Ten-year performance of Influenzanet: ILI time series, risks, vaccine effects, and care-seeking behaviour

(1)

ContentslistsavailableatScienceDirect

Epidemics

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

Ten-year

performance

of

Inﬂuenzanet:

ILI

time

series,

risks,

vaccine

effects,

and

care-seeking

behaviour

Sander

P. van

Noort

1,∗

_,

_Cláudia

_T.

_Codec¸

_o

1,2

_,

_Carl

_E.

_Koppeschaar

3

_,

_Marc

_van

_Ranst

4

_,

Daniela

Paolotti

5

_,

_M.

_Gabriela

_M.

_Gomes

1

1_Instituto_Gulbenkian_de_Ciência,_Oeiras,_Portugal 2_Oswaldo_Cruz_Foundation,_Rio_de_Janeiro,_Brazil

3_De_Grote_{Griepmeting/Science}_in_Action,_Amsterdam,_The_Netherlands 4_Rega_Institute_for_Medical_Research,_University_of_Leuven,_Leuven,_Belgium 5_ISI_Foundation,_Turin,_Italy

a

r

t

i

c

l

e

i

n

f

o

Articlehistory: Received4January2014

Receivedinrevisedform27April2015 Accepted31May2015

Availableonline15June2015

Keywords: Syndromicsurveillance Inﬂuenza-likeillness Timeseries Riskfactors Vaccineeffectiveness

a

b

s

t

r

a

c

t

Recent public health threats havepropelled major innovationson infectious diseasemonitoring, culminatingin thedevelopmentof innovativesyndromicsurveillancemethods. Influenzanetis an internet-basedsystemthatmonitorsinfluenza-likeillness(ILI)incohortsofself-reportingvolunteers inEuropeancountriessince2003.Weinvestigateandconfirmcoherencethroughthefirsttenyearsin comparisonwithILIdatafromtheEuropeanInfluenzaSurveillanceNetworkanddemonstrate country-specificbehaviourofparticipantswithILIregardingmedicalcareseeking.Usingregressionanalysis,we determinethatchronicdiseases,beingachild,livingwithchildren,beingfemale,smokingandpetsat home,areallindependentpredictorsofILIrisk,whereaspracticingsportsandwalkingorbicyclingfor locomotionareassociatedwithasmallriskreduction.Noeffectforusingpublictransportationorliving alonewasfound.Furthermore,wedeterminethevaccineeffectivenessforILIforeachseason.

1. Introduction

Recent concerns with emerging infectious diseases have exposeddeﬁcienciesindiseasesurveillancesystemsandimpelled radical rethinking on how to monitor population health and detectanomaliesinrealtime(Butler,2006).Inthiscontext,new approachesinsyndromicsurveillance–thecollectionand inter-pretationof data for publichealth beforelaboratoryor clinical conﬁrmationisavailable(Lazarusetal.,2001;Mandletal.,2004) – have emerged. Several systems are in evaluation, showing a largediversityofdatasourcesandmethodologiesemployed,such as telephone-based health information services (Cooper et al., 2008),automatedmedicalrecords(Lazarusetal.,2001;vanden Wijngaardetal.,2008),pharmacysalesandabsenteeism(Chretien et al., 2008), queries to online search engines (Ginsberg et al., 2009),andtelephone-basedself-reportingincohortsofrandomly selectedparticipants(Merketal.,2013;Rehnetal.,2014). Syn-dromicsurveillanceiscomplementarytotraditionalpublichealth

∗ Correspondingauthorat:InstitutodeMatemáticaeEstatísticadaUniversidade deSãoPaulo,RuadoMatão,1010,CEP05508-090SãoPaulo,SP,Brazil.

E-mailaddress:[email protected](S.P.vanNoort).

surveillanceindiseasereporting(Henning,2004;Lipsitchetal., 2009).

Influenzanetisamonitoringsystemforinfluenza-likeillness (ILI)in voluntarycohortsofinternet users.It wasinitially con-ceivedtomakescientificinformationaccessibletoabroadpublic andtokindlestudents’enthusiasmforscience,andwaslaunched intheNetherlandsandBelgium(www.degrotegriepmeting.nl)in 2003andinPortugal(www.gripenet.pt)in2005.Thesystemwas consecutively implemented in Italy (www.influweb.it)in 2008, UnitedKingdom(www.flusurvey.org.uk)in2009,Sweden(www. halsorapport.se) in 2011, France (www.grippenet.fr) and Spain (www.gripenet.es)in 2012, and Denmark (www.influmeter.dk) andIreland(www.flusurvey.ie)in2013.Similarsystemshavebeen implemented outside Europe,most notably in Australia (www. flutracking.net)in2007,Mexico(reporta.c3.org.mx)in2009,and theUnitedStates(www.flunearyou.com)in2011.

Basedonsingle-seasonanalysis,previousstudiesestablished good correlations between ILI incidences as determined by InﬂuenzanetandbytheclinicalsurveillancebysentinelGeneral Practitioners (GPs) as coordinated by the European Centre for Disease Prevention and Control (ECDC) (Friesema et al., 2009; Marquetetal.,2006;vanNoortetal.,2007;Paolottietal.,2014; Vandendijcketal.,2013).TheabsoluteILIincidenceasreportedby http://dx.doi.org/10.1016/j.epidem.2015.05.001

(2)

Influenzanetis,however,muchmoreconsistentacrosscountries accordingtoInfluenzanetthanreportedbytheECDC,dueto coun-tryspecificmedicalcareseekingratesanddisparitiesinILIcase definitions usedbyGPs in differentcountries(vanNoort etal., 2007).ThisuniformityinratesreportedacrossEuropeancountries facilitatesthegeographicalanalysisand modellingofepidemics (vanNoortetal.,2012).IntheUnitedKingdom(Eamesetal.,2012a) andFrance(Debinetal.,2014),thevaccineeffectivenessforILIhas beendeterminedforasingleseason.Byintegratingserologicaldata sources,ILIratesreportedbyInfluenzanethavebeenconvertedto estimatesofinfluenzaattackrates(Patterson-Lombaetal.,2014).

HereweaimtofurtherestablishtheInfluenzanetsystemasa validsentinelforILIsurveillance,byconfirmingthatboththe tim-ingandrelativeintensitiesofepidemicsareconsistentwiththose reportedbyECDC,andtheidentifiedriskfactorsforILIare consis-tentwiththoseinthepublishedliterature.Theanalysisisbased ondatacollectedoverthefirst10seasons(2003–2013)fromthe countriesinwhich Influenzanetwasimplementedforatleast5 seasons:theNetherlands,Belgium,Portugal,andItaly.Timeseries analysesareappliedtocompareILIincidencesfromInfluenzanet andECDC,whereasregressionanalysisisusedtodetermine indi-vidualriskfactorsbasedonpersonalcharacteristicsandvaccination status.Furthermore,basedonthehealthcareseekingbehaviouras reportedtoInfluenzanet,differencesinILIincidenceby Influen-zanetandECDCareexplained.

2. Method

2.1. Datacollection

Inﬂuenzanetparticipantsarerecruitedfromthegeneral popu-lationbycompletinganintakequestionnaireononeofthenational websites,containingvariousdemographicandlifestylequestions. Duringtheinﬂuenzaseason,participantsreceiveaweekly newslet-terbye-mailinwhichtheyaredirectedtoanonlinequestionnaire aboutanumberofsymptomsthattheymighthaveexperienced sincetheirlastreport.TheEthicsCommitteeofInstitutoGulbenkian deCiênciaapprovedthestudy.

2.2. Participants

Anactivepopulationofparticipantsisessentialforthe consis-tencyofthesystem.Animportantcornerstoneforsuccessisthe feedbackofinformationtokeeptheparticipantsinvolvedand moti-vated.Althoughthespecificrecruitmentstrategiesvarybetween countries(Bajardi etal., 2014), theytend tobebased onmass communication.Thewebsitescontainawealthofinformationon influenza,ILIandcommoncold,whiletheeducationalandscientific aimsoftheprojectareexplainedindirectmailingstoschools,in repeatedinterviewsontelevisionandradio,andinnewspapers. Schools areprovided witheducationalmaterial oninfluenzato promoteincorporationofideasofdiseasesurveillanceinscience classes.Atthebeginningofeachseason,allparticipantsfrom pre-viousseasonsaresentanemailinvitingthemtoparticipateagain bycompletinganintakequestionnaireforthenewseason.Basedon auniqueuserid,participantscanbetrackedovermultipleseasons. 2.3. Bias

Publichealthstatisticssuchasasthma,diabetesandinfluenza vaccinationratesintheInfluenzanetparticipantshavebeenshown tobe similar for the Dutch(Marquet et al.,2006) and Belgian (Vandendijck et al., 2013) populations. Although younger and olderagegroupsareunderrepresentedinInfluenzanet(Cantarelli etal., 2014),these differencesdidnot seemtohave animpact ontheobservedILItrends(vanNoortetal.,2007).Tominimize

theselectionbiasinrecruitingparticipantswhoalreadyhaveILI, anysymptomsthatstartedbeforeorontheregistrationdateare excludedfromtheanalysis.Onlyparticipantswhoparticipateat least3timesduringaseasonareincludedintheanalyses.

2.4. ILIincidence

ILIisdefinedastheacuteonset(withinafewhours)offever (ameasuredtemperatureof atleast38◦C),togetherwith mus-clepainorheadache,andcoughorsorethroat.Thedayoffever onsetdeterminesthedayofILIonset.Participantsareconsidered activebetweenregistrationdateandthedateoftheirlast com-pleted symptoms’questionnaire.ILIincidence isdeterminedby dividingthenumberofILIonsetsperweekbythenumberofactive participants.IfparticipantsfittheILIcasedefinitioninconsecutive questionnaires,thisisconsideredasasingleILIepisode.

2.5. EuropeanInﬂuenzaSurveillanceNetwork

TheclinicalsurveillanceofinfluenzaintheEuropeanInfluenza Surveillance Network (EISN, formally EISS), coordinated by the ECDC,isgenerallybasedonreportsmadebysentinelGPs.TheILI incidenceforeachcountryisdeterminedbythenumberofpatients whovisittheir(sentinel)GPandfitthe(country-specific)ILIcase definition(Aguileraetal.,2003),dividedbythetotalnumberof peopleassignedtotheparticipatingGPs.

2.6. Crosscorrelation

For each country, thecrosscorrelationbetweenILIincidence ratesasreportedbyECDCandInfluenzanetisdetermined.Since bothtimeseriesareautocorrelatedandshareacommonseasonal trend,thisdirectcrosscorrelationcouldgiveamisleading indica-tionoftheirrelationship(Bloometal.,2007).Therefore,bothtime seriesarealsoprewhitenedbyfittingseasonalARIMAmodelsusing theBox–Jenkinsapproach(Allard,1998),wherethemodelwith thelowestAkaikeinformationcriterionisselected(Hyndmanand Athanasopoulos,2014).Thedetrendedtimeseriesareobtainedby filteringeachtimeseriesbytheselectedmodel,andforeach coun-trythecrosscorrelationbetweenthedetrendedtimeseriesfrom InfluenzanetandECDCisdetermined.

Since the ILI incidence from Influenzanet is based on the reporteddayofonsetandtheILIincidencefromECDCis deter-minedbytheweekapatientvisitedtheirGP,itwouldbeexpected thatthereportedILIincidencefromInfluenzanetprecedestheILI incidenceasdeterminedbyECDC.SinceinInfluenzanetnotonly theweekofonsetbuttheactualdayisrecorded,theweeklyILI incidencefromInfluenzanetcanbeshiftedbysingledays,wherea shiftofzerodaysindicatesthattheILIincidencefrombothsystems iscomparedfortheperiodMonday-Sunday.

2.7. Medicalcareseekingbehaviour

Eachparticipant who reports (ILI)symptoms, isasked some follow-up questions, such as whetherthe participant visited a medicaldoctor.Thisallowsthedeterminationofthepercentage ofparticipantswithILIwhoseekmedicalcare.Sinceparticipants couldseekmedicalcareaftertheyhavereportedtheirsymptoms toInﬂuenzanet,a reportedvisit within15 daysaftera reported ILIonsetisstillconsidered.Sinceseason2011–2012,Inﬂuenzanet participantswhoreportedtohavevisitedamedicaldoctorarealso askedhowmanydayselapsedbetweentheonsetofsymptomsand thevisit.

(3)

2.8. Riskfactoranalysis

Weapplyalogistic regressionmodeltoexplorethe associa-tionbetweenseveralindividualcovariatesandtheoccurrenceofat leastoneILIepisodeduringaseason.Thesecovariatesareselected beforehand,consistingmostlyofcharacteristicswhichhavebeen identiﬁedinotherstudiesregardinginﬂuenzarisk,andsomeextra characteristicswhicharenotnormallyanalyzed.Mostcovariates areconsideredequalacrossallseasons:agegroup(<15,15–49, 50–64,65+),householdsituation(alone,withchildrenorwithonly adults),gender,chronicdisease(asthma,diabetes,heartdisease, and/orimmuno-compromised),smoking,sports(atleast1hper week),petsathome(dogs,cats,and/orbirds),andprimarymodeof dailylocomotion(bicycle/foot,car,orpublictransport).The covari-ate“riskgroup(others)”includesthoseparticipantswhoreportto belongtoariskgroup,butareyoungerthan65years(60inthe Netherlandssince2008)anddidnotreportanyofthechronic dis-eases.Theeffectofvaccinationisconsideredasaseason-dependent covariate.Fortheseason2009–2010,thevaccinestatusisbasedon thepandemicvaccine.Countryofresidenceandseason(indirectly) aretwoextracovariates.

OnlyILIonsetsduringtheweekswheninfluenzastrainswere circulatingin the populationare considered.These periods are definedforeachseasonandcountryastheweekswhenthe num-berofinfluenza-confirmedsamplesasreportedbyECDCwasat least15%ofthemaximumforthatseason(movingaverageover3 weeks)(SupplementaryFigureS13).Allparticipantsareconsidered independentbetweentwodifferentseasons,andparticipantswho werenotactiveforthecompleteinfluenzaperiodareexcluded. SinceInfluenzanetisacohortstudyinwhichhealthyindividuals (withoutILI)arerecruitedandthepossibleonsetofILIis moni-toredoverafixedperiodoftime,Influenzanetcandeterminethe riskratiosforallthecovariates.Foreachcovariateaunivariaterisk ratioisdetermined.Adjustedriskratiosaredeterminedbya mul-tivariatelog-binomialregressionmodel includingall globaland season-dependentcovariatesanalyzedbythegenerallinearmodel inRsoftware(RDevelopmentCoreTeam,2014).Thevariance infla-tionaryfactor(VIF)foreachcovariateisdeterminedtocheckfor collinearityinthemultivariateregressionmodel.

3. Results 3.1. ILIincidence

The ILI incidence as determined by Inﬂuenzanet correlates wellovermultipleseasonswiththeILIincidenceasreportedby ECDC(Fig.1).However,InﬂuenzanetmeasuresILIincidenceinall countriesonthesamescale,whiletheincidencesreportedbyECDC areingenerallowerandvaryinscalebetweencountries.

The crosscorrelation between the raw ILI incidences from InfluenzanetandECDCissignificant(Fig.2A,C,E,andG).TheILI incidencesshowahighlevelofautocorrelationandsomedegree ofseasonality(SupplementaryFiguresS5–S12A).Wefittedtoeach timeseriesaseasonalARIMAmodel(SupplementaryTableS1),and filteredthetimeseriesbyeachmodeltoobtainadetrendedtime series(SupplementaryFiguresS1–S4).Thedetrendedtimeseries arenolongerautocorrelated(SupplementaryFiguresS5–S12B),as confirmedbytheLjung–Boxtest(SupplementaryTableS1).The detrendedtimeseriesofInfluenzanetandECDCalsoshowa signif-icantlevelofcrosscorrelationatalagofzeroweeks(Fig.2B,D,F, andH),fortheNetherlands(0.38),Belgium(0.53),andItaly(0.29). 3.2. Medicalcareseekingbehaviour

ThepercentageofparticipantswithILIwhoseekedmedicalcare variesgreatlybycountry(Fig.3).Similardifferencesareobserved

Table1

Crosscorrelationbetween Inﬂuenzanetand ECDCfor daily lags basedonthe detrendedtimeseries(2003–2013).Valuesinboldindicatethecrosscorrelationat alagequaltothemediannumberofdaysbetweenILIonsetandvisittoamedical doctor.

Lag(days) Netherlands Belgium Portugal Italy

−7 0.33 0.34 0.27 −0.06 −6 0.40 0.42 0.23 −0.02 −5 0.40 0.45 0.32 0.09 −4 0.47 0.52 0.32 0.19 −3 0.46 0.50 0.35 0.24 −2 0.39 0.57 0.35 0.22 −1 0.32 0.57 0.42 0.24 0 0.38 0.53 0.38 0.29 1 0.34 0.46 0.22 0.24 2 0.35 0.44 0.22 0.20 3 0.34 0.37 0.24 0.18 4 0.31 0.40 0.13 0.16 5 0.35 0.32 0.17 0.17 6 0.34 0.24 0.06 0.16 7 0.19 0.24 0.15 0.18

inthenumberofdaysbetweentheonsetofsymptomsandvisiting

thedoctor (Fig.4).Theobservedpatternsdo notchangeifonly

workingadultparticipantsareconsideredintheanalysis. Thecrosscorrelationbetweenthedetrendedtimeseriesfrom InﬂuenzanetandECDCismaximumwhenashiftof4daysisapplied intheNetherlands,1–2daysinBelgium,1dayinPortugal,andno shiftinItaly(Table1).Thiscorrespondswellwiththemediandelay betweenILIonsetandseekingmedicalcareasreportedduringthe seasons2011–2013:4daysintheNetherlands,2daysinBelgium, 1dayinPortugal,and1dayinItaly.

3.3. Participation

TheNetherlandshasmostparticipants(onaverage19,491per season,ofwhich16,481completedatleast3symptoms’ question-naires),followedbyBelgium(6001;5072),Portugal(2871;1894), andItaly(1882;1219)(Fig.5).Thiscorrespondsto0.1%ofthe pop-ulationintheNetherlandsandBelgium(Flanders,sincebasically allparticipantsarefromFlanders),0.02%inPortugal,and0.003% inItaly. Ofallparticipantswho completedatleast3symptoms questionnairesduringaseason,intheNetherlands76±8% partici-patedagaininthefollowingseason,74±12%inBelgium,69±12% inPortugal,and70±4%inItaly.

3.4. Riskfactors

The univariate risk ratios are listed in the Supplementary material (Table S2), whereas the adjusted risk ratiosfrom the multivariateregressionarelistedinTable2.Thevariance infla-tionaryfactor(VIF)forthecovariatesvariesbetween1.0and2.7 (SupplementaryTableS2),reassuringthatmodelspecificationis not compromised by undesirable collinearities (O’Brien, 2007). McFadden’spseudoR-squaredforthefinalmodelfit(R2₌_0.035)

isrelativelylow,indicatingthatonlyasmallpartofthevariation inILIinfectionsisexplainedbythepresentedcovariates.The pri-maryriskfactorforacquiringaninfectionishavingcontactwithan infectiouspersonandthisisabsentfromtheseanalyses.

Accordingtotheadjustedriskratios,havingachronicdisease (asthma, diabetes, heart disease and/or immune-compromising condition), living with children, being female, belonging to a youngeragegroup,petsathome(catsand/ordogs),andbeinga smoker,wereallindependentpredictorsoftheriskofhavingat leastoneILIepisode duringa ﬂuseason(Table2).A smallrisk reductionwasobservedinparticipantswhoprimarilyusebicycle orfootforlocomotion(comparedtoacar)andparticipantswho practicemorethan1hofsportsperweek.Nosigniﬁcanteffectwas

(4)

Fig.1. ILIincidenceforInfluenzanetandECDCfor(A)theNetherlands,(B)Belgium,(C)Portugaland(D)Italy.TheILIincidencebyInfluenzanetisdefinedasthenumberof ILIonsetsper100,000participant-weeks.TheverticalaxesfortheILIincidencearescaledbasedonalinearregressionbetweentheILIincidencesofInfluenzanetandECDC.

observedforparticipantswholivewithotheradults(comparedto livingalone),participantswhohavebirdsathome,andparticipants whousepublictransportation(comparedtousingacar).

Thevaccineeffectivenessforinfluenza-likeillnessvariesfrom seasontoseason.AsignificantreductioninILIduetovaccination wasobservedintheseasons2007–2008,2008–2009,2010–2011, and2012–2013,whilenosignificanteffectwasobservedinother seasons.Thevaccineeffectivenessforseason2009–2010ispossibly underestimated,sincethevaccineonlybecameavailablewhenthe ILIactivitywasalreadyepidemic.

4. Discussion

Basedonsingle-seasonanalysis,previousstudiesestablished excellent correlations betweenILI incidences as determinedby InﬂuenzanetandbyECDC(Friesemaetal.,2009;Marquetetal., 2006;vanNoortetal.,2007).Aquestionremainedonwhetherthis consistencywouldpersistformultiple-seasondatastreams.We showedthat during10 seasonsintheNetherlandsandBelgium (2003–2013),8seasonsinPortugal(2005–2013),and5seasonsin Italy(2008–2013),theILItrendsfromInﬂuenzanetandECDCare

consistentinbothtimingandrelativemagnitude,withasignificant crosscorrelationbetweenbothtimeseriesaslagsofzeroweeks.The signalfromInfluenzanetprecedesECDCbyafewdays, correspond-ingapproximatelytothemediannumberofdaysbetweenILIonset andseekingmedicalcare.However,thisdoesnotnecessaryindicate thatinreal-timemonitoringInfluenzanetwoulddetectILItrends earlier,sincethisdependsonwhenthedatabecomesavailableand thestatisticaluncertaintiesinthedata(vanNoort,2014).

Althoughbothtimeseriesarecorrelatedoverthefull10-year period,therearelocalizeddiscrepanciesbetweenthedatastreams, whichcouldbeattributedtothedifferentmethodologyand com-positionof thecohorts in both systems.As anexample, young childrenarelargelyunderrepresentedinInfluenzanet(vanNoort etal.,2007),whereasyoungchildrenvisitrelativelymoreoftena medicaldoctor.Thiscouldexplainwhyfortheseason2007–2008in Portugal,dominatedbytheinfluenzaBstrainaffectingmostly chil-dren,asmallepidemicwasreportedbyECDCwhichwentmostly undetectedbyInfluenzanet.Anotherlocaldiscrepancyisthe rel-ativelyhighILIincidenceasreportedbyECDCintheNetherlands duringthemonthsprecedingthe2009ILIpandemic,whichmight beattributedtoanincreaseinawarenessbymedicaldoctorsand

(5)

Fig.2.Cross-correlationbetweentheILIincidencesofInﬂuenzanetandECDCusingeithertherawILIincidences(leftpanel)orthedetrendedtimeseries(rightpanel)in(A andB)Netherlands,(CandD)Belgium,(EandF)Portugaland(GandH)Italy(2003–2013).

patientsduetoaglobalconcernaboutthenewH1N1inﬂuenza strain(Keramarouetal.,2011).

Thepresenceofmultipleindependentsourcesencouragesthe development of integrative methods that explore the speciﬁc strengthsofeachsystem(Reisetal.,2007).Havingmultiple inde-pendentsystemscoulduncoveraspectsofinﬂuenzatransmission

thatwould gounnoticedifonly onedatastreamwasavailable. Anothercross-countrydatasourceforILIincidencesisGoogleFlu Trends(Ginsbergetal.,2009),whichdeterminesILIincidencebased onthefrequencyofILI-relatedsearchterms.However,GoogleFlu Trendsisnotastrictlyindependentdatasource,sincetheir algo-rithmsrelyontheECDCdatastreamsforcalibration.

Fig.3.InfluenzanetparticipantswithILIwhovisitedamedicaldoctorspecifiedbycountry(2011–2013).ILIisdefinedastheacuteonset(withinafewhours)offever(a measuredtemperatureofatleast38◦_C),_together_with_muscle_pain_or_headache,_and_cough_or_sore_throat.

(6)

Fig.4.TimelagbetweenILIonsetandvisittomedicaldoctorin(A)Sweden,(B)UnitedKingdom,(C)theNetherlands,(D)Belgium,(E)France,(F)Spain,(G)Portugaland (H)Italy(2011–2013).ILIisdeﬁnedastheacuteonset(withinafewhours)offever(ameasuredtemperatureofatleast38◦_C),_together_with_muscle_pain_or_headache,_and

coughorsorethroat.

Patterns in medical care seeking behavioursuggest cultural differencebetween northernand southern Europe. In southern Europe(France,Italy,Portugal,andSpain)participantsgenerally visitamedical doctorwithin1–2daysaftertheonset of symp-toms,whereasinnorthernEurope(Sweden,UnitedKingdom,and theNetherlands)participantsseekmedicalcaregenerallyonly5–7 daysaftertheonsetofsymptoms.Belgium (Flanders)seemsan exceptiontothesuggestedpattern, mostlikely because accord-ingtoBelgianlaw,anemployercanrequirefromtheiremployee amedical statementwithin24htojustifyworkabsenteeism.A similarpatternisobservedinthepercentageofparticipantswith ILIwhoseekmedicalcare,whichislowerinthenorthernEurope (exceptBelgium)thaninsouthernEurope.Thetwopatternscould beassociatedbyconsideringthatincountrieswhereparticipants waitlongerbeforeseekingmedicalcare,manyparticipantswould nolongerfeelsufﬁcientlyilltowarrantavisittoamedicaldoctor. Thisvariationinmedicalcareseekingratesacrosscountriesis oneofthereasonswhyILIincidences reportedbyECDCcannot becompareddirectly(vanNoortetal.,2007).Variationsin med-icalcare seekingcouldalsoaffectthedetermined ILIincidence byECDCwithinacountry,ifcertainsubgroupsofthepopulation visitadoctoratdifferentrates.Inﬂuenzanetdoesnotonlyserve asanindependentsourceforILIactivity,butcouldalsobeusedto calibrateILIdataascollectedbyGPsentinelsystems(vanNoort, 2014).

AcrucialelementinthesuccessofInfluenzanet,ishavinga suffi-cientlylargecohortofparticipants.IntheNetherlands(onaverage 16,481activeparticipants),Belgium(5072),Portugal(1894),and Italy(1219)theInfluenzanetcohortwaslargeenoughtodetect similarILIepidemicsasECDCinallseasons,withtheexception of season2007–2008in Portugal.Larger cohortswould leadto lowerstatisticalnoisesuchthatepidemicscouldbedetectedearlier andevensmallILIepidemicscouldbedistinguishedfrom base-lineILIactivity.Furthermore,inlargercohorts,differentsubgroups, forexamplebasedonageorvaccinestatus,couldbemonitored

separately. Of all active participants during a certain season, 73±11%participatedagaininthefollowingseason.Althoughthis shows impressive loyalty of participants,each seasonan effort shouldbemadetorecruitnewparticipantstoatleastreplacethose whohaveleft.

RiskfactorsestimatedfromtheInfluenzanetcohort are con-sistentwiththeinfluenzaliterature.HigherriskofILIinchildren andinthoselivingwithchildrenwasobserved,inconsistencywith observationalstudies(Cauchemezetal.,2009;Montoand Ross, 1977;Monto,2004;Viboudetal.,2004).TheincreasedILIriskin womencomparedtomen,whichmaybeduetomoreintensive contactbetweenwomenandchildren,hasalsobeenpreviously rec-ognized(MontoandRoss,1977).Wefoundasignificantlyreduced riskofILIamongparticipantsover65.Thisisnotduethehigher vac-cineuptakeinseniors,sincevaccinestatusisalreadyincludedasa separatecovariateinthismultivariateanalysis.Seniorsare gener-allyconsideredariskgroupforinfluenza,notbecauseofahigher probabilityforinfection,butduetotheirgreaterriskfor compli-cations(Monto,2004).Havingachronicdisease,suchasasthma, diabetes, heart diseaseor an immune-compromising condition, wasastrongpredictorofILIintheInfluenzanetcohort.Peoplewith thesechronicdiseasesaregenerallyadvisedtotakeaninfluenza vaccine.Increasedriskofinfluenzahasbeenobservedinchildren withasthmainclinicalcohortstudies(Gordonetal.,2009),while diabetesisknowntobestronglyassociatedwithcomplicationsdue toinfluenzainfections(Irwinetal.,2001).AnincreasedriskofILI wasobservedamongtheInfluenzanetparticipantswhosmoke,as hasbeenconfirmedbyotherstudies(ArcaviandBenowitz,2004). TheInfluenzanetsystemisflexibletotheextentthatquestions ofinterestcaneasilybeaddedorremoved,allowingforthe esti-mationof riskfactorswhich arenot usuallyconsidered.In this study,wefoundasmallbutsignificantprotectiveeffectof walk-ingorbicyclingasaprimarymeansoflocomotionincomparison withtravellingbycar,whilenosignificantriskoftravellingby pub-lictransportationwasobserved,norinparticipantswholivewith

(7)

Fig.5. NumberofInﬂuenzanetparticipantswhocompletedatleast3symptoms’questionnairesin(A)theNetherlands,(B)Belgium,(C)Portugal,and(D)Italy.Theparticipants aredifferentiatedbytheseasoninwhichtheyregistered.InPortugalinseason2006–2007,inItalyinseason2010–2011,andinallcountriesinseason2011–2012anew databasestructurewasintroduced,providingaresetofparticipantids.

otheradultsincomparison withadultswholivealone. Asmall increase inrisk wasobserved in participantswhohave pets at home.Practicingsportsforatleastonehourperweekwas asso-ciatedwithasmallbutsigniﬁcantdecreaseontheILIrisk.

Notonlyextraquestionscouldbeincludedintheintake ques-tionnaire,entirenewquestionnairescouldbeaddedinparticular seasonsenabling further studies.A stress-related questionnaire releasedin theNetherlands in season2004–2005revealed sig-niﬁcanttrendsbetweenstress/personalityand ILIself-reporting (Smolderen etal., 2007), and a simple questionnairerelated to contactbehaviour,showedthatchangesincontactpatternscould explainchangesindiseaseincidence(Eamesetal.,2012b).

In4outof10seasonsInfluenzanetestimatedasignificant reduc-tionin ILIdue tovaccination,whereas intheother seasonsno significanteffectwasobserved.Thedirecteffectivenessof vacci-nationvariedbetween 33%(22–42%) inseason2010–2011 and −10% (−28 to6%) in season 2004–2005. A relatively low vac-cineeffectivenessagainstILIistobeexpected,sincevaccination targetsspecificallytheinfluenzavirus, andnot other influenza-likeillnesses.Adouble-blind,randomized,placebo-controlledtrial

measuredwithinthesamecohortavaccineefficacyfor serologi-callyconfirmedinfluenzaofrespectively50%(1997/98)and86% (1998/99),butavaccineeffectivenessforILIof−10%(1997–1998) and33%(1998–1999)(Bridgesetal.,2000).Accordingtoalarge meta-studybasedon48reportsonvaccineeffectivenessinhealthy adults,inactivatedparenteralvaccineswere30%effectiveagainst ILI,and80%efficaciousagainstinfluenzawhenthevaccinematched thecirculatingstrainandcirculationwashigh,butthisdecreased toaneffectivenessagainstILIof12%andefficacyagainstinfluenza of50%whenitdidnot(Demichelietal.,2009).

Fortwoseasons(2003–2004and2004–2005)Influenzanet esti-matedanegativealthoughnon-significantvaccineeffectiveness. Bothseasonswerecharacterizedbyapoorvaccinematch(Belongia etal.,2009;Jinetal.,2005).Anegativevaccineeffectcanbedue tooriginalantigenicsin,thetendencyforantibodiesproducedin response toexposuretoinfluenzavaccineantigensto suppress thematurationofantibodieswithhighaffinitytotheactualvirus (Guptaetal.,2006).

Inanobservationalstudyofvaccineeffectiveness,any preexis-tingbiasbetweenvaccinatedandunvaccinatedparticipantscould

(8)

Table2

AdjustedriskfactorsandvaccineeffectivenessforILIintheNetherlands,Belgium, Portugal,andItaly(2003–2013).

Question Answer RR(adjusted)

Age <18 1.59(1.46–1.74)

18–49 *

50–64 0.82(0.78–0.86)

65+ 0.46(0.41–0.51)

Household Alone *

Onlywithadults 0.99(0.92–1.05)

Withchildren 1.31(1.22–1.40)

Gender Male *

Female 1.22(1.17–1.28)

Chronic diseases

Asthmaorlungdisease 1.58(1.47–1.69)

Diabetes 1.27(1.15–1.41)

Heartdisease 1.29(1.13–1.47)

Kidneydisorder 1.23(0.80–1.90)

Immune-compromised 1.23(1.02–1.49)

Riskgroup(other)a _1.23_{(1.06–1.41)}

Smoking 1.16(1.10–1.22)

Pets Dogs 1.15(1.09–1.22)

Cats 1.07(1.02–1.12) Birds 1.03(0.94–1.13) Sports >1hperweek 0.95(0.90–1.00) Primarymeansof locomotion Bicycle/foot 0.95(0.90–1.00) Car * Publictransport 0.97(0.89–1.05) Vaccination 2003/04 1.07(0.78–1.47) 2004/05 1.10(0.94–1.28) 2005/06 0.97(0.83–1.12) 2006/07 1.00(0.86–1.15) 2007/08 0.81(0.70–0.94) 2008/09 0.80(0.71–0.90) 2009/10b _0.87_{(0.74–1.03)} 2010/11 0.67(0.58–0.78) 2011/12 0.97(0.82–1.16) 2012/13 0.80(0.71–0.91) Dependentvariable:havingatleastoneILIepisodeduringaseason(inﬂuenza period).

a_Classiﬁed_by_GP_as_belonging_to_risk_group_due_to_factors_not_speciﬁed_on_this

table.

b_Vaccination_for_the_new_H1N1pdm_strain.

distorttheresults.Aunivariateanalysis(SupplementaryTableS2) indicatesa10%higherILIreductioninvaccinatedparticipantsthan themultivariatestudy.However,withparticipantsover65yearsof agehavingalowerILIrateandarelativelyhighvaccinationrate,the multivariatemodelestimatesthatapartofthereductioninILIin vaccinatedparticipantsisduetotheirage.Althoughthe multivari-atelogisticregressionaimstocorrectforthesebiases,itispossible thatotherbiasesnotrepresentedbyanyoftheriskfactorslistedin Table2exist.

Acohortstudyof72,527seniorsover65yearsofagefollowed duringan8yearperiod,foundthatvaccinatedseniorsalreadyhada reducedriskofdeathandpneumoniahospitalizationintheperiods beforetheinfluenzaseason,andthattheriskreductionactually decreasedduringtheinfluenzaseason(Jacksonetal.,2006).Such apreferentialreceiptofvaccinebyrelativelyhealthyseniorscould leadtooverestimationofthevaccineeffectivenessinobservational studies.ItisplausiblethatmostelderlyInfluenzanetparticipants arerelativelyhealthyand thatthisselectionbiasislesspresent inInfluenzanet,leading torelativelylower estimatesofvaccine effectivenessthanintheaverageliterature.Becauseofglobal rec-ommendationsforinfluenzavaccination,placebo-controlledtrials, whichcouldclarifytheeffectsofinfluenzavaccinesinindividuals, arenolongerconsidered possibleonethicalgrounds(Jefferson etal.,2010).

SincetheInfluenzanetparticipantsarenotarandomsample oftheoverall population,careshouldbetakeninextrapolating the estimated risks tothe overall population in the respective countries.However,theobservedconsistencyinriskfactorsforILI betweenInfluenzanetandthosereportedbystudiesincommunity settingsfurtherestablishesthattheInfluenzanetpopulationisa valuablesentinelforILIsurveillanceinthepopulation,inaddition tothemeritsofengagingtheparticipantsinpublichealthresearch andpromotingriskawareness.

Thesystempresentedherestandsonaconceptforsyndromic surveillance that depends on intense activity in science com-munication, public awareness and sufficient levels of Internet penetration. It hasreportedILI activityin a consistent wayfor over 10 seasons in multiple countries. Influenzanet reports ILI trendsconsistentwithGPsentinelsurveillance(ECDC),andcan complement these systems by providing valuable information aboutmedicalcareseekingbehaviour.Basedonreported symp-toms,Influenzanetcanbeextendedtodetectdiseasesotherthan influenza,includingthoseindevelopingsettings.Influenzanetasan Internetmonitoringsystembasedonvoluntaryparticipantsmight thereforedevelopintoanimportantweapontofightinfluenzaas wellasothercontagiousdiseasesglobally.

Acknowledgements

WethankthenationalInfluenzanetteamsfromtheNetherlands and Belgium (Grote Griepmeting), Portugal (Gripenet), Italy (Influweb),UnitedKingdom(Flusurvey),Sweden(Halsörapport), France(Grippenet),andSpain(Gripenet)forprovidingdataforthis study.ThisresearchwasfundedbytheEuropeanCommission (EC-ICT-231807).MVRacknowledgesthesupportoftheInstituteforthe PromotionofInnovationbyScienceandTechnology(IWT)in Flan-ders(strategicbasicresearchprojectSIMID).MGMGthanksICyT projects60/2010and343/2010.CTCwassupportedbya fellow-shipfromtheBrazilianResearchCouncil(CNPq)duringthisstudy. SPNwassupportedbyCAPESinthefinalstagesofthiswork. AppendixA. Supplementarydata

Supplementarydataassociatedwiththisarticlecanbefound,in theonlineversion,atdoi:10.1016/j.epidem.2015.05.001

References

Aguilera,J.F.,Paget,W.J.,Mosnier,A.,Heijnen,M.L.,Uphoff,H.,etal.,2003. Hetero-geneouscasedeﬁnitionsusedforthesurveillanceofinﬂuenzainEurope.Eur.J. Epidemiol.18(8),751–754.

Allard,R.,1998.Useoftime-seriesanalysisininfectiousdiseasesurveillance.Bull. WorldHealthOrgan.76(4),327–333.

Arcavi,L.,Benowitz,N.L.,2004.Cigarettesmokingandinfection.Arch.Intern.Med. 164(20),2206–2216.

Bajardi,P.,Paolotti,D.,Vespignani,A.,Eames,K.,Funk,S.,etal.,2014.Association betweenrecruitmentmethodsandattritioninInternet-basedstudies.PLoSONE 9(12),e114925.

Belongia,E.A.,Kieke,B.A.,Donahue,J.G.,Greenlee,R.T.,Balish,A.,etal.,2009. Effec-tivenessofinactivatedinﬂuenzavaccinesvariedsubstantiallywithantigenic matchfromthe2004–2005seasontothe2006–2007season.J.Infect.Dis.199 (2),159–167.

Bloom,R.M.,Buckeridge,D.L.,Cheng,K.E.,2007.Findingleadingindicatorsfor dis-easeoutbreaks:ﬁltering,cross-correlation,andcaveats.J.Am.Med.Inform. Assoc.14(1),76–85.

Bridges,C.B.,Thompson,W.W.,Meltzer,M.I.,Reeve,G.R.,Talamonti,W.J.,etal., 2000.Effectivenessandcost-beneﬁtofinﬂuenzavaccinationofhealthy work-ingadults:arandomizedcontrolledtrial.J.Am.Med.Inform.Assoc.284(13), 1655–1663.

Butler,D.,2006.Diseasesurveillanceneedsarevolution.Nature440(7080),6–7.

Cantarelli,P.,Debin,M.,Turbelin,C.,Poletto,C.,Blanchon,T.,etal.,2014.The rep-resentativenessofaEuropeanmulti-centernetworkforinﬂuenza-like-illness participatorysurveillance.BMCPublicHealth14,984.

Cauchemez,S.,Donnelly,C.A.,Reed,C.,Ghani,A.C.,Fraser,C.,etal.,2009.Household transmissionof2009pandemicinﬂuenzaA(H1N1)virusintheUnitedStates. N.Engl.J.Med.361(27),2619–2627.

(9)

Chretien,J.P.,Burkom,H.S.,Sedyaningsih,E.R.,Larasati,R.P.,Lescano,A.G.,etal., 2008.Syndromicsurveillance:adaptinginnovationstodevelopingsettings.PLoS Med.5(3),e72.

Cooper,D.L.,Smith,G.E.,Regan,M.,Large,S.,Groenewegen,P.P.,2008.Trackingthe spatialdiffusionofinﬂuenzaandnorovirususingtelehealthdata:a spatiotem-poralanalysisofsyndromicdata.BMCMed.6,16.

Debin, M.,Colizza, V., Blanchon, T.,Hanslik, T., Turbelin,C., Falchi, A., 2014.

Effectivenessof2012–2013inﬂuenzavaccineagainstinﬂuenza-likeillnessin generalpopulation:estimationinaFrenchweb-basedcohort.Hum.Vaccine Immunother.10(3),536–543.

Demicheli,V.,Jefferson,T.,Al-Ansary,L.A.,Ferroni,E.,Rivetti,A.,etal.,2009.Vaccines forpreventinginﬂuenzainhealthyadults.CochraneDatabaseSyst.Rev.3.

Eames,K.T.,Brooks-Pollock,E.,Paolotti,D.,Perosa,M.,Gioannini,C.,Edmunds, W.J.,2012a.Rapidassessmentofinﬂuenzavaccineeffectiveness:analysisofan internet-basedcohort.Epidemiol.Infect.140(7),1309–1315.

Eames,K.T.D.,Tilston,N.L.,Brooks-Pollock,E.,Edmunds,W.J.,2012b.Measured dynamicsocialcontactpatternsexplainthespreadofH1N1vinﬂuenza.PLoS Comput.Biol.8(3),e1002425.

Friesema,I.H.M.,Koppeschaar,C.E.,Donker,G.A.,Dijkstra,F.,vanNoort,S.P.,etal., 2009.Internet-basedmonitoringofinfluenza-likeillnessinthegeneral popula-tion:experienceoffiveinfluenzaseasonsintheNetherlands.Vaccine27(45), 6353–6357.

Ginsberg,J.,Mohebbi,M.H.,Patel,R.S.,Brammer,L.,Smolinski,M.S.,etal.,2009.

Detectinginﬂuenzaepidemicsusingsearchenginequerydata.Nature 457 (7232),1012–1014.

Gordon,A.,Ortega,O.,Kuan,G.,Reingold,A.,Saborio,S.,etal.,2009.Prevalenceand seasonalityofinﬂuenza-likeillnessinchildren,Nicaragua,2005–2007.Emerg. Infect.Dis.15(3),408–414.

Gupta,V.,Earl,D.J.,Deem,M.W.,2006.Quantifyinginﬂuenzavaccineefﬁcacyand antigenicdistance.Vaccine24(18),3881–3888.

Henning,K.J.,2004.Whatissyndromicsurveillance?MMWRMorb.Mortal.Wkly. Rep.53(Suppl.),5–11.

Hyndman,R.,Athanasopoulos,G.,2014.Forecasting:PrinciplesandPractice,http:// otexts.org/fpp/(accessedApril2014).

Irwin,D.E.,Weatherby,L.B.,Huang,W.Y.,Rosenberg,D.M.,Cook,S.F.,etal.,2001.

Impactofpatientcharacteristicsontheriskofinﬂuenza/ILI-related complica-tions.BMCHealthServ.Res.1(1),8.

Jackson,L.A.,Jackson,M.L.,Nelson,J.C.,Neuzil,K.M.,Weiss,N.S.,2006.Evidenceof biasinestimatesofinﬂuenzavaccineeffectivenessinseniors.Int.J.Epidemiol. 35(2),337–344.

Jefferson,T.,DiPietrantonj,C.,Al-Ansary,L.A.,Ferroni,E.,Thorning,S.,etal.,2010.

Vaccinesforpreventinginﬂuenzaintheelderly.CochraneDatabaseSyst.Rev. (2),CD004876.

Jin,H.,Zhou,H.,Liu,H.,Chan,W.,Adhikary,L.,Mahmood,K.,Lee,M.S.,Kemble, G.,2005.TworesiduesinthehemagglutininofA/Fujian/411/02-likeinﬂuenza virusesareresponsibleforantigenicdriftfromA/Panama/2007/99.Virology336 (1),113–119.

Keramarou,M.,Cottrell,S.,Evans,M.R.,Moore,C.,Stiff,R.E.,etal.,2011.Twowaves ofpandemicinﬂuenzaA(H1N1)2009inWales—thepossibleimpactofmedia coverageonconsultationrates,April–December2009.EuroSurveill.16(3).

Lazarus,R.,Kleinman,K.P.,Dashevsky,I.,DeMaria,A.,Platt,R.,2001.Using auto-matedmedicalrecordsforrapididentiﬁcationofillnesssyndromes(syndromic surveillance):theexampleoflowerrespiratoryinfection.BMCPublicHealth1, 9.

Lipsitch,M.,Hayden,F.G.,Cowling,B.J.,Leung,G.M.,2009.Howtomaintain surveil-lancefornovelinﬂuenzaaH1N1whentherearetoomanycasestocount.Lancet 374(9696),1209–1211.

Mandl,K.D.,Overhage,J.M.,Wagner,M.M.,Lober,W.B.,Sebastiani,P.,etal.,2004.

Implementingsyndromicsurveillance:apracticalguideinformedbytheearly experience.J.Am.Med.Inform.Assoc.11(2),141–150.

Marquet,R.L.,Bartelds,A.I.M.,vanNoort,S.P.,Koppeschaar,C.E.,Paget,J.,etal.,2006.

Internet-basedmonitoringofinﬂuenza-likeillness(ILI)inthegeneral popula-tionoftheNetherlandsduringthe2003–2004inﬂuenzaseason.BMCPublic Health6,242.

Merk,H.,Kuhlmann-Berenzon,S.,Bexelius,C.,Sandin,S.,Litton,J.E.,etal.,2013.The validityofself-initiated,event-driveninfectiousdiseasereportingingeneral populationcohorts.PLOSONE8(4),e61644.

Monto,A.S.,2004.Occurrenceofrespiratoryvirus:time,placeandperson.Pediatr. Infect.Dis.J.23(1Suppl.),S58–S64.

Monto,A.S.,Ross,H.,1977.Acuterespiratoryillnessinthecommunity:effectof familycomposition,smoking,andchronicsymptoms.Br.J.Prev.Soc.Med.31 (2),101–108.

O’Brien,R.M.,2007.Acautionregardingrulesofthumbforvarianceinﬂationfactors. Qual.Quant.41(5),673–690,ISSN0033-5177.

Paolotti,D.,Carnahan,A.,Colizza,V.,Eames,K.,Edmunds,J.,etal.,2014.Web-based participatorysurveillanceofinfectiousdiseases:theInﬂuenzanetparticipatory surveillanceexperience.Clin.Microbiol.Infect.20(1),17–21.

Patterson-Lomba,O.,vanNoort,S.P.,Cowling,B.J.,Wallinga,J.,Gomes,M.G.M.,etal., 2014.Utilizingsyndromicsurveillancedataforestimatinglevelsofinﬂuenza circulation.Am.J.Epidemiol.179(11),1394–1401.

RDevelopmentCoreTeam,2014.R:ALanguageandEnvironmentforStatistical Computing.RFoundationforStatisticalComputing,Vienna,Austria.

Rehn,M.,Carnahan,A.,Merk,H.,ouml,K.,hlmann-Berenzon,S.,Galanis,I.,etal., 2014.EvaluationofanInternet-basedmonitoringsystemforinﬂuenza-like ill-nessinSweden.PLoSONE9(5),e96740.

Reis,B.Y.,Kohane,I.S.,Mandl,K.D.,2007.Anepidemiologicalnetworkmodelfor diseaseoutbreakdetection.PLoSMed4(6),e210.

Smolderen,K.G.E.,Vingerhoets,A.J.J.M.,Croon,M.A.,Denollet,J.,2007.Personality, psychologicalstress,andself-reportedinﬂuenzasymptomatology.BMCPublic Health7,339.

vanNoort,S.P.,Muehlen,M.,deAndrade,H.R.,Koppeschaar,C.,Lourenc¸o,J.M.L., etal.,2007.Gripenet:aninternet-basedsystemtomonitorinﬂuenza-likeillness uniformlyacrossEurope.EuroSurveill.12(7),E5–E6.

vanNoort,S.P.,Águas,R.,Ballesteros,S.,Gomes,M.G.M.,2012.Theroleofweather ontherelationbetweeninﬂuenzaandinﬂuenza-likeillness.J.Theor.Biol.298, 131–137.

vanNoort,S.P.,(PhDthesis)2014.ParticipatorySurveillanceandMathematical Mod-elsinEpidemiologicResearch:SuccessesandChallenges.UniversidadeNovade Lisboa.

Vandendijck,Y.,Faes,C.,Hens,N.,2013.Eightyearsofthegreatinﬂuenzasurveyto monitorinﬂuenza-likeillnessinFlanders.PLOSONE8(5),e64156.

Viboud,C.,Boëlle,P.Y.,Cauchemez,S.,Lavenu,A.,Valleron,A.J.,etal.,2004.Risk factorsofinﬂuenzatransmissioninhouseholds.Br.J.Gen.Pract.54(506), 684–689.

vandenWijngaard,C.,vanAsten,L.,vanPelt,W.,Nagelkerke,N.J.D.,Verheij,R.,etal., 2008.Validationofsyndromicsurveillanceforrespiratorypathogenactivity. Emerg.Infect.Dis.14(6),917–925.