1 Douglas Mental Health Research Institute, McGill University, Montreal, QC, Canada; 2 Northwestern University, Chicago, US; 3 Knight Alzheimer's Disease Research Center, Washington University School of Medicine, St. Louis, US * These authors contributed equally to the work

Prediction of brain age using resting-state functional connectivity reveals accelerated aging in the preclinical phase of autosomal dominant Alzheimer’s disease, irrespectively of amyloid pathology

Usingresting-statefMRIgraphmetrics,wedevelopedamodelthatcanpredictbrainageacrossthewholelifespan.ApplyingthismodeltopredictbrainaginginthecontextofpreclinicalADrevealedthatthepresymptomaticphaseofADADischaracterizedbyacceleratedfunctionalbrainaging.Thisphenomenonisindependentfrom,andmightthereforeprecede,Aβaccumulation.

InindividualsatriskofsporadicAD,neitherAPOE4genotypeorAβburdenwasassociatedwithacceleratedbrainaging.

FurtherstudieswillbeneededtobetterunderstandthefactorsmodulatingacceleratedfunctionalbrainaginginthecontextofpreclinicalAD.

Overlaps exist between the neural systems vulnerable to aging andAlzheimer's disease(AD). It is a matter of debate whether aging and AD progression are independentphenomena.We aimed at developing amodel able to predict brain aging from resting-statefunctionalconnectivity(rsfMRI).Wethenusedthedifferencebetweenthepredictedageand thechronologicalage to testwhetherpresymptomaticautosomaldominantAD(ADAD)mutationcarriershaveprematureaging(DIANcohort).Wealsotestedifthebeta-amyloid(Aβ)status(positiveornegative)contributestothediscrepancybetweentheageestimated from brain functional connectivity and the actual age. We repeated theseanalyses inasymptomatic individualsat riskof sporadicAD,assessingboth theeffectofAPOE4genotypeandAβburden(PREVENT-ADcohort).

Funding

Julie Gonneaud,1* Etienne Vachon-Presseau,1,2* Alexis Baria,1 Alexa Pichet Binette,1 Tammie Benzinger,3 John Morris,3 Randall Bateman,3 John Breitner,1 Judes Poirier,1 and Sylvia Villeneuve1 for the Alzheimer’s Disease Neuroimaging Initiative, the Dominantly Inherited Alzheimer Network and the PREVENT-AD Research Group

NeuralNetworkdevelopment

Modelperformance–BrainAgepredictionagainstactualageacrossdatasets

Background and objectives Brain Aging in the preclinical phase of AD

Summary and conclusions

Brain Age Predictive Model

Increase in the number of features andarchitecture complexity introducedoverfitting(i.e. lowestrootmeansquareerrorinthetrainingsetbuthighesterrorinthevalidationset).Themodelusing10featuresand2layersof 5 and 2 nodes was the one showingthebetter generalizability (i.e. providingthe lowest error in the validation set)andwasselectedforthefinalmodel.

Regression tree ensembles1.00.80.60.40.20.00

Sup

port

vec

tor

mac

hine

1.00

.80

.60

.40

.20

.00

Page 1

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

1617

18

1920

21

2223242526

Rootmeansquareerrorinthedifferentsetsasfunctionofthenumberoffeaturesandnetworkarchitecture

Features(i.e.graphmetrics)rankedbyimportance

Cohorts Trainingset

Validationset

Testingset

DIANmut.noncarriers 105[19-69yo] - 30

[18-61yo]

DIANmutationcarriers - - 128[20-58yo]

PREVENT-AD 36[55-78yo] - 257

[55-84yo]

CamCAN 602[18-87yo] - 100

[18-88yo]

ADNI 30[65-90yo] - 15

[66-94yo]

ICBM - 47[19-79yo] -

Totalsample 773[18-90yo]

47[19-79yo]

530[18-94yo]

Datasets.Samplesize[age-range]

Actual Age (years)100806040200

Pre

dict

ed A

ge (

year

s)

100

80

60

40

20

0

Page 1

Actual Age (years)100806040200

Pre

dict

ed A

ge (

year

s)

100

80

60

40

20

0

Page 1

Trainingset Validationset(ICBM)

1)Sub-graphcentrality2)Positive-weightedclusteringcoefficient3)Binarizedassortativity4)Weightedmodularitycoefficient5)Smallworldness6)Flowcoefficient7)Betweennesscentrality8)Negative-weightedparticipationcoefficient9)Positive-weighteddiversitycoefficient10)Negative-weightedclusteringcoefficient11)Eigenvectorcentrality12)Binarizedefficiency13)Weightedassortativity14)Negative-weightedgatewaycoefficient15)Binarizedclusteringcoefficient16)Resilience17)Binarizedmodularitycoefficient18)Participationcoefficient19)Negative-weighteddiversitycoefficient20)Positive-weightedparticipationcoefficient21)Characteristicpathdiameter22)Characteristicpatheccentricity23)Characteristicpathradius24)Positive-weightedgatewaycoefficient25)Characteristicpathlambda26)Characteristicpathefficiency

DIANmutationnoncarriers

DIANmutationcarriers

Samplesize 30 128

ActualAge 39.07±11.39 34.27±9.56*

PredictedAge 36.92±15.61 42.36±14.43t

DIANmutationnoncarriersAβ-

DIANmutationcarriersAβ-

DIANmutationcarriersAβ+

Samplesize 28 77 40

ActualAge 38.86±11.74 33.01±8.96** 38.20±9.39††

PredictedAge 37.89–15.43 40.83±12.89 45.68±16.30*†

PREVENT-ADAβ-

PREVENT-ADAβ+

Samplesize 51 11

ActualAge 62.82±4.25 65.45±6.22

PredictedAge 68.89±6.07 70.09±3.09

PREVENT-ADAPOE4noncarriers

PREVENT-ADAPOE4carriers

Samplesize 148 108

ActualAge 63.93±5.75 63.05±4.78

PredictedAge 68.96±5.14 68.46±5.01

Authors have no disclosure. Acknowledgments: We would like to thank all the participants, the collaborators from the DIAN and PREVENT-AD studies and all members of the Villeneuve lab who contributed to this study.

contact : julie.gonneaud@gmail.com | poster downloadable @ villeneuvelab.com

Model was built on the training set, optimized based on itsgeneralizability in the validation set and, finally, the model’spredictionswereanalyzedinthetestingset.

AregeneticmutationandAβburdenassociatedwithacceleratedbrainaginginpreclinicalADAD?

AregeneticriskfactorandAβburdenassociatedwithacceleratedbrainaginginindividualsatriskofsporadicAD?

R2=0.53rmse=14.01

R2=0.49rmse=13.84

R2=-.23rmse=16.88

Acutal Age (years)100806040200

Pre

dict

ed A

ge (

year

s)

100

80

60

40

20

0

Page 1

R2=0.36rmse=13.02

Predicted age was overestimated inDIAN mutation carriers while it wasunder-estimatedinnoncarriers.

The overestimation inmutation carrierswas not associatedwithAβstatusorload.

Predicted age was overestimated in thePREVENT-ADwithnodifferencebetweenAPOE4 carriers (the main genetic riskfactorofsporadicAD)andnoncarriers.

Differentfrommutationnoncarriersattp<.10,*p<.05,**p<.01;DifferentfrommutationcarriersAβ-at†p<.05,††p<.01,ns:notsignificant

amyloid_positivity10

dif'

3 0

20

10

0

- 1 0

- 2 0

- 3 0

Page 1

mutation/apoe10

dif'

3 0

20

10

0

- 1 0

- 2 0

- 3 0

Page 1

Group_amyloid_mutMUT_posMUT_neg1_NC_neg

dif'

6 0

40

20

0

- 2 0

- 4 0

- 6 0

Page 1

Predicted age overestimation in the PREVENT-AD did notdiffereitherbetweenAβpositiveandnegativeindividuals.

MutationarriersAβ-

groupDIAN_mut1_DIAN_n

dif'

6 0

40

20

0

- 2 0

- 4 0

- 6 0

Page 1

PIB_fSUVR_TOT_CORTMEAN2,22,01,81,61,41,21,0

dif'

6 0

40

20

0

- 2 0

- 4 0

- 6 0

Page 1

amyloid_index3,02,52,01,51,0

dif'

3 0

20

10

0

- 1 0

- 2 0

- 3 0

Page 1

F2,142=8.44p=.004

p=.06

p=.92

F1,60<1P=0.57

Aβ- Aβ+ APOE4carriers

Mutationnoncarriers

Aβ-

MutationcarriersAβ+

Mutationnoncarriers

Mutationcarriers

60

40

20

0

-20

-40

-60

Pred

ictedAg

eDifferen

ce p p=.03 r=-.12

p=0.22

r=-.11p=0.45

F1,254<1p=.97

APOE4noncarriers

Pred

ictedAg

eDifferen

ce

30

20

10

0

-10

-20

-30

Pred

ictedAg

eDifferen

ce

30

20

10

0

-10

-20

-30

30

20

10

0

-10

-20

-30

60

40

20

0

-20

-40

-60

Pred

ictedAg

eDifferen

ce

Pred

ictedAg

eDifferen

ce

60

40

20

0

-20

-40

-60

Pred

ictedAg

eDifferen

ce

MeanneocorticalAβ[PIB-PET]

MeanneocorticalAβ[NAV4964-PET]

1.01.21.41.61.82.02.2

1.01.52.02.53.0

DIANmutationnoncarriersDIANmutationcarriersPREVENT-ADCamCANADNI

Testset

Training set

5 10 15 20 25

5

5 2

10

10 2

10 5

15

15 2

15 5

Net

wor

k ar

chite

ctur

e

Validation set

5 10 15 20 25

Number of features

5

5 2

10

10 2

10 5

15

15 2

15 5

Null set

5 10 15 20 25

5

5 2

10

10 2

10 5

15

15 2

15 5

10

11

12

13

14

15

16

17

18

Neu

ralNetworkArchite

cture

Rootmeansquareerror(rm

se)

Participants and Methods Resting-state functional magnetic resonance imaging (rsfMRI) scans were collected in 1,350cognitivelynormalparticipantsfrom18to94yearsoldprovidedbytheDIAN,PREVENT-AD,Cam-CAN,ADNI,andICBMcohortstotrainandtesta“BrainAge”predictivemodel.

CohortsDominantlyInheritedAlzheimerNetwork isamultisite longitudinalstudywhichenrolls individualsaged18andolderwhohaveabiologicalparentthatcarriesageneticmutationresponsibleforautosomaldominantAD(ADAD).Cognitivelynormalmutationcarriersandnoncarrierswereincludedinthepresentstudy.Pre-symptomaticEvaluationofExperimentalorNovelTreatmentsforAlzheimer’sDiseaseisamonocentriclongitudinalcohortwhichincludescognitivelynormalolderindividualsaged55andolderwithafamilyhistoryofsporadicAD.Cambridge Centre for Ageing and Neuroscience is a large-scale monocentric research project includingcognitivelynormalindividualsaged18to88yearsold.Alzheimer’sDiseaseNeuroimagingInitiativeisamultisitelongitudinalstudywhichenrollscognitivelynormalandimpairedolderindividuals.Onlycognitivelynormalolderadultswereincludedinthepresentstudy.InternationalConsortiumforBrainMappingisamultisitestudy.Cognitivelynormalindividualsaged19to85formtheMontreal’ssitearchivedinthe1000FunctionalConnectomesProject’srepositorywereincludedinthepresentstudy.

Resting-statefunctionalMRI(rsfMRI)

Resting-statescanswereallpreprocessedwithNIAK(http://niak.simexp-lab.org/)

Averaged BOLD signals were extracted using thePowerandPetersonparcellation(Poweretal.,2011).Some ROIs were removed due to low coverageresultingina238x238Pearsoncorrelationmatrix.

26graphmetricswerethenextractedusingtheBrainConnectivityToolbox(Rubinov&Sporn,2010;https://sites.google.com/site/bctnet/)

Amyloid(Aβ)Positronemissiontomography(PET)

AβscanswereacquiredusingC11-PIBinDIAN(N=145)andF18-NAV4694inPREVENT-AD(N=62).Standardized uptake value ratios (SUVR; reference region:cerebellum grey matter) were averaged across frontal,temporal,parietalandposteriorcingulatecorticestoobtainaglobalindexofAβburden.

NeuralNetworkwastrainedtopredictage

NeuralnetworkswereconstructedusingMatlab(https://www.mathworks.com/products/deep-learning.html)

PIB-PETNAV4964-PET

MeanneocorticalAβextraction