Time-resolved functional brain networks:Dynamics of human brain connectivity at rest

Andrew Zalesky

azalesky@unimelb.edu.au

ICON 2014

Background

Resting-state functional connectivity is dynamic

Time

BOLDSignal

Region 1

Region 2Theoretical:

• Breakspear, 2004

• Deco et al, 2008

• Friston, 1997

• Honey et al, 2007

Empirical:

• Chang & Glover, 2010

• Hutchison et al, 2013

• Kitzbichler et al, 2009

• Smith et al, 2012

Background

Resting-state functional connectivity is dynamic

Time

BOLDSignal

Region 1

Region 2Theoretical:

• Breakspear, 2004

• Deco et al, 2008

• Friston, 1997

• Honey et al, 2007

Empirical:

• Chang & Glover, 2010

• Hutchison et al, 2013

• Kitzbichler et al, 2009

• Smith et al, 2012

However , resting-state networks are studied as static networks...

Time-Average

Time

NetworkProperty

Motivation for studying time-resolved brain connectivity

• Test predictions from global workspace theory(Dehaene et al, 2014)

Inattentional Blindness

Wyart & Tallon-Baudry, 2009

Motivation for studying time-resolved brain connectivity

• Better characterize the “resting-state”

50% of participants are asleep for some interval during dataacquisition

Tagliazucchi & Laufs, 2014

Methods

• 10+10 healthy, young adults analysed• 15 mins rsfMRI with 0.72 s TR

Methods

• 10+10 healthy, young adults analysed• 15 mins rsfMRI with 0.72 s TR

Region

Region

Tim

e

Ghosh et al, 2008

Sliding WindowConnectivity estimated usingcorrelation in regionally averagedrsfMRI time series data fallingwithin tapered windows of length60 seconds (83 TRs).

ThresholdingConnectivity matricesthresholded to ensure a fixedconnection density for all timepoints.

Fluctuations in connectivity are coordinated among themost dynamic connections

−1

−0.5

0

0.5

1

0

10

20

30

HCP 105115

0.01FWER

Cor

rela

tion

Coe

ffici

ent

Tran

sitio

nC

ount

51%

1 min.

data1

−1

−0.5

0

0.5

1

−1

0

20

40data1

Sample Null Data

data1

data2

Actual Data

Surrogate

Median

Transition

• Null data was stationary, butotherwise matched to statisticalproperties of BOLD (i.e. auto- &cross-spectrum).

Network Efficiency

Low Efficiency (Low Cost) High Efficiency (High Cost)

Bullmore & Sporns, 2012

Sporadic emergence of brief, high efficiency states

Reg

ion

Time 1 min.

Sample Null DataHCP 105115

Simulated rsfMRI Data(Macaque)

HCP 111312

0.40

Sporadic emergence of brief, high efficiency states

Reg

ion

Time 1 min.

Sample Null DataHCP 105115

Simulated rsfMRI Data(Macaque)

HCP 111312

Low Efficiency State

High Efficiency State

0.40

High efficiency states are metabolically costly

• High efficiency states comprise longer connections than lowefficiency states (p < 0.01).

State Average Connection LengthLow Efficiency 67.1 ± 2.1 mmHigh Efficiency 73.0 ± 1.9 mm

High efficiency states are metabolically costly

• High efficiency states comprise longer connections than lowefficiency states (p < 0.01).

State Average Connection LengthLow Efficiency 67.1 ± 2.1 mmHigh Efficiency 73.0 ± 1.9 mm

• Long connections are metabolically costly (Liang et al, 2013).

• Hypothesis: Intermittent periods of high efficiency may be adynamic strategy that has evolved to minimize metabolicrequirements, while maintaining the connectome in a responsivestate.

Network Modules

Bullmore & Sporns, 2012

Dynamic connections interconnect distinct modules,whereas static connections are intra-modular

100

200

300

400

Somatomotor

Visual

Default Mode

OrbitofrontalLimbic

Static

Dynamic

Test

Sta

tistic

Dynamic connections have time-averaged correlationsnear zero

0 0.5 10

100

200

300

400

500

600

700

0 0.5 10

100

200

300

400

500

600

700

0 0.5 10

100

200

300

400

500

600

700

Test

Sta

tistic

Time-Averaged Functional Connectivity (Pearson Correlation Coefficient)

HCP 105115 HCP 118932 HCP 111312

r = −0.44, p < 0.001 r = −0.42, p < 0.001 r = −0.55, p < 0.001

• Hypothesis: Modular decomposition algorithms are effectivelydelineating communities of regions that are interconnected bystatic connections.

Summary

Multiple regions briefly increase their topological efficiency,and by inference, their capacity to transfer information.

But these intervals of high efficiency are supported by longanatomical connections and thus likely carry an extrametabolic cost.

Therefore , intermittent periods of high efficiency may be adynamic strategy that has evolved to minimize metabolicrequirements.

Food foraging in animalsEye tracking

Summary

Multiple regions briefly increase their topological efficiency,and by inference, their capacity to transfer information.

But these intervals of high efficiency are supported by longanatomical connections and thus likely carry an extrametabolic cost.

Therefore , intermittent periods of high efficiency may be adynamic strategy that has evolved to minimize metabolicrequirements.

Food foraging in animalsEye tracking

Time-averaged modular decompositions can be explained bythe layout of static and dynamic connections.

Reproducibility

Regional DefinitionsAAL-116

Random-90

Craddock-200

Confound CleanupCompCor

CompCor + Scrubbing

ICA-based X-Noiseifier

Simulated BOLD Data

MacaqueAxonal

Network

Morris-LecarNeural

Dynamics

BalloonWindkessel

BOLD Model

SimulatedBOLDData

No motion or physiological confounds in simulated BOLD

Honey et al, 2009

Further details:Zalesky A, Fornito A, Cocchi L, Gollo L, Breakspear M.Time-resolved functional brain networks reveal dynamics of humanbrain connectivity at rest.Proc Natl Acad Sci U S A. In Press.