Dysconnectivity of neurocognitive networks at rest in very-preterm born adults☆.

1.6 MB·English

by  White, Thomas P.

NeuroImage:Clinical4(2014)352–365
ContentslistsavailableatScienceDirect
NeuroImage: Clinical
journal homepage: www.elsevier.com/locate/ynicl
Dysconnectivity of neurocognitive networks at rest in very-preterm
☆
born adults
ThomasP.Whitea,IonaSymingtona,NazarethP.Castellanosa,PhilipJ.Brittaina,SeánFroudistWalsha,
Kie-WooNama,JoãoR.Satob,MatthewP.G.Allina,SukhiS.Shergilla,RobinM.Murraya,
SteveC.R.Williamsc,ChiaraNosartia,⁎
aDepartmentofPsychosisStudies,InstituteofPsychiatry,King'sCollegeLondon,deCrespignyPark,LondonSE58AF,UK
bCentreofMathematics,ComputationandCognition,UniversidadeFederaldoABC,Av.dosEstados,5001BairroBangu,SantoAndré,SPCEP09210-580,Brazil
cDepartmentofNeuroimaging,InstituteofPsychiatry,King'sCollegeLondon,deCrespignyPark,LondonSE58AF,UK
a r t i c l e i n f o a b s t r a c t
Articlehistory: Advancesinneonatalmedicinehaveresultedinalargerproportionofpreterm-bornindividualsreachingadult-
Received23September2013 hood.Theirincreasedliabilitytopsychiatricillnessandimpairmentsofcognitionandbehaviourintimatelasting
Receivedinrevisedform28November2013 cerebralconsequences;however,thecentralphysiologicaldisturbancesremainunclear.Offundamentalimpor-
Accepted12January2014 tancetoefficientbrainfunctionisthecoordinationandcontextually-relevantrecruitmentofneuralnetworks.
Availableonline18January2014
Large-scaledistributednetworksemergeperinatallyandincreaseinhierarchicalcomplexitythroughdevelop-
ment.Preterm-bornindividualsexhibitsystematicreductionsincorrelationstrengthwithinthesenetworksdur-
Keywords:
inginfancy.Here,weinvestigateresting-statefunctionalconnectivityinfunctionalmagneticresonanceimaging
Pretermbirth
Resting-state datafrom29very-preterm(VPT)-bornadultsand23term-borncontrols.Neurocognitivenetworkswereidenti-
Functionalconnectivity fiedwithspatialindependentcomponentanalysisconductedusingtheInfomaxalgorithmandemployingIcasso
Neurocognitivenetworks procedurestoenhancecomponentrobustness.Networkspatialfocusandspectralpowerwerenotgenerallysig-
Executivefunction nificantlyaffectedbypretermbirth.Bycontrast,Granger-causalityanalysisofthetimecoursesofnetworkactiv-
ityrevealedwidespreadreductionsinbetween-networkconnectivityinthepretermgroup,particularlyalong
pathsincludingsalience-networkfeatures.ThepotentialclinicalrelevanceoftheseGranger-causalmeasure-
mentswassuggestedbylineardiscriminantanalysisoftopologicalrepresentationsofconnectionstrength,
whichclassifiedindividualsbygroupwithamaximalaccuracyof86%.Functionalconnectionsfromthestriatal
saliencenetworktotheposteriordefaultmodenetworkinformedthisclassificationmostpowerfully.Inthe
VPT-borngroupitwasadditionallyfoundthatperinatalfactorssignificantlymoderatedtherelationshipbetween
executivefunction(whichwasreducedintheVPT-bornascomparedwiththeterm-borngroup)andgeneralised
partialdirectedcoherence.Togetherthesefindingsshowthatresting-statefunctionalconnectivityofpreterm-
bornindividualsremainscompromisedinadulthood;andpresentconsistentevidencethatthestriatalsalience
networkispreferentiallyaffected.Therapeuticpracticesdirectedatstrengtheningwithin-networkcohesion
andfine-tuningbetween-networkinter-relationsmayhavethepotentialtomitigatethecognitive,behavioural
andpsychiatricrepercussionsofpretermbirth.
1.Introduction disorders(Boyleetal.,2011;Dalzieletal.,2007;Hacketal.,2004;
Nosartietal.,2012;Walsheetal.,2008;Wilesetal.,2005).
Infants born at or before a gestational age of 33weeks (very- OurunderstandingofthecerebralsequelaeofVPT-birthisconsider-
preterm;VPT)arelikelytoexhibitcognitive,educationalandbehav- ableforearlystagesoflife.ThebrainsofVPT-borninfantsoftenexhibit
iouralproblemsinchildhood(Boyleetal.,2011;JohnsonandMarlow, haemorrhagicandhypoxic–ischaemicdamageleadingtoventriculardi-
2011),whichpersistthroughadolescenceintoadulthood(Allinetal., latation,whitematterabnormality,anenlargedsubarachnoidspace
2008;Hack,2009;Hilleetal.,2007;SaigalandDoyle,2008).Preterm (Inderetal.,2003)andregionalmyelindamage(Sieetal.,1997).Fur-
birthhasfurtherbeenassociatedwithanincreasedriskofpsychiatric ther,itappearsthatratesofcorticalgrowthandrelatedmicrostructure
developmentarereducedduringinfancyinlinewithprematurity(Ball
etal.,2013).Anatomicalabnormalitiesobservedduringchildhoodin-
☆ Thisisanopen-accessarticledistributedunderthetermsoftheCreativeCommons cludehemispherically-asymmetricreductionsingrey-mattervolume
AttributionLicense,whichpermitsunrestricteduse,distribution,andreproductionin
intemporalandperi-Rolandiccortex(Petersonetal.,2000),increases
anymedium,providedtheoriginalauthorandsourcearecredited.
⁎ Correspondingauthorat:DepartmentofPsychosisStudies,InstituteofPsychiatry,de inparietalandfrontalcorticeswhicharepredictedbygestationalage
CrespignyPark,LondonSE58AF,UK.Tel.:[phone];fax:[phone]. atbirth(Kesleretal.,2004);andbilateralincreasesingyrificationin
2213-1582/$–seefrontmatter
http://dx.doi.org/10.1016/j.nicl.2014.01.005
T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365 353
thetemporalcortex(Kesleretal.,2006).Togetherthesefindingssug- stronglywarranted.Here,weapplyspatialindependentcomponent
gestwidespreaddisruptiontoprototypicpatternsofcerebraldevelop- analysis(ICA)(BellandSejnowski,1995)–adata-drivenblind-source
ment.Similarly,inadolescence,intricateanddistributedanomaliesof separationmethodthatidentifiesspatialpatternsonthebasisoftheir
grey-mattervolumeareexhibitedbyVPT-bornindividuals.Inaddition maximalspatialindependence–toresting-statefMRIdatacollectedin
tovolumedecreasesinregionsincludingthefrontal,temporal,insular VPT-born adults and in term-born controls. In line with Menon's
andoccipitalcortices,caudatenucleusandputamen,increasesarealso (2011) unifying triple-network model of psychopathology and the
seeninfrontalandtemporalregions,cerebellumandcingulategyrus established cognitive impairments demonstrated by preterm-born
(Nosartietal.,2008).Thesefindingsarecompatiblewithasystematic adults,thisinvestigationisfocusedonthreeneurocognitiveRSNs:the
andfunctionally-influentialstructuralremodellingofcerebralarchitec- defaultmodenetwork(DMN)(Raichleetal.,2001),whichcomprises
tureinassociationwithprematurebirth(Allinetal.,2010).Consistent themedialprefrontal,posteriorcingulate,precuneusandbilateralangu-
withtheanatomicalliteraturearereportsthatbloodoxygenation-level largyrus,andwhoseactivityiselevatedduringintrospectivetasks
dependent(BOLD)activationresponsesduringlanguage,inhibitory (Gusnardetal.,2001);thecentralexecutivenetwork(CEN),whichin-
control,attentionallocationandassociativelearningimplythatVPT- cludesdorsolateralprefrontalandparietalregions,andisactivatedby
bornadolescentsadoptpossiblycompromisedandalternativefunction- tasksinvolvingexternally-focusedattention,workingmemoryandre-
alpathways(Gimenezetal.,2005;Kalpakidouetal.,2012;Lawrence sponseselection(Corbettaetal.,2002;Hesteretal.,2007;Medaetal.,
etal.,2009;Petersonetal.,2002). 2008);andthesaliencenetwork(SN),whichisfocusedinfrontoinsular
AccordingtoHebbianprinciples,itislikelythattheseobserveddis- cortex,dorsalanteriorcingulatecortex(ACC)andsubcorticalstructures
turbancesofneuronalinfrastructureareatleastinpartfunctionally includingthestriatum(Seeleyetal.,2007),andwhoseactivityisdriven
driven. Sinceboth endogenous and event-related neural activity is bycognitive,emotionalorhomeostaticsalience.
vitalfortherefinementofnervous-systemcircuitry(PennandShatz, Theprincipalaimsofthisworkarethreefold:first,toinvestigate
1999),theemergenceofneuralnetworksislikelytoplayacriticalrole whetherVPTbirthisassociatedwithsystematicchangesinthespatial
incorticaldevelopment;andrelatedalterationisimpliedinthepreterm andspectralcharacteristicsoftheselarge-scaleneurocognitivenet-
brain.Thestudyofresting-statefunctionalconnectivity(RSFC)has worksinearlyadulthood;second,toevaluatewhetherthecausalrela-
demonstratedintrinsicandcoherentBOLDsignalfluctuationsindisso- tionshipsbetweenthetimecoursesofthesenetworksarealteredin
ciable networks involving large-scale, spatially-disparate regions VPT-borncomparedtoterm-bornindividualsusingGrangercausality;
(Damoiseauxetal.,2006).Thefunctionalrelevanceoftheseresting- andthird–inordertotestthepotentialclinicalrelevanceoftheseindi-
statenetworks(RSNs)canbeinferredfromtheirspatialcongruence ces–toexaminetheextenttowhichthesenetwork-focusedmeasures
withnetworksofregionsco-activatedinassociationwithspecificcogni- areassociatedwithfunctionalalterationsobservedinVPTsamples.
tiveandsensorytasks(Biswaletal.,1995;Foxetal.,2005;Hampson
etal.,2002);aswellasinter-regionanatomicalpathways(Greicius 2.Method
etal.,2009;Skudlarskietal.,2008;Vincentetal.,2007).
Animportantdevelopmentwasmarkedbyrecentstudiesoftheon- 2.1.Participants
tologicaldevelopmentofRSNs.Inapioneeringstudy,Franssonetal.
(2007)demonstratedthatRSNsareevidentatterm-equivalentage 29VPT-bornindividuals(b33weeksofgestation)wererecruitedto
(around40weeksofgestation)inVPT-borninfants,albeitwithlesser takepartinthisstudyfromalargecohortofindividualswhohadbeen
complexitythanadultnetworks;term-equivalentnetworksarefewer admittedtotheNeonatalUnitatUniversityCollegeHospital,London
in number, non-lateralised and restricted to hemispherically- between1979and1984,whohavebeenfolloweduptoinvestigate
homologousstructures.Doriaetal.(2010)furtherdemonstratedthat thelong-termconsequencesofprematurebirth.VPT-bornparticipants'
evencomplexRSNs,suchastheexecutivecontrolnetwork,arealso neonatalultrasound(US)scanswereclassifiedas:normal(n=16);
presentatterm,andbeforetheacquisitionofrelatedcognitivefunctions exhibiting uncomplicated periventricular haemorrhage (n = 4);
laterinchildhood.Supekaretal.(2009)complementarilyobservedthat orexhibitingperiventricularhaemorrhageandventriculardilatation
thehierarchicalorganisationofRSNsincreasesfromchildhoodtoyoung (n=9).PleaserefertoNosartietal.(2008)forfurtherdetailsonUS
adulthood;andsuggestedthatwhilsthighly-hierarchicalbrainsafford classification.
benefitsintermsofmaximisingrapidtop-downcommunicationand 23age-matchedterm-bornindividuals(37–42weeksofgestation)
minimisingwiringcostsinadults,theless-hierarchicalarrangementin were also assessed. The inclusion criterion for this group was full-
childrenmayservetoprotectfromthepotentialofpervasiveeffects termbirth(37–42completedweeksofgestation);exclusioncriteria
resultingfromdamagetohubs.Together,thesefindingsillustratethe werebirthcomplications(forexample,lowbirthweightdefinedas
non-stationarityofRSNsthroughdevelopment. b2500 g, endotracheal mechanical ventilation, prolongedgestation
ThereisdirectevidencethatpretermbirthisassociatedwithRSN (N42weeks)).
abnormality.Smyseretal.(2011)investigatedRSFCinVPT-andterm- Exclusioncriteriaforallparticipantswereseverehearingandmotor
borninfantsusingseed-regionconnectivityandacrosswide-ranging impairment,oranymentalretardation.Twoparticipantsfromeach
seedsfoundconsistentlylowercorrelations,amorelimiteddistribution grouphadapersonalhistoryofpsychiatricillness(VPT-borngroup:bi-
ofsignificantcorrelationanddecreasedlong-rangefunctionalconnec- polardisorder(n=1);borderlinepersonalitydisorder(n=1);term-
tivityinVPT-borninfants.Furthermore,longitudinalinvestigationin born group: depression (n = 2)). One member of the term-born
the same group revealed age-dependent increases in correlation group was prescribed an antidepressant medication (sertraline) at
strength.ThesefindingsdemonstratetheadvantageouseffectsofRSN timeofstudy.AllparticipantswereEnglishnativespeakers.Allpartici-
developmentinthesheltereduterineenvironmentduringthefinalges- pantsgaveinformedwrittenconsentandwerereimbursedfortravel
tationalweeks;andareconsistentwithstructuralobservationsthat expensesandreceivedanominalremunerationforparticipationinthe
therearesignificantincreasesinlong-rangecortico-corticalconnectivi- study.ThestudywasgivenethicalapprovalbytheSouthLondonand
tyandnoteworthyresolutionofdistinctcorticallaminainthisperiod MaudsleyResearchandEthicsCommitteeandbythePsychiatry,Nurs-
(KostovicandJovanov-Milosevic,2006). ingandMidwiferyResearchEthicsSubcommittee(PNMRESC),King's
SincelargecohortsofVPT-bornindividualsarenowreachingadult- CollegeLondon.
hood,inlightofsubstantialimprovementsinneonatalcareoverthe Intelligencequotient(IQ)wasassessedforeachindividualbyarater
pastthreedecades,ithasbecomemorefeasibletoinvestigatethelast- blindtogroupmembershipusingtheWechslerAbbreviatedScaleofIn-
ingeffectsofVPT-birthontheadultbrain.Furthermore,inviewofthe telligence (WASI) (Wechsler, 1999); and the Hayling sentence
observedeffectsoninfantbrains,studyofRSFCinVPT-bornadultsis completion test (Burgess and Shallice, 1997) was used to measure
354 T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365
executivefunction.Table1displayssamplecharacteristicsforboth 2.4.Between-groupdifferencesinheadmotion
studygroups.Chi-squaredtestingdemonstratedthatthegroupsdid
notdifferbysex.IndependentsampleT-testsshowednon-significant Inlightofrecentreportsthatheadmotioncaninfluencepatternsof
between-group differences in age and IQ, although the VPT-born intrinsicfunctionalconnectivityasmeasuredbyfMRI(Poweretal.,
group exhibited significantly reduced executive function scores 2012),possiblesystematicbetween-groupdifferencesinheadmotion
(Haylingscaledvalues)ascomparedwiththeterm-bornindividuals inthecurrentdatasetwerecalculatedusingthetechniquessetoutby
(T(46)=2.12,P=.032).WithintheVPT-borngroup,aSpearman's vanDijketal.(2012),whichexaminethemovementparameterscalcu-
ranktestperformedtoinvestigatetherelationshipbetweengestational latedduringrealignment.Fourmetricswerecalculatedforeachstudy
age(GA)atbirthandneonatalUSclassification(codedasnormal(0), participant:meanmotion;maximummotion;numberofmovements;
uncomplicatedperiventricularhaemorrhage(1),andperiventricular androtation.Meanmotionrepresentedthemeanabsolutedisplace-
haemorrhage and ventricular dilatation (2)) revealed a significant mentofeachbrainvolumeascomparedwithitsprecedingvolume
negativerelationshipbetweentheseperinatalriskfactors(Spearman's andusedthetranslationparametersinthex,yandzplanes.Displace-
ρ=−.545,P=.002). mentwascomputedastherootmeansquareofthesevalues.Maximum
motionwasthegreatestdisplacementvalueacrossthedataset.Number
2.2.fMRIdatacollection ofmovementswascalculatedbyevaluatingthenumberofbrainvol-
umesinwhichdisplacementexceeded0.1mm,andwasthusbounded
256gradient-echoecho-planar images(TR/TE:2000/30ms, flip betweenaminimumvalueof0andamaximumvalueofn−1wheren
angle:75°,matrix:64×64)wereacquiredona3TeslaGESignaMR was the number of brain volumes in the time series. Rotation was
scanner(GEHealthcare,USA)attheMaudsleyHospital,London.Each calculatedaccordingtoEuler'srotationtheoremthatexpressesany
whole-brain image contained 37 non-contiguous slices of 2.4-mm three-dimensionalrotationasasingleangleandcorrespondingaxisof
thicknessseparatedbyadistanceof1mm,andwithin-planeisotropic rotation,andcomputedforeachbrainvolumeincomparisonwithits
voxelresolutionof3.4mm.Participantswereinstructedtoremainstill precedingvolume.TheEuleranglewascomputedusingthefollowing
withgazefixedonacentralcrossforthedurationofthisnine-minute formula:arccos((cos(phi)cos(theta)+cos(phi)cos(psi)+cos(theta)
resting-statescan. cos(psi)+sin(phi)sin(psi)sin(theta)−1)/2),wherephi,theta,and
psi are the respective rotational parameters around the x, y and z
2.3.fMRIpreprocessingandindependentcomponentanalysis axes.Between-groupdifferencesinthesemetricswereassessedusing
independentsampleT-testson1000bootstrappedsamples,withthe
fMRIdatawerepreprocessedusingSPM8(StatisticalParametric α-threshold for the bootstrapped P-value Bonferroni corrected to
Mapping,WellcomeDepartmentofImagingNeuroscience,University .0125toreflecttheexaminationofthefourhead-motioncharacteristics.
ofLondon,UK).Datawerespatiallyrealignedtothefirstimageofthese- Theapproachofcombiningleave-one-outpermutationwithBonferroni
ries,time-correctedtothefirstsliceofeachimage,normalisedtoastan- correction of the significance threshold was used where feasible
dardtemplateoftheMontrealNeurologicalInstitute(MNI)brain,and throughoutthisworkasameansofenhancingthedependabilityofre-
smoothedusingan8-mmfull-widthathalf-maximumGaussiankernel. portedfindings.
SpatialICAwasperformedonthepreprocesseddatausingtheGroup
ICAfMRIToolbox(GIFT;http://icatb.sourceforge.net)withinMATLAB
7.8(MathWorks,USA).GIFTusesatemporalconcatenationapproach, 2.5.Componentselection
duringwhichdatareductionisperformedviamulti-stagedprincipal
component analysis and aggregation to generate common-group Agoodness-of-fit(GOF)procedurewasusedtoassessthespatial
maps.Thesecomponentsarethenback-projectedontoeachindividual's correspondence between each of the 44 whole-sample component
datatocreatesubject-specificspatialmapswithcorrespondingtime mapsandbinarymasksconstructedfrompreviousworktocharacterise
courses(Calhounetal.,2008;Calhounetal.,2009). featuresofthethreenetworksuponwhichthisworkisfocused(http://
PriortoICA,datadimensionalitywasestimated(usingMinimum findlab.stanford.edu/research;Shireretal.,2012).Eachnetworkwas
DescriptionLengthcriteria)tobe44.Sincemodel-orderdetermines splitintotwosubnetworkstopermitinvestigationofregion-specific
networkspatialcharacteristicsincludingsubnetworkparcellation,ICA phenomena.ThesemasksaredepictedinFig.1.Thesaliencenetwork
wasconstrainedtoproduce44components.ICAwasperformedusing maskscomprised:(i)bilateralanteriorinsulaanddorsalanteriorcingu-
the Infomax algorithm (Bell and Sejnowski, 1995), and repeated 5 latecortex;and(ii)basalgangliaregions.Itisreasonabletodissociate
timeswithIcasso(Himbergetal.,2004)tomaximisethestabilityof thelatteronaccountofthefundamentalrolethatthestriatumisbe-
thederivedcomponents.Componentswerealsoscaledaccordingto lievedtoplayinsalienceassignmentandtheinclusionofstriataland
percentsignalchangetofacilitateinter-subjectcomparisonsoftheir thalamicstructuresintheoriginaldescriptionofthisnetwork(Seeley
timecourses.Back-reconstructionwascarriedoutusingGICA3onthe etal.,2007).TheCENmaskssplittheCENintoitsleftandrighthemi-
basisofpreviousempiricalsupportfortheaccuracyofthismethod sphericnodes(asperpreviousfMRIinvestigationsofthisnetwork;
(Erhardtetal.,2011). Sridharanetal.,2008;Whiteetal.,2010).TheDMNmasksincluded:
(i)theanteriormedialprefrontalcortexnodeoftheDMN;and(ii)the
posterior midline structures of the posterior cingulate gyrus and
Table1
Samplecharacteristics.Frequenciesandmeanvalueswithbracketedstandarddeviations. precuneus.Again,thesesub-featureswereinvestigatedindependently
inthecurrentstudyonthebasisoftheirdivisioninpreviousICAinves-
Term-borngroup VPT-borngroup
tigations(forexample,Sridharanetal.,2008;Whiteetal.,2010).GOF
(N=23) (N=29)
scoreswerecalculatedbyZ-scoringeachcomponentmapandthen
Demographicinformation
subtracting the mean voxel value outside a binary mask from the
Ageattesting(years) 27.55(2.22) 28.59(2.04)
Sex(males/females) 10/13 12/17 meanvoxelvaluewithinit(Seeleyetal.,2007;Whiteetal.,2013).Com-
ponentselectionwasperformedaccordingtorankedGOFscores.
Cognitiveassessment
OnesampleT-testswereconductedonback-reconstructedimages
Full-scaleIQ 116.00(12.20) 106.25(11.59)
VerbalIQ 114.70(12.16) 105.90(12.74) ofcomponentloadingsforeachcomponentofinteresttoassesscompo-
PerformanceIQ 114.70(13.06) 105.35(12.45) nentspatialrobustnesswithinthesampleasawhole.Clustersofsignif-
Executivefunction(Haylingscaled) 6.60(1.05) 5.78(1.40) icantpositivitywereascribedforeachcomponentusingafamily-wise
VPT,verypreterm;IQ,intelligencequotient. errorcorrectedP-thresholdof.05onthebasisofthenumberandspatial
T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365 355
Fig.1.Rankedgoodness-of-fit(GOF)scoresforeachcomponentwithpre-specifiedfunctionally-derivedmasks,for:(A)insularsaliencenetwork;(B)striatalsaliencenetwork;(C)left
centralexecutivenetwork;(D)rightcentralexecutivenetwork;(E)frontaldefaultmodenetwork;and(F)posteriordefaultmodenetwork.Ineachsub-figure,thecomponentchosen
forsubsequentextendedanalysisisdepictedbyablackcircleandallothercomponentsarerepresentedbywhitecircles.Theinsetsdepictthebinarymaskforeachnetworkinyellow
overlaidonsectionsofastandardisedT1-weightedimage.
extent of significant voxels at an uncorrected P-threshold of .001 evaluatedacrossthewholebrain.Asinthepreviousanalysis,thesignif-
(Fristonetal.,1994). icanceofbetween-groupdifferenceswasascribedaccordingtoavoxel-
levelinclusionofPb.001uncorrectedandacluster-levelsignificanceof
Pb.05family-wiseerrorcorrected(Fristonetal.,1994).
2.6. Spatial differences between VPT-born and term-born group
components
2.7.Assessingbetween-groupdifferencesincomponentpowerspectra
To assess between-group differences in the whole-brain spatial
focus,extentandamplitudeofeachcomponent,twosampleT-tests Toinvestigatebetween-groupdifferencesinthespectralcharacter-
wereconductedontheback-reconstructedcomponentloadingmaps. isticsofactivity,powerspectraldensity(PSD)wasevaluatedforthe
Effects were investigated in all six components independently and individual-specifictimecoursesforthesixcomponentsofinterest.
356 T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365
MeanPSDwasthencalculatedbetween0.05and0.15Hzinlinewith andtheelementa(l)(i=1,…,k;j=1,…,k)isthecausalitycoeffi-
ij
subsequentGranger-causalitymeasures.Foreachcomponentapermu- cientfromthetimeseriesy totheseriesy .Thevectorε hasaco-
jt it t
tationanalysiswasconductedinwhichbetween-groupeffectswere variancematrixgivenby
testedusingmultipleKruskal–Wallistestsandaleave-one-outstrategy 2 3
forallpossiblecombinationsofparticipants,withbetween-groupef- σ2 σ2 ⋯ σ2
6 11 12 1k7
fectsjudgedsignificantinthisanalysisifreliableinmorethan99.17% X 6σ2 σ2 ⋯ σ2 7
ofpermutations.ThisthresholdrepresentsBonferroni-correctionof ¼666σ221 σ222 ⋯ σ22k777: ð4Þ
theconventional95%thresholdonaccountofthesixtestsperformed. 4 ⋮31 ⋮32 ⋱ ⋮3k5
Kruskal–Wallis tests were used as a non-parametric alternative σ2 σ2 ⋯ σ2
k1 k2 kk
hereandinallotheranalyseswhereskewnessandkurtosisoftherele-
vantdatafeaturesprecludedtheassumptionofnormality.Cut-offsof2 Undertheassumptionthatε iszero,usingallinformationuntiltime
t
and7wereusedforskewnessandkurtosisrespectively(Curranetal., (t−1),Y canbepredictedby
t
1996).
Xp
2.8.Granger-causalityanalysisandclassification Y^t¼vþ AlYt−1: ð5Þ
l¼1
Causalrelationshipsbetweentimeseriescanbeinferredfromtheir
Inthisvectorautoregressivemodel,y canbeconsideredtoGranger-
jt
temporalrelationships(Granger,1969).Ifpastvaluesofonetimeseries causey ifthecoefficienta(l)isnonzeroforaparticularvalueofl.
it ij
improvepredictionsofcurrentandfuturevaluesofanother,thereis
GPDCisthefrequencydomainrepresentationofGrangercausality
Grangercausality(GC)fromthefirsttimeseriestothesecond.Itisim-
(BaccalaandSameshima,2001;Takahashietal.,2010),whichincorpo-
portanttonotethatsignificantcausalityinonedirectiondoesnotassure
ratesvariancestabilisationwiththeeffectofimprovingthedecision
causalityintheother. errorrateofclassifiedcausalinfluence(Baccalaetal.,2006;Satoetal.,
WhilstGCisprimarilyatime-domainconcept(Goebeletal.,2003), 2009).GPDCfromtimeseriesjtotimeseriesiatfrequencyλisdefined
assessingconnectivityinthefrequencydomainhasseveraladvantages:
by:
physiologicalandnon-physiologicalnoisefactorshavecharacteristic
spectralcharacteristicswhoseeffectscanbeminimised;andendog- 1
a ðλÞ
enous BOLD oscillations are maximal within distinct frequency ij σ
bands(Cordesetal.,2000).Withthisinmind,thetimeseriesofall πijðλÞ¼sffiXffiffiffiffiffiffiffiffiffiffiffiffi(cid:3)ffiffiffiffiffiffiffiffiffiffiiffiffiffi(cid:3)ffiffiffiffiffiffiffiffiffiffi ð6Þ
six components of interest were bandpassed between 0.05 and k (cid:3)(cid:3)a ðλÞ(cid:3)(cid:3)2 1
i¼1 ij σ2
0.15Hz. i
Generalisedpartialdirectedcoherence(GPDC)isafrequency-based
where
measureoffunctionalconnectivity,whichhasrecentlybeenappliedto
fMRIdata(Havliceketal.,2010;Satoetal.,2009;Shimetal.,2013; (cid:4) (cid:5)
Silfverhuthetal.,2011).GPDCcanevaluatethedirectionofinformation a ðλÞ¼δ −Xp aðlÞexp − 2pπffiffiffi ð7Þ
flowsimilarlytoalternativefrequency-basedapproaches,suchasthe ij ij l¼1 ij λ 1
directedtransferfunctionandrelativepowercontribution,andisequiv-
alenttothesemeasuresinbivariatecases,butholdsasignificantadvan- forδ =1ifi=jand0otherwise.ThesquaremodulusoftheGPDC
ij
tage when assessing multivariate data, as it can discern directand valuefromthej-thtimeseriesbythei-thseriescanbeunderstoodas
indirectinfluencesoncausalrelationshipsbypartiallingouttheeffects theproportionofthepowerspectraofthej-thtimeseriessenttothe
ofadditionaltimeseries(Satoetal.,2009). i-th series after accounting for the effects of the other series (Sato
TheformaliseddescriptionoftheGPDCapproachbelowfollows et al., 2009). GPDC is bounded between 0 and 1, with the former
the elegant previous conceptualisations laid out elsewhere representinganabsenceoffunctionalconnectivityfromthej-thtime
(Baccala and Sameshima, 2001; Sato et al., 2009). Inferences of seriestothei-thtimeseriesandthelatterindicatingstrongconnectivity
Grangercausalityaremadeusingvectorautoregressivemodelling. betweenthesesignals.
SupposingY isamultidimensionaltimeseriesmadeupofksignals Here,GPDCwasevaluatedbetweenall-pairwisecomponenttime
t
suchthat coursesusingtheFunctionalNetworkConnectivitytoolbox(FNC;ver-
sion2.3;http://mialab.mrn.org/software/fnc/).Crucialtotheaccuracy
⌊ ⌋
y ofmultivariateautoregressiveprocedures,suchasGCanalysisusing
1t
y GPDC,isthegenerationofthetransfermatrixtowhichthedataisfitted
Y ¼ 2t ; t¼1;2;…T ð1Þ
t ⋮ (Bianchietal.,2013).Inlightofrecentdemonstrationsofadvantagesfor
ykt BayesianInformationCriterion(BIC)overAkaikeInformationCriterion
(AIC)approaches(Hemmelmannetal.,2009;Porcaroetal.,2009)due
toorder-overestimationinAICmodels,modelorderwasestimatedin
thevectorautoregressivemodelcanbecalculatedby
thecurrentanalysisusingagroup-levelBICestimate.
X
This analysis produced bi-directional indices of GPDC for each
Yt¼vþ pt¼1AlYt−1þεt ð2Þ pairwisecombinationofcomponentsin0.002Hzfrequencybinsbe-
tween0.05and0.15Hz,withthisbinresolutiondeterminedbyFNC
inwhichvisavectorofconstantsandεtisavectorofrandomdistur- toolboxalgorithmsonaccountoftherepetitiontimeofthefMRIdata
bances.ThematricesAl(l=1,…,p)inturnrepresent andtheextentofthefrequencybandofinterest.Decomposingthe
datainthiswaywasdonetoaffordgreaterspectralspecificitytothere-
2aðlÞ aðlÞ ⋯ aðlÞ3 sultsinlightofpreviousobservationsthatfMRIfunctionalconnectivity
6 11 12 1k7 varies across frequencies (Salvador et al., 2005). To investigate
Al¼6666aað2ðll1ÞÞ aað2ðll2ÞÞ ⋯⋯ aað2ðllkÞÞ7777 ð3Þ bmeutwltiedeimn-egnrsoiuonpadlisfefetroefnicnetesridnepGePnDdCenincieasm,saenvenrearlwanhailcyhseustwiliesreedctohnis-
4 31 32 3k5
⋮ ⋮ ⋱ ⋮ ducted.Inaninitialexploratoryanalysis,Kruskal–Wallistestswerecon-
aðkl1Þ aðkl2Þ ⋯ aðklkÞ ductedusinga95%thresholdtoassessbetween-groupdifferencesin
T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365 357
GPDCmagnitudeforeachpairwisecombinationofcomponentswithin pathsastheindependentvariable.Next,toinvestigatetheputativerela-
eachfrequencybin.Aspectral-clusteringcriterionwasusedtominimise tionshipbetweenthesevariablesinVPT-bornindividuals,acorrespond-
Type-1errors,wherebybetween-groupdifferenceswerereportedas inglinearregressionanalysiswasconductedinthisgroup.Finally,to
significantonlywhendifferencesatthe95%thresholdwereobserved testthehypothesisthatperinatalfactorsplayamodulatoryroleon
forfiveneighbouringfrequencybins. thestrengthofthisrelationship,themoderatingeffectsofGAandneo-
ToassesswhetheroverallGPDC-connectionstrengthdistribution natalUSclassificationonthepathwayfromGPDCtoexecutivefunction
differedbystudygroup,thepercentageofconnectionswascalculated wereassessedusingthePROCESSmacroexpansion(Hayes,2013)for
asafunctionofGPDCmagnitudeforVPT-bornindividualsandcontrols SPSS(SPSSInc.,USA).ThereportedcollinearitybetweenGAandUSclas-
independently.Onthebasisoftheresultanthistograms(seeResults), sificationprecludedconcomitantinvestigationoftheireffectsusingre-
connectionswerecategorisedinthreeconnectivitywindows(CWs) gression. Instead, two independent moderation analyses were
as:lowstrength(0.05bGPDCb0.20); mid-strength(0.20 bGPDC therefore performed. Moderation analysis within PROCESS utilises
b0.35);orhighstrength(GPDCN0.35).Between-groupdifferences ordinary-leastsquaresregressionandpermitstheflexibleevaluation
werethenevaluatedusingaleave-one-outpermutationapproachby ofdiversestatisticalinterdependencies.Here,thepredictiveeffectson
whichmultipleKruskal–Wallistestswereusedtoassessdifferencesin executivefunctionof:(i)GPDC(theindependentvariable);(ii)USclas-
percentageofconnectionsandmeanGPDCateachconnectionstrength. sificationorGA(themoderatingvariable);and(iii)theinteractionof
Effectsreliableinmorethan98.3%ofpermutationswereconsideredsig- GPDCandtheperinatalmoderatorweremodelled.Hypothesistesting
nificant;thisthresholdsignifiesBonferronicorrectionoftheconven- wasperformedonthebasisofeachregressioncoefficientanditsstan-
tional 95% threshold, to take into account the three windows of darderrorcalculatedover1000randomlybootstrappedsamples,with
investigation. significanceascribedusingpercentile-basedbootstrapconfidenceinter-
Toinvestigatebetween-groupdifferencesinthetopologyofthenet- valsandrelatedα-levelsBonferronicorrectedto.025toreflectthe
workofneurocognitivenetworks,lineardiscriminantanalysis(LDA) numberofperinatalfactorsinvestigated.
wasconducted.LDAcanbeusedtoidentifyfeaturesetswhichpermit (Inanattempttoelucidatetheprimacyoftheseperinatalfactorson
categoricaldissociationbydefiningalinearsummaryofmultidimen- GPDC-derivednetworktopologyaseriesofcomplementaryLDAinves-
sionaldata,andhasbeenusedpreviouslytoclassifyindividualsonthe tigationswereconductedwiththerationalethatthefactorshowntobe
basisoftraumaticbraininjury(Castellanosetal.,2010).Here,LDA associatedwiththemostrobustdeviationfromthenormativedata
wasconductedonbinarisedmatricesdenotingconnectiontopologyac- could be reasonably adjudged to exert the most deleterious effect.
cordingtothresholdedGPDC.Eachmatrixfeaturedbi-directionalcon- TheseanalysesarepresentedasAppendixB.)
nections between all components. Accuracy of this procedure at
classifyingindividualsbystudygroupwasassessedatvaryingGPDC
3.Results
thresholds, with the rationale that the identification of a GPDC
thresholdwhichsuccessfullyclassifiesindividualsbygroupdenotesa
3.1.Headmovement
clinicallyusefulmarkerofVPT-birthassociateddysfunction.Correct
classificationonthebasisofFisherdiscriminantvaluecomparedto
Table2presentsthegroup-averagedhead-movementcharacteris-
theprojectedhyperplanewascalculatedforeachparticipant;andvali- tics.Thedegreeofheadmovementdidnotsignificantlydifferbetween
datedusingaleave-one-outstrategyforallpossibleparticipantpermu-
thestudygroupsinanycalculatedmetric.
tations. Discriminant accuracy is defined here as the product of
population accuracy and Mahalonobis distance (Castellanos et al.,
3.2.Components
2010).Theprojectionweightscanbereasonablyinterpretedastherel-
ativecontributionofeachpathtothediscriminantfunctionthatbest
Fig.1illustratesGOFscoresforall44componentswitheachofthe
separates the groups (Ecker et al., 2010), permitting us to identify
sixsubnetworksofinterest.Forfiveofthese,thecomponentwiththe
pathsparticularlycontributingtothediscriminationofthephysiology
greatestGOFscorewasselectedforsubsequentconnectivityanalysis.
of term-born and VPT-born individuals. However, it should be
Visualinspectionofthebest-fitstriatalsaliencenetworkcomponent
emphasisedthatduetothemultivariatecharacteristicofclassifying
suggestedthatthiscomponentwasprimarilyfocusedinventricular
methods,eachpathintheprojectingspaceshouldbeinterpretedin
spaces.ThiswasconfirmedbyanadditionalGOFevaluation,whereby
thecontextoftheentirediscriminatingpatternandshouldnotbecon-
thiscomponentwasfoundtobethebest-fitcomponentwithacerebro-
sideredinisolation(Eckeretal.,2010).
spinalfluidmask.(PleaserefertoAppendixCforfurtherdetails.)The
TocomplementthecurrentGCanalyses,afunctionalconnectivity
componentwiththesecond-highestGOFscorewiththestriatalsalience
analysiswasadditionallyperformedbymeansofzero-laggedcorrela-
networkmaskwasthereforeselectedforconnectivityanalysis.
tionbetweenthetimecoursesofactivityoftheneurocognitivenet-
Thesixcomponentsidentifiedonthebasisoftheirspatialconcor-
worksandispresentedasAppendixA.
dancewithwell-studiedlarge-scaledistributednetworksthereforein-
cluded:(A)theinsularportionofthesaliencenetwork,focusedon
2.9.GPDC,executivefunctionandperinatalriskfactors
bilateralinferiorfrontalgyrusandinsula;(B)thestriatalcomponent
ofthesaliencenetwork,focusedontheputamenbilaterally;(C)the
Ithasrecentlybeenobservedthatpreterm-birthrelatedimpair-
left-hemisphericconstituentsoftheCEN,withstrongloadingsinleft-
ments in executive function preferentially persist into adulthood
hemisphericinferiorparietallobuleanddorsolateralprefrontalcortex
(Allinetal.,2011;Nosartietal.,2007);andthatimpairmentsinexecu-
tivefunctionresultfromdisruptionofthecoordinationofactivityacross
Table2
large-scalefunctionalbrainnetworks(Repovsetal.,2011).Withthe
Group-averagedcharacteristicsofheadmotion.Standarddeviationsaregiveninbrackets.
motivationoffurtherexaminingtheputativerelationshipbetweenex-
ecutivefunctionandnetworkcoordination,whilstassessingthemoder- Metric Group
atingeffectsoffactorsassociatedwithpretermbirth,severaladditional Term-bornindividuals VPT-bornindividuals
analyseswereconducted.
Meanmotion 0.060(0.025) 0.078(0.034)
First,toevaluatewhethernetworkcoordinationpredictedexecutive Maximummotion 0.341(0.245) 0.474(0.434)
functionaftertimelyanduncomplicatedbirth,alinearregressionanal- Numberofmovements 34.56(37.38) 60.16(46.85)
ysiswasperformedfortheterm-borngroupdatausingexecutivefunc- Rotation 8×10−3(5×10−3) 0.001(3×10−3)
tionasthedependentvariableandmeanGPDCinLDA-determined VPT,very-preterm.
358 T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365
(PFC);(D)theright-hemisphericCEN,withstrongloadingsinright-
hemisphericdorsolateralparietalandfrontalregions;(E)thefrontal
portionoftheDMN,withmaximalloadingsinmedialandsuperior
PFC;and(F)theposteriorportiionoftheDMN,focusedonposteriorcin-
gulate gyrus and bilateral angular gyrus. These components are
depictedinFig.2andtheirgrey-matterfocipresentedinTable3.
3.3. Spatial differences between VPT-born and term-born group
components
Analysisofbetween-groupwithin-componenteffectsrevealedthat
fortherightCENcomponentonlyaregionoftherightanteriorinsula
exhibitedsignificantlyreducedloadingsintheVPT-bornindividuals
comparedtocontrols.CharacteristicsofthiseffectareshowninTable4.
3.4. Spectral differences between VPT-born and term-born group
components
Broadbandpower(between0.05and0.15Hz)foreachcomponent
isdisplayedinFig.3.Meanpowerwasnotsignificantlydifferentbe-
tweengroupsforanycomponent.
3.5.Generalisedpartialdirectedcoherence
Between-groupcomparisonsofGPDCwithineachfrequencybinand
for each pairwise combination of components revealed path- and
frequency-specificreductionsinGPDCintheVPT-borncomparedto
theterm-bornindividuals.ThesefindingsarepresentedinTable5.No
significantincreasesinGPDCwereobservedintheVPT-borngroup
comparedtotheterm-borncontrols.
Fig. 4 presents GPDC distributions for both study groups. Low-
strengthconnectionswerethemostnumerousinbothgroups(VPT:
43.4%;term-born:43.6%);andnon-significantbetween-groupdiffer-
enceswerenotedinthepercentageoflow-ormid-strengthconnec-
tions or the mean GPDC within these windows. By contrast, high-
strengthconnectionswererobustlymorenumerousintheterm-born
group as compared with the VPT-born group (VPT: 22.02%; term-
born:26.2%,P=.0016).Furthermore,meanGPDCwithinthehigh-
strengthconnectionwindowwassignificantlygreaterinterm-born
as compared with VPT-born individuals (VPT: 0.43 ± 0.05 (mean
GPDC±standarddeviation);term-born:0.45±0.05,P=0.0097).
3.6.Lineardiscriminantanalysis
Fig.5presentstheresultsoftheLDAperformedtoassessgroup-
classificationaccuracyonthebasisofnetworktopologyasafunction
ofweightingthreshold.Classificationaccuracywascriticallyimproved
byvaryingtheGPDCthreshold,reachingamaximumaccuracy(86%)
at0.35(Fig.5A);here,allcontrolparticipantsarecorrectlyclassified
andonly4of29VPT-bornindividualsaremisclassified(Fig.5B).LDA
also provides path-specific weights for the discrimination. Fig. 5C
showstheweightsforall30pathsfortheLDAusingaGPDC-threshold
of 0.35. These weights permit identification of the most pertinent
Fig.2.Componentsofinterest,showing(A)insularsaliencenetwork(B)striatalsalience
paths for the topological discrimination of VPT- and term-born
network;(C)leftcentralexecutivenetwork;(D)rightcentralexecutivenetwork;(E)fron-
individuals.
taldefaultmodenetwork;and(F)posteriordefaultmodenetwork.Resultsdepictedrep-
Fig.6showsthepath-specificweightsassociatedwiththemostac- resentclusterswithsignificantpositiveloadings(Pb.05,family-wiseerrorcorrected)on
curateclassification,andsuggeststhatthepathfromstriatalSNtopos- thebasisofone-sampleT-testsincludingallstudyparticipants.Resultsareoverlaidon
teriorDMNwasparticularlyusefulindifferentiatingthegroups.Also standardisedT1-weightedimageandscaledaccordingtotheT-valuecolour-barshown.
showntobeimportantwerereciprocalconnectionsbetweenfrontal
DMNandrightCEN. predictedbyUSclassification(β=−1.70±0.69,T=−2.49,P=.020)
andfurtherthattherewasasignificantinteractiononexecutivefunction
3.7.TherelationshipbetweenGPDCandexecutivefunction betweenGPDCandUSclassification(β=6.72±2.74,T=2.45,P=
.023). Similarly, in the analysis assessing the moderating effects of
ExecutivefunctionwasnotsignificantlypredictedbyGPDCineither GA,GAwasobservedtosignificantlypredictexecutivefunction(β=
studygroupconsideredasawhole.However,themoderationanalysis 0.67±0.23,T=3.00,P=.007)andaninteractiononexecutivefunction
revealedthatintheVPT-borngroupexecutivefunctionwassignificantly wasobservedbetweenGAandGPDC(β=−2.80±0.97,T=−2.88,
T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365 359
Table3
Grey-matterfociforthesixcomponentsofinterest.Reportedvoxelsarewithinclusterssignificantatafamily-wiseerrorcorrectedP-thresholdof.05.
Component Location(BrodmannArea) Talairachcoordinates PeakvoxelT-statistic
x y z
InsularSN Inferiorfrontalgyrus(13) −32 14 −12 6.10
Inferiorfrontalgyrus(13) −40 26 10 5.73
Inferiorfrontalgyrus(13) −44 26 −2 5.57
Insula(13) −40 6 2 5.38
Insula(13) 44 16 −2 4.90
Superiortemporalgyrus(38) 48 14 −10 5.77
Anteriorcingulategyrus(24) 0 22 36 5.40
StriatalSN Putamen 32 −14 6 7.31
Putamen −24 22 −10 7.18
Insula(13) −46 12 0 5.33
Middleoccipitalgyrus(37) 56 −70 0 5.73
Fusiformgyrus(27) 38 −56 −16 5.52
LeftCEN Inferiorparietallobule(40) −46 −52 46 6.41
Inferiorparietallobule(40) 40 −60 54 6.56
Superiorfrontalgyrus(8) −34 16 56 6.13
Superiorfrontalgyrus(8) 34 22 54 5.35
Middlefrontalgyrus(10) −38 46 −6 6.26
Middlefrontalgyrus(10) 40 22 52 5.33
Medialfrontalgyrus(8) 2 40 48 5.33
Cingulategyrus(31) −2 −28 44 6.61
Cingulategyrus(31) −4 −46 38 5.40
Inferiortemporalgyrus(20) −62 −28 −20 5.34
Middletemporalgyrus(21) 50 −64 26 5.34
Cerebellum:posteriorlobe 32 −68 −38 5.69
Cerebellum:anteriorlobe 16 −56 −30 5.59
Cerebellum:posteriorlobe 20 −82 −32 5.49
RightCEN Superiorparietallobule(7) 32 −66 58 7.23
Middlefrontalgyrus(8) 36 26 52 7.17
Inferiorfrontalgyrus(9) 56 6 30 5.33
Precuneus(7) 6 −54 40 5.80
Parahippocampalgyrus(36) 14 −36 2 5.41
Middletemporalgyrus(21) 62 −50 −10 5.85
Middletemporalgyrus(21) −58 −2 −10 5.83
Cerebellum:posteriorlobe −10 −84 −30 6.36
Cerebellum:posteriorlobe −30 −64 −40 5.82
FrontalDMN Superiorfrontalgyrus(8) 8 54 44 6.78
Superiorfrontalgyrus(9) 28 48 38 5.48
Medialfrontalgyrus(9) −2 46 32 6.37
Medialfrontalgyrus(9) 10 58 20 6.23
Middlefrontalgyrus(10) −26 58 26 5.54
Anteriorcingulategyrus(32) 4 32 22 5.39
Orbitalgyrus(11) 0 44 −20 5.50
PosteriorDMN Posteriorcingulategyrus(30) 22 −68 10 5.38
Posteriorcingulategyrus(31) −4 −58 22 6.68
Posteriorcingulategyrus(23) 2 −36 40 5.81
Medialfrontalgyrus(11) 4 56 −10 6.11
Angulargyrus(39) −42 −76 30 6.48
Angulargyrus(39) 40 −70 30 5.95
Parahippocampalgyrus(36) −26 −16 −26 6.68
Parahippocampalgyrus(36) 34 −24 −24 6.81
Precuneus(19) 36 −72 38 6.07
Precuneus(19) −30 −76 42 5.71
Superioroccipitalgyrus(19) −36 −82 30 6.54
Thalamus 10 −10 3 5.52
Middlefrontalgyrus(8) 30 20 48 5.49
SN,saliencenetwork;CEN,centralexecutivenetwork;DMN,defaultmodenetwork.
P=.008).Furthermore,accountingforGAproducedasignificanteffect 4.Discussion
ofGPDConexecutivefunctionintheVPT-borngroup(β=83.45±
28.53,T=2.92,P=.008).StratificationbyGArevealedthatthemost Theprimaryaimofthecurrentstudywastodelineatedifferencesin
robustrelationshipbetweenGPDCandexecutivefunctionwasobserved resting-statefunctionalconnectivitywithinandbetweenthreeimpor-
inthelowGAindividuals(β=8.20±3.40,T=2.41,P=.024). tant and robust neurocognitive brain networks in VPT-born adults
Table4
Between-groupcomponentspatialdifferences.
Component Contrast Location(BrodmannArea) Talairachcoordinates PeakvoxelT-statistic
x y z
RightCEN TermNVPT Insula(13) 48 14 5 4.89
360 T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365
Fig.3.Group-averagedperiodograms,depictingpowerspectrumdensityforthesixcomponentsofinterest.SN,saliencenetwork;CEN,centralexecutivenetwork;DMN,defaultmode
network;VPT,verypreterm.
relativetoterm-borncontrols.Nosignificantbetween-groupdiffer- between-networkratherthanwithin-networkdifference.Thegeneral
ences were noted in broadband component power. Analysis of consistencyincross-groupcomponentfocusindicatesthattheselected
regionally-specificcomponentloadingsrevealednosignificantdiffer- componentswerereliablyevidentinbothgroupsandservesasuseful
encesbetweenVPT-andterm-bornadultsintermsofnetworkspatial validationfortheirsubsequenttargetedevaluation.Furthermore,it
focus for five of thesix components.The sole significant between- demonstratesthatinadulthoodRSNsarenotforthemostpartdistin-
group differencewas observed in the rightCEN componentwhere guishableintheirspatialcharacteristicsbetweenpreterm-andterm-
term-bornindividualsexhibitedsignificantlygreaterloadingsinthe bornindividuals.Thissupportsthefindingsofpreviousinvestigation
rightanteriorinsula.However,sincethisregionwasnotasignificant ofRSNspatialfocusininfancy,wheredifferencesinRSNlocationwere
focusoftheright-CENcomponent,butwasinsteadincludedwiththe smallandfew(Damarajuetal.,2010).
insular-SN component (Table 3), this finding is likely to reflect a Despitethegeneralconsistencyofwithin-networkfocusandspec-
tralpower,severalimportantdifferenceswereobservedinbetween-
Table5 networkconnectivityasindexedbyGPDC.First,viewingallfrequency
Significantpath-andfrequency-specificbetween-groupdifferencesinGPDC. binsforallpairwisecombinationofcomponents,therewasarobustre-
ductioninthenumberofhigh-strength(GPDCN0.35)connectionsin
Pathdirection Band(Hz) Effectdirection Significance
VPT-bornindividuals;thatis,fewernetworkswerestrongpredictors
RightCEN→ 0.1221–0.1298 Term-bornNVPT-born P=.018
offutureactivityinothernetworksintheVPT-borngroupcompared
StriatalSN
StriatalSN→ 0.1124–0.1221 Term-bornNVPT-born P=.030 tocontrols.ThisimpliesafunctionaldisconnectintheadultVPT-born
InsularSN brain.Inotherwords,thesefindingssuggestlessrobustinter-network
InsularSN→ 0.1124–0.1221 Term-bornNVPT-born P=.019 coordinationintheseindividuals.Reducedtemporalanti-correlation
PosteriorDMN betweenDMNandexogenously-directedattentionnetworkshasbeen
InsularSN→ 0.0736–0.0814 Term-bornNVPT-born P=.026
suggestedasausefulindexofcompromisedbrainfunctioninassocia-
FrontalDMN
PosteriorDMN→ 0.0911–0.0988 Term-bornNVPT-born P=.003 tionwithpsychiatricillness(Broydetal.,2009;Buckneretal.,2008;
LeftCEN Whiteetal.,2010)andsleepdeprivation(DeHavasetal.,2012).Our
currentfindingsareconsistentwithVPT-birthrelateddisruptionto
CEN,centralexecutivenetwork;SN,saliencenetwork;DMN,defaultmodenetwork;VPT,
very-preterm. this mechanism. This may provide explanation for the increased
T.P.Whiteetal./NeuroImage:Clinical4(2014)352–365 361
Fig.4.Group-averagedhistogramsshowingthepercentageofconnectionsasafunctionofgeneralisedpartialdirectedcoherence(GPDC).Dottedblacklineshowsterm-groupdistribution
whenhiddenbyVPT-groupresults.Otherverticallinesrepresentboundariesforlow(0.05–0.20),mid(0.20–0.35)andhigh(0.35–0.55)GPDC.Inset:Bardiagramshowsgrand-average
connectionstrengthforeachconnectivitywindow(CW).Asteriskdenotessignificantbetween-groupdifferenceinCW3.
Fig.5.Classificationbywhole-networktopology,usinglineardiscriminantanalysisofbinarisednetworksaccordingtovariablegeneralisedpartialdirectcoherence(GPDC)thresholds,
showing:(A)classificationaccuracyasafunctionofGPDCthreshold;(B)scatterplotofclassificationataGPDCthresholdof0.35;and(C)rankedLDAweightsforeachpathforclassification
ataGPDCthresholdof0.35.