Risk Screening of Obstructive Sleep Apnea Syndrome by Body Profiles via Random Forests Model

Abstract

Background
Obstructive Sleep Apnea Syndrome (OSAS) is a major global health concern and is typically diagnosed by in-lab polysomnography (PSG). This examination though has high medical manpower costs and alternative portable methods have further limitations. This paper develops a new model for screening the risk of OSAS in different age groups and gender by using body profiles. The effects of body profiles for different subgroups in sleep stage alteration and OSAS severity are also investigated.

Methods
The data is derived from 6614 Han-Taiwanese subjects who have previously undergone PSG in order to assess the severity of OSAS in the sleep center of Taipei Medical University Shuang-Ho Hospital between March 2015 and October 2019. Characteristics of subjects, including age, gender, body mass index (BMI), neck circumference, and waist circumference, were obtained from a questionnaire. Pearson regression was used to evaluate the correlations between body profiles and sleep stages as well as sleep disorder indexes. To develop an age and gender independent model, random forests (RF), which is an ensemble learning method with high explainability, were trained by the four groups by gender and age (older or younger than 50 years old) with ratios of 70% (training dataset) and 30% (testing dataset), respectively. Prediction performance was evaluated by sensitivity, specificity and accuracy. Variable importance was assessed by averaging the impurity decrease to account for the effect of different factors.

Results
Results indicate that high BMI, neck circumference and waist circumference decreased the duration of slow-wave sleep and increased the sleep disorder indices and the percentage of wake and N1. Additionally, screening models for different gender and age utilizing anthropometric features as predictors via RF were established and demonstrated to have high accuracy (75.63% for younger males, 74.72% for elder males, 78.81% for younger females, and 72.10% for elder females). Feature importance indicated that waist circumference was the highest contributing factor in females and elder males, whereas the BMI was the highest contribution in younger males.

Conclusions
The authors recommend the use of the prediction models for those with Han-Taiwanese craniofacial features.

Full text

1
Preprint:Pleasenotethatthisarticlehasnotcompletedpeerreview.
RiskScreeningofObstructiveSleepApneaSyndrome
byBodyProfilesviaRandomForestsModel
CURRENTSTATUS:UNDERREVIEW
Cheng-YuTsai
ImperialCollegeLondonDepartmentofCivilandEnvironmentalEngineering
Wen-TeLiu
ShuangHoHospital
Yin-TzuLin
ShuangHoHospital
Shang-YangLin
TaipeiMedicalUniversityCollegeofMedicine
ArnabMajumdar
ImperialCollegeLondon
a.majumdar@imperial.ac.ukCorrespondingAuthor
ORCiD:https://orcid.org/0000-0002-6332-7858
RobertHoughton
ImperialCollegeLondonDepartmentofCivilandEnvironmentalEngineering
DeanWu
ShuangHoHospital
Hsin-ChienLee
TaipeiMedicalUniversityCollegeofMedicine
Cheng-JungWu
ShuangHoHospital
LokYeeJoyceLi
ShinKongWuHoSuMemorialHospital
Jer-NanJuang
NationalChengKungUniversity

2
Yi-ShanTsai
TaipeiMedicalUniversityCollegeofPharmacy
Shin-MeiHsu
ShuangHoHospital
Chen-ChenLo
ShuangHoHospital
KangLo
ShuangHoHospital
You-RongChen
ShuangHoHospital
Feng-ChingLin
NationalTaiwanUniversity
DOI:
10.21203/rs.3.rs-22545/v1
SUBJECTAREAS
MedicalInformatics
KEYWORDS
Obstructivesleepapneasyndrome,Polysomnography,anthropometricfeatures,
sleepdisorderindexes,randomforestsmodel

3
Abstract
Background
ObstructiveSleepApneaSyndrome(OSAS)isamajorglobalhealthconcernandistypicallydiagnosed
byin-labpolysomnography(PSG).Thisexaminationthoughhashighmedicalmanpowercostsand
alternativeportablemethodshavefurtherlimitations.Thispaperdevelopsanewmodelforscreening
theriskofOSASindifferentagegroupsandgenderbyusingbodyprofiles.Theeffectsofbody
profilesfordifferentsubgroupsinsleepstagealterationandOSASseverityarealsoinvestigated.
Methods
Thedataisderivedfrom6614Han-TaiwanesesubjectswhohavepreviouslyundergonePSGinorder
toassesstheseverityofOSASinthesleepcenterofTaipeiMedicalUniversityShuang-HoHospital
betweenMarch2015andOctober2019.Characteristicsofsubjects,includingage,gender,bodymass
index(BMI),neckcircumference,andwaistcircumference,wereobtainedfromaquestionnaire.
Pearsonregressionwasusedtoevaluatethecorrelationsbetweenbodyprofilesandsleepstagesas
wellassleepdisorderindexes.Todevelopanageandgenderindependentmodel,randomforests
(RF),whichisanensemblelearningmethodwithhighexplainability,weretrainedbythefourgroups
bygenderandage(olderoryoungerthan50yearsold)withratiosof70%(trainingdataset)and30%
(testingdataset),respectively.Predictionperformancewasevaluatedbysensitivity,specificityand
accuracy.Variableimportancewasassessedbyaveragingtheimpuritydecreasetoaccountforthe
effectofdifferentfactors.
Results
ResultsindicatethathighBMI,neckcircumferenceandwaistcircumferencedecreasedthedurationof
slow-wavesleepandincreasedthesleepdisorderindicesandthepercentageofwakeandN1.
Additionally,screeningmodelsfordifferentgenderandageutilizinganthropometricfeaturesas
predictorsviaRFwereestablishedanddemonstratedtohavehighaccuracy(75.63%foryounger
males,74.72%foreldermales,78.81%foryoungerfemales,and72.10%forelderfemales).Feature
importanceindicatedthatwaistcircumferencewasthehighestcontributingfactorinfemalesand
eldermales,whereastheBMIwasthehighestcontributioninyoungermales.
Conclusions
TheauthorsrecommendtheuseofthepredictionmodelsforthosewithHan-Taiwanesecraniofacial

4
features.
Background
Inrecentyears,ObstructiveSleepApneaSyndrome(OSAS)hasbecomeasourceofmajorhealth
concernsglobally(1).Astudyby(2)reportedthattheestimatedprevalenceofOSAS(moderateto
severedegree)intheUnitedStateswas10%inmalesbetween30and49yearsofage,risingto17%
inelderlymales(between50and70yearsold).Thesamestudynotedthatforfemales,theestimated
prevalencewas3%ofmoderate-to-severeOSASbetweentheage30and49years,againincreasing
to9%between50and70years.Withrespecttocomorbidity,OSASisalsoconsideredasan
independentriskfactorforawiderangeofailments,including:cardiovasculardiseases,systemic
hypertension,stroke,abnormalglucosemetabolismandevencancer(3,4).Furthermore,previous
studieshaveobservedthatOSAScorrelatesto:braindamage,cognitiveimpairment,anddementia
(5–7).Therefore,withoutdoubt,OSASsignificantlyaffectsanindividual’squalityoflife.
TodiagnosetheseverityofOSASandtherebydeveloptherapeuticstrategies,thein-lab
polysomnography(PSG)isthestandardexamination(8).However,thisexaminationhasassociated
highresourcecostsintermsofmedicalmanpowerforcontinuoussleepmonitoring(9).Givenboth
thattheexaminationmodalitiesareexpensiveandthatthereisoftenalackofspaceinasleep
laboratory,thewaitinglistsforanindividualtoreceiveaPSGareusuallylong,e.g.,typically,the
averagewaittimeforreceivingmedicaltherapyafteraPSGintheUnitedStatesis11.6months(10).
Thislimitedavailabilityresultsindelaysinthetimerequiredtodiagnosesleepdisorders(11).To
overcometheselimitations,theHomeSleepTest(HST)hasbeenconsideredasanalternative
portableexaminationfordiagnosingtheseverityofOSAS.However,thistesttoohasnumerous
limitationsregardingitsuse.Forinstance,theaccuratediagnosisofOSASthroughaHSTiscurtailedif
thepatientsuffersfromothercomorbidities(12).Forexample,forsubjectswhosebodymassindex
(BMI)aremorethan40andelderlysubjects(overthe65yearsofage),therearenoestablishedHST
clinicalguidelines(13).Inadditiontothesetests,thereareavarietyofquestionnairesdesignedon
thebasisoftheclinicalpredictionrules,suchas:theBerlinquestionnaire,EpworthSleepinessScale,
andtheSTOP-Bangquestionnaire(14).Althoughthesequestionnaireshavebeenvalidatedashigh

5
sensitivitytoolsforscreeningOSAS,lowspecificityhasbeenobservedineachoftheseveritygroups
(15).Hence,giventheseshortcomingsofthecurrentmethodsofevaluatingOSAS,thereisaneedto
developanewmethodforanaccurateexaminationoftheriskofOSAS.
Inordertodothis,itisworthwhiletoinvestigatethesignificantpredictorsrelatedtoOSASseverity.
Onesuchpredictorisgender,e.g.(16)notedthattheprevalenceofOSASinsouthernPennsylvaniais
nearlythreetimeshigherinmalescomparedtofemales.ThesamestudyreportedthatOSAS
prevalenceinpostmenopausalfemaleswas2.7%andsignificantlyhigherthanthe0.6%prevalencein
premenopausalfemales.Withrespecttoanthropometricprofiles,ahighermeanBMIandalarger
meanneckcircumferencehasbeenobservedinsevereOSASsubjects(N = 25)comparedwithnormal
subjects(N = 14)inTurkey(BMI:34.55kg/m2versus29.83kg/m2,p = 0.021;Neck:40.84cmversus
36.11cm,p < 0.001)(17).AnotherstudyindicatedthatforTurkishadults,theoddsratioforOSAS
was1.09(95%confidenceinterval(CI):1.014–1.17,p < 0.05)witheachincreaseof3.5cminneck
circumference(18).Inanotherstudy,thesignificantmeanlargerwaistcircumferencewasobservedin
asevereOSASgroup(N = 437)comparedwithacontrolgroup(N = 72)intheTurkishsubjects(OSAS
group:111.74 ± 12.47cm;controlgroup:91.67 ± 12.00cm,p < 0.001)(19).Inafurtherstudy,which
involvedOSASsubjects(N = 59)recruitedintheUSA,significantassociationswereobservedbetween
theapnea-hypopneaindex(AHI)andBMI(r = 0.349,p = 0.008),neckcircumference(r = 0.276,p =
0.038),andwaistcircumference(r = 0.459,p < 0.001)(20).Despiteepidemiologicalreportsshowing
thatexcessbodyprofilesareassociatedwiththeseverityofOSAS,tothebestoftheauthors’
knowledge,thereisstillnoapplicablescreeningmodelforaccessingOSASriskbyconsideringthe
anthropometricdata,genderandageeffects.Furthermore,theassociationsbetweenanthropometric
features,alternationsofsleepstructureandsleep-disorderedindexesfordifferentagegroupsand
gendersalsoremainunclear.
Therefore,thispaperexaminesthehypothesisthatbodyprofiles,asindicatorsofOSASseverity,can
beusedtoperformOSASriskscreening.TheprimaryobjectiveistoestablishOSASriskscreening
modelsfordifferentageandgendergroupsbasedonbodyprofiles.Thesemodelsareestablished
usingtherandomforests(RF)methodwhichhasanumberofadvantageswhencomparedtocurrent

6
methodsofanalysis.Furthermore,thispaperinvestigatedtheeffectsofanthropometricfeaturesin
relationtosleepstagealternationsandsleep-disorderedindexesinsubgroupsbyusingthe
regressionmodel.Thepaperisorganizedasfollows.Section2describesthecollectionofthedataset,
statisticalanalysisandtheestablishmentofthescreeningmodel.Section3presentsthebaseline
characteristicsofthesubjects,thestatisticaloutcomesbetweenbodyprofilesandPSGparameters
andtheclassifyingperformanceofthetrainedmodel.Section4discussestheresultsandcompares
thefindingswithotherrelatedstudies.ThepaperconcludesinSect.5.
Methods
Ethics
ThestudyprotocolwasapprovedbytheEthicsCommitteeoftheTaipeiMedicalUniversity-Joint
InstitutionalReviewBoard(SHH:N201911007).Theexaminationinstitution(SleepcenterofShuang-
HoHospital)wasqualifiedbytheTaiwanSocietyofSleepMedicine.Themethodswereconductedin
accordancewiththeapprovedguidelines.
Studypopulation
ThedataisderivedfromsubjectswhohavepreviouslyundergonePSGinordertoassesstheseverity
ofOSASinthesleepcenterofTaipeiMedicalUniversityShuang-HoHospital(SHH,NewTaipeiCity,
Taiwan)betweenMarch2015andOctober2019.Thecriteriaforparticipantselectionwereasfollows:
(1)theywerebetween18yearsand80yearsofage(2)nonehadreceivedanyinvasivesurgeryfor
OSAStreatment(3)nonehadregularlytakenhypnoticorpsychotropicmedications,andfinally(4)the
totalrecordingtimeofthePSGwasmorethansixhours.Abaselinescreeningquestionnairewas
administeredtoassessthefollowing:age,gender,BMI,neckcircumference,andwaistcircumference.
Additionally,theusageofmedicationandthesurgicalhistoryofeachparticipantwereobtainedfrom
theirclinicalregistration.ItisworthnotingthatallparticipantswereHan-Taiwaneseinethnicorigin,
andcraniofacialfeaturesareconsideredastheeffectivefactorsrelatedtotheOSASseverityforHan
ethnicitycomparedwithotherethnicities(21).However,thesefactorswerenotusedaspredictors
sincethispaperonlyrecruitedHan-Taiwanesesubjects.
Polysomnography
Thefull-nightPSGexaminationwasperformedbyusingResMedEmblaN7000andEmblaMPRinthe

7
sleepcenterofSHH.TherecordedPSGdatainvolvedthefollowing:electroencephalography(EEG),
electrooculography(EOG),chinandlegelectromyography(EMG),electrocardiography(ECG),nasal
andoralairflow,thoracicandabdominalbands,snoringsensor,bodypositionmeter,andoxygen
saturation.ThesedatawerescoredbycertifiedpolysomnographictechnologistsusingRemLogic
(version3.4.1)software.ThesleepstagesandrespiratoryeventswerescoredusingtheAmerican
AcademyofSleepMedicinescoringmanualfor2017(22).ThediagnosisofOSASwasdeterminedby
thefrequencyofapneaandhypopneaevents(23).TheAHIforeachsubjectwascalculatedbythe
totaleventnumbersofapneaandhypopneadividedbythetotalsleepingtime(TST).Infollowingwith
clinicalpractice,subjectswererecommendedtoundertaketheactiveinterventionwhentheirAHIwas
higherthan15timesperhour,whichistheclinicalthresholdformoderatetosevereOSAS(24).
StatisticalAnalysis
AllthestatisticalanalyseswereconductedbyusingPythonstatisticsmodule:Scikit-learn(version:
0.21.2).ThecollectedPSGdata,whichqualifiedwithinclusioncriteria,weredividedintotwogroups
bygender.Thecharacteristicsofthetwogroupswerecomparedbyusingtheindependentstudentt-
testforcontinuousvariablesorthechi-squaretestforcategoricalvariables.Inordertodeterminethe
correlationsbetweenanthropometricfeatures,sleepstructurealternationsandsleep-disordered
indicesconsideringgenderandmenopauseeffects,themaleandfemalegroupsweredividedinto
subgroupsbytheage(overtheageof50yearsorunder)(25).Linearregressionmodelswereusedto
associatethebodyprofilestotheparametersofPSGreportamongfourgroups.Thelevelof
significancewassettop < 0.05.
RandomForests
Previousstudieshaveusedclassificationmethodssuchasgeneticalgorithm(26)andsupportvector
machine(27).Inthispaper,weusedRF,whichisanensemblelearningmodelthatcanbeusedto
performclassification(28).Comparedwithotherclassificationmethods,thismethodhasthefollowing
advantages:fastresultcomputation,highexplainabilityoffeatureimportance,betteranti-noise
ability,stableperformancewithhighaccuracyandtheavoidanceofoverfitting.Therefore,theRF
methodhasbeenwidelyusedtoperformdiagnosisordecisionsupportinthemedicalfield(29,30).In

8
thisstudy,consideringtheamountofdatacollected,whichiseasilyoverfitted,andtoinvestigatethe
importanceofpredictorseffectingtheseverityofOSAS,theRFwasusedtodevelopscreeningmodels
foreachsubgroup.
TheprocedurefortrainingandtestingmodelisillustratedinFig.1.ThePSGdataofthefourgroups
weredividedintheratioof70–30%topreparethetrainingandtestingdatasetrespectively.The
structureofthemodelconsistsofanumberofclassificationandregressiontrees(CART)whichare
trainedonselecteddatausingthebootstraptechnique(28).Thistechniquerandomlysamplesa
subsetateachinternalpartfromthetrainingdatasettodecreasetrainingtimeandtoprevent
overfitting.ThetrainingdatawasinputinamodeliterativelyfortrainingeachCARTinRFwith
bootstrapping.ThenumberofCARTwasdecidedbyout-of-bag(OOB)samplesestimationwhichcan
beusedtoevaluatetheconvergenceofthepredictionerror(31,32).Inthisstudy,themaximum
numberofCARTwasdefinedas800forstabilityandresourceimplication.SinceeachCARTwas
trainedbyavarietyofsubsetdata,overfittingsituationscanbeavoidedandcanbeassembledfor
votingtoperformtheclassification.Furthermore,featureimportancewascomputedbyaveragingthe
impuritydecreasefordeterminingtheeffectbydifferentfactors(33–35).
AccuracyEvaluation
Uponcompletionofthetrainingprocess,thetestingdatasetwasinputintotheRFmodeltoaccess
themodelsensitivity,specificityanditsaccuracy.Theconfusionmatrix,whichisameasurementfor
theperformanceoftheclassification,wascomputed.Thevaluesoftrue-positive(TP),true-negative
(TN),false-positive(FP),andfalse-negative(FN)weredetermined.Inaddition,severalindexes,
includingsensitivity(TP/(TP + FN)),specificity(TN/(TP + FN)),andaccuracy((TP + TN)/(TP + TN + FP
+ FN))werecalculatedtovalidatetheaccuracyofthetrainedmodel.TheReceiverOperating
Characteristic(ROC)curvewascalculatedtodeterminetheoptimalpoint,whichservedasthemost
balancedsensitivityandspecificitypoint.TheAreaUndertheROCcurve(AUC)measuresthe
separabilityofthemodel.Thecloserthisvalueisto1,thebettertheseparabilityofthemodel.In
addition,thepositivelikelihoodratioandthenegativelikelihoodratiowerecalculatedtovalidatethe
performance.

9
Results
Characterizationofstudysubjects
Atotalof6614subjects,ofwhom69.9%(4632)weremale,wereenrolledinthisstudy,andtheir
baselinecharacteristicsaredemonstratedinTable1.Severalfeaturesareworthyofnoteinthis
sample.TherewasasignificantdifferenceinnumbersofdifferentOSASseveritiesbetweenthemale
andfemalegroups(p < 0.01).Theaverageagesofthesubjectsinthemalegroups(48.47 ±
12.81years)weresignificantlylowerthanthevalueofthesubjectsinthenormalgroup(51.94 ±
12.67years;p < 0.01).Foranthropometricfeatures,themeanBMI(26.97 ± 3.96kg/m2),themean
neckcircumference(39.25 ± 3.14cm),andthemeanwaistcircumference(93.73 ± 10.46cm)inthe
malegroupweresignificantlyhigherthanthemeanvaluesinthenormalgroup(p < 0.01).
Table1
Baselinecharacteristicsofsubjectsdividedbygender.
CategoricalVariables Male Female Total P
Number(total) 4623 1991 6614 
Number(age ≥ 50) 2096 1205 3141 
Number(age < 50) 2527 786 3473 
Age(years) 48.47 ± 12.81 51.94 ± 12.67 49.52 ± 12.87 < 0.01
BMI(kg/m2)26.97 ± 3.96 25.46 ± 4.72 26.52 ± 4.26 < 0.01
Neckcircumference
(cm) 39.25 ± 3.14 34.13 ± 3.09 37.71 ± 3.91 < 0.01
Waistcircumference
(cm) 93.73 ± 10.46 84.38 ± 11.74 90.92 ± 11.68 < 0.01
AHI(events/hour) 32.18 ± 23.01 17.29 ± 18.94 27.70 ± 22.91 < 0.01
OSASSeveritya   < 0.01
AHI < 15,n(%) 1399(30.26%) 1199(60.22%) 2598(39.28%) 
AHI ≥ 15,n(%) 3224(69.74%) 792(39.78%) 4016(60.72%) 
Dataareexpressedasameanwithstandarddeviation
BMI:Bodymassindex;OSAS:Obstructivesleepapneasyndrome
aChi-squaredtest
PSGresultsofstudysubjects
ThemeanvaluesofPSGresultsamongthesubjectsarelistedinTable2.Inthemales,theaverageof
sleepefficiency(SE)intheyoungergroups(9.68 ± 15.54%)weresignificantlyhigherthantheelderly
groups(73.00 ± 16.37%;p < 0.01).Similarresultswereobservedinthefemalegroups(younger:
79.92 ± 15.20%;elder:74.41 ± 16.37%;p < 0.01).Foreachsleepstage,themeanpercentageofthe
wakeandthestageone(N1)inbothgenders,thevaluesoftheyounggroupsweresignificantlylower
thantheelderlygroupsrespectively(allp < 0.01).Incontrast,themeanpercentageofthestagetwo
(N2)andtherapid-eye-movement(REM)stageintheyoungergroupswerehigherthanthevaluesof
elderlygroups(allp < 0.01).Theaverageminutesofthewakeaftersleeponset(WASO)inthe
youngergroupsofbothgenders(male:53.33 ± 48.43min;female:46.13 ± 43.18min)were

10
significantlylowerthanthevaluesoftheelderlygroups(male:76.41 ± 50.33min;female:65.92 ±
47.89min;p < 0.01).
Table2
PSGresultsofsubjectsdividedbygender.
CategoricalVariables Male ≤ 50y/o Male > 50y/o Female ≤ 50y/o Female > 50y/o
SE(%) 79.68 ± 15.54%*73.00 ± 16.37%*$ 79.92 ± 15.20%#74.41 ± 16.37%#$
Wake(%ofSPT) 15.71 ± 14.65%*& 22.67 ± 15.46%*$ 13.96 ± 13.67%#& 19.86 ± 14.85%#$
N1(%ofSPT) 11.75 ± 8.18%*& 14.10 ± 10.18%*$ 8.40 ± 5.88%#& 9.34 ± 6.85%#$
N2(%ofSPT) 57.88 ± 14.81%*52.89 ± 16.15%*$ 59.03 ± 14.56%#57.50 ± 15.06%#$
REM(%ofSPT) 10.99 ± 6.28%*& 9.41 ± 6.26%*$ 12.31 ± 6.71%#& 10.35 ± 6.44%#$
WASO(mins) 53.33 ± 48.43*& 76.41 ± 50.33*$ 46.13 ± 43.18#& 65.92 ± 47.89#$
AHI(events/hour) 31.06 ± 23.71*& 33.53 ± 22.09*$ 11.34 ± 16.16#& 21.16 ± 19.62#$
Desaturationsindex
(events/hour) 29.92 ± 23.98*& 31.97 ± 22.40*$ 10.48 ± 16.05#& 20.29 ± 19.62#$
Snoringindex
(events/hour) 259.62 ± 212.69&253.51 ± 217.72$149.04 ± 193.74#& 176.49 ± 202.58#$
Arousalsindex
(events/hour) 25.47 ± 16.78*& 28.89 ± 16.56*$ 16.42 ± 10.80#& 20.72 ± 13.01#$
Dataareexpressedasameanwithstandarddeviation
SE:Sleepefficiency;SPT:Sleepperiodtime;N1:Sleepstageone;N2:Sleepstagetwo;REM:Rapideyemovement
stage;WASO:Wakeaftersleeponset;AHI:Apneahypopneaindex.
*:thep-valueislessthan0.05betweenyoungandelderwithinmales.
#:thep-valueislessthan0.05betweenyoungandelderwithinfemales.
&:thep-valueislessthan0.05betweenmalesandfemaleswithinsubjectsyoungerthan50yearsold.
$:thep-valueislessthan0.05betweenmalesandfemaleswithinsubjectsolderthan50yearsold.
Intermsofthesleepqualityindexes,themeanvaluesofAHIanddesaturationindexintheelderly
groupsinbothgendersweresignificantlyhigherthanthevaluesoftheyoungergroups(AHI:p < 0.01;
desaturationindex:p < 0.05).Therewasnosignificantdifferenceinthesnoringindexinmalegroups,
whereasasignificantdifferencewasfoundinthefemalegroups(male:149.04 ± 193.74events/hour;
female:176.49 ± 202.58events/hour,p < 0.01).Also,thereweresignificantdifferencesinarousal
indexbetweenyoungerandeldergroupsinbothgenders(allp < 0.01).Additionally,thepercentage
ofwakeandN1,WASO,AHI,desaturationindex,snoringindex,andarousalindexinmaleswas
significantlyhigherthanvaluesinfemalesinbothyoungerandelderagegroups(allp < 0.05).
Conversely,thepercentageofREMinyoungermalesweresignificantlylowerthanvaluesinyounger
females(allp < 0.05).ThelowerSE,thelowerpercentageofN2,andthelowerpercentageofREMin
eldermaleswereobservedcomparedwithelderfemales(allp < 0.05).
PSGresultsandanthropometricfeatures
TheassociationsbetweentheparametersofPSGresultsandbodyprofilesinthedifferentagegroups
andgendersareillustratedinTable3andTable4respectively.Thesleepefficiency(SE)inelder

11
malesassociatednegativelywiththewaistcircumference,whereasthepercentageofwakestage
associatedpositively(p < 0.05).Similarly,thepercentageofwakestageassociatedpositivelywith
neckandwaistcircumferenceinyoungermales(p < 0.05).Thereweresignificantpositivecorrelations
betweenthepercentageofN1stageandallbodyprofilesamongallgroupsexcepttheBMIofyounger
females.Additionally,significantnegativecorrelationsbetweenthepercentageofN2,neck
circumference,andwaistcircumferenceinmaleswereobserved.ThepercentageofREMnegatively
correlatedtoallbodyprofilesinmales(p < 0.05)andnegativelycorrelatedtotheBMIaswellaswaist
circumferenceinelderfemales(p < 0.05).WhiletherewerepositivecorrelationsbetweenWASOand
allbodyprofilesamongallgroup,theylackedstatisticalsignificance.TheindexesofOSASseverity,
AHI,desaturationindex,snoringindexandarousalsindexallcorrelatedpositivelywithBMI,neck
circumferenceaswellaswaistcircumferenceinallgroups(allp < 0.05).
Table3
Pearson’scorrelationcoefficientsforvariablesofPSGresultsandanthropometricfeaturesinmales
Categorical
Variables Male ≤ 50y/o Male > 50y/o
 BMI NC(cm) WC(cm) BMI NC(cm) WC(cm)
Sleepefficiency
(%) 0.0001 -0.0013 -0.0003 -0.0001 -0.0005 -0.0007*
Wake(%of
SPT) 0.0005 0.0020*0.0005*0.0004 0.0013 0.0007*
N1(%ofSPT) 0.0032*0.0047*0.0013*0.0036*0.0061*0.0014*
N2(%ofSPT) -0.0007 -0.0027*-0.0006*-0.0019 -0.0051*-0.0014*
REM(%ofSPT) -0.0017*-0.0014*-0.0006*-0.0017*-0.0015*-0.0006*
WASO(mins) 0.1346 0.5788 0.1553 0.0670 0.3069 0.2041
AHI
(events/hour) 2.7538*3.3725*1.0994*2.5695*2.8574*0.9283*
Desaturations
index
(events/hour)
2.8501*3.4735*1.1196*2.6869*2.9548*0.9587*
Snoringindex
(events/hour) 12.8627*18.0209*5.5982*18.5562*21.5843*6.6324*
Arousalsindex
(events/hour) 0.9191*1.5281*0.3946*0.9420*1.3266*0.3601*
SE:Sleepefficiency;SPT:Sleepperiodtime;N1:Sleepstageone;N2:Sleepstagetwo;REM:Rapideyemovement
stage;WASO:Wakeaftersleeponset;AHI:Apneahypopneaindex;BMI:Bodymassindex;NC:Neck
circumference;WC:Waistcircumference.
*:thep-valueislessthan0.05.

12
Table4
Pearson’scorrelationcoefficientsforvariablesofPSGresultsandanthropometricfeaturesinfemales
Categorical
Variables Female ≤ 50y/o Female > 50y/o
 BMI NC(cm) WC(cm) BMI NC(cm) WC(cm)
SE(%) < 0.0001 -0.001 -0.0003 -0.0009 -0.0017 -0.0005
Wake(%of
SPT) 0.0006 0.0019 0.0005 0.001 0.0024 0.0005
N1(%ofSPT) 0.0005 0.0016*0.0004*0.0018*0.0030*0.0007*
N2(%ofSPT) -0.0001 -0.0004 -0.0002 -0.0003 -0.0027 -0.0004
REM(%ofSPT) -0.0001 < 0.0001 -0.0003 -0.0014*-0.0012 -0.0004*
WASO(mins) 0.2924 0.7094 0.1976 0.1876 0.7309 0.11
AHI
(events/hour) 1.4414*2.3677*0.6746*2.0423*2.6475*0.7909*
Desaturations
index
(events/hour)
1.4971*2.4165*0.6904*2.1067*2.6857*0.8024*
Snoringindex
(events/hour) 10.4552*17.5522*4.5551*14.8561*18.7307*5.9572*
Arousalsindex
(events/hour) 0.2342*0.4529*0.1350*0.5117*0.9050*0.2279*
SE:Sleepefficiency;SPT:Sleepperiodtime;N1:Sleepstageone;N2:Sleepstagetwo;REM:Rapideyemovement
stage;WASO:Wakeaftersleeponset;AHI:Apneahypopneaindex;BMI:Bodymassindex;NC:Neck
circumference;WC:Waistcircumference.
*:thep-valueislessthan0.05
AccuracyPerformanceandparameterimportance
TheresultsoftheclassificationandvariableimportanceforeachmodelarepresentedinTable5.The
predictionaccuraciesfordifferentgroupswererangedfrom72.10–78.81%.TheAUCofROCwas
arrangedfrom76.80–81.61%.Forvariableimportance,theBMIcontributedthehighestimpact,of
38.10%inthemodelforyoungermales.However,themostimpactingfactorinothergroups,wasthe
waistcircumference:39.16%foreldermales,39.90%foryoungerfemales,and39.08%forelder
females.
Table5
Theclassificationresultsoftheage-dependentbodyprofilemodelonfoursubgroups.
CategoricalVariables Male ≤ 50y/o Male > 50y/o Female ≤ 50y/o Female > 50y/o
Casenumber(AHI ≥
15vsAHI < 15) 1676vs851 1548vs548 596vs190 603vs602
Sensitivity,% 66.84 ± 6.54 62.90 ± 13.21 83.96 ± 6.06 80.12 ± 5.87
Specificity,% 78.69 ± 3.62 76.01 ± 3.73 59.18 ± 14.97 65.67 ± 7.01
Accuracy,% 75.63 ± 3.22 74.72 ± 3.59 78.81 ± 5.77 72.10 ± 4.56
LR+ 3.14 ± 0.54 2.62 ± 0.56 2.06 ± 0.84 2.33 ± 0.43
LR- 0.42 ± 0.08 0.49 ± 0.14 0.27 ± 0.14 0.27 ± 0.12
AUC,% 81.61 ± 2.7 77.11 ± 3.01 79.30 ± 5.86 76.80 ± 4.19
Featureimportance
(100%)    
BMI(%) 38.10% 38.84% 33.71% 37.00%
Neckcircumference
(%) 24.87% 23.00% 26.39% 23.92%
Waistcircumference
(%) 37.03% 39.16% 39.90% 39.08%
Dataareexpressedasameanwithstandarddeviation
AHI:Apneahypopneaindex;LR+:Positivelikelihoodratio;LR-:Negativelikelihoodratio;AUC:Areaunderthe
curve.
Discussion
Inthisstudy,theassociationsbetweenPSGparametersandanthropometricfeatureshavebeen

13
determinedconsideringageandgendereffect.Thegenderandage-independentmodelsbasedon
theanthropometricfeatureswereestablishedsuccessfullytoassesstheriskofOSAS.Theapplicable
modelsweredemonstratedtopossesshighpredictionaccuracyforclassifyingAHIhigherorlower
than15,inparticularforHan-Taiwanesesubjects.Theanthropometricfeatureimportancefor
effectingOSASseveritywasobtainedforeachsubgroup.Thelarge-scalestatistics,includingHan-
Taiwaneseanthropometricfeatures,sleepstagedetailsandPSGresultswerealsoprovided.
Forbothgenders,betteraccuracycanbeobservedintheyoungergroups,andthewaist
circumferenceshowedthehighestimportanceforitsaffectsontheAHIexceptforthecaseof
youngermales.Itisknownthatvisceralfat,whichisatypeofbodyfatdepositedaroundinternal
organs,isrelatedtotheAHIandwaistcircumferenceisauniqueindicatorforindicatingvisceralfat
distribution.Similarly,apriorstudy,whichusedtraditionalstatisticalmethods,reportedthe
observationthatwaistcircumferencewasabetterpredictorfortheseverityofOSAScomparedwith
theBMIandneckcircumference(36).AnotherstudyrevealedthatOSASprevalencewasexacerbated
inmenopausalfemalesandwaistcircumferenceservedasthemainfactor(37).Additionally,theBMI
andwaistcircumferenceshowedsimilarimportanceforeffectingtheAHIinthemalesandelderly
femalegroups,butthereisadifferenceinyoungerfemales.Therewerestatisticallysignificant
correlationsbetweentheBMI,waistcircumferenceandsleepstagepercentage,forthemalesand
elderlyfemalegroups,butnotforyoungerfemales.Theseresultsmaybeinducedbythemenopause
effect.Thiseffectmaynotleadtoweightgaindirectly,butitmaybecorrelatedtothefatdistribution
changes.Inperimenopausefemales,theincreasedabdominaladipositydepositionanddecreased
leanbodymasswereobserved.Thischangeissimilartothefatdistributionofmaleswhichtendsto
developagreaterdegreeofupperbodyobesity.
Withrespecttoneckcircumference,whichisanindicatoroffatdistributionintheupperairway,there
arethesignificantcorrelationsbetweenthesnoringindexandarousalindex.Increasedneck
circumferencemaynarrowtheupperairwayandincreasetheturbulenceofairflow,therebycausing
snoring.Besides,thecollapsibleupperairwayandtheincreasedupperairwayresistancemay
increaseintherespiratoryeffortinordertokeepventilating.Thisisthemechanismbywhich

14
increasingrespiratoryeffortmaystimulatetransientarousalinanindividualandleadtosleep
fragmentation.Similarfindingsweredemonstratedinpreviousstudies.Largeneckcircumferencehas
beenreportedasafactorinsnoring(38).
Furthermore,forbothgenders,theyoungergroupsreportedbetteroverallsleepqualityandlower
sleepdisorderindices.Theseresultsmaybeduetothecollagenlossofupperairwayconnective
tissuewithageing.Withthedecreaseinneuromusculartension,theupperairwayeasilycollapses
duringsleep.Forelderlyfemales,thedecreasedhormonelevelsduringmenopausealsoresult
increasedsleepdisorderindexesandaffectsleepquality.Similarly,numerousstudieshaveshown
thatobstructioneventsandtheseverityofOSASaresignificantlymorelikelyinelderlypatients(39,
40).Collectively,theobservationsofthisstudyareconsistentwithpreviousstudiesdemonstrating
thattheriskofOSASwasaffectedbyage,gender,andbodyprofiles.
Therearesomelimitationstothisstudy,whichshouldbeaddressedinthefuture.Firstofall,inthis
study,thedatasetwaslimitedtoaSouth-EastAsianpopulation,withcraniofacialfactors,ratherthan
fromdiversebodyprofileswithawidergeographicaldistribution.Hencetheresultsofthispaper
shouldbeviewedwiththisinmind,sincecraniofacialfactors,whichalsoaffectssleep-disordered
breathingshouldbeconsideredaspredictors(41).Next,theclinicalstandardforclassifyingOSAS
severityrequiresaPSGtodeterminetheAHI.Thissleepexaminationisstillconductedbymanual
interpretation,andsincethePSGresultswerescoredbydifferenttechnologists,thescoringvariability
canaffecttheaccuracy(42).Althoughthedatawasderivedfromonesleepcenter,whichregularly
performedinter-scoringtraining,scoringvariabilitycouldstillhaveaffectedtheresults.Furthermore,
thefirstnighteffect,whichisaphenomenononthefirstnightoftestingcharacterizedbyanaltered
sleepcycleandimpactedsleepphysiology,canalsocauseinaccuraciesofPSGresults(43).To
minimizethiseffect,somePSGparameters,suchassleepefficiency,shouldbeusedtoruleout
subjectsandrearrangethePSGforavoidingthebias.
AnotherlimitationconcernsthelackofsomeinteractingfactorsofOSAS,whileOSAShasbeen
recognizedasmultifactorialsleep-disorderedbreathing.Somebehaviors,includingsmoking,alcohol
use,environmentalparameters,andmenopausalstatusarehighlyassociatedwithOSAS(44,45).To

15
understandinfluenceofbackgrounddetails,thequestionnairecanbeusedtoobtainpersonalhabits.
OSASalsoinfluencedbydifferentdiseases(44).Thesituationofcomorbidityalsoaffectstheresultsof
PSG.Thedisease-relatedparametersthatarealreadyavailablefromclinicalinformationcanbe
obtainedandserveassignificantvariablesforpreformingpre-screenclassification.
Infuturework,adatasetwithcomprehensivedimensions,whichincludepersonalhabits,personal
comorbidity,moreanthropometricfeatures,andbodycompositions,willbecollectedfortraininga
novelmodel.
Conclusion
GiventheconcernsgloballyabouttheimpactsofOSAS,thispapernotedtheneedforamethodof
measuringOSASgiventhelimitationsofcurrentmethods.Inordertodothis,OSASriskscreening
modelsfordifferentageandgendergroupsbasedonbodyprofilesweredevelopedbasedupondata
from6614participantsfromTaiwan.
ResultsindicatethathighBMI,neckcircumferenceandwaistcircumferencedecreasedthedurationof
slow-wavesleepandincreasedthesleepdisorderindicesandthepercentageofwakeandN1.
Additionally,predictionmodelsfordifferentgenderandageutilizinganthropometricfeaturesas
predictorsviaRFwereestablishedanddemonstratedtohavehighaccuracy.Featureimportancewas
explored,withwaistcircumferencethehighestcontributingfactorinfemalesandeldermales,
whereastheBMIwasthehighestcontributioninyoungermales.
TheauthorsrecommendtheuseofthepredictionmodelsforTaiwanandindeedforthosewithHan-
Taiwanesecraniofacialfeatures.
Abbreviations
OSAS:Obstructivesleepapneasyndrome;PSG:polysomnography;HST:HomeSleepTest;AHI:
Apnea-hypopneaindex;BMI:Bodymassindex;NC:Neckcircumference;WC:Waistcircumference;
SE:Sleepefficiency;SPT:Sleepperiodtime;N1:Sleepstageone;N2:Sleepstagetwo;REM:Rapid
eyemovementstage;WASO:Wakeaftersleeponset;AUC:Areaunderthecurve;RF:Randomforest
Declarations
Availabilityofdataandmaterials
Alldatageneratedoranalyzedduringthisstudyareincludedinthispublishedarticleand

16
supplementaryinformationfiles.
Ethicsapprovalandconsenttoparticipate
Ourstudyusedexistingrecordstoconductaretrospectivestudy.Allnecessaryapprovalsanda
waiverofinformedconsentwereobtainedfromtheEthicsCommitteeoftheTaipeiMedicalUniversity-
JointInstitutionalReviewBoard(SHH:N201911007).
Acknowledgments
Wewouldliketoappreciatealltheparticipantsfortheircontributiontothisresearch.
Authors’contributions
AMandWTLguidedthewholeresearchwork;CYTandWTLmadesubstantialcontributionsto
conceptionanddesignofthestudy;LYJL,YST,SMH,CCL,KL,YRC,andFCLcollectedandorganized
thedata;CYT,YTL,SYLandRHperformedtheanalysisandinterpretationofdata;CYTdraftedthe
article;WTL,AM,DW,HCL,CJW,andJNJrevisedthemanuscriptcriticallyforimportantintellectual
content.Allauthorshavereadandapprovedthefinalversionofthemanuscriptforpublication.
Funding
ThisstudywasfundedbytheMinistryofScienceandTechnologyofTaiwan(MOST108-2634-F-038-
003).Thefunderhadnoroleinthestudydesign,datacollectionandanalysis,orinwritingofthe
manuscript.
Consenttopublication
Notapplicable.
Competinginterest
Theauthorsdeclarethattheyhavenocompetinginterest.
References
1.  G i b sonG.Obstructivesleepapn o e a syndrome:underestimate d  a n dundertreated.
Britishmedicalbulletin.2004 ; 7 2 (1):49–64.
2.  P e ppardPE,YoungT,Barnet  J H ,  P altaM,HagenEW,HlaKM .  I n c r easedprevalenceof
sleep-disorderedbreathing i n  a d ults.AmJEpidemiol.2013;1 7 7 (9):1006–14.
3.  G o nzagaC,BertolamiA,Berto l a m i M,AmodeoC,CalhounD.O b s t r uctivesleep

17
apnea,hypertensionandc a r d io vasculardiseases.JHumH y p e rtens.2015;29(12):705.
4.  B r ennerR,KivityS,PekerM,R e i n h ornD,Keinan-BokerL,Silver m a nB,etal.
Increasedriskforcancerin y o u n gpatientswithsevereobstr u c t i v e sleepapnea.
Respiration.2019;97(1):1 5 – 2 3.
5.  P e ter-DerexL,YammineP,Ba s t u j i  H,CroisileB.SleepandAlzheim e r 's disease.Sleep
medicinereviews.2015;1 9 : 2 9–38.
6.  A l v arez-SabínJ,RomeroO,De l g a d oP,QuintanaM,Santamar i n a  E ,FerréA,etal.
Obstructivesleepapneaan d  s i l e ntcerebralinfarctioninhyper t e n s i v eindividuals.J
SleepRes.2018;27(2):2 3 2 – 9.
7.  B u rattiL,LuzziS,PetrelliC,Baldine l l i  S ,  V it i c chiG,ProvincialiL,etal.Obstru c t i v e
sleepapneasyndrome:a n  e m ergingriskfactorfordem e n t i a .CNS&Neurological
Disorders-DrugTargets(Fo r m e rlyCurrentDrugTargets-CN S  &  Neurological
Disorders).2016;15(6):6 7 8 – 82.
8.  K u shidaCA,LittnerMR,Morge n t h a le rT,AlessiCA,BaileyD,Colem a n  J J r ,etal.
Practiceparametersforthe  i n d ic ationsforpolysomnograph y  a n drelatedprocedures:
anupdatefor2005.Slee p .  2 005;28(4):499–523.
9.  A n coli-IsraelS,ColeR,AlessiC,Ch a m bersM,MoorcroftW,Pollak  C P .  Theroleof
actigraphyinthestudyofsle e p  a ndcircadianrhythms.Sleep .  2 0 03;26(3):342–92.
10. R o t e n bergBW,GeorgeCF,Sulliv a n  K M,WongE.Waittimesfors l e e p apneacarein
Ontario:amultidisciplinaryas s e s s ment.Canadianrespiratory  j o u r n al.
2010;17(4):170–4.
11. S t e w artSA,SkomroR,ReidJ,Pe n z  E ,FentonM,GjevreJ,etal.Im p r o v ementin
obstructivesleepapneadia g n o s isandmanagementwait t i m e s:Aretrospective
analysisofahomemana g e m entpathwayforobstructive  s l e e papnea.Canadian
respiratoryjournal.2015;2 2 ( 3 ):167–70.

18
12. C o l l o p NA,AndersonWM,Boehle c k e  B,ClamanD,GoldbergR,G o t t l i e bDJ,etal.
Clinicalguidelinesfortheuseo f  u n a ttendedportablemonitors in  t h e diagnosisof
obstructivesleepapneaina d u l t  p atients.JClinSleepMed.200 7 ; 3 (7):737–47.
13. K u n d elV,ShahN.Impactofport a b l e  sleeptesting.SleepMedClin.  2 0 1 7;12(1):137–
47.
14. M i r r a khimovAE,SooronbaevT,  M i r r akhimovEM.Prevalenceo f  o b structivesleep
apneainAsianadults:asys t e m aticreviewoftheliterature.B M C  p ulmonary
medicine.2013;13(1):10 .
15. E l - S a y edIH.Comparisonoffour s l e e p questionnairesforscreening  o b structivesleep
apnea.EgyptianJournalofC h e s tDiseasesTuberculosis.201 2 ; 6 1(4):433–41.
16. B i x l e r EO,VgontzasAN,LinH-M,T e n  H aveT,ReinJ,Vela-BuenoA, e t  a l .  Prevalence
ofsleep-disorderedbreathin g  i n  women:effectsofgender .  A m JRespirCritCare
Med.2001;163(3):608–13.
17. A h b a bS,AtaoğluHE,TunaM,K a r a s uluL,ÇetinF,TemizL,etal.N e c k  circumference,
metabolicsyndromeand o b s t r uctivesleepapneasyndrom e ;  e valuationofpossible
linkage.Medicalsciencemo n i t o r :internationalmedicaljournal o f  e x perimental
clinicalresearch.2013;19: 1 1 1 .
18. O n a t A,HergençG,YükselH,Ca n  G ,  AyhanE,KayaZ,etal.Neck  c i r c umferenceasa
measureofcentralobesity :  a s s ociationswithmetabolicsynd r o m eandobstructive
sleepapneasyndromebe y o n dwaistcircumference.Clinica l  n u t rition.2009;28(1):46–
51.
19. U n a l  Y,OzturkDA,TosunK,KutluG .  A s s ociationbetweenobstructive  s l e e p apnea
syndromeandwaist-to-heig h t  r a tio.SleepBreathing.2019; 2 3 ( 2 ):523–9.
20. T o m C,RoyB,VigR,KangDW, A y s o la RS,WooMA,etal.Correla t i o n s BetweenWaist
andNeckCircumferences  a n d ObstructiveSleepApneaCh a r a cteristics.Sleep

19
vigilance.2018;2(2):111– 8 .
21. S u t h e rlandK,LeeRW,CistulliPA.Ob e s i t y andcraniofacialstructurea s  r i s k factors
forobstructivesleepapnoe a :  i m pactofethnicity.Respirology.  2 0 12;17(2):213–22.
22. B e r r yRB,BrooksR,Gamaldo C ,  H ardingSM,LloydRM,Quan  S F , etal.AASMscoring
manualupdatesfor2017  ( v e rsion2.4).JClinSleepMed.2 0 1 7;13(05):665–6.
23. F o r c e AAoSMT.Sleep-relatedbre a t h i n gdisordersinadults:recom m e ndationsfor
syndromedefinitionandm e a s urementtechniquesinclinicalr e s e a rch.Sleep.
1999;22:667–89.
24. K a p u rVK,BaldwinCM,ResnickH E ,  G ottliebDJ,NietoFJ.Sleepinessin  p a ti e ntswith
moderatetoseveresleep- d i s o r deredbreathing.Sleep.20 0 5 ; 2 8(4):472–8.
25. P o l o - K antolaP.Sleepandmenop a u s e .Women’sHealth.2007 ; 3 ( 1 ): 99–106.
26. S u n  L M,ChiuH-W,ChuangCY,Liu  L .  A predictionmodelbasedon  a n  a rtificial
intelligencesystemformod e r a t e tosevereobstructivesleep a p n ea.SleepBreathing.
2011;15(3):317–23.
27. L i u  W -T ,WuH-t,JuangJ-N,Wisniews k i  A ,  LeeH-C,WuD,etal.Predictio n  o f the
severityofobstructivesleep a p n e abyanthropometricfeatu r e s  v iasupportvector
machine.PloSone.2017; 1 2 ( 5 ).
28. B r e i m anL.Randomforests.M a c h i n elearning.2001;45(1):5– 3 2 .
29. Y a n g F,WangH-z,MiH,CaiW.- w .  U singrandomforestforrelia b l e  c la ssificationand
cost-sensitivelearningforme d i c a l diagnosis.BMCBioinform.2 0 0 9;10(1):22.
30. A l i c k o vicE,SubasiA.Medicaldecis i o n  s u pportsystemfordiagnosis  o f  h eart
arrhythmiausingDWTand  r a n domforestsclassifier.Journa l  o f  m edicalsystems.
2016;40(4):108.
31. X i a o  M,YanH,SongJ,YangY,Ya n g  X.Sleepstagesclassificationb a s e donheartrate
variabilityandrandomfore s t .  B io medSignalProcessContro l.  2 0 13;8(6):624–33.

20
32. M i t c h ellMW.BiasoftheRandom  F o r estout-of-bag(OOB)errorfo r  c e rtaininput
parameters.OpenJourna l o f  S ta tistics.2011;1(03):205.
33. A l t m a nnA,ToloşiL,SanderO,Len g a u erT.Permutationimportan c e :  a corrected
featureimportancemeas u r e .  Bioinformatics.2010;26(1 0 ) : 1 340–7.
34. A r c h e rKJ,KimesRV.Empiricalcha r a c terizationofrandomforestv a r i a b leimportance
measures.ComputStatD a t a  A nal.2008;52(4):2249–6 0 .
35. P a l  M .Randomforestclassifierfo r  r e motesensingclassification.Int J  R e m oteSens.
2005;26(1):217–22.
36. D a v i d sonTM,PatelMR.Waistcirc u m ferenceandsleepdisordere d  b reathing.
Laryngoscope.2008;11 8 ( 2 ) :3 39–47.
37. P o l e s e lDN,HirotsuC,NozoeKT,B o i n  A C,BittencourtL,TufikS,etal.W a i s t
circumferenceandpostm e n o p ausestagesasthemaina s s o c iatedfactorsforsleep
apneainwomen:across - s e c t i o nalpopulation-basedstudy. M e n opause.
2015;22(8):835–44.
38. F a z l e eKibriaChowdhuryMd
BiswasR,RahmanM.Fa z l e e  KibriaChowdhuryMd(201 8 )  R elationofSnoringHabits
withBodyMassIndexand  N e c kCircumferenceamongA d u l t  P opulation.JSleep
DisordTher.2018;7(293):2167-0277.1000293.
39. K i r a l  N ,SalepçiB,FidanA,TorunE ,  C ö m ertS,SaraçG,etal.Relat i o n s h ip between
obstructivesleepapneasyn d r o meseverityandage.EurR e s p ir atorySoc;2011.
40. L e p p änenT,TöyräsJ,Mervaala E ,  P e nzelT,KulkasA.Severityofin d i v i d ual
obstructioneventsincreases  w i t h ageinpatientswithobstructiv e  s l e epapnea.Sleep
medicine.2017;37:32–7 .
41. L e e  R W,VasudavanS,HuiDS, P r v a nT,PetoczP,DarendelilerMA ,  e t  al.Differences
incraniofacialstructuresand  o b e sityinCaucasianandChinese  p a ti e ntswith

21
obstructivesleepapnea.Sle e p .  2010;33(8):1075–80.
42. C o l l o p NA.Scoringvariabilitybetwe e n  p olysomnographytechnolog i s t s  i n d ifferent
sleeplaboratories.Sleepm e d i c i n e.2002;3(1):43–7.
43. R i e d e lBW,WinfieldCF,LichsteinKL .  F i r s t nighteffectandreversefirst  n i g h teffect
inolderadultswithprimary in s o m nia:doesanxietyplayarole ?  S l e epmedicine.
2001;2(2):125–33.
44. H ı z l ı  Ö , ÖzcanM,ÜnalA.Evaluation  o f  c omorbiditiesinpatientswithO S A S and
simplesnoring.TheScientific W o r ld Journal.2013;2013.
45. V e l d i  M ,AniR,VaherH,EllerT,Hio n  T ,  AluojaA,etal.Obstructiveslee p  a p nea
syndrome(OSAS):Pathop h y s i o lo gyinEstonians.Pathophysio lo g y .  2010;17(3):219–
23.
SupplementaryInformation
Additionalfile1.TheDatasetfromSHHSleepCenter
Figures
Figure1
ThetrainingprocessofRFmodelwithbootstrappingtechniqueToincreasetheefficiency
andpreventtheoverfittinghappened,bootstrappingcanbeappliedtosamplethesubset
fromtheoriginaldataset.Thetechniqueistorepeatedlysampletoacquiresub-training
datawithreplacementfromtheoriginaldatasettogeneratemultipleseparatedatasetsfor
trainingeachclassificationandregressiontrees(CART).Abbreviations:RF,randomforests

22
SupplementaryFiles
Thisisalistofsupplementaryfilesassociatedwiththispreprint.Clicktodownload.
Additionalfile1.TheDatasetfromSHHSleepCenter.csv