Cardiovascular Risk of Remote Indigenous Australians

(1)

Development and Use of Prediction Models for Classification of

Cardiovascular Risk of Remote Indigenous Australians

An Tran-Duy, PhD

^a*

, Robyn McDermott, PhD

^b,c

, Josh Knight, PhD

^a

, Xinyang Hua, PhD

^a

, Elizabeth L.M. Barr, PhD

^d,e

, Kerry Arabena, PhD

^f

, Andrew Palmer, MD, PhD

^a,g

, Philip M. Clarke, PhD

^a,h

aCentreforHealthPolicy,MelbourneSchoolofPopulationandGlobalHealth,UniversityofMelbourne,Melbourne,Vic,Australia

bCentreforChronicDiseasePrevention,JamesCookUniversity,Cairns,Qld,Australia

cSchoolofHealthSciences,UniversityofSouthAustralia,Adelaide,SA,Australia

dMenziesSchoolofHealthResearch,Darwin,NT,Australia

eBakerHeartandDiabetesInstitute,Melbourne,Vic,Australia

fCentreforHealthEquity,MelbourneSchoolofPopulationandGlobalHealth,UniversityofMelbourne,Melbourne,Vic,Australia

gMenziesInstituteforMedicalResearch,UniversityofTasmania,Hobart,Tas,Australia

hHealthEconomicsResearchCentre,NuffieldDepartmentofPopulationHealth,UniversityofOxford,Oxford,UK

Received2September2018;receivedinrevisedform5December2018;accepted5February2019;onlinepublished-ahead-of-print22February2019

Background Cardiovasculardisease(CVD)istheleadingcauseofdeathforIndigenousAustralians.Thereiswidespread beliefthatcurrenttoolshavedeficienciesforassessingCVDriskinthishigh-riskpopulation.Wesoughtto developa5-yearCVDriskscoreusingawiderangeofknownriskfactorstofurtherimproveCVDrisk predictioninthispopulation.

Methods WeusedclinicalanddemographicinformationonIndigenouspeopleagedbetween30and74yearswithout ahistoryofCVDeventswhoparticipatedintheWellPerson’sHealthCheck(WPHC),acommunity-based survey.Baselineassessmentswereconductedbetween1998and2000,anddatawerelinkedtoadminis- trativehospitalisationanddeathrecordsforidentificationofCVDevents.WeusedCoxproportionalhazard modelstoestimatethe5-yearCVDrisk,andtheHarrell’sc-statisticandthemodifiedHosmer-Lemeshow (mH-L)

x

²^statistic^to^assess^the^modeldiscriminationandcalibration,respectively.

Results Thestudysampleconsistedof1,583individuals(48.1%male;meanage45.0year).Theriskscoreconsistedof sex,age,systolicbloodpressure,diabetesmellitus,waistcircumference,triglycerides,andalbumincreati- nineratio.Thebias-correctedc-statisticwas0.72andthebias-correctedmH-L

x

²^statistic^was^12.01^(p-value,

0.212),indicatinggooddiscriminationandcalibration,respectively.Usingourriskscore,theCVDriskofthe IndigenousAustralianscouldbestratifiedtoagreaterdegreecomparedtoarecalibratedFraminghamrisk score.

Conclusions Aseven-factorriskscorecouldsatisfactorilystratify5-yearriskofCVDinanIndigenousAustraliancohort.

ThesefindingsinformfutureresearchtargetingCVDriskinIndigenousAustralians.

Keywords Aboriginal TorresStraitIslanders Predictionmodel Riskscore Cardiovasculardisease Coronaryheartdisease

PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.0/).

*Correspondingauthorat:CentreforHealthPolicy,SchoolofPopulationandGlobalHealthTheUniversityofMelbourne,Level4,207BouverieStreet,Carlton, Victoria3053,Australia.Tel.:+61390356387;Fax:+61393481174.,Email:[email protected]

https://doi.org/10.1016/j.hlc.2019.02.005

(2)

Introduction

Earlypreventionofcardiovasculardisease (CVD)isrecog- nised tobe crucial forindigenous populations given their substantiallyhigherCVDriskcomparedtonon-indigenous people[1].TheexcessCVDriskamongIndigenousAustral- iansislargelyduetotheveryhighprevalenceoftraditional CVDriskfactorssuchashypertensionanddiabetes.Inaddi- tion,otherhighlyprevalentriskfactorssuchasalbuminuria havealsobeenshowntobeassociatedwithCVDinIndige- nousAustralians[2,3].

InCVDprimaryprevention,absoluteCVDriskassessment isanimportantcomponentthathasbeenwidelyused and recommended[4,5].Thishelpstoensurethathigher-riskindi- vidualsare appropriatelytargetedfor initiationofprimary preventionstrategies[6].WhiletheannualCVDriskassess- mentusingavalidatedabsoluteriskalgorithmhasbeenrec- ommended forIndigenousAustralians[7],thereis widespread beliefthatcurrenttoolshavedeficienciesinassessingCVDrisk inthishighriskpopulation[6,8,9].Severalstudieshaveshown thattheequationsderivedfromtheFraminghamstudyunder- estimatetheCVDriskofIndigenousAustralians[10,11].One explanationforthisunderestimationmaybethatcertainCVD riskfactors forIndigenouspeoplesuchasalbuminuriaand waistcircumferencearenotincludedintheFraminghamCVD equations [2,3,12,13]. While it ispossible torecalibrate the Framinghammodelstocorrectforthisunderestimation[11], thisdoesnotaccountforabroaderrangeofriskfactorswhich arepotentiallyusefulinfurtherstratifyingrisk.

Ontheotherhand,therearealsocircumstancesinwhichitis usefultostratifyCVDriskusingarestrictedsetofavailablerisk factors [14].Forexample,sometimesIndigenousparticipants wereexcludedfromlaboratorytestsaccordingtodifferentrules [15].Wheninformationontraditionalriskfactorsisnotavail- able,astandardapproachforriskstratificationistoproducea modelorriskchartbasedonareducedsetoffactors[16].

Inthepresent study, weaimedtodevelopand validate predictionmodelsfor5-yearriskofCVDusingdatafroman IndigenousAustraliancohort.Wedevelopedtwomodels:a primary model that included any relevant factors and a reducedinformationmodelthatdoesnotcontainlaboratory variables.Afterinternallyvalidatingthesemodels,weused themtoclassifythestudycohortasgroupswithhigh,mod- erateandlowCVDrisksalongsideclassificationbasedonthe AustralianguidelinesforCVDprevention[4]andtherecalibrated 2008Framingham equation[11],and comparedthe averagerisksineachpredictedcategorywiththeobserved risks.Then,weexaminedtheextenttowhichtheAustralian guidelinesmaymisclassifyIndigenousAustralianswithand withouthigh5-yearCVDrisk.

Methods

Study Cohort

Thecohortusedinthepresentstudywasobtainedfromthe WellPerson’sHealthCheck(WPHC)study[17]andhasbeen

describedindetailinthestudybyHuaetal.[11].Briefly,the WPHCwasconductedbetween1998and2000andincluded 3,508peoplein26ruralandremoteIndigenouscommunities inNorthQueensland.Forouranalysis,weincludedIndige- nousmenandwomenagedbetween30 and74yearswho could be linked to hospitalisation and death records for identificationof theCVDevents occurring after thestudy entry.WeexcludedpeoplewithahistoryofCVDevents.In contrasttothestudybyHuaetal.[11],wedidnotexclude peoplewithmissingriskfactorsatbaseline.Instead,weused imputationstoreplacethemissingvalueswiththeestimates.

Asaresult,thestudycohortconsistedof1,583individuals (seesupplementarydocument-FigureS1foraflowchartof thecohortselection).

Outcome Events and Follow-Up Time

WeusedthedefinitionofCVDfromtheFraminghamHeart Studyforouroutcomeevents,whichincludedacompositeof coronary heartdisease (coronaryinsufficiency, myocardial infarction, and angina pectoris), cerebrovascular events (ischaemicstroke,haemorrhagicstrokeandtransientischae- mic attack), congestive heart failure, peripheral vascular disease,andcoronarydeaths[18].Theseeventswereidenti- fied using theInternational Classificationof Diseases and procedure codes(versions ICD-9-CM and ICD-10-AM; for details,seeHuaetal.)[11].Thefollow-uptimewascalculated asthedurationbetweenbaselinemeasurementandhospital- isationofaCVDevent,coronarydeath,non-coronarydeath orlastdateofdatacollection(1December2014),whichever camefirst.TomaintainalignmentwiththeCVDriskcalcu- lators recommended bytheAustralian guidelines for pre- ventionofCVD[4],a5-yearfollow-upperiodwasgenerated forallindividuals.This wasdonebyconsidering anindi- vidualasrightcensorediftherewasnorecordedCVDevent withinthe5-yearfollow-upperiod.

Statistical Analysis

MissingDataandMultipleImputation

Thepercentage ofobservations with missing data ranged from0% to15.8%dependingonvariables recordedin the dataset.Weusedmultipleimputation[19]toreplacemissing data with appropriate estimates (see detailed methods in supplementarydocument-section2).

ModelSelection

WeusedCoxproportionalhazardmodelstoestimatethe5- yearCVDrisk.Twomodelsweredeveloped:(i)theprimary modelwithvariablesselectedfromriskfactorsincludedin theFraminghamCVDmodels[18] as wellas otherfactors suchas waistcircumference[12]andalbumin-to-creatinine ratio(ACR)[2,3,20]whichhavebeenshowntobeassociated with CVD risk in Indigenous populations; and (ii) the reducedinformationmodelthatexcludedriskfactorsrequir- inglaboratorymeasurements.Initialcovariatesconsideredin theprimarymodelincludedsex,age,totalcholesterol,high- densitylipoprotein(HDL)cholesterol,low-densitylipoprotein (LDL) cholesterol, total cholesterol to HDL ratio,

(3)

triglycerides,systolic bloodpressure (SBP),diastolicblood pressure (DBP), current smoking, diabetes status, fasting bloodglucose,bodymassindex(BMI),waistcircumference andACR(seesupplementarydocument-section3formeth- odsofriskfactormeasurements).Ofnote,thediabetescon- ditionwasidentifiedbasedonthepatientself-reportsand confirmed by the primary care medical records. In the reducedinformationmodel,onlysex,age,SBP,DBP,current smoking,diabetesstatus,BMIandwaistcircumferencewere considered.Toselecttheoptimalsetofpredictors,weper- formedbothstepwiseandmanualprocedures,andincluded variableswithap-valuesmallerthan0.15inthefinalmodels (seesupplementarydocument-section4fordetailedmeth- ods).WeusedthescaledSchoenfeldresiduals[21]totestthe proportionalhazardassumption.

RiskEquationDevelopmentandPerformance Assessment

Theequationfor5-yearCVDriskbasedontheCoxpropor- tionalhazardmodeltakesthefollowingform[22]:

p¼1Sð5Þ^exp Pⁿ

i¼1

b^iðxixiÞ

wherepisthe5-yearCVDrisk,Sð5Þisthesurvivalrateat5 yearsofapersonwithmeanvaluesofriskfactors,

b

^i isthe estimatedcoefficientforthei-thriskfactor,andxiandxiare the actual and mean values, respectively, of the i-th risk factor.WeestimatedSð5Þand

b

^iandtheir95%CIsbyfitting thefinalmodeltoeach oftheimputeddatasetsandcom- binedtheestimatesusingRubin’srule(seesupplementary document-section5fordetailedmethods).

WeusedtheHarrell’sc-statistic[23]toevaluatethemodel discrimination, i.e. themodel’s ability toseparate persons with events from those without events. The c-statistic is defined as theproportion of allpossiblepairs ofpersons, atleastoneofwhomhasanevent,inwhichthepredicted survivaltimeislargerforthepersonwholivedlonger.Ac indexof0.7orlargerindicatesgooddiscrimination.

To quantify the model calibration, i.e. the agreement between the observed and predicted risks, we used the modifiedHosmer-Lemeshow (mH-L)

x

² statistic proposed byD’AgostinoandNamforsurvivalmodels[24].Thepar- ticipantsweredividedintodecilesbasedontheirpredicted CVDrisks,andtheKaplan-Meierestimateforfailurewithin eachdecilewascomparedwiththemeanpredictedriskinthe samedecile. AmH-L

x

²statisticlargerthan20indicatesa poorcalibration[25].

Tointernallyvalidatethemodels,wecalculatedtheappar- ent c and mH-L

x

² ^statistics ^and ^corrected ^these ^for ^the

optimismthat mightarise fromoverfittingusingthealgo- rithmdevelopedbyHarrell[23](seesupplementary document-section6fordetailedmethods).Thecorrectedcand mH-L

x

²statisticswereshowntorepresentnearlyunbiased assessmentofthemodelperformance[23].

RiskClassification

We classified 1,583 individuals as either a high (>15%), moderate(10–15%) or low 5-year CVD risk (<10%) using

thealgorithmprovidedbytheAustralianguidelinesforCVD prevention[4],therecalibrated2008Framinghamequation [11] and thenew primary andreduced informationequa- tions, thencomputed meanrisks andtheir95%CIswithin each category using Rubin’s rule [19,26]. We used the Kaplan-Meier estimator to compute 5-year CVD risks of people withinthesamecategoryindifferent imputeddata setsandcombinedthemtoestimatetheobservedrisks.The Australianguidelinesautomaticallyclassifypeoplewithone or more of the following conditions as high-risk people:

diabetesandage>60years,diabeteswithmicroalbuminuria, moderate or severe chronic kidney disease, a history of familial hypercholesterolaemia,aSBP180mmHg,serum totalcholesterol>7.5mmol/L,orage>74.Forpeople not automatically classifiedashighrisk,the1991Framingham equation[27]isrecommendedbytheAustralianguidelines forcalculatingtheabsoluterisks,whichareusedtoclassify thesepeopleintohigh(>15%),moderate(10–15%)andlow (<10%)risks.Therefore,peopleclassifiedashighriskbythe guidelinesmayhaveornothaveanyoftheabove-mentioned conditions.Weusedthesamecut-offpointsforriskcatego- ries as thoseintheguidelinestoclassifypeople usingour new equations. Because the Australian guidelines do not recommend anyequation forthe calculation ofthe actual riskofthoseautomaticallyclassifiedashighrisk,weusedthe 1991 Framingham equation [27] to examine their absolute risksinreferencetotheobservedrisksandtheriskspredicted byournew modelsandtherecalibrated2008Framingham model.

Toexaminethepotentialmisclassificationof5-yearCVD riskusingtheAustralianguidelines,wecalculatedthepro- portionsofpeopleclassifiedas highriskandmedium-low risk bytheAustralianguidelines,andwithineachofthese groups determined the proportionsof people with a con- trastingriskcategorybasedonournewmodels.

Results

Of1,583people,142developedCVDwithin5yearsafterthe firstscreening.Wefoundthatsex,age,SBP,waistcircumference,diabetesstatus,triglyceridesandACRratiowereasso- ciated with 5-year CVD risk. Table 1 provides summary statistics of theserisk factors andthe 5-year survival rate of an average person. Current smoking and cholesterol- relatedvariablesdidnotsignificantlycontributetothepre- diction(p-values>0.5;seesupplementarydocument- sec- tion 7 for coefficients of all variables includedin the full model).Table2showstheregressioncoefficientsandhazard ratios(HRs)forCVDeventsassociatedwiththeriskfactorsin theprimaryandreducedinformationmodels.Thepropor- tionalhazardassumptionwassatisfiedinfittingthemodels toanyoftheimputeddatasets.

Table3providesperformanceindicesofournewmodels andtherecalibrated2008Framinghammodel[11].Thebias- correctedc-statisticsoftheprimaryandreducedinformation modelswere0.722and0.719,respectively,indicatinggood

(4)

discrimination.Thebias-correctedmH-L

x

²statistic(p-values)oftheprimaryandreducedinformationmodelswere 12.01(0.212)and12.49(0.188),respectively,indicatinggood calibration.Comparedtotherecalibrated2008Framingham model[11],whichhadabias-correctedc-statisticof0.727and abias-correctedmH-L

x

²statistic(p-value)of29.47(0.0005), ournewmodelsperformsimilarlyintermsofdiscrimination, but far better in terms of calibration. The observed and predictedrisksindecilesoftheprimaryandreducedinfor- mationmodelsweresimilar(Figure1).

Table4showsmeanobservedandpredictedabsoluterisks forthewholecohortanddifferent groupsstratifiedbysex, age and predicted risk levels. The predicted mean 5-year risks for the whole cohort, women and men were 9.4%, 8.8% and 10.1%, respectively, using the primary model, and were 9.4%, 8.6% and 10.2%, respectively, using the reduced information model. These predicted values were almost identical to theobserved 5-years risks (9.4%, 8.7%

and 10.1% for the whole cohort, women and men, respectively).

Whenthecohort wasstratified intohigh, moderateand lowrisksbasedonthepredictedvalues,the1991Framing- hamequation[27]slightlyoverestimatedthemeanriskofthe moderate-riskgroup (12.4% against 11.7%) andunderesti- mated the mean risk of thelow-risk group (3.1% against 5.7%). In the high-risk group classified by theguidelines’

algorithm, the 1991 Framingham equation [27] underesti- mated the mean risk of the whole group (13.8% against 17.1%), and overestimated the mean risk of persons who didnothaveanyoftheconditionspre-definedintheguide- lines(20.9%against17.1%).Therecalibrated2008Framing- hammodel[11]overestimatedthemeanriskinthehigh-risk Table1 Baselineriskfactorsand5-yearsurvivalrateof

thestudycohort(N=1,583).

Riskfactor Value^*

Men,%(SE)^y,z 48.1(1.3)

Age,mean(SE)^y,z 45.0(0.3)

SystolicBP,mean(SE),mmHg^y,z 134.5(0.5) DiastolicBP,mean(SE),mmHg 75.6(0.3)

Currentsmoking,%(SE) 53.4(1.2)

BMI,mean(SE),kg/m² 28.8(0.18)

Waistcircumference,mean(SE),cm^y,z 99.2(0.4)

Diabetes,%(SE)^y,z 23.2(1.1)

Fastingbloodglucose,mean(SE),mg/dL 114.4(1.4) Totalcholesterol,mean(SE),mg/dL 198.9(1.0) HDLcholesterol,mean(SE),mg/dL 44.5(0.3) LDLcholesterol,mean(SE),mg/dL 121.2(0.9)

Totalcholesterol:HDLratio 4.7(0.04)

Triglycerides,mean(SE),mg/dL^y 179.8(3.7)

ACR,mean(SE)^y 19.2(1.6)

Underlying5-yearsurvivalrate^§

Primaryequation 0.9298

Reducedinformationequation 0.9344

Abbreviations:ACR,Albumincreatinineratio;BMI,bodymassindex;BP, bloodpressure;HDL,high-densitylipoprotein;LDL,low-densitylipoprotein;SE,standarddeviationofthemean.

*PooledvaluesusingRubin’srule[19]fromestimatesfor20imputeddata sets.

yRiskfactorsintheprimarymodel.

zRiskfactorsinthereducedinformationmodel.

§EstimatedusingthefittedCoxproportionalhazardmodelsforaperson withmeanriskfactors.

Table2 EstimatedcoefficientsandhazardratiosfromCoxproportionalhazardmodels.

Variable Primarymodel Reducedinformationmodel

Coefficient(p-value) Hazardratio Coefficient(p-value) Hazardratio

[95%CI] [95%CI] [95%CI] [95%CI]

Men 0.272(0.074) 1.328 0.261(0.076) 1.312

[0.029,0.573] [0.925,1.731] [0.031,0.552] [0.927,1.697]

Age,years 0.050(<0.001) 1.052 0.048(<0.001) 1.050

[0.038,0.063] [1.039,1.065] [0.036,0.061] [1.037,1.062]

SystolicBP,mmHg 0.010(0.012) 1.010 0.014(<0.001) 1.014

[0.002,0.018] [1.002,1.018] [0.006,0.022] [1.006,1.022]

Waistcircumference,cm 0.009(0.110) 1.010 0.009(0.09) 1.010

[-0.002,0.020] [0.999,1.020] [0.001,0.020] [0.998,1.020]

Diabetes^* 0.268(0.104) 1.325 0.405(0.009) 1.517

[0.015,0.592] [0.974,1.755] [0.101,0.709] [1.055,1.980]

Triglycerides,mg/dL 0.001(0.142) 1.001

– –

[0.000,0.001] [1.000,1.001]

ACR 0.003(0.001) 1.003

– –

[0.001,0.005] [1.001,1.004]

Abbreviations:ACR,albumincreatinineratio;BP,bloodpressure;CI,confidenceinterval.

*Self-reporteddiabetesconfirmedbytheprimarycaremedicalrecords,usedinbothprimaryandreducedinformationmodels.

(5)

group(26.4%against19.5%),butslightlyunderestimatedthe meanriskinthemoderate-(12.3%against13.2%)andlow- risk (4.5% against 5.1%) groups. In contrast, mean risks predicted by the primary model were relatively close to theobservedrisksinthehigh-(25.0%vs25.4%),moderate- (12.3%vs12.6%)andlow-risk(4.9%vs4.6%)groups.Similar meanpredictedandobservedrisksinthreeriskcategories wereobservedusingthereducedinformationmodel.

Whenriskwascalculatedseparatelyforwomenandmen (seeTablesS5andS6inthesupplementarydocument),we foundthat,inwomen,theCVDriskpredictedbythe1991 Framinghammodel[27]wasonlyonethirdoftheobserved CVDriskintheyoungestgroup(30–34yearsold),whilethis ratiowasmuchlargerinotheragegroupsofwomen(0.53–

0.86) and in any age groups of men (0.6–1.1). When the cohort was classified intomajor CVD riskcategories, the 1991Framinghamequation[27]underestimatedCVDrisk

inallcategories inwomen, whilethis wasnotthecase in men with moderate risk. In both women and men, the agreementbetweenthe observedand predictedCVDrisk usingtherecalibrated2008Framinghammodel[11]andour new models was much better compared tothe 1991 Fra- mingham model [27] in every age group and major risk category.

Using the primary model, the reduced information model,the recalibrated2008 Framinghammodel [11]and the guidelines for risk classification, the proportions of cohort (i) with high risk were 17.1%, 18.3%, 22.0% and 28.4%, respectively, (ii) with moderate risk were 13.8%, 11.6%,13.4%and6.8%,respectively,and(iii)withlowrisk were69.1%,70.2%,64.7%and64.8%,respectively(Figure2).

Given thegood calibration and the predicted mean risks almost identical to the observed mean risks in any risk categories of the new models, these figures suggest that Table3 Performanceofthenewmodelsandtherecalibrated2008Framinghammodelfor5-yearCVDriskprediction.

Newmodel Recalibrated2008

Framinghammodel Primary Reducedinformation

Discrimination

Apparentc-statistic 0.735 0.731 0.728

[95%CI] [0.698,0.782] [0.701,0.783] [0.681–0.776]

Optimism 0.013 0.012 0.001

Bias-correctedc-statistic 0.722 0.719 0.727^*

Calibration

ApparentmH-Lx²^statistic^(p-value) ^5.70^(0.770) ^7.46^(0.589) ^19.17^(0.024)^*

Optimism 6.31 5.03 10.30

Biased-correctedmH-Lx²^statistic^(p-value) ^12.01^(0.212) ^12.49^(0.188) ^29.47(<0.001)

*ThesevaluesareslightlydifferentfromthoseinthestudybyHuaetal.[11]becauseinthepresentstudyweusedmultipleimputationtoreplacemissingvalues withtheestimatesandevaluatedperformanceoftherecalibrated2008Framinghammodelonthe20imputeddatasets,whileinthepreviousstudy[11]the recalibrated2008Framinghammodelwasevaluatedonthecohortwithcompletecases.

Figure1 Meanobservedandpredicted5-yearcardiovascularrisksindecilesofthepredictedrisksusing(A)theprimary modeland(B)thereducedinformationmodel.Theerrorbarsrepresent95%confidenceintervals.

(6)

theguidelinesoverestimate theproportionofpeoplewith high risk, underestimate the proportion of people with moderateandlowrisk,andpotentiallymisclassifyasmall proportionofpeoplewithactualhigh,moderateorlowrisk.

Theagreementbetweentheguidelinesandthenewmodels inclassificationofproportionswithdifferentriskcategories

washigherinthelow-riskcategorythaninthemoderate- andhigh-riskcategories.

Usingournewmodelsasastandardtoolforidentification ofapersonwithahigh5-yearCVDrisk,wefoundpotential misclassification of the high risk and moderate-low risk people by the Australian guidelines (Figure 3). Sixteen Table4 Mean5-yearcardiovascularriskestimatedusingKaplan-Meiermethod,1991Framinghamequation[27], recalibrated2008Framinghammodel[11]andthenewmodelsforthewholecohortandgroupsclassifiedbysex,ageand predictedrisklevels.

Sample size

Observedrisk basedon Kaplan-Meier method,%[95%CI]^y

Predictedrisk,%[95%CI]^y

1991Framingham equation[27]

Recalibrated2008 Framinghammodel[11]

Newmodel

Primary Reduced information Total 1583^* 9.4[7.9–10.8] 6.7[6.4–7.1]^{ 10.4[9.8–10.9] 9.4[8.9–9.8] 9.4[8.9–9.8]

Sex

Women 822^* 8.7[6.7–10.7] 5.7[5.2–6.2]^{ 9.6[8.9–10.3] 8.8[8.1–9.4] 8.6[8.1–9.3]

Men 761^* 10.1[7.9–12.2] 7.9[7.3–8.4]^{ 11.2[10.4–11.9] 10.1[9.4–10.6] 10.2[9.5–10.8]

Agegroups(years)

30–34 327^* 3.2[1.2–5.2] 1.3[1.2–1.5]^{ 2.9[2.6–3.1] 3.3[3.1–3.5] 2.8[2.7–3.0]

35–44 541^* 3.7[2.1–5.3] 3.7[3.4–4.0]^{ 6.1[5.7–6.5] 5.6[5.4–5.9] 5.1[4.9–5.3]

45–54 377^* 13.0[9.5–16.4] 7.9[7.3–8.5]^{ 11.8[11.0–12.5] 9.9[9.2–10.5] 9.3[8.9–9.8]

55–74 338^* 20.3[15.8–24.6] 15.5[14.5–16.5]^{ 22.8[21.4–24.3] 20.6[19.5–21.7] 22.7[21.623.8]

Riskcategoriesbasedontheguidelines[4]

High(>15%) 450^z 17.1[11.0–23.2] 20.9[19.6–22.2]^§/ 13.8[12.8–14.8]^{

20.4[18.9–21.9] 16.5[15.4–17.6] 16.3[15.2–17.4]

Moderate(10–15%) 108^z 11.7[1.9–22.8] 12.4[11.5–13.2] 17.1[16.1–18.1] 14.1[13.0–15.2] 15.8[14.3–17.3]

Low(<10%) 1025^z 5.7[3.0–8.5] 3.1[2.9–3.2] 5.3[5.0–5.5] 5.7[5.0–6.0] 5.7[5.4–6.0]

Riskcategoriesbasedontherecalibrated2008Framinghamequation[11]

High(>15%) 348^z 19.5[13.5–25.5] 15.8[14.9–16.7]^§/ 18.1[17.1–19.1]^{

26.4[24.9–27.9] 20.4[19.2–21.6] 21.4[20.2–22.5]

Moderate(10–15%) 212^z 13.7[1.1–25.3] 8.4[7.6–9.2] 12.3[11.3–13.3] 11.6[11.0–12.3] 11.8[11.0k12.6]

Low(<10%) 1023^z 5.1[1.4–9.3] 2.4[2.2–2.6] 4.5[4.2–4.8] 5.2[4.7–5.3] 4.8[4.6–5.0]

Riskcategoriesbasedontheprimaryequation

High(>15%) 271^z 25.4[15.8–35.1] 14.7[13.3–16.1]^§/ 18.0[17.1-19.2]^{

26.6[25.0–28.2] 25.0[23.1–25.7] 25.9[24.7–27.1]

Moderate(10–15%) 218^z 12.6[6.6–19.9] 10.3[9.7–10.9] 14.9[14.1–15.8] 12.3[12.1–12.5] 12.5[12.1–12.8]

Low(<10%) 1094^z 4.6[2.3–7.0] 3.2[3.0–3.4] 5.4[5.2–5.6] 4.9[4.9–5.1] 4.7[4.5–4.8]

Riskcategoriesbasedonthereducedinformationequation

High(>15%) 289^z 25.2[14.9–32.9] 13.5[12.3–14.7]^§/ 17.0[16.4–18.5]^{

25.7[24.1–27.2] 23.4[22.1–24.6] 25.6[24.5–26.7]

Moderate(10–15%) 183^z 12.9[5.2–20.6] 10.4[9.7–11.1] 15.0[14.0–15.9] 12.3[11.8–12.8] 12.3[12.1k12.5]

Low(<10%) 1111^z 5.0[3.0–7.1] 3.3[3.1–3.5] 5.6[5.4–5.9] 5.2[5.0–5.4] 4.7[4.5–4.8]

*Numberofpersonsineachof20imputeddatasets.

yPooledvaluesusingRubin’srule[19]fromestimatesfor20imputeddatasets.

zAveragenumberofpersonsacrossimputeddatasets.

§EstimatedforpersonsnotautomaticallyclassifiedasthosewithhighCVDriskbytheguidelines[4],i.e.personsdonothaveanyofthefollowingconditions:

(1)Diabetesandage>60years,(2)Diabeteswithalbumincreatinineratio>2.5mg/mmolformalesor>3.5mg/mmolforfemales,(3)moderateorseverechronic kidneydisease,(4)apreviousdiagnosisoffamilialhypercholesterolaemia,(5)Systolicbloodpressure(BP)180mmHgordiastolicBP>110mmHg,and(6)total cholesterol>7.5mmol/L.

{Estimatedforallpeopleincludingpersonsautomaticallyclassifiedasthosewithahighcardiovascularriskbytheguidelines.Seecriteriainthepreviousnote.

(7)

percent(16%)ofthepopulationwerefalsepositives,i.e.the individualsclassifiedashighriskbytheguidelinesbutas moderate-lowriskbyeithertheprimaryorreducedinfor- mationmodel.Fourpercent(4%)or6%ofthepopulation werefalsenegativesonthebasisoftheprimaryorreduced informationmodel,respectively,i.e.theindividualsclassi- fiedasmoderate-lowriskbytheguidelinesbutashighrisk by the primary or reduced information model. The esti- mated sensitivity and specificity of guidelines were 74%

and 81%, respectively, based on the primary model, and were 67% and 80%, respectively, based on the reduced informationmodel.

Discussion

Inthisstudy,wedevelopedthefirst5-yearCVDriskpredic- tionmodelsbasedonIndigenousAustraliandata.Boththe primary model and the reduced information model per- formed well, and had a far better calibration (mH-L

x

²

statistic) than the recalibrated 2008 Framingham model [11]. This wasexpected as inthe recalibratedmodel only the baseline survival rate and means of risk factors were derived from the Indigenous data, while the predictors andtheircoefficientswereobtained fromtheFramingham equation[18].Asthereducedinformationmodelperformed similarlytotheprimarymodel,itprovidesaconvenienttool foraquickassessmentof5-yearCVDriskwherenolabora- toryvariablesareneeded.Usingmultipleimputationcom- binedwithbootstrapping,weconfirmedthefindingsinthe studybyHuaetal.[11]thatthe1991Framinghammodel[27]

underestimatedmean5-yearriskoftheIndigenouspeople, andtherecalibrated2008Framinghammodel[11]corrected theunderpredictionoftheoriginalmodel.Wealsoshowed thatthe1991Framinghamequation[27]underestimatedthe CVD risk more severely in younger women (aged 30–34 years)comparedtoolderwomenandtomen.Despitethis, theobservedCVDriskinthisagegroupislow(mean,3.2;

95%CI,1.2–5.2)and,therefore,thisunderestimationwould not have impacted the cardiovascular health of younger womenmorestronglycomparedtoolderwomenandmen.

In terms of discrimination, the primary model and the reducedinformationmodelperformedsimilarlytothereca- librated2008Framinghammodel[11].Thismightbebecause the c-statistic is not always sensitive to model fit [28,29].

Anotherexplanationcouldbethatanumberofriskfactors Figure2Percentageofcohortwithhigh(>15%),mod-

erate (10–15%) and low 5-year cardiovascular risk (<10%) classifiedbythenew models, therecalibrated 2008Framinghammodelandtheguidelines.

Figure3 Potentialmisclassificationofthehighriskandmoderate-lowriskbytheAustralianguidelinesforpreventionof CVDbasedonthe(A)primarymodeland(B)reducedinformationmodelfor5-yearCVDriskprediction.Thebarsrepresent proportionsofpeopleclassifiedashighrisk(>15%)andmoderate-lowriskclassifiedbytheguidelines,withineachofwhich theupperpartrepresentstheproportionofpeoplethataremisclassifiedonthebasisoftheclassificationbythenewmodels.

(8)

thatareassociatedwithCVDsuchasfamilyhistoryofCVD [30], depression [31] and history ofacute rheumaticfever [32,33] (which isstillcommon inmanyAustralian Indige- nous communities [34]) were not collected and therefore couldnotbecontrolledforinouranalysis.Ourdatasetalso didnotcontaininformationonsocioeconomicdeprivation,a risk factorfor CVDmortalityinanAustraliancohort [35].

Therefore, socioeconomicdeprivationmight be apotential risk factorforCVDincluding bothfatalandnon-fatalcar- diovasculareventsasdefinedinourstudy.Itisimportantto notethatthedistributionsofriskfactorsinourstudysample aredifferentfromthoseintheFraminghamcohort[18].For example,participantsinourstudysamplehavealowermean totalcholesterol(198.9vs213.9mg/dL),buthighermeanSBP (134.5vs127.6mmHg)andsubstantiallyhighermeanratesof smoking (53.4% vs 34.7%) and diabetes (23.2% vs 5.0%) comparedtothoseintheFraminghamstudy[18].Ourmod- elsincludeseveralfactorssuchastriglyceridesandACRthat arenot includedintheFraminghamriskequations[18,27].

Theresultsfromouranalysisconfirmed previousfindings that ACR is an important CVD risk factor in Indigenous populations [2,3]. Two studies of Indigenous Australians showed that microalbuminuria (ACR between 3 and 30mg/mmol)andmacroalbuminuria(ACR30mg/mmol) increasedtheriskofcoronaryheartdiseasebytwotothree times comparedtonormalalbuminuria, afteradjustingfor other CVD risk factors[3,20]. We did not categorise ACR becausethisreducedthemodelcalibration.

Theoutcomesfromourmodelfittingshowedthat,inthe presence ofwaistcircumference,BMI wasnotsignificantly associatedwithCVDrisk.Thisisinlinewiththeresultsfrom thestudy byWang andHoy[12],which foundthat waist circumferencewasabetterpredictorforCVDriskthanBMI inIndigenousAustralians.Thismaybeexplainedbythefact that Indigenous Australians have a greater amount of abdominal fatcompared totheEuropean Australians and theuseofBMIforweightstatusclassificationhasbeenfound tobeinappropriateinIndigenouspopulations[36,37].

Surprisingly,smokingwasnotasignificantpredictorof5- yearCVDriskinourstudy.Anexplanationforthismightbe thatsmokingwasmaskedbySBPandtriglycerides,assmok- ingwasstronglycorrelatedwiththelatter.Inthestudyby Lukeetal.[38],smokingwasalsofoundtobesignificantly associatedwithdyslipidaemiaandSBP.Althoughsmoking did not significantly contribute totheprediction of5-year CVD risk, we found it significant when fitting a model withouttheconstraintonthe5-yearfollow-uptime.Hence, thefactthesmokingisnotincludedinourmodelsfor5-year CVDriskpredictiondoesnotmeanthatreducingsmokingis notanimportantcomponentoftheCVDpreventionstrategy.

Theagreementbetween theAustralianguidelines’ algorithm[4]andournewmodelsinriskclassificationwereonly moderate. The false positives represented the largest dis- agreement, which can lead to unnecessary treatment and increasedhealthcarecosts. Theconsequences offalseneg- atives are potentially severe with individuals who would otherwise bemedicated not receivingtreatment toreduce

CVD risk. Although the proportion of people with false negatives wasrelativelysmall (about 4–6%ofthepopulation),thisisroughlyequivalenttoalargenumberof31,000–

47,000 Indigenous people with high CVD risk but could potentiallybeuntreated,assumingthatthedistributionsof the risk factors in the WPHC was similar to those in the IndigenousAustralianpopulation[39].

AsourmodelswereabletobetterclassifytheCVDriskof Indigenouspeoplecomparedtothealgorithmrecommended bytheAustralianguidelines[4],theyshouldbeconsideredas acontemporarilyevidence-basedtoolforCVDriskassess- mentin remoteIndigenous Australians. Since CVDis the leadingcauseofdeathinIndigenousAustralians[40],trans- lationofourfindingstoclinicalpracticeisanimportantstep towardsaddressingthetargetsetbytheCommonwealthof Australia,thatistoclosethegapinlifeexpectancybetween Indigenousandnon-IndigenousAustralianswithinagener- ation [41,42]. Detailing strategies for this translation is beyondthescopeofthepresentstudy,butitisworthnoting that developing a tool for risk assessment that is readily availableandacceptabletohealthprofessionalsandIndige- nouscommunitiesiscrucial.Communicationofthistooland ourfindings with physicians canbedone incollaboration withtheHeartFoundationandotherrelevantorganisations viatheinternet,scientificmeetingsandmedia.

Strengths and Limitations

Inthepresentstudy,weusedarelativelylargecommunity- basedcohortandhigh-qualitydataformodeldevelopment, multipleimputationtodealwithmissingdata,andrigorous methodstoselectthepredictorsandassessthemodelper- formance. However, our study has some limitations. The cohort used for model development might not represent the Indigenous population in North Queensland because theparticipantsinthescreeningprogramwere volunteers.

BecauseourmodelsweredevelopedforremoteIndigenous peopleandhavenotbeentestedinotherpopulations,their generalisabilityiscurrentlyinconclusive.However,theypro- videanadditionalevidence-based toolforriskassessment thatcanbeusedalongsideexistingriskscoressuchas the recalibratedFraminghammodel[11].Asdataontreatment for hypertension were not ascertained in our study, our modelsdid not includetreatmentas a covariate,whichis alignedwiththe1991Framinghamequation[27].However, wenotethatthe2008Framinghamequation[18]accounted fortheeffectofanti-hypertensivetreatmentandthereforeitis importanttoexplorethistreatmenteffectinfuturestudiesfor riskstratificationinIndigenouspopulations.Otherpotential riskfactorssuchasfamilyhistoryofCVD[30],socioeconomic deprivation[35]andrheumaticheartdisease[32,33]werenot collected,andthusadditional relevantpredictorsmightbe missedto beincluded inthemodels. Ourequations were designed as a risk prediction tooland thus they may not representallkeyriskfactorsassociatedwiththeaetiologyof CVD.Anexampleofthisistheinclusionoftriglyceridesas thesolelipidfractioninourprimarymodelwhilethereisa well-establishedstrongcorrelationbetweenCVDandother

(9)

lipidfractionssuchastotalandHDLcholesterol[30].Inthis respecttherole oftriglycerides inourequationrepresents boththedirectimpactaswellastheindirectimpactonCVD via itscorrelation with other lipid fractionsin Indigenous Australians[43,44].Althoughwerigorouslyassessthemodel performance internally, external validation of the models usingotherdatasourcesisneededandthiswarrantsfurther studyinthefuture.

Conclusions

Thealgorithmforclassificationofpeopleashigh,moderate andlow 5-yearCVDrisk recommendedbytheAustralian guidelinesforpreventionofCVD[4]issubjecttomisclassi- ficationofIndigenouspeopleinthesemajorriskcategories, andisnotaccurateinestimatingmeanabsoluteriskineachof these categories. The prediction models we developed improvedthese.Thereducedinformationmodel,withgood discrimination andcalibration, providesa convenient and reasonabletoolfor aquickassessmentof5-year CVDrisk wherenolaboratoryvariablesarerequired.

Funding Source

ThisstudyissupportedbytheAustraliaNationalHealthand Medical Research Council (project number NHMRC/

1107140). Thesponsorhasnoroleinstudydesign;orcollection, analysisandinterpretationofdata;orwritingofthemanu- script;orthedecisiontosubmitthearticleforpublication.

Conflicts of Interest

Theauthors declared that there is no potential conflictof interest.

Appendix A. Supplementary data

Supplementarymaterialrelatedtothisarticlecanbefound, in theonline version, at doi:https://doi.org/10.1016/j.hlc.

2019.02.005.

References

[1] InternationalDiabetesFederation.Expertmeetingonindigenouspeo- ples,diabetesanddevelopmentreport.Copenhagen:InternationalDia- betesFederation; 2012 .Accessed 9 August 2018https://www.

worlddiabetesfoundation.org/sites/default/files/Indigenous%

20Peoples%2C%20Diabetes%20and%20Development_IP%20Expert%

20Meeting%202012.pdf.

[2] McDermottRA,McCullochB, LiM.Glycaemia andalbuminuriaas predictorsof coronaryheartdiseasein AboriginalandTorresStrait Islanderadults:anorthQueenslandcohort.MedJAust2011;194:514–8.

[3] BarrE,CunninghamJ,TatipataS,DunbarT,KangaharanN,GuthridgeS, etal.Associationsofmortalityandcardiovasculardiseaseriskswith diabetesandalbuminuriainurbanIndigenousAustralians:theDRUID follow-upstudy.DiabetMed2017;34:946–57.

[4]NationalVascularDiseasePreventionAlliance.Guidelinesfortheman- agement of absolute cardiovascular diseaserisk; 2012 . Accessed9 August2018https://www.heartfoundation.org.au/images/uploads/

publications/Absolute-CVD-Risk-Full-Guidelines.pdf.

[5]GoffDC,Lloyd-JonesDM,BennettG,CoadyS,D’agostinoRB,Gibbons R,etal.ACC/AHAguidelineontheassessmentofcardiovascularrisk:a reportoftheAmericanCollegeofCardiology/AmericanHeartAssocia- tionTaskForceonPracticeGuidelines.JAmCollCardiol2013;2014 (63):2935–9.

[6]ReathJ,BrownN.ManagingcardiovasculardiseaseinAboriginaland TorresStraitIslanderpeople.AustPrescr2010;33:72–5.

[7]BrownA,O’SheaRL,MottK,McBrideKF,LawsonT,JenningsGL.

Essentialservicestandardsforequitablenationalcardiovascularcare for Aboriginal and Torres Strait Islander people. Heart Lung Circ 2015;24:126–41.

[8]ChanA.Makingthecaseforamoreaccuratecardiovasculardisease risk assessment tool for Indigenous Australians. Contemp Nurse 2015;50:92–3.

[9]VasantBR,Matthews V,BurgessCP, ConnorsCM,BailieRS. Wide variationinabsolutecardiovascularriskassessmentinAboriginaland TorresstraitIslanderpeoplewithtype2diabetes.FrontPublicHealth 2016;4:37.

[10]WangZ,HoyWE.IstheFraminghamcoronaryheartdiseaseabsolute riskfunctionapplicabletoAboriginalpeople?MedJAust2005;182:66–9.

[11]HuaX,McDermottR,LungT,WenitongM,Tran-DuyA,LiM,etal.

ValidationandrecalibrationoftheFraminghamcardiovasculardisease risk modelsinan AustralianIndigenouscohort.EurJ Prev Cardiol 2017;24:1660–9.

[12]WangZ,HoyW.Waistcircumference,bodymassindex,hipcircumfer- enceandwaist-to-hipratioaspredictorsofcardiovasculardiseasein Aboriginalpeople.EurJClinNutr2004;58:888–93.

[13]AdegbijaO,HoyW,WangZ.Predictionofcardiovasculardiseaserisk usingwaistcircumferenceamongAboriginalsinaremoteAustralian community.BMCPublicHealth2015;15:57.

[14]ModestiPA,AgostoniP,AgyemangC,BasuS,BenetosA,CappuccioFP, etal.Cardiovascularriskassessmentinlow-resourcesettings:aconsen- susdocumentoftheEuropeanSocietyofHypertensionWorkingGroup onHypertensionandCardiovascularRiskinLowResourceSettings.J Hypertens2014;32:951–60.

[15]Rowley KG,ThorpeAH. Research,information andconsent forthe AustralianHealthSurvey:aseparatestandardforIndigenouspeople?

MedJAust2011;195:158–9.

[16]World Health Organization. Prevention of cardiovascular disease : guidelines for assessment and management of cardiovascular risk.

Geneva: World Health Organization; 2007 Accessed 9 August 2018http://www.who.int/cardiovascular_diseases/guidelines/Full%

20text.pdf.

[17]MillerG,McDermottR,McCullochB,LeonardD,ArabenaK,MullerR.

TheWellPerson’sHealthCheck:apopulationscreeningprogramin indigenous communities in north Queensland. Aust Health Rev 2002;25:136–47.

[18]D’AgostinoRB,VasanRS,PencinaMJ,WolfPA,CobainM,MassaroJM, etal.Generalcardiovascularriskprofileforuseinprimarycare—the FraminghamHeartStudy.Circulation2008;117:743–53.

[19]RubinD.Multipleimputationfornonresponseinsurveys.NewYork:

JohnWiley&Son;2004.

[20]WangZ,HoyWE.Albuminuriaandincidentcoronaryheartdiseasein AustralianAboriginalpeople.KidneyInt2005;68:1289–93.

[21]GrambschPM,TherneauTM.Proportionalhazardstestsanddiagnostics basedonweightedresiduals.Biometrika1994;515–26.

[22]WoodwardM.Epidemiology:studydesignanddataanalysis, 3rded.

BocaRaton:CRCPress;2014.

[23]HarrellFE,LeeKL,MarkDB.Multivariableprognosticmodels:issuesin developingmodels,evaluatingassumptionsandadequacy,andmeasur- ingandreducingerrors.StatMed1996;15:361–87.

[24]D’Agostino R,Nam B-H.Evaluation ofthe performanceof survival analysis models:discriminationandcalibration measures.Handbook ofstatistics,23.2003;p.1–25.

[25]D’AgostinoSrRB,GrundyS,SullivanLM,WilsonP,CRPGroup.Vali- dationof the Framinghamcoronary heartdiseaseprediction scores:

resultsofamultipleethnicgroupsinvestigation.JAMA2001;286:180–7.

[26]RubinDB,SchenkerN.Multipleimputationforintervalestimationfrom simplerandomsampleswithignorablenonresponse.JAmStatAssoc 1986;81:366–74.

[27]AndersonK,OdellP,WilsonP,KannelW.Cardiovasculardiseaserisk profiles.AmHeartJ1991;121:293–8.