Network modelling using resting-state fMRI:
effects of age and APOE
Lars T. Westlye
University of Oslo
CAS kickoff meeting 23/8-2011
Patterns of brain activation during rest
The brain is not primarily reflexive
Whilst part of what we perceive comes through our senses from the object before us, another part (and
it may be the larger part) always comes out of our own head (William James, 1890)
Functional networks
Hierarchical clustering
The resting brain is highly organized into functional hierarchical networks
Independent component analysis (ICA)
A computational method for separating a multivariate signal into
additive and statistically independent (though not necessarily
orthogonal) subcomponents.
Originally proposed to solve blind source separation or so-called
cocktail-party problems:
Allows blind separation of N sound sources summed in recordings at N
microphones, without relying on a detailed model of the sound
characteristics of each source or the mixing process.
Example: Speech Separation
ICA
Courtesy of dr. Arno Delorme, UCSD
Typically, brain imaging data are high-dimensional and multivariate in
nature, i.e. the estimated signal could be regarded as a mixture of various
independent sources
The case of EEG
Spatial group ICA on temporally concatenated FMRI data
Multivariate Exploratory Linear Optimized Decomposition
into Independent Components (MELODIC)
Beckmann et al.
The various IC spatial maps reflect
intrinsic patterns of functional
organization across subjects and
correspond with known neuroanatomical
and functional brain ”networks”
Veer et al., 2010
How do we get from the group level to the subject level?
Dual regression allows for estimations of subect-specific spatial maps and
corresponding time courses
Spatiotemporal regression in two steps:
A) Use the group-level spatial maps as spatial
regressors to estimate the temporal dynamics (time
courses) associated with each gICA map
B) Use time courses (after optional
normalization to unit variance) spatial
regressors to find subject-specific
maps associated with the group-level
maps.
Spatiotemporal regression:
Yielding n by d time courses, where n=number of subjects and d=model order
(number of ICs).
Imaging phenotype: The covariance of the time courses reflect the large-scale
functional connectivity of the brain, and can be submitted to various connectivity
analysis - including graph theoretical approaches and other varietis of network
modeling - and subsequent analysis with relevant demographic, cognitive and
genetic data.
Application: Modelling the effects of age and APOE
N
Mea
n
Medi Min
an
Max
SD
Ap3
148
49.2
51.5
21.1
81.2
17.2
Ap4
74
52.2
55.4
21.7
79.1
16.1
Total
222
50.2
52.1
21.1
81.2
168
Imaging data: 1.5 T Siemens Avanto, 10 min resting state fMRI (200 TRs)
Conventional preprosessing including motion correction, filtering etc
Group ICA (on 94 subjects to avoid bias due to age and genotype) using temporal
concatenation in melodic (d=80) and dual regression in order to estimate subjectspecific time courses of each IC.
Exclusion of 43 ICs reflecting motion artefacts, pulsation etc yielded 36 resting state
networks (RSNs)
Questions:
1) Is the covariance between the time courses influenced by age?
2) Is the covariance between the time courses influenced by APOE status?
Select group spatial maps:
Main effects (Ap3 vs Ap4) and group by age interactions (are the age slopes comparable
across groups?) modelled using ANCOVAs
Hierarchical clustering of the connectivity matrix across subjects
Hierarchical clustering of the connectivity matrix across subjects
DMN
Motor
Visual
Hierarchical data-driven clustering reveals/recovers largescale brain networks
Visual
Motor
The connectivity matrix (across subjects)
Full correlations
Partial correlations (ICOV, lambda=10) (see Smith et al., 2010, NeuroImage)
The connectivity matrix (across subjects)
Partial correlations (ICOV, lambda=10) (see Smith et al., 2010, NeuroImage)
Full correlations
Direct links
Direct + indirect links
The connectivity matrix (across subjects)
Partial correlations (ICOV, lambda=10)
Full correlations
Direct links
Full correlation
-50
T-
Direct + indirect links
L=0 (no regularization)
50
L=5
L=10
The connectivity matrix (across subjects)
Network modelling using full correlation (strongest edges shown only)
The connectivity matrix (across subjects)
Network modelling using icov (L=10) (strongest edges shown only)
Modelling effects of age and APOE
Age
-10
0
T values
0
T values
Ap3>Ap4 (separate slopes)
Ap3>Ap4 (parallell slopes)
Ap3>Ap4 (separate slopes)
Ap3Age > Ap4Age
10
Age
10
Ap3>Ap4 (parallell slopes)
10
Ap3Age > Ap4Age
Modelling effects of age
Direct links
Direct + indirect links
Modelling effects of age (edges showing abs(tage>7))
Network modelling
Modelling effects of APOE
Ap3 > Ap4
Small effects compared to the age effects!
Ap3Age > Ap4Age
Modelling effects of APOE (edges showing abs(tgroup>2.5))
Network modelling (Ap3>Ap4) – NB! Multiple comparisons
Modelling effects of APOE by age interactions (edges showing abs(tgroup_by_age>2.5))
Network modelling (Ap3Age>Ap4Age) – NB! Multiple comparisons
Modelling effects of APOE by age interactions (edges showing abs(tgroup_by_age>2.5))
Although small effect sizes, all
”significant” edges point point
in the same direction
Alternative to the univariate edge-analyses:
Edge-ICA
Alternative to the univariate edge-analyses: Edge-ICA
Perform temporal ICA on the edges (one connectivity matrix per subject)
ICA
Transpose matrices
Correlate with age/APOE
Alternative to the univariate edge-analyses: Edge-ICA
Edge-ICA #2
Future possibilities?
Integrating measures of structural connectivity (DTI) and integrity (cortical thickness,
surface area) e.g. using linked ICA (Groves et al., 2010)
Assessing the between subject/sample reliability of the various measures
Implementing graph-theoretical procedures (Sporns et al)
etc
Thanks!