AutomatedDetectionofOff-LabelDrugUse
Table2.Selectedpredictednoveloff-labelusages.DrugSimvastatinTacrolimusPregabalinEtanerceptLamotrigineAdalimumabRituximabDaptomycinFludarabineInfliximabErlotinibIndicationdiabetesmellitusrheumatoidarthritismigrainedisorderslupuserythematosus,systemicmigrainedisorderslupuserythematosus,systemichodgkindiseaseosteomyelitiswaldenstrommacroglobulinemiapyodermagangrenosummalignantneoplasmofovaryFAERSSupport13694041527975715145393428MEDLINESupport334535634820666816Predicted,noveldrugusageswithsubstantialsupportinFAERS.FAERSSupportforeachdrug-indicationpairisthenumberofdistinctcasereportsinFAERSinwhichthedrugwasexplicitlylistedasbeingusedtotreattheindication.AcompletelistingisavailableinTableS1.doi:10.1371/journal.pone.0089324.t002normalizedtoliebetween0and1,withavalueof0fordrugswithnoadverseeventassociationsinMedi-Span(811outof1,602drugs)and1fordrugsassociatedwithmanyseriousadverseevents.Notsurprisingly,drugswiththehighestriskindiceswereimmunosupressants,suchasmycophenalatemofetil,andanti-tumoragents,suchasgemtuzumab,clofarabine,bevacizumab,andfludarabine.Well-supportednoveloff-labelusageshadriskindicesrangingfrom0.002foramphotericintoamaximumof0.995forclofarabine.
ThedrugcostindexisbasedonthemeanunitpriceforthedruginMedi-Spanandisalsonormalizedtoliebetween0and1,withavalueof1forthedrugwiththehighestmeanunitcostinMedi-Span.Theunitcostisanimperfectmeasureofactualtreatmentcost—forinstance,itmaybeforaquantitythatissufficientformultipletreatments.Nevertheless,thecostindexprovidesapartialorderingthatisusefulforrelativerankingbecausethedrugswiththehighestcostindexareexpensive,targetedtherapiessuchasranibizumab,whilethedrugswithlowcostindexvaluesareoverthecounteragentssuchasmagnesiumchlorideandiodine.
Weusedtheriskandcostindicestogroupwell-supportednoveloff-labelusagesintohighrisk,highcostandlowrisk,lowcostusages,resultingin28and51usages,respectively(thetop5usagesineachgrouparelistedinTable3;TableS3containsthefulllists).Wedefinedthresholdsforhighsandlowsbylookingatthedistributionoftheriskandcostindicesforthe403well-supportedusagesandchoosingtheupperandlowerquartilesascutoffs.Forexample,theupperquartileforthe403well-supportedusageshadriskindexvalue0.828,whichdefinesthethresholdforthehigh-riskgroup.Forthe403well-supportedusages,Figure1showsthehigh–high(28drug-indicationpairs)andlow–lowgroups(51drug-indicationpairs).Many(16of28)ofthehighrisk,highcostusagesinvolvedanti-tumoragentsbeingusedtotreatunapprovedtumortypes.Incontrast,thelowcost,lowriskusagescontainmanyoverthecounterdrugssuchasvitaminE,aswouldbeexpected.
Discussion
Off-labelusageofdrugsisanimportantenoughaspectofdrugsafetytowarrantafullissue(May2012)ofNatureClinicalTherapeuticsandPharmacologydevotedtothetopic[7].Currentlythemostcomprehensiveinformationaboutoff-labeldrugusageisfromtheNationalDiseaseandTherapeuticIndex
PLOSONE|www.plosone.org
5
(IMSHealth,PlymouthMeeting,PA),whichreliesonperiodicsurveysofoffice-basedphysicians.Webelievethatoff-labelusecanbelearnedsystematically,inadata-drivenmannerdirectlyfromelectronicmedicalrecords.Ourworkrepresentsthefirstefforttodetectnoveloff-labelusagefromclinicalfreetextovertheentirerangeofdrugsandindicationsobservedinthemedicalrecord.Wealsodevelopedquantitativeriskandcostindicesasawaytoprioritizethenovelusagesforfurtherinvestigation.
Inthepast,NLPhasbeenappliedtotheproblemofdetectingused-to-treatrelationshipsbetweendrugsandindicationsinclinicaltext.StateoftheartNLPapproachesrequiretrainingtextinwhichdrugandindicationmentionsarelabeled,alongwiththerelationshipsbetweenthem.Incontrast,associationbasedapproachesthatusecountsofdrugandindicationmentionsaremorescalable,butlimitedbyconfoundingcausalandindirectrelationships.Wehavedevelopedanautomatedmethodfordetectingnoveloff-labelusagesfromclinicaltextthatdoesnotrequiretrainingtextandaddressesconfoundingrelationshipsbyincorporatingpriorknowledgeaboutdrugusage.Weappliedthismethodto1,602drugsand1,475indicationstoidentify6,142noveloff-labelusages,403ofwhicharewellsupportedbyevidenceinindependentandcomplementarydatasets.
Ourmethodshaveimportantlimitations.First,ourworkfocusesononeformofoff-labeluse—theuseofdrugstotreatunapprovedindications—anddoesnotdetectoff-labelusewithrespecttoage,gender,dosageandcontraindications.Second,co-morbiditiesanddrugadverseeventsmaystillleadtospuriousused-to-treatrelationshipsdespiteoureffortstoreducetheirimpactonourresults.Third,althoughourmethodcandetectused-to-treatrelationshipsbetweendrugsandindicationswithhighspecificityandgoodsensitivity,thetaskofrecognizingwhethertheknowledgeisalreadyknownismoredifficultthanmightbeexpected.ThisdifficultywasnotduetoerrorsinrecognizingtermsinclinicaltextbutratherduetomismatchesinthelanguageusedtodescribeindicationsinMedi-SpanandtheNDF-RTversusclinicaltextandFAERS.Asystematiclistingofsuchindicationmismatchescouldidentifyareasinontologiesandterminologiesthatneedimprovement—andwouldbeadata-drivenwaytoidentifyportionsofterminologiesforreview.Fourth,theriskandcostindiceshavesomeshortcomings.Forinstance,thecostindexignoresthefactthatdosageanddurationoftreatmentforoff-labelusagesmaydifferfromapproveduses,andourrisk
February2014|Volume9|Issue2|e89324
AutomatedDetectionofOff-LabelDrugUse
Table3.Predictedoff-labelusagesbinnedbyriskandcostandrankedbysupportinFAERS.
Drug
Highrisk,highcostusagesDocetaxelClofarabineRituximabBevacizumabPaclitaxel
Lowrisk,lowcostusagesFolicacidMethadoneFolicacidMegestrolFolicacid
IndicationFAERSSupportMEDLINESupportRiskIndexCostIndex
MalignantneoplasmofprostateLeukemia,myelocytic,acute
Purpura,thrombocytopenic,idiopathicMalignantneoplasmofovaryMalignantneoplasmofstomach
604341259170122
6403716989421
0.9640.9950.9400.9910.956
0.9490.8690.8210.8790.776
mentaldepressiondepressivedisorderhyperlipidemia
carcinoma,non-smallcelllungdiarrhea
4931621387967
183310410
0.0820.0020.0820.0020.082
0.1020.1670.1020.2380.102
Werankedpredicted,noveloff-labelusagesonthebasisofriskandcost,asrepresentedbyourriskandcostindicesforeachdrug.FAERSSupportforeachdrug-indicationpairisthenumberofdistinctcasereportsinFAERSinwhichthedrugwasexplicitlylistedasbeingusedtotreattheindication.Theriskindexisaquantitativescorethatrepresentstheexpecteddisutilityofadverseeventsrelatedtotheuseofthedruginquestion,normalizedtotherange[0,1]sothatdrugsthathaveahigherriskofcausingseriousadverseeventshavehighervalues.ThecostindexisbasedonthemeanunitcostofthedruginquestioninMedi-Span,normalizedtotherange[0,1]withmoreexpensivedrugshavingahighervalue.doi:10.1371/journal.pone.0089324.t003
indexdoesnottakethedependenceofadverseeventsondosage,co-morbidities,andpoly-pharmacyintoaccount.Finally,inthisworkwehaveaimedtoproducealistofhighlyconfidentpredictionsofnoveloff-labelusagessowerequirecorroborationofpredictionsinFAERS,whichhasmuchlowerrecallinthetestsetthantheclassifier.Thustheoverallmethodsacrificessensitivityforgreaterspecificity.Thisisappropriateforouraiminthiswork,butotherstudiesmayrequireadifferenttradeoffbetweensensitivityandspecificity.Forinstance,ifwewereconcernedexclusivelywithpotentiallyriskyusages,wemightnotrequiresupportinFAERSandinsteadfilterforusagesinvolvingdrugsassociatedwithknownserioussideeffectsthatdon’talwaysgetreported.Wenotethatourmethodcanbemodifiedforsuchusecases.
Theselimitationsnotwithstanding,ourstudyisthefirstlarge-scalecharacterizationofoff-labelusageusingfullyautomatedmethodstocombineinformationfromclinicalnoteswithpriorknowledgeandtoprovidearankingofthelearnedusagesonriskandcost.Itisasteptowardssystematic,data-drivenmonitoringofoff-labelusage.ThemethodhascharacteristicsthatallowittogeneralizetositesbeyondStanford.First,thesystemdoesnotrequiretrainingtextlabeledwithmentionsofdrugsandindications,andtherelationshipsbetweenthem.Second,ourmethodisveryflexiblewithrespecttothetargetdrugandindicationvocabulary.Third,thesystemisveryfast—annotationof9.5millionclinicalnotestakesonlytwohoursonasinglemachine;constructingfeatures,trainingaclassifierandmakingpredictionstakesanadditionalfewhours.Itisthusconceivabletoprocessclinicaltextfromalargenumberofsites,providingapictureofoff-labelusageacrossawidespectrumofinstitutions.Mostimportantly,ourmethodwasabletodetectusagesthatweredocumentedinthebiomedicalliterature,andinonecaseapprovedintheEU,despitenotappearinginanyofourcuratedsourcesofknownusage.Thissuggeststhatsuchsystemscouldpotentiallyprovideanautomatedlearningsystemforoff-labelusage.Suchassystemcouldflagemergingusagesbeforetheycometotheattentionofthebroadermedicalcommunity,regulatoryagenciesanddrugmanufacturers,inmuchthesamewaythatGoogleFluTrendscanprovideanearlywarningofflu
PLOSONE|www.plosone.org
6
trendsinadvanceofCDCdata[45].Wespeculatethatapplyingourmethodtoawiderrangeofclinicaltextfrommultiplesitescanprovideatimelierandmorecomprehensivepictureofoff-labelusagethaniscurrentlypossible[46,47].
MaterialsandMethodsConstructingagoldstandard
WeconstructedagoldstandardofpositiveandnegativeexamplesofdrugusageusingknownusagesfromMedi-Span.Medi-Spancontains13,453drug-indicationpairscomprising1,642uniquedrugsand2,313uniqueindications.Ofthese,1,602ofthedrugsand1,475oftheindicationsoccurinSTRIDEatleastonce,yieldingasetof8,861testabledrug-indicationpairs.Toconstructnegativeexamples,wesampledknownusagesfromMedi-Spanwithreplacementandthensamplednewdrugsandindicationsthatoccurinthedatawithapproximatelythesamefrequency.Forinstance,giventheknownusage‘‘dexamethasoneforsystemiclupuserythematosus’’,wesampleanewdrugfromthesetofdrugsthatoccurwithintenitemsofdexamethasoneinalistofdrugssortedbyoverallfrequencyinthedata.Anewindicationissimilarlygeneratedfromsystemiclupuserythematosus.Frequencymatchingwasdonebecausepreviousworksuggestedthatfrequenciescanhelpdistinguishbetweendrugassociatedadverseeventsandtreatmentrelationships[48].The‘‘negative’’pairswerefilteredtoremoveinadvertentknownusages.Thefinalgoldstandardconsistedof34,974negativeand8,861positiveexamples.
AnnotationofclinicaltextfromSTRIDE
WeusedtheNCBOAnnotatoronfreetextof9.5millionclinicalnotesfromSTRIDEtoannotatetheeachnotewithmentionsofdrugsandindicationsintermsofUMLS[49]uniqueconceptidentifiers(CUI’s).Negatedmentions(e.g.,‘‘MIwasruledout’’)orthosereferringtootherpeople(e.g.,‘‘fatherhadastroke’’)wereremovedusingNegEx[50]andConText[51],respectively.Drugswerenormalizedto1,602uniqueactiveingredients(e.g.,Excedrinwasrewrittenintoacetaminophen,aspirinandcaffeine)
February2014|Volume9|Issue2|e89324
AutomatedDetectionofOff-LabelDrugUse
usingRxNorm[52].Indicationswerenormalizedtothesetof1,475indicationsusedinMedi-SpanbyrecursivelyrewritingtheindicationasitsparentsintheSNOMEDCThierarchyuntilwereachedanindicationusedbyMedi-Span.Forinstance,‘amok’isnotintheMedi-Spantargetvocabularysoitisrewrittenasitsparentterm,‘mania.’Wenotethatifthementionedindicationisanancestoroftheknownindication,itmaybecountedasanoveloff-labelusagelateron.Weconsiderthistobereasonablebecauseifthedetectedusageisbroaderthantheknown,approvedusage,itisindeedoff-labelprovidedthetermsareusedpreciselyasintended.Inreality,termsarenotusedsoprecisely,soweallowforsomeimprecisionintheusageoftermswhenfilteringoutknownusagesfrompredictedusagesasdescribedbelow.Theclinicalnotescovered1.6millionpatientsandspanned18yearsofdata,andincludedallclinicalnotesgeneratedforthesepatientsatStanfordHospitalduringthattime.
FeatureConstruction
Foreachpatient,adrugorindicationiscountedaspresentiftheyappearinanyofthepatient’snotes.Theycountasco-occurringiftheyarebothmentionedinthepatient’snotesandthereisnootherindicationmentionedintherecordthatisaknownusageforthedrug;allco-occurrencesofknownindicationsarealsocounted.Doingsoensuresthatadrug(e.g.Lisinopril)doesnotgetassociatedwithadisease(e.g.Diabetes)justbecausethediseaseisacommonco-morbidityofthedrug’sactualindication(e.g.Hypertension).Inthisprocess,knownusageisdefinedasappearingineitherMedi-SpanorNDF-RT.Thesecounts,alongwithderivedassociationmeasures(chisquaredstatistic,oddsratioandconditionalprobabilityofdrugmentiongivenindicationmention),wereusedasfeatures.Thefractionofpatientsinwhichthedrugoccursbeforetheindication(drugfirstfraction)wasalsoincluded,alongwithdrugfirstfractionsadjustedforfrequencyofthedrugsandindications[48].Overall,weusedninefeaturesencodingthepatternofmentionsofthedrugsandindicationsinclinicaltext.
Wealsousedfeaturesthatencodepriorknowledgeofthedrugs,indicationsandknownusage.Thesefeaturesweremotivatedbytheintuitionthatdrugsaretypicallyusedoff-labelbecauseofsomesimilaritywithanapproveddrug,suchasasharedmoleculartarget,pathwayordrugclass[7].WeusedtheMedi-SpanandDrugBankdatabasestoconstructfeaturesforeachdrug-indicationpair.ForMedi-Span,theseincludedthenumberofdrugsapprovedorknowntobeusedfortheindication,thefractionofknowntreatmentsfortheindicationthatareapproved,thesimilarityofthedrugtodrugsknowntobeusedfortheindication,andthesimilarityoftheindicationtootherindicationstreatedbythedrug.Drug-drugsimilarityfeatureswerecalculatedasdescribedinFigure4.Indication-indicationsimilaritieswerecalculatedsimilarly,withtheroleofthedrugsandindicationsreversed.Whencalculatingthesefeatures,weignoredknownusagesthatwereinthetestsettoavoidcontaminatingthetrainingdatawithknowledgeoftestusages.
TheDrugBank3.0[53]databaseprovidesinformationon6,711drugsandtheirmoleculartargets,pathways,andindications.TheannotatorwasusedtomapDrugBankdrugnamesandindicationstoourtargetsetsofdrugsandindications.Moleculartargets,pathways,anddrugcategorieswerealsoextractedforeachdrug.WecalculatedsimilarityfeaturesanalogoustotheMedi-Spansimilarityfeatures,alongwithotherfeaturesthatcapturesimilaritywithrespecttomoleculartargets,pathways,anddrugcategories.AswiththeMedi-Spanderivedfeatures,weremovedtestusagesfromDrugBankbeforecalculatingfeatures.SeeTableS4foracompletelistoffeatures.
PLOSONE|www.plosone.org
7
Figure4.Usingpriorknowledgetocalculatedrug-drugandindication-indicationsimilarity.Werepresentknownusageasamatrixwhererowirepresentsdrugiandcolumnjrepresentsindicationj.Acheckinentry(i,j)indicatesthatthedrugiisusedtotreattheindicationj,whileacrossindicatestheconverse.Weareinterestedinwhetheragivendrug,lamotrigine,isusedtotreatmigrainedisorders.Wethusask—howsimilaristheknownusageoflamotriginetootherdrugsweknowareusedtotreatmigrainedisorders?Topirimateisusedtotreatmigrainedisorders,andlamotrigineissimilartoitinthatbothareusedtotreattonic-clonicseizuresandmyoclonicepilepsies,butnotnon-Hodgkin’slymphoma.Thissimilarityinusageprofilesuggeststhatitismorelikelytobeusedtotreatmigrainedisordersthan,say,Rituximab.WemeasuredthissimilarityusingthemaximumcosineandJaccardsimilarityoflamotrigineversusalldrugsknowntotreattheindication.Wecalculatethesimilaritybetweenindicationsbasedonknownusageusingthesamedata,withtherolesofdrugsandindicationsreversed.
doi:10.1371/journal.pone.0089324.g004
Trainingapredictivemodel
Thegoldstandarddatasetwasrandomlysplitinto35,050trainingand8,784testexamples.WetrainedanSVMclassifierusingradialbasisfunctionkernelsonthetrainingexamplesusingthee1071libraryinR.Theperformanceoftheclassifierwastestedonthetestexamples.Wealsotrainedandtestedclassifiersusingsubsetsofthefeaturestoassessthecontributionofdifferentgroupsoffeatures.Wethentrainedaclassifierontheentiregoldstandardandappliedittoall2,362,950possibledrug-indicationpairs.Inallcases,weusedten-foldcrossvalidationonthetrainingdataandthe‘‘1-se’’ruletoselectthecosthyperparameterfortheSVMmodels.Estimatesofeachprediction’sclassmembershipprobabilitieswereobtainedvialogisticregression[26].Weusedaprobabilitythresholdof0.99inordertolimitthesetofpredictedusagestothemostconfidentpredictions.Thishardthresholdwasnottunedinanyway;thusourfinalsetofpredictednoveloff-labelusagescouldpossiblybeimprovedbyadjustingthisthreshold.However,thisisaverysimplewaytorestrictourattentiontothemostconfidentpredictions.
Identifyingknownusages
KnownusagewasdeterminedbypresenceinMedi-SpanortheNDF-RT–drug-indicationpairsabsentfrombothareassumedtobenovelusages.Reviewoftheseusagesrevealedthatindicationsweresometimescloselyrelatedtoknownusages—e.g.,glaucomaandopenangleglaucoma.Weaddressedthisproblemusingbiomedicalontologies,whichorganizebiomedicalconceptsintohierarchies—e.g.,amokisasubtypeoftheparentconceptmania.
February2014|Volume9|Issue2|e89324
AutomatedDetectionofOff-LabelDrugUse
Specifically,weremovedpredictedusagesinwhichtheindicationisasubtypeofaknownusageindicationinSNOMED-CT,oristhedirectparentofaknownusageindicationinSNOMED-CT.AsafinalcheckagainstMedi-SpanandNDF-RT,wemanuallyreviewedpredictedusagesremainingaftervalidationinFAERS,MEDLINEandSIDER2(describedbelow),removing63usagesthatwerenotdetectedasknownusagesbythemethodsdescribedabove.
ValidationbyFAERS,MEDLINEandmechanisticplausibility
FAERScasereportscontainexplicitused-to-treatlinksbetweendrugsandindications.WevalidatedpredictedusagesusingtheselinksusingpublicdomaincasereportsfromQ32007throughQ22012.FAERSdrugsandindicationsweremappedtoUMLSCUI’s,yieldingasetof3milliondrug-indicationreportscovering160,989uniquepairs.Only3,756outof8,861(43%)positiveexamplesofdrugusagesinourgoldstandarddatasetappearatleastonceinFAERS.Werequiredatleast10suchreportsbecausethisthresholdresultsinafalsepositiverateoflessthan0.005whenappliedtothegoldstandarddataset.
MEDLINEentriesaremanuallyannotatedwithMeSHterms,SupplementaryConceptsfordrugs,andsubheadingsthatprovidefurthercontextfortheannotation.Forinstance,anarticleabouttreatmentofwetmaculardegenerationbybevacizumabwouldbeannotatedwith‘‘wetmaculardegeneration/drugtherapy*’’and,separately,‘‘bevacizumab.’’WedownloadedthecompletesetofannotationsforMEDLINEentriesfrom2002–2012.MeSHannotationswerefilteredforindicationswiththedrugtherapysub-headingandmappedtoUMLSCUI’susingtheNCBOAnnotator.TheSubstanceannotationswerealsomappedtoUMLSCUI’susingtheAnnotator.IfnoMeSHtermcorrespond-edexactlytotheindication,weexpandedtheindicationtothemoregeneralMeSHterm,e.g.,‘malignantneoplasmoftheovary’wasinterpretedas‘ovarianneoplasm’.AsinAvillachetal[28],weconsideredusageswithatleastthreearticlesannotatedwithboththeindicationandthedrugtobewell-supportedbyMEDLINE.Weassessedthemechanisticplausibilityofpredictedusagesbyexaminingpatternsofgeneexpressioninducedbythedrugandindication.Briefly,weperformedgenesetenrichmentanalysisongeneexpressiondatafromtheNCBIGeneExpressionOmnibus(GEO)[41,54]andtheConnectivityMap[40]toidentifybiologicalpathwaysandexpressionmodulesthatareinverselyregulatedbetweenpairsofdiseasesanddrugs,suggestingapossiblebasisforatherapeuticassociation[24].DetailsofthismethodareinMethodsS1.
investigation.Decisionanalysissuggeststhatwerankusagesbasedontheirexpectedutility—i.e.,thedesirabilityofpossibleoutcomesoftheuse,weightedbytheprobabilityeachoutcome[55].Forexample,theuseofacheapantibioticwithfewsideeffectstotreatarareconditionhasalowerurgencyforfollow-upthantheuseofanexpensivedrug,withseveresideeffects,totreatacommondisorder.Weapproximatedthisapproachbydevelopingquantitativeindicesofdrugcostandriskassociatedwithdrugusagebasedonknownadverseevents.
ThecostindexwascalculatedbyrankingdrugsbytheirmeanunitcostinMedi-Span(adrugmayhavemultipleunitcostsduetodifferentformulations,etc.).Therankswerenormalizedtoliebetween0and1,withthemostexpensivedrughavingascoreof1.TheriskindexforeachdrugwasbasedonanestimateoftheexpecteddisutilityofadverseeventsassociatedwithusingthatdruginMedi-Span,describedindetailinMethodsS1.Briefly,weassignedquantitativedisutilityvaluestoadverseeventsassociatedwithdrugsinMedi-Span.Theexpecteddisutilityofdruguseduetoadverseeventswasthenestimatedastheweightedsumofthedisutilitiesforassociatedadverseevents,withtheweightsgivenbyprobabilitiesestimatedfromMedi-Span’sestimatesofthefrequencyoftheadverseevents.Drugswererankedbyexpecteddisutilityandtheranksnormalizedtoliebetween0and1suchthattheriskiestdrughadavalueof1.Thelowerandupperquartilesofthecostandriskindexvaluesobservedinthe403well-supportednovelusageswereusedasthresholdsfordefininghighandlowriskorcostgroups.
SupportingInformation
TableS1Highconfidencepredictedoff-labeldrug
usagesvalidatedinFAERSandMEDLINE,withcostandriskindexvalues.(PDF)
TableS2Molecularplausibilityofoff-labelusages
evaluatedusingGSEAandmicroarraygeneexpressiondatafromGEO.(PDF)
TableS3Highcost,highriskandlowcost,lowrisk
usages.(PDF)
TableS4Featuresusedintheclassifierfortheused-to-
treatrelationship.(PDF)
MethodsS1
Removalofdrugadverseevents
WeuseddrugadverseeventslistedintheSIDER2resourcetominimizetheimpactofconfoundingcausalrelationshipsonourresults.67outof406novel,well-supportedoff-labelusagesmatchedSIDER2entries.However,manualreviewofthesematchesrevealedthatonly28drug-indicationpairswerelikelytobebonafidedrugrelatedadverseevents.ThisisduetothefactthatSIDER2isnotcuratedandthusincludesmanyspuriousresultssuchasindicationsbeinglistedasadverseevents.Afterremovalofthetrueadverseevents,403off-labelusagesremained.
Additionaldetailsaboutmethodsusedin
thiswork.(ZIP)
Acknowledgments
WegratefullyacknowledgeRandyStafford,MarkMusen,ChrisManning,andHowardStrasburgforfeedbackonthemanuscriptandoff-labeldrugusage.
AuthorContributions
Conceivedanddesignedtheexperiments:KJNHSJD.Performedtheexperiments:KJWCBR.Analyzedthedata:KJWCBRJD.Contributedreagents/materials/analysistools:KJPLWCSVIBRJD.Wrotethepaper:KJNHSJDPL.
Calculationofriskandcostindices
Thecostandriskindicesaremotivatedbytheobservationthatoff-labelusagesdonotallhavethesameurgencyforfurther
PLOSONE|www.plosone.org8February2014|Volume9|Issue2|e89324
AutomatedDetectionofOff-LabelDrugUse
References
1.StaffordRS(2012)Off-labeluseofdrugsandmedicaldevices:areviewofpolicyimplications.ClinPharmacolTher91:920–925.
2.DalPanGJ(2012)Monitoringthesafetyofmedicinesusedoff-label.ClinPharmacolTher91:787–795.
3.RadleyDC,FinkelsteinSN,StaffordRS(2006)Off-labelprescribingamongoffice-basedphysicians.ArchInternMed166:1021–1026.
4.ChenDT,WyniaMK,MoloneyRM,AlexanderGC(2009)U.S.physicianknowledgeoftheFDA-approvedindicationsandevidencebaseforcommonlyprescribeddrugs:resultsofanationalsurvey.PharmacoepidemiolDrugSaf18:1094–1100.
5.FlowersCM,RacoosinJA,KortepeterC(2006)Seizureactivityandoff-labeluseoftiagabine.NEnglJMed354:773–774.
6.KimlandE,OdlindV(2012)Off-labeldruguseinpediatricpatients.ClinPharmacolTher91:796–801.
7.EpsteinRS,HuangSM(2012)Themanysidesofoff-labelprescribing.ClinPharmacolTher91:755–758.
8.MorrisJ(2012)Theuseofobservationalhealth-caredatatoidentifyandreportonoff-labeluseofbiopharmaceuticalproducts.ClinPharmacolTher91:937–942.
9.PlattR,CarnahanRM,BrownJS,ChrischillesE,CurtisLH,etal.(2012)TheU.S.FoodandDrugAdministration’sMini-Sentinelprogram:statusanddirection.PharmacoepidemiolDrugSaf21Suppl1:1–8.
10.Wei-QiW,CroninR,XuH,LaskoT,BastaracheL,etal.(2013)Development
ofanensembleresourcelinkingMEDicationstotheirIndications(MEDI).AMIASummitonTranslationalBioinformatics2013.SanFrancisco,CA.pp.172–176.
11.XuR,WangQ(2013)Large-scaleextractionofaccuratedrug-diseasetreatment
pairsfrombiomedicalliteraturefordrugrepurposing.BMCBioinformatics14:181.
12.PoissantL,TaylorL,HuangA,TamblynR(2010)Assessingtheaccuracyofan
inter-institutionalautomatedpatient-specifichealthproblemlist.BMCMedInformDecisMak10:10.
13.UzunerO,SouthBR,ShenS,DuVallSL(2011)2010i2b2/VAchallengeon
concepts,assertions,andrelationsinclinicaltext.JAmMedInformAssoc18:552–556.
14.ChapmanWW,NadkarniPM,HirschmanL,D’AvolioLW,SavovaGK,etal.
(2011)OvercomingbarrierstoNLPforclinicaltext:theroleofsharedtasksandtheneedforadditionalcreativesolutions.JAmMedInformAssoc18:540–543.15.ChenES,HripcsakG,XuH,MarkatouM,FriedmanC(2008)Automated
acquisitionofdiseasedrugknowledgefrombiomedicalandclinicaldocuments:aninitialstudy.JAmMedInformAssoc15:87–98.
16.RindfleschTC,PakhomovSV,FiszmanM,KilicogluH,SanchezVR(2005)
Medicalfactstosupportinferencinginnaturallanguageprocessing.AMIAAnnuSympProc:634–638.
17.LePenduP,LiuY,IyerS,UdellM,ShahNH(2012)AnalyzingPatternsofDrug
UseinClinicalNotesforPatientSafety;2012March21–23,2012;SanFrancisco,CA.
18.LependuP,IyerSV,Bauer-MehrenA,HarpazR,MortensenJM,etal.(2013)
PharmacovigilanceUsingClinicalNotes.ClinPharmacolTher.
19.IyerSV,LePenduP,HarpazR,Bauer-MehrenA,ShahNH(2013)Learning
SignalsofAdverseDrug-DrugInteractionsfromtheUnstructuredTextofElectronicHealthRecords.AMIASummitsTranslSciProc2013.
20.JungK,LependuP,ShahNH(2013)AutomatedDetectionofSystematicOff-labelDrugUseInFreeTextofElectronicMedicalRecords;SanFrancisco,CA.AmericanMedicalInformaticsAssociation.pp.94–99.
21.LiY,SalmasianH,HarpazR,ChaseH,FriedmanC(2011)Determiningthe
reasonsformedicationprescriptionsintheEHRusingknowledgeandnaturallanguageprocessing.AMIAAnnuSympProc2011:768–776.
22.KuhnM,CampillosM,LetunicI,JensenLJ,BorkP(2010)Asideeffect
resourcetocapturephenotypiceffectsofdrugs.Molecularsystemsbiology6:343.
23.BrownSH,ElkinPL,RosenbloomST,HusserC,BauerBA,etal.(2004)VA
NationalDrugFileReferenceTerminology:across-institutionalcontentcoveragestudy.StudHealthTechnolInform107:477–481.
24.SirotaM,DudleyJT,KimJ,ChiangAP,MorganAA,etal.(2011)Discovery
andpreclinicalvalidationofdrugindicationsusingcompendiaofpublicgeneexpressiondata.SciTranslMed3:96ra77.
25.HastieT,TibshiraniR,FriedmanJ(2009)TheElementsofStatisticalLearning:
Springer-Verlag.763p.
26.PlattJ(1999)ProbabilisticOutputsforSupportVectorMachinesand
ComparisonstoRegularizedLikelihoodMethods.AdvancesInLargeMarginClassifiers:61–74.
27.Weiss-SmithS,DeshpandeG,ChungS,GogolakV(2011)TheFDAdrugsafety
surveillanceprogram:adverseeventreportingtrends.Archivesofinternalmedicine171:591–593.
28.AvillachP,DufourJC,DialloG,SalvoF,JoubertM,etal.(2013)Designand
validationofanautomatedmethodtodetectknownadversedrugreactionsinMEDLINE:acontributionfromtheEU-ADRproject.JAmMedInformAssoc20:446–452.
29.PooleSG,DooleyMJ(2004)Off-labelprescribinginoncology.SupportCare
Cancer12:302–305.
30.(2005)‘‘Off-label’’indicationsforoncologydruguseanddrugcompendia:
historyandcurrentstatus.JOncolPract1:102–105.
31.HagenbeekA,GadebergO,JohnsonP,PedersenLM,WalewskiJ,etal.(2008)
Firstclinicaluseofofatumumab,anovelfullyhumananti-CD20monoclonalantibodyinrelapsedorrefractoryfollicularlymphoma:resultsofaphase1/2trial.Blood111:5486–5495.
32.OrR,ShapiraMY,ResnickI,AmarA,AckersteinA,etal.(2003)
Nonmyeloablativeallogeneicstemcelltransplantationforthetreatmentofchronicmyeloidleukemiainfirstchronicphase.Blood101:441–445.
33.LamplC,KatsaravaZ,DienerHC,LimmrothV(2005)Lamotriginereduces
migraineauraandmigraineattacksinpatientswithmigrainewithaura.JNeurolNeurosurgPsychiatry76:1730–1732.
34.PizzolatoR,VillaniV,ProsperiniL,CiuffoliA,SetteG(2011)Efficacyand
tolerabilityofpregabalinaspreventivetreatmentformigraine:a3-monthfollow-upstudy.JHeadachePain12:521–525.
35.MerrillJT,NeuweltCM,WallaceDJ,ShanahanJC,LatinisKM,etal.(2010)
Efficacyandsafetyofrituximabinmoderately-to-severelyactivesystemiclupuserythematosus:therandomized,double-blind,phaseII/IIIsystemiclupuserythematosusevaluationofrituximabtrial.ArthritisRheum62:222–233.36.MerrillJT,Burgos-VargasR,WesthovensR,ChalmersA,D’CruzD,etal.
(2010)Theefficacyandsafetyofabataceptinpatientswithnon-life-threateningmanifestationsofsystemiclupuserythematosus:resultsofatwelve-month,multicenter,exploratory,phaseIIb,randomized,double-blind,placebo-controlledtrial.ArthritisRheum62:3077–3087.
37.HayatSJ,UppalSS,NarayananNampooryMR,JohnyKV,GuptaR,etal.
(2007)SafetyandefficacyofinfliximabinapatientwithactiveWHOclassIVlupusnephritis.ClinRheumatol26:973–975.
38.ShakoorN,MichalskaM,HarrisCA,BlockJA(2002)Drug-inducedsystemic
lupuserythematosusassociatedwithetanercepttherapy.Lancet359:579–580.39.MartinDF,MaguireMG,YingGS,GrunwaldJE,FineSL,etal.(2011)
Ranibizumabandbevacizumabforneovascularage-relatedmaculardegener-ation.NEnglJMed364:1897–1908.
40.LambJ,CrawfordED,PeckD,ModellJW,BlatIC,etal.(2006)The
ConnectivityMap:usinggene-expressionsignaturestoconnectsmallmolecules,genes,anddisease.Science313:1929–1935.
41.EdgarR,DomrachevM,LashAE(2002)GeneExpressionOmnibus:NCBI
geneexpressionandhybridizationarraydatarepository.NucleicAcidsRes30:207–210.
42.GripO,JanciauskieneS,LindgrenS(2002)AtorvastatinactivatesPPAR-gammaandattenuatestheinflammatoryresponseinhumanmonocytes.InflammRes51:58–62.
43.AltshulerD,HirschhornJN,KlannemarkM,LindgrenCM,VohlMC,etal.
(2000)ThecommonPPARgammaPro12Alapolymorphismisassociatedwithdecreasedriskoftype2diabetes.NatGenet26:76–80.
44.PerrenTJ,SwartAM,PfistererJ,LedermannJA,Pujade-LauraineE,etal.
(2011)Aphase3trialofbevacizumabinovariancancer.NEnglJMed365:2484–2496.
45.GinsbergJ,MohebbiMH,PatelRS,BrammerL,SmolinskiMS,etal.(2009)
Detectinginfluenzaepidemicsusingsearchenginequerydata.Nature457:1012–1014.
46.HalevyA,NorvigP,PereiraF(2009)TheUnreasonableEffectivenessofData.
IeeeIntelligentSystems24:8–12.
47.BankoM,BrillE(2001)Scalingtoveryverylargecorporafornaturallanguage
disambiguation.39thAnnualMeetingoftheAssociationforComputationalLinguistics,ProceedingsoftheConference:26–33.
48.LiuY,LependuP,IyerS,ShahNH(2012)Usingtemporalpatternsinmedical
recordstodiscernadversedrugeventsfromindications.AMIASummitsTranslSciProc2012:47–56.
49.BodenreiderO(2004)TheUnifiedMedicalLanguageSystem(UMLS):
integratingbiomedicalterminology.NucleicAcidsRes32:D267–270.
50.ChapmanWW,BridewellW,HanburyP,CooperGF,BuchananBG(2001)A
simplealgorithmforidentifyingnegatedfindingsanddiseasesindischargesummaries.JBiomedInform34:301–310.
51.ChapmanWW,ChuD,DowningJN(2007)ConText:analgorithmfor
identifyingcontextualfeaturesfromclinicaltext.ProceedingsoftheWorkshoponBioNLP:81–88.
52.NelsonSJ,ZengK,KilbourneJ,PowellT,MooreR(2011)Normalizednames
forclinicaldrugs:RxNormat6years.JAmMedInformAssoc18:441–448.53.KnoxC,LawV,JewisonT,LiuP,LyS,etal.(2011)DrugBank3.0:a
comprehensiveresourcefor’omics’researchondrugs.NucleicAcidsRes39:D1035–1041.
54.BarrettT,EdgarR(2006)Geneexpressionomnibus:microarraydatastorage,
submission,retrieval,andanalysis.MethodsEnzymol411:352–369.
55.MeltzerDO,HoomansT,ChungJW,BasuA(2011)Minimalmodeling
approachestovalueofinformationanalysisforhealthresearch.MedDecisMaking31:E1–E22.
PLOSONE|www.plosone.org9February2014|Volume9|Issue2|e89324
相关推荐：
loading