; 21: 1036–1044. doi:10.1002/pds.2225.
70. Shih Y‐CT, Smieliauskas F, Geynisman DM, Kelly RJ, Smith TJ. Trends in the Cost and Use of Targeted Cancer Therapies for the Privately Insured Nonelderly: 2001 to 2011. J Clin Oncol 2015; 33: 2190–6. doi:10.1200/JCO.2014.58.2320.
71. Turesson I, Velez R, Kristinsson SY, Landgren O. Patterns of improved survival in patients with multiple myeloma in the twenty‐first century: a population‐based study. J Clin Oncol 2010; 28: 830–834. doi:10.1200/JCO.2009.25.4177.
72. Cid Ruzafa J, Merinopoulou E, Baggaley RF, et al. Patient population with multiple myeloma and transitions across different lines of therapy in the USA: an epidemiologic model. Pharmacoepidemiol Drug Saf 2016; 25: 871–879. doi:10.1002/pds.3927.
73. Defossez G, Rollet A, Dameron O, Ingrand P. Temporal representation of care trajectories of cancer patients using data from a regional information system: an
312
application in breast cancer. BMC Med Inform Decis Mak 2014; 14: 24. doi:10.1186/1472‐6947‐14‐24.
74. Boinot L, Gautreau G, Defossez G, et al. Trajectoires hospitalières des patientes atteintes de cancer du sein en Poitou‐Charentes. Rev Épidémiol Santé Publique 2007; 55: 142–148. doi:10.1016/j.respe.2006.11.003.
75. Gray SL, LaCroix AZ, Larson J, et al. Proton pump inhibitor use, hip fracture, and change in bone mineral density in postmenopausal women: results from the Women’s Health Initiative. Arch Intern Med 2010; 170: 765–771. doi:10.1001/archinternmed.2010.94.
76. Kaye JA, Jick H. Proton pump inhibitor use and risk of hip fractures in patients without major risk factors. Pharmacotherapy 2008; 28: 951–959. doi:10.1592/phco.28.8.951.
77. Dimopoulos MA, Richardson PG, Brandenburg N, et al. A review of second primary malignancy in patients with relapsed or refractory multiple myeloma treated with lenalidomide. Blood 2012; 119: 2764–7. doi:10.1182/blood‐2011‐08‐373514.
78. Risque potentiel de seconds cancers primitifs chez les patients traités par Revlimid® (lénalidomide) ‐ Lettre aux professionnels de santé (02/05/2011). Available at: http://ansm.sante.fr/content/download/33391/438020/version/1/file/lp‐110502‐Revlimid.pdf. Accessed September 18, 2014.
79. Abbing‐Karahagopian V, Souverein PC, Korevaar JC, et al. Concomitant medication use and its implications on the hazard pattern in pharmacoepidemiological studies : Example of antidepressants, benzodiazepines and fracture risk. Epidemiol Biostat Public Health 2015; 12: e11273‐1–10.
80. Mini‐Sentinel. Mini Sentinel Methods Development. Case‐Based Methods Workgroup Report. Available at: http://www.mini‐sentinel.org/work_products/Statistical_Methods/Mini‐Sentinel_Methods_Case‐Based‐Report.pdf. Accessed August 9, 2016.
81. Ryan P. Using Exploratory Visualization in the Analysis of Medical Product Safety in Observational Healthcare Data. In: Krause A, O’Connell M (eds.) A Picture is Worth a Thousand Tables. Boston, MA: Springer US, 2012; 391–413. Available at: http://link.springer.com/10.1007/978‐1‐4614‐5329‐1_21. Accessed December 28, 2015.
82. Eco‐Santé France. Couverture de base par les principaux régimes d’Assurance maladie. Available at: http://www.ecosante.fr/index2.php?base=DEPA&langh=FRA&langs=FRA&sessionid=. Accessed February 16, 2016.
313
83. Tuppin P, de Roquefeuil L, Weill A, Ricordeau P, Merliere Y. French national health insurance information system and the permanent beneficiaries sample. Rev Epidemiol Sante Publique 2010; 58: 286–90. doi:10.1016/j.respe.2010.04.005.
84. Loi n° 2016‐41 du 26 janvier 2016 de modernisation de notre système de santé. 2016. Available at: https://www.legifrance.gouv.fr/affichTexte.do?cidTexte=JORFTEXT000031912641. Accessed February 16, 2016.
85. Alert generation using the case‐population approach in the French claims databases. Available at: http://www.encepp.eu/encepp/viewResource.htm;jsessionid=a_u1bazqEhl4aO4ky9q4vMx30ZVZyUo1rn4R6CeEeR2M5WMgK0xe!‐1357411575?id=13079. Accessed August 5, 2016.
86. Bannay A, Chaignot C, Blotière P‐O, et al. The Best Use of the Charlson Comorbidity Index With Electronic Health Care Database to Predict Mortality. Med Care 2016; 54: 188–194. doi:10.1097/MLR.0000000000000471.
87. Bras P‐L. La gouvernance et l’utilisation des données de santé. Available at: http://www.ladocumentationfrancaise.fr/rapports‐publics/134000670/. Accessed April 24, 2015.
88. Cours des comptes. Les données personnelles de santé gérées par l’assurance maladie (2016). Available at: https://www.ccomptes.fr/Accueil/Publications/Publications/Les‐donnees‐personnelles‐de‐sante‐gerees‐par‐l‐assurance‐maladie. Accessed August 20, 2016.
89. Horsfall L, Walters K, Petersen I. Identifying periods of acceptable computer usage in primary care research databases. Pharmacoepidemiol Drug Saf 22: 64–9. doi:10.1002/pds.3368.
90. Kripke DF, Langer RD, Kline LE. Hypnotics’ Association with Mortality or Cancer: A Matched Cohort Study. BMJ Open 2012; 2. doi:10.1136/bmjopen‐2012‐000850.
91. Feskanich D, Hastrup JL, Marshall JR, et al. Stress and suicide in the Nurses’ Health Study. J Epidemiol Community Health 2002; 56: 95–98.
92. Obiora E, Hubbard R, Sanders RD, Myles PR. The impact of benzodiazepines on occurrence of pneumonia and mortality from pneumonia: a nested case‐control and survival analysis in a population‐based cohort. Thorax 2013; 68: 163–70. doi:10.1136/thoraxjnl‐2012‐202374.
93. Tiihonen J, Suokas JT, Suvisaari JM, Haukka J, Korhonen P. Polypharmacy with antipsychotics, antidepressants, or benzodiazepines and mortality in schizophrenia. Arch Gen Psychiatry 2012; 69: 476–83. doi:10.1001/archgenpsychiatry.2011.1532.
314
94. Weich S, Pearce HL, Croft P, et al. Effect of anxiolytic and hypnotic drug prescriptions on mortality hazards: retrospective cohort study. BMJ 2014; 348: g1996. doi:10.1136/bmj.g1996.
95. Kojima M, Wakai K, Kawamura T, et al. Sleep patterns and total mortality: a 12‐year follow‐up study in Japan. J Epidemiol Jpn Epidemiol Assoc 2000; 10: 87–93.
96. Baandrup L, Gasse C, Jensen VD, et al. Antipsychotic polypharmacy and risk of death from natural causes in patients with schizophrenia: a population‐based nested case‐control study. J Clin Psychiatry 2010; 71: 103–108. doi:10.4088/JCP.08m04818yel.
97. Vinkers DJ, Gussekloo J, van der Mast RC, Zitman FG, Westendorp RGJ. Benzodiazepine use and risk of mortality in individuals aged 85 years or older. JAMA 2003; 290: 2942–2943. doi:10.1001/jama.290.22.2942.
98. Winkelmayer WC, Mehta J, Wang PS. Benzodiazepine use and mortality of incident dialysis patients in the United States. Kidney Int 2007; 72: 1388–93. doi:10.1038/sj.ki.5002548.
99. Requena G, Logie J, Martin E, et al. Do case‐only designs yield consistent results across design and different databases? A case study of hip fractures and benzodiazepines: A Comparison across Case‐Only Designs and Databases. Pharmacoepidemiol Drug Saf 2016; 25: 79–87. doi:10.1002/pds.3822.
100. Ryan PB, Madigan D, Stang PE, Marc Overhage J, Racoosin JA, Hartzema AG. Empirical assessment of methods for risk identification in healthcare data: results from the experiments of the Observational Medical Outcomes Partnership. Stat Med 2012; 31: 4401–4415. doi:10.1002/sim.5620.
101. Huybrechts KF, Rothman KJ, Silliman RA, Brookhart MA, Schneeweiss S. Risk of death and hospital admission for major medical events after initiation of psychotropic medications in older adults admitted to nursing homes. Cmaj 2011; 183: E411‐9. doi:10.1503/cmaj.101406.
102. Palmaro A, Dupouy J, Lapeyre‐Mestre M. Benzodiazepines and risk of death: Results from two large cohort studies in France and UK. Eur Neuropsychopharmacol 2015; 25: 1566–77. doi:10.1016/j.euroneuro.2015.07.006.
103. Phadnis MA, Shireman TI, Wetmore JB, et al. Estimation of Drug Effectiveness by Modeling Three Time‐dependent Covariates: An Application to Data on Cardioprotective Medications in the Chronic Dialysis Population. Stat Biopharm Res 2014; 6: 229–240. doi:10.1080/19466315.2014.920275.
315
104. Warren JL, Harlan LC, Fahey A, et al. Utility of the SEER‐Medicare data to identify chemotherapy use. Med Care 2002; 40: IV‐55‐61. doi:10.1097/01.MLR.0000020944.17670.D7.
105. Lamont EB, Lauderdale DS, Schilsky RL, Christakis NA. Construct validity of medicare chemotherapy claims: the case of 5FU. Med Care 2002; 40: 201–211.
106. Bikov KA, Mullins CD, Seal B, Onukwugha E, Hanna N. Algorithm for Identifying Chemotherapy/Biological Regimens for Metastatic Colon Cancer in SEER‐Medicare. Med Care 2013; 53: e58‐64. doi:10.1097/MLR.0b013e31828fad9f.
107. Byfield SD, Yu E, Morlock R, Evans D, Teitelbaum A. Corroboration of claims algorithm for second‐line costs of metastatic colorectal cancer treatment with targeted agents. J Med Econ 2013; 16: 1071–1081. doi:10.3111/13696998.2013.813513.
108. Andersson ML, Böttiger Y, Lindh JD, Wettermark B, Eiermann B. Impact of the drug‐drug interaction database SFINX on prevalence of potentially serious drug‐drug interactions in primary health care. Eur J Clin Pharmacol 2013; 69: 565–571. doi:10.1007/s00228‐012‐1338‐y.
109. Palmaro A, Conte C, Lagadic C, et al. Identifying multiple myeloma patients using data from the SNIIRAM and PMSI: validation using the Tarn cancer registry. Congress of the Group of Registries of the Latin language countries (GRELL), 4‐6 May 2016, Albi, France.
110. Palmaro A, Gauthier M, Despas F, Lapeyre‐Mestre M. Identifying cancer treatment regimens using SNIIRAM and PMSI databases: an application in multiple myeloma. 20th Annual Meeting of French Society of Pharmacology and Therapeutics, 37th Pharmacovigilance Meeting, 17th APNET Seminar, 14th CHU CIC Meeting, 19–21 April 2016, Nancy, France. Fundam Clin Pharmacol 2016; 30: 23–23.
111. Tobi H, Faber A, van den Berg PB, Drane JW, de Jong‐van den Berg LTW. Studying co‐medication patterns: the impact of definitions. Pharmacoepidemiol Drug Saf 2007; 16: 405–411. doi:10.1002/pds.1304.
112. Ryan PB, Madigan D, Stang PE, Schuemie MJ, Hripcsak G. Medication‐wide association studies. CPT Pharmacomet Syst Pharmacol 2013; 2: e76. doi:10.1038/psp.2013.52.
113. Oncomip. Thésaurus de Chimiothérapie. 2013. Available at: http://oncomip.org/fr/dldoc/?t=recommandations&f=doc1&d=69&h=354a47eeb5eb872bc7605d6ccfd8033c. Accessed August 20, 2016.
114. Oncomip. Référentiel Traitement Myélome Multiple. 2011. Available at: http://www.oncomip.org/fr/espace‐professionnel/referentiels/myelome‐multiple‐67.html. Accessed August 20, 2016.
316
115. Revlimid. Summary of product characteristics. Available at: http://www.ema.europa.eu/ema/index.jsp?curl=pages/medicines/human/medicines/000717/human_med_001034.jsp&mid=WC0b01ac058001d124. Accessed January 10, 2017.
116. Palumbo A, Bringhen S, Ludwig H, et al. Personalized therapy in multiple myeloma according to patient age and vulnerability: a report of the European Myeloma Network (EMN). Blood 2011; 118: 4519–4529. doi:10.1182/blood‐2011‐06‐358812.
117. Palumbo A, Mateos M‐V, Bringhen S, San Miguel JF. Practical management of adverse events in multiple myeloma: can therapy be attenuated in older patients? Blood Rev 2011; 25: 181–191. doi:10.1016/j.blre.2011.03.005.
118. Tsiropoulos I, Andersen M, Hallas J. Adverse events with use of antiepileptic drugs: a prescription and event symmetry analysis. Pharmacoepidemiol Drug Saf 2009; 18: 483–491. doi:10.1002/pds.1736.
119. Hallas J. Evidence of depression provoked by cardiovascular medication: a prescription sequence symmetry analysis. Epidemiology 1996; 7: 478–484.
120. Hersom K, Neary MP, Levaux HP, Klaskala W, Strauss JS. Isotretinoin and antidepressant pharmacotherapy: A prescription sequence symmetry analysis. J Am Acad Dermatol 2003; 49: 424–432. doi:10.1067/S0190‐9622(03)02087‐5.
121. Hallas J, Bytzer P. Screening for drug related dyspepsia: an analysis of prescription symmetry. Eur J Gastroenterol Hepatol 1998; 10: 27–32.
122. Nicole L Pratt IAW. Sequence Symmetry Analysis and Disproportionality Analyses: What Percentage of Adverse Drug Reaction do they Signal? Adv Pharmacoepidemiol Drug Saf 2013; 02. doi:10.4172/2167‐1052.1000140.
123. Allen NB, Siddique J, Wilkins JT, et al. BLood pressure trajectories in early adulthood and subclinical atherosclerosis in middle age. JAMA 2014; 311: 490–497. doi:10.1001/jama.2013.285122.
124. Jensen AB, Moseley PL, Oprea TI, et al. Temporal disease trajectories condensed from population‐wide registry data covering 6.2 million patients. Nat Commun 2014; 5: 4022. doi:10.1038/ncomms5022.
125. Gilbertson DT, Bradbury BD, Wetmore JB, et al. Controlling confounding of treatment effects in administrative data in the presence of time‐varying baseline confounders. Pharmacoepidemiol Drug Saf 2015; 25: 269–277. doi:10.1002/pds.3922.
317
126. Tukey JW. Exploratory data analysis. 1977. Available at: http://xa.yimg.com/kq/groups/16412409/1159714453/name/exploratorydataanalysis.pdf. Accessed July 9, 2016.
127. van der Corput P, Arends J, van Wijk JJ. Visualization of Medicine Prescription Behavior. Comput Graph Forum 2014; 33: 161–170. doi:10.1111/cgf.12372.
128. Rhee S‐Y, Blanco JL, Liu TF, et al. Standardized representation, visualization and searchable repository of antiretroviral treatment‐change episodes. AIDS Res Ther 2012; 9: 1–8. doi:10.1186/1742‐6405‐9‐13.
129. Wang TD, Plaisant C, Quinn AJ, Stanchak R, Murphy S, Shneiderman B. Aligning Temporal Data by Sentinel Events: Discovering Patterns in Electronic Health Records. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. CHI ’08. New York, NY, USA: ACM, 2008; 457–466. doi:10.1145/1357054.1357129.
130. Hallgreen CE, van den Ham HA, Mt‐Isa S, et al. Benefit–risk assessment in a post‐market setting: a case study integrating real‐life experience into benefit–risk methodology. Pharmacoepidemiol Drug Saf 2014; 23: 974–983. doi:10.1002/pds.3676.
131. Hallgreen CE, Mt‐Isa S, Lieftucht A, et al. Literature review of visual representation of the results of benefit‐risk assessments of medicinal products: Visuals for Benefit‐Risk Representation. Pharmacoepidemiol Drug Saf 2016; 25: 238–250. doi:10.1002/pds.3880.
132. Avillach P, Coloma PM, Gini R, et al. Harmonization process for the identification of medical events in eight European healthcare databases: the experience from the EU‐ADR project. J Am Med Inform Assoc 2013; 20: 184–192. doi:10.1136/amiajnl‐2012‐000933.
133. Herman RA, Gilchrist B, Link BK, Carnahan R. A systematic review of validated methods for identifying lymphoma using administrative data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 203–212. doi:10.1002/pds.2315.
134. Andrade SE, Harrold LR, Tjia J, et al. A systematic review of validated methods for identifying cerebrovascular accident or transient ischemic attack using administrative data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 100–128. doi:10.1002/pds.2312.
135. Schneider G, Kachroo S, Jones N, et al. A systematic review of validated methods for identifying erythema multiforme major/minor/not otherwise specified, Stevens‐Johnson Syndrome, or toxic epidermal necrolysis using administrative and claims data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 236–239. doi:10.1002/pds.2331.
318
136. Schneider G, Kachroo S, Jones N, et al. A systematic review of validated methods for identifying hypersensitivity reactions other than anaphylaxis (fever, rash, and lymphadenopathy), using administrative and claims data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 248–255. doi:10.1002/pds.2333.
137. Carnahan RM, Moores KG, Perencevich EN. A systematic review of validated methods for identifying infection related to blood products, tissue grafts, or organ transplants using administrative data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 213–221. doi:10.1002/pds.2332.
138. Kee VR, Gilchrist B, Granner MA, Sarrazin NR, Carnahan RM. A systematic review of validated methods for identifying seizures, convulsions, or epilepsy using administrative and claims data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 183–193. doi:10.1002/pds.2329.
139. Walkup JT, Townsend L, Crystal S, Olfson M. A systematic review of validated methods for identifying suicide or suicidal ideation using administrative or claims data. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 174–182. doi:10.1002/pds.2335.
140. Williams SE, Carnahan R, Krishnaswami S, McPheeters ML. A systematic review of validated methods for identifying transverse myelitis using administrative or claims data. Vaccine 2013; 31 Suppl 10: K83‐87. doi:10.1016/j.vaccine.2013.03.074.
141. Institut National Du Cancer. Algorithme de sélection des hospitalisations liées à la prise en charge du cancer dans les bases nationales d’activité hospitalière de court séjour (2013). Available at: www.e‐cancer.fr%2Fcontent%2Fdownload%2F63161%2F568597%2Ffile%2FAlgorithme‐cancer‐2013‐V2.pdf. Accessed June 23, 2016.
142. Schulman KL, Berenson K, (Tina) Shih Y‐C, et al. A Checklist for Ascertaining Study Cohorts in Oncology Health Services Research Using Secondary Data: Report of the ISPOR Oncology Good Outcomes Research Practices Working Group. Value Health 2013; 16: 655–669. doi:10.1016/j.jval.2013.02.006.
143. Manuel DG, Rosella LC, Stukel TA. Importance of accurately identifying disease in studies using electronic health records. BMJ 2010; 341: c4226. doi:10.1136/bmj.c4226.
144. Ehrenstein V, Petersen I, Smeeth L, et al. Helping everyone do better: a call for validation studies of routinely recorded health data. Clin Epidemiol 2016; 8: 49–51. doi:10.2147/CLEP.S104448.
145. Hafdi‐Nejjari Z, Couris C‐M, Schott A‐M, et al. Role of hospital claims databases from care units for estimating thyroid cancer incidence in the Rhône‐Alpes region of France. Rev Epidemiol Sante Publique 2006; 54: 391–398.
319
146. Coureau G, Baldi I, Savès M, et al. Performance evaluation of hospital claims database for the identification of incident central nervous system tumors compared with a cancer registry in Gironde, France, 2004. Rev Epidemiol Sante Publique 2012; 60: 295–304. doi:10.1016/j.respe.2012.02.003.
147. Quantin C, Benzenine E, Hagi M, et al. Estimation of National Colorectal‐Cancer Incidence Using Claims Databases. J Cancer Epidemiol 2012: e298369. doi:10.1155/2012/298369, 10.1155/2012/298369.
148. Carnahan RM, Moores KG. Mini‐Sentinel’s systematic reviews of validated methods for identifying health outcomes using administrative and claims data: methods and lessons learned. Pharmacoepidemiol Drug Saf 2012; 21 Suppl 1: 82–89. doi:10.1002/pds.2321.
149. Princic N, Chris G, Willson T, et al. Development of an Algorithm to Identify Patients with Multiple Myeloma Using Administrative Claims Data. 57th American Society of Hematology Annual Meeting, United States of America, Orlando, 5 ‐ 8 December 2015. Available at: https://ash.confex.com/ash/2015/webprogram/Paper85570.html. Accessed June 23, 2016.
150. Quantin C, Bouzelat H, Allaert FA, Benhamiche AM, Faivre J, Dusserre L. Automatic record hash coding and linkage for epidemiological follow‐up data confidentiality. Methods Arch 1998; 37: 271–277.
151. Cleveland WS. Visualizing Data. 1 edition. Murray Hill, N.J. : Summit, N.J: Hobart Press, 1993.
152. Gray J, Bounegru L, Chambers L, Kayser‐Bril N, Robert C. Guide du datajournalisme : Collecter, analyser et visualiser les données. Paris: Eyrolles, 2013.
153. Belleville G. Mortality hazard associated with anxiolytic and hypnotic drug use in the National Population Health Survey. Can J Psychiatry 2010; 55: 558–67.
154. Hays JC, Blazer DG, Foley DJ. Risk of napping: excessive daytime sleepiness and mortality in an older community population. J Am Geriatr Soc 1996; 44: 693–698.
155. Kripke DF, Garfinkel L, Wingard DL, Klauber MR, Marler MR. Mortality associated with sleep duration and insomnia. Arch Gen Psychiatry 2002; 59: 131–6.
156. Gandrud C. Reproducible Research with R and RStudio. Boca Raton: Chapman and Hall/CRC, 2013.
157. Stodden V, Leisch F, Peng RD. Implementing Reproducible Research. Boca Raton: Chapman and Hall/CRC, 2014.
320
158. Moulis G, Palmaro A, Sailler L, Lapeyre‐Mestre M. Corticosteroid Risk Function of Severe Infection in Primary Immune Thrombocytopenia Adults. A Nationwide Nested Case‐Control Study. PLoS One 2015; 10: e0142217. doi:10.1371/journal.pone.0142217.
159. Fairman KA. Health Care Research Done Right: A Journal Editor Shares Practical Tips and Techniques for High Quality and Efficiency. 1 edition. Denver: Outskirts Press, 2012.
160. REDSIAM (Réseau données SNIIRAM). Available at: http://www.redsiam.fr/. Accessed September 2, 2016.
161. Enquête nationale : Besoins des chercheurs sur les Données de Santé ‐ ITMO Santé Publique. Portail Epidemiologie ‐ France. Available at: https://epidemiologie‐france.aviesan.fr/epidemiologie‐france/toute‐l‐actualite/enquete‐nationale‐besoins‐des‐chercheurs‐sur‐les‐donnees‐de‐sante‐itmo‐sante‐publique. Accessed September 2, 2016.
162. Odejide OO, Cronin AM, Davidoff AJ, LaCasce AS, Abel GA. Limited stage diffuse large B‐cell lymphoma: comparative effectiveness of treatment strategies in a large cohort of elderly patients. Leuk Lymphoma 2015; 56: 716–724. doi:10.3109/10428194.2014.930853.
163. Neuman HB, Greenberg CC. Comparative effectiveness research: opportunities in surgical oncology. Semin Radiat Oncol 2014; 24: 43–48. doi:10.1016/j.semradonc.2013.09.003.
164. Belhassen M, Tamberou C, Laigle V, et al. Effectiveness of Montelukast on Asthma Control in Infants: Methodology of a Claims Data Study. Value Health 2013; 16: A367. doi:10.1016/j.jval.2013.08.259.
165. Blommestein HM, Verelst SGR, de Groot S, Huijgens PC, Sonneveld P, Uyl‐de Groot CA. A cost‐effectiveness analysis of real‐world treatment for elderly patients with multiple myeloma using a full disease model. Eur J Haematol 2016; 96: 198–208. doi:10.1111/ejh.12571.
166. Kalinjuma AV. Adjusting for crossover bias in an observational study for patients with multiple myeloma. 2012. Available at: https://uhdspace.uhasselt.be/dspace/handle/1942/14157. Accessed April 1, 2016.
167. Center for Medical Technology Policy. Recommendations for Comparing Therapeutic Sequences of Patients with Breast, Kidney, and other Advanced Cancers. Version 1.1, August 4 , 2015. Available at: http://www.cmtpnet.org/docs/resources/Recommendations_for_Comparing_Cancer_Therapy_Sequences_v1.1_August_2015.pdf. Accessed August 18, 2016.
321
168. Indicateurs de polymédication pour le Sniiram. GitHub. Available at: https://github.com/pierucci/polymed. Accessed September 2, 2016.
169. Le Cossec C, Sermet C. Mesurer la polymédication chez les personnes âgées : impact de la méthode sur la prévalence et les classes thérapeutiques. IRDES. Questions d’économie de la santé n° 213 ‐ octobre 2015. Available at: http://www.irdes.fr/recherche/questions‐d‐economie‐de‐la‐sante/213‐mesurer‐la‐polymedication‐chez‐les‐personnes‐agees.pdf.
170. PharmD SH, Bilker WB, Weber A, Strom BL. Descriptive analyses of the integrity of a US Medicaid claims database. Pharmacoepidemiol Drug Saf 2003; 12: 103–111. doi:10.1002/pds.765.
171. Palmaro A, Moulis G, Despas F, Dupouy J, Lapeyre‐Mestre M. Overview of drug data within french health insurance databases and implications for pharmacoepidemiological studies. Fundam Clin Pharmacol 2016; 30: 616–624. doi:10.1111/fcp.12214.
322
323
XIII. Appendices
1. Othercontributionsduringthethesiscourse
Anne‐LaureBourgeois,PascalAuriche,AurorePalmaro,Jean‐LouisMontastruc.Riskof hormonotherapy in transgender people: literature review and data from theFrenchPharmacoVigilanceDatabase.Annalesd'Endocrinologie,2016,77(1),14–21.DOI:10.1016/j.ando.2015.12.001
Guillaume Moulis, Aurore Palmaro, Laurent Sailler, Maryse Lapeyre‐Mestre.Corticosteroid Risk Function of Severe Infection in Primary ImmuneThrombocytopeniaAdults.ANationwideNestedCase‐ControlStudy.PloSOne,2015,10(11):e0142217.DOI:10.1371/journal.pone.0142217
Montastruc, François, Aurore Palmaro, Haleh Bagheri, Laurent Schmitt, Jean‐LouisMontastruc,Maryse Lapeyre‐Mestre. Role of Serotonin 5‐HT2C and Histamine H1Receptors in Antipsychotic‐Induced Diabetes: A Pharmacoepidemiological‐Pharmacodynamic Study in VigiBase. European Neuropsychopharmacology,2015;25(10):1556‐65.DOI:10.1016/j.euroneuro.2015.07.010.
Loan Nguyen Thi‐Thanh, Aurore Palmaro, François Montastruc, Maryse Lapeyre‐Mestre, Guillaume Moulis. 2015. Signal for Thrombosis with Eltrombopag andRomiplostim: A Disproportionality Analysis of Spontaneous Reports WithinVigiBase®.DrugSafety,2015;38(12):1179‐86.DOI:10.1007/s40264‐015‐0337‐1
JoëlleMicallef,EdithFrauger,AurorePalmaro,QuentinBoucherie,MaryseLapeyre‐Mestre. Un exemple d'investigation d'un phénomène « émergent » enaddictovigilance : à propos duméthylphénidate. Thérapie, 2015;70(2):191‐6. doi:10.2515/therapie/2015012.
GuillaumeMoulis,Maryse Lapeyre‐Mestre, Aurore Palmaro, Grégory Pugnet, Jean‐LouisMontastruc,LaurentSailler.Frenchhealthinsurancedatabases:Whatinterestfor medical research? Revue de Médecine Interne, 2015;36(6):411‐7. DOI:10.1016/j.revmed.2014.11.009
Aurore Palmaro, Raphael Bissuel, Nicholas Renaud, Geneviève Durrieu, BrigitteEscourrou, Jean‐LouisMontastruc,MaryseLapeyre‐Mestre.Adversedrugreactionsand off‐label drug prescription in paediatric outpatients. Pediatrics, 2015Jan;135(1):49‐58.DOI:10.1542/peds.2014‐0764
GuillaumeMoulis,AurorePalmaro,Jean‐LouisMontastruc,BernardGodeau,MaryseLapeyre‐Mestre, Laurent Sailler. Epidemiology of incident immunethrombocytopenia: a nationwide population‐based study in France. Blood,2014;124(22):3308‐15.DOI:10.1182/blood‐2014‐05‐578336
324
François Montastruc, Guillaume Moulis, Aurore Palmaro, Virginie Gardette,Geneviève Durrieu, Jean‐Louis Montastruc. Relationships between MedicalResidentsandDrugCompanies:ANationalSurvey.PLoSOne9(10):e104828.
Régis Fuzier, Isabelle Serres, Robert Bourrel, Aurore Palmaro, Jean‐LouisMontastruc, Maryse Lapeyre‐Mestre. Analgesic drug consumption increases afterknee arthroplasty: a pharmacoepidemiological study investigating postoperativepain.Pain,2014;155(7):1339‐45.DOI:10.1016/j.pain.2014.04.010.
VincentBounes,AurorePalmaro,AnneRoussin,MaryseLapeyre‐Mestre.Longtermconsequences of acute pain on patients under methadone or buprenorphinemaintenance treatment: a prospective multicenter cohort study. Pain Physician,2013;16(6):E739‐47
Julie Dupouy, Jean‐Pascal Fournier, Émilie Jouanjus, Aurore Palmaro, Jean‐ChristophePoutrain,StéphaneOustric,MaryseLapeyre‐Mestre.Baclofenforalcoholdependence in France: incidence of treated patients and prescription patterns. Acohort study. European Neuropsychopharmacology, 2014;24(2):192‐9. DOI:10.1016/j.euroneuro.2013.09.008.
Geneviève Durrieu, Aurore Palmaro, Laure Pourcel, Céline Caillet, AngéliqueFaucher, Alexis Jacquet, Shéhérazade Ouaret, Marie‐Christine Perault‐Pochat,CarmenKreft‐Jais,AnneCastot,MaryseLapeyre‐Mestre,Jean‐LouisMontastrucandtheFrenchNetworkofPharmacovigilanceCentres.FirstFrenchexperienceofADRsreporting bypatients aftermass immunization campaignwithA(H1N1) pandemicvaccines: a comparison of reports submitted by patients and HealthcareProfessionals.DrugSafety,2012;35(10):845‐54.
Communicationsoralesouaffichéeslorsdecongrès
2015‐2016
Aurore Palmaro, Régis Fuzier, Isabelle Serres, Robert Bourrel, MaryseLapeyre‐Mestre. Analgesic drug consumption increases after carpal tunnelsurgery: apharmacoepidemiological study investigatingpostoperativepain.XIèmeCongrèsdePhysiologie,dePharmacologieetdeThérapeutique,Nancy,19‐21avril2016(communicationorale)
Maryse Lapeyre‐Mestre, Aurore Palmaro, Camille Ponte, Emilie Jouanjus.EarlysignalofdiverteduseoftropicamideeyedropsinFrance:anexampleofinvestigation from the addictovigilance point of view. XIème Congrès dePhysiologie,dePharmacologieetdeThérapeutique,Nancy,19‐21avril2016.Fundamental and Clinical Pharmacology, 30 (Suppl. 1), 72 (PS‐109)(communicationaffichée)
325
MartinGauthier,SandraDeBarros,AurorePalmaro,CécileConte,FrançoiseHuguet,RobertBourrel,GuyLaurent,MaryseLapeyre‐Mestre,FabienDespas.Initiation de psychotropes chez les patients diagnostiqués porteurs d’uneLMC : étude de population, Congrès de la Société Française d’Hématologie,Paris,23‐25Mars2016(communicationaffichée)
Aurore Palmaro, Marie‐Eve Rougé‐Bugat, Martin Gauthier, Fabien Despas,Maryse Lapeyre‐Mestre. Real‐life practices for preventing venousthromboembolism in multiple myeloma patients: a cohort study from theFrenchhealth insurancedatabase.10thCongressofGeneralPractice,Paris,31mars‐02avril2016(communicationorale)
2014‐2015
AurorePalmaro,MaryseLapeyre‐Mestre.Trends inopioidanalgesicsuse inEurope: a ten‐year perspective. 12th European Association for ClinicalPharmacology & Therapeutics Congress (EACPT 2015), Madrid, 27‐30 juin2015(communicationorale)
Aurore Palmaro, Fabien Despas, Maryse Lapeyre‐Mestre.Thromboprophylaxisinmultiplemyelomapatientstreatedwithlenalidomideor thalidomide. 12th European Association for Clinical Pharmacology &Therapeutics Congress (EACPT 2015), Madrid, 27‐30 juin 2015(communicationaffichée)
Aurore Palmaro, Régis Fuzier, Isabelle Serres, Robert Bourrel, MaryseLapeyre‐Mestre. Analgesic drug consumption increases after carpal tunnelsurgery: apharmacoepidemiological study investigatingpostoperativepain.12th European Association for Clinical Pharmacology & TherapeuticsCongress(EACPT2015),Madrid,27‐30juin2015(communicationaffichée)
JulienGredin,AurorePalmaro,RaphaelBissuel,BrigitteEscourrou,StéphaneOustric,MaryseLapeyre‐Mestre,MichelBismuth.Off‐label drugprescribingin paediatric outpatients with ear, nose, and throat conditions: a surveyamong General Practitioners in south western France. Xème Congrès dePhysiologie, de Pharmacologie et de Thérapeutique, P2T, Caen, 21‐23 avril2015(communicationaffichée)
326
2. PhDtrainingGeneralandSpecificcourses
Scientificcommunicationanddatavisualization
2015 CursusPhotoshop&IllustratorÉcoleDoctoraleBiologieSantéBiotechnologies,Toulouse
2015 PublicationmultimédiaavecIndesignUniversitéFédéraledeToulouseMidi‐Pyrénées
2015 PROTECT symposium. Pre‐symposium training. Structuredmethodologies for the assessment and visualization of thebenefit‐riskofmedicinesEuropeanMedicinesAgency,London,(1day)
2014 Lesoutilsenlignepourcartographierouprésentervisuellementvosdonnées(1day)PRESUniversitédeToulouse
2014 Data‐sharinginbiomedicalandhealthresearch:legalprotection,ethicalissuesandgovernance(INSERMWorkshop)
Bordeaux,France(3days)2013 Techniquesdevulgarisationscientifiqueàl'écrit(1day)
PRESUniversitédeToulouse2013 Techniquesdevulgarisationscientifiqueàl'oral(1day)
PRESUniversitédeToulouse
Dataanalyticsandstatistics
2016 TraitementdeMassedeDonnéesScientifiques(4days)EcoleDoctoraleSystèmes
2015 Networkmeta‐analysis(INSERMWorkshop234)Bordeaux,France
2014 EPIMIX–AnalysededonnéeslongitudinalesougroupéesenépidémiologieEcoled'été2014Méthodesettechniquesenépidémiologie(4days).InstitutdeSantéPublique,d'ÉpidémiologieetdeDéveloppement(ISPED),universitédeBordeaux
327
Trainingforthedatabasesused
2014 EGBS–ÉchantillonGénéralistedesBénéficiairesSimplifié(3days).CNAM‐TS/DirectiondelaStratégie,desÉtudesetdesStatistiques,CRFdeTours
2013 ArchitectureetdonnéesduSNIIRAM(1day)CNAM‐TS,Paris
2013 CPRDonlinetraining
Congress,conferenceandworkshopsattendance
24‐26November2016
16èmecongrèsdelaSFETD(SociétéFrançaised'EtudeetdeTraitementdelaDouleur),Bordeaux
11‐12October2016
e‐HealthResearch2016.Howdigitaltechnologiesdisruptepidemiologyandmedicalresearch.Paris
4‐6may2016 CongrèsduGroupedesRegistresdeLangueLatine(GRELL),Albi
19‐21April2016 XIèmeCongrèsdePhysiologie,dePharmacologieetdeThérapeutique,Nancy
31March‐02April2016
10thCongressofGeneralPractice,Paris
10‐11March2016
CongrèsADELF‐EMOIS,Dijon
24November2015
ENCePPPlenaryMeetingEuropeanMedicinesAgency,London,UK
27‐30June,2015 12thEuropeanAssociationforClinicalPharmacology&TherapeuticsCongress(EACPT2015),Madrid,Spain
19‐20February2015
PROTECTsymposiumEuropeanMedicinesAgency,London,UK
13‐15April2015 5thBordeauxPharmacoepifestivalBordeaux,France
20‐22November2014
14èmecongrèsdelaSFETD(SociétéFrançaised'EtudeetdeTraitementdelaDouleur),Toulouse
328
26‐27mai2014 FirstmeetingofEPICHRONIC
CenterforBiomedicalResearchofLaRioja(CIBIR)Logroño,Spain
22‐24April2014 IXèmeCongrèsdePhysiologie,dePharmacologieetde
Thérapeutique,P2T,Poitiers
14‐15March2014
8èmeCongrèsinterrégionalDevenirJeuneChercheurenMédecineGénérale,Toulouse
9‐11April2014 4thBordeauxPharmacoepifestivalBordeaux,France
Teachingactivity
Master2Professionnel«Métiersdumédicament»,UniversitéPaulSabatier,ToulouseIII
Populationsofanalysis(1hour) Sub‐groupanalysis(1hour) Missingdata(1hour) Riskassociatedwithdruguse(3hours) Meta‐analysis(2hours)
Master 2 Recherche «Épidémiologie clinique», Université Paul Sabatier,ToulouseIII
Clinicalresearchmethodology(3hours) Meta‐analysis(3hours)
Master 1 Santé Publique, «Méthodologie de la recherche clinique etépidémiologique»,UniversitéPaulSabatier,ToulouseIII
Choiceofendpointsinclinicalstudies(2hours) Interimanalysis(2hours) Noninferioritytrials(2hours) Powerandsamplesize(2hours) Pharmacoepidemiology(2hours)
Master2MékongPharma(Masterensciencespharmaceutiques),UniversitédessciencesdelasantéduLaos,UniversitéPaulSabatier,ToulouseIII
Biostatisticsinclinicalpharmacology(25hours)
329
Titre:Méthodesdemesuredel’expositionmédicamenteusediscontinueàpartirdesgrandesbasesdedonnéesensanté
Directeurdethèse:MaryseLapeyre‐Mestre
Co‐directeurdethèse:FabienDespas
Lieuetdatedesoutenance:Toulouse,le20janvier2017
Résumé:Lecontexte internationalde lapharmacoépidémiologie,marquépar lamiseenœuvred’unnombrecroissantd’étudesmulti‐sources,afaitémergeruncertainnombredequestionnementsautourdelagestiondedonnéesconflictuellesoudel’impactdeschoixméthodologiquessurlesrésultats.Accroîtrelaconfiancedansces études observationnelles et renforcer leur crédibilité face aux données issues des essais cliniquesreprésente un enjeu majeur, qui dépend étroitement de la robustesse des conclusions produites. Dans cedomaine, lamesurede l’expositionmédicamenteuserevêtdoncune importance touteparticulière, tantpourdes études portant sur l’estimation d’un risque ou d’un critère d’efficacité, que lors de la description desmodalités d’utilisation en vie réelle. L’exposition médicamenteuse reste un phénomène complexe qui secaractériselaplupartdutempspardescyclesdiscontinus,marquéspardesévolutionsdedosesetlaprésencedemédicamentsconcomitants.Comptetenudescaractéristiquespharmacodynamiquesetpharmacocinétiquespropres à chaque médicament, cette mesure d’exposition revêt un caractère majeur. Cependant, la façond’appréhenderlescyclesd’expositionauseindesbasesdedonnées‐médico‐administrativespeutvarierselonlesétudes.Or,onconnaîtpeul’impactdecesméthodesdemesuresurlesestimationsderisqueobtenues.Deplus, elles sont parfois peu adaptées à la prise en compte d’expositions concomitantes multiples, d’où lanécessitédedévelopperdenouvellesapproches.Aprèsavoirréaliséunerevuedesdonnéessurlemédicamentcontenuesdans lesbasesdedonnéesde l’assurancemaladie française,en insistantplusparticulièrementsurlesrupturesdansladisponibilitédesdonnées,desétudesdecasontétémenéesafind’explorercesquestionsdansdifférentscontextes.Dansunpremiertemps,unmodèlegénériqueaétéemployécommeprototyped’uneexposition discontinue, celui de la population générale utilisatrice de benzodiazépines anxiolytiques ethypnotiques,médicaments très répandus.Cette étude explorant lamortalité associéeauxbenzodiazépines aégalementétéutiliséepourévaluerl’impactdespériodesd’expositioninobservableslorsdeshospitalisations.Dansun second temps, des travauxont étémenésdans le champde l’onco‐hématologie, enprenant commemodèled’expositioncomplexe,àlafoisdiscontinueetmultiple,lesprotocolesdechimiothérapiedumyélomemultiple.Enfin,undernierprojet a étudié l’apportpotentiel desméthodesdevisualisationdedonnéespouraméliorer la description de l’exposition longitudinale au médicament et des situations de concomitance, etrendrepluspertinenteleurmodélisation.Cestravauxméthodologiquesontainsicherchéàaméliorerlavaliditéet la robustesse de lamesurede l’expositionmédicamenteusedansdes contextes d’expositionsmultiples etdiscontinues.
Titreenanglais:Measurementofdiscontinuousdrugexposureinlargehealthcaredatabases
Disciplineadministrative:Pharmacologie
Intituléetadressedulaboratoire:LaboratoiredePharmacologieMédicaleetCliniqueInserm1027,Équipe6–Pharmacoépidémiologie,évaluationdel'utilisationetdurisquemédicamenteuxUniversitédeToulouse,FacultédeMédecine,37,alléesJulesGuesde‐31000Toulouse