Title: The General Linear Model and Statistical Parametric Mapping
1MEG Analysis in SPM Rik Henson (MRC CBU,
Cambridge) Jeremie Mattout, Christophe
Phillips, Stefan Kiebel, Olivier David, Vladimir
Litvak, ... Karl Friston (UCL, London)
2Overview
- 0. Standard preprocessing (filtering,
epoching...) - Random Field Theory for Space-Time images
- Empirical Bayesian approach to the Inverse
Problem - A Canonical Cortical mesh and Group Analyses
- Dynamic Causal Modelling (DCM)
3(No Transcript)
4Overview
- Random Field Theory for Space-Time images
- Empirical Bayesian approach to the Inverse
Problem - A Canonical Cortical mesh and Group Analyses
- Dynamic Causal Modelling (DCM)
5(No Transcript)
61. Localising in Space/Freq/Time
- Random Field Theory is a method for correcting
for multiple statistical comparisons with
N-dimensional spaces (for parametric statistics,
eg Z-, T-, F- statistics) - When is there an effect in time, eg GFP (1D)?
- Where is there an effect in time-frequency space
(2D)? - Where is there an effect in time-sensor space
(3D)? - Where is there an effect in time-source space
(4D)?
Worsley Et Al (1996). Human Brain Mapping, 458-73
7Where is an effect in time-sensor space (3D)?
- F-statistic for ANOVA across EEG subjects (Henson
et al, 2008, Neuroimage) - MEG data first requires sensor-level realignment,
using e.g, SSS...
Taylor Henson (2008) Biomag
81. Localising in Space/Freq/Time
- Extended to 2D cortical mesh surface
- RFT generally requires Gaussian smoothing, but
exerts exact FWE control for sufficient smoothing
- Nonparametric (permutation) methods of FWE
control make fewer distributional assumptions (do
not require smoothing), but do require
exchangeability
Pantazis Et Al (2005) NeuroImage, 25383-394
9Overview
- Random Field Theory for Space-Time images
- Empirical Bayesian approach to the Inverse
Problem - A Canonical Cortical mesh and Group Analyses
- Dynamic Causal Modelling (DCM)
102. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
11Weighted Minimum Norm, Regularisation
Linear system to be inverted
Y Data, n sensors x t1 time-samples J
Sources, p sources x t time-samples L Forward
model, n sensors x p sources E Multivariate
Gaussian noise, n x t Ce error covariance
over sensors
Since nltp, need to regularise, eg weighted
minimum (L2) norm (WMN)
W Weighting matrix W I
minimum norm W DDT coherent W
diag(LTL)-1 depth-weighted Wp
(LpTCy-1Lp)-1 SAM W
.
Phillips Et Al (2002) Neuroimage, 17, 287301
12Equivalent Bayesian Formulation
Equivalent Parametric Empirical Bayes
formulation
Y Data, n sensors x t1 time-samples J
Sources, p sources x t time-samples L Forward
model, n sensors x p sources C(e) covariance
over sensors C(j) covariance over sources
Posterior is product of likelihood and prior
W Weighting matrix W I
minimum norm W DDT coherent W
diag(LTL)-1 depth-weighted Wp
(LpTCy-1Lp)-1 SAM W
.
Maximal A Posteriori (MAP) estimate is
(Contrasting with Tikhonov)
Phillips Et Al (2005) Neuroimage, 997-1011
13Covariance Constraints (Priors)
How parameterise C(e) and C(j)?
Q (co)variance components (Priors) ?
estimated hyperparameters
142. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
15Expectation-Maximisation (EM)
How estimate ?? . Use EM algorithm
to maximise the (negative) free energy (F)
(Note estimation in nxn sensor space)
Once estimated hyperparameters (iterated
M-steps), get MAP for parameters (single E-step)
(Can also estimate conditional covariance of
parameters, allowing inference)
Phillips et al (2005) Neuroimage
16Multiple Constraints (Priors)
Multiple constraints Smooth sources (Qs), plus
valid (Qv) or invalid (Qi) focal prior
Mattout Et Al (2006) Neuroimage, 753-767
172. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
18Model Evidence
A (generative) model, M, is defined by the set of
Q(e), Q(j), L
The model log-evidence is bounded by the free
energy
Friston Et Al (2007) Neuroimage, 34, 220-34
(F can also be viewed the difference of an
accuracy term and a complexity term)
Two models can be compared using the Bayes
factor
Also useful when comparing different forward
models, ie Ls, Henson et al (submitted-b)
19Model Comparison (Bayes Factors)
Multiple constraints Smooth sources (Qs), plus
valid (Qv) or invalid (Qi) focal prior
Mattout Et Al (2006) Neuroimage, 753-767
202. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
21Temporal Correlations
To handle temporally-extended solutions, first
assume temporal-spatial factorisation
In general, temporal correlation of signal
(sources) and noise (sensors) will differ, but
can project onto a temporal subspace (via S) such
that
Friston Et Al (2006) Human Brain Mapping,
27722735
V typically Gaussian autocorrelations
S typically an SVD into Nr temporal modes
Then turns out that EM can simply operate on
prewhitened data (covariance), where Y size n x
t
22Localising Power (eg induced)
Friston Et Al (2006) Human Brain Mapping,
27722735
232. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
24Automatic Relevance Detection (ARD)
When have many constraints (Qs), pairwise model
comparison becomes arduous
Moreover, when Qs are correlated, F-maximisation
can be difficult (eg local maxima), and
hyperparameters can become negative (improper for
covariances)
Note Even though Qs may be uncorrelated in
source space, they can become correlated when
projected through L to sensor space (where F is
optimised)
Henson Et Al (2007) Neuroimage, 38, 422-38
25Automatic Relevance Detection (ARD)
When have many constraints (Qs), pairwise model
comparison becomes arduous
Moreover, when Qs are correlated, F-maximisation
can be difficult (eg local maxima), and
hyperparameters can become negative (improper for
covariances)
To overcome this, one can
1) impose positivity constraint on
hyperparameters
2) impose (sparse) hyperpriors on the
(log-normal) hyperparameters
Complexity
(where ? and S? are the posterior mean and
covariance of hyperparameters)
26Automatic Relevance Detection (ARD)
When have many constraints (Qs), pairwise model
comparison becomes arduous
Moreover, when Qs are correlated, F-maximisation
can be difficult (eg local maxima), and
hyperparameters can become negative (improper for
covariances)
Henson Et Al (2007) Neuroimage, 38, 422-38
272. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
28Multiple Sparse Priors (MSP)
So why not use ARD to select from a large number
of sparse source priors.!?
Friston Et Al (2008) Neuroimage
29Multiple Sparse Priors (MSP)
So why not use ARD to select from a large number
of sparse source priors.!
Friston Et Al (2008) Neuroimage
No depth bias!
302. Parametric Empirical Bayes (PEB)
- Weighted Minimum Norm Bayesian equivalent
- EM estimation of hyperparameters
(regularisation) - Model evidence and Model Comparison
- Spatiotemporal factorisation and Induced Power
- Automatic Relevance Detection (hyperpriors)
- Multiple Sparse Priors
- MEG and EEG fusion (simultaneous inversion)
31Fusion of MEG/EEG
Separate Error Covariance components for each of
i1..M modalities (Ci(e))
Data and leadfields scaled (with mi spatial
modes)
Remember, EM returns conditional precisions (S)
of sources (J), which can be used to compare
separate vs fused inversions
Henson Et Al (submitted-a)
32Fusion of MEG/EEG
Magnetometers (MEG)
Gradiometers (MEG)
Electrodes (EEG)
Fused
Henson Et Al (submitted-a)
33Overview
- Random Field Theory for Space-Time images
- Empirical Bayesian approach to the Inverse
Problem - A Canonical Cortical mesh and Group Analyses
- Dynamic Causal Modelling (DCM)
343. Canonical Mesh Group Analyses
- A canonical (Inverse-normalised) cortical mesh
- Group analyses in 3D
- Use of fMRI spatial priors (in MNI space)
- Group-based inversions
35A Canonical Cortical Mesh
Given the difficulty in (automatically)
creating accurate cortical meshes from MRIs, how
about inverse-normalising a (quality) template
mesh in MNI space?
Ashburner Friston (2005) Neuroimage
36A Canonical Cortical Mesh
N1
Apply inverse of warps from spatial
normalisation of whole MRI to a template cortical
mesh
Mattout Et Al (2007) Comp. Intelligence
Neuroscience
37A Canonical Cortical Mesh
But warps from cortex not appropriate to
skull/scalp, so use individually (and easily)
defined skull/scalp meshes
N9
- Statistical tests of model evidence over N9 MEG
subjects show - MSP gt MMN
- BEMs gt Spheres (for CanInd)
- (7000 gt 3000 dipoles)
- (Normal gt Free for MSP)
Henson Et Al (submitted-b)
383. Canonical Mesh Group Analyses
- A canonical (Inverse-normalised) cortical mesh
- Group analyses in 3D
- Use of fMRI spatial priors (in MNI space)
- Group-based inversions
39Group Analyses in 3D
Once have a 1-to-1 mapping from M/EEG source to
MNI space, can create 3D normalised images (like
fMRI) and use SPM machinery to perform
group-level classical inference
N19, MNI space, PseudowordsgtWords 300-400ms with
gt95 probability
Smoothed, Interpolated J
Taylor Henson (2008), Biomag
403. Canonical Mesh Group Analyses
- A canonical (Inverse-normalised) cortical mesh
- Group analyses in 3D
- Use of fMRI spatial priors (in MNI space)
- Group-based inversions
41fMRI spatial priors
Group fMRI results in MNI space can be used as
spatial priors on individual source
space... ...importantly each fMRI cluster is
separate prior, so is weighted independently
Flandin Et Al (in prep)
423. Canonical Mesh Group Analyses
- A canonical (Inverse-normalised) cortical mesh
- Group analyses in 3D
- Use of fMRI spatial priors (in MNI space)
- Group-based inversions
43Group-based source priors
Concatenate data and leadfields over i1..N
subjects
projecting data and leadfields to a reference
subject (0)
Common source-level priors
Subject-specific sensor-level priors
Litvak Friston (2008), Neuroimage
44Group-based source priors
N19, MNI space, PseudowordsgtWords, 300-400ms
with gt95 probability
Group Inversion
Individual Inversions
Taylor Henson (in prep)
45Summary
- SPM also implements Random Field Theory for
principled correction of multiple comparisons
over space/time/freq - SPM implements a variant of the L2-distributed
norm that - effectively automatically regularises in
principled fashion - allows for multiple constraints (priors), valid
invalid - allows model comparison, or automatic relevance
detection - to the extent that multiple (100s) of sparse
priors possible - also offers a framework for MEGEEG fusion
- SPM can also inverse-normalise a template
cortical mesh that - obviates manual cortex meshing
- allows use of fMRI priors in MNI space
- allows using group constraints on individual
inversions
46Overview
- Random Field Theory for Space-Time images
- Empirical Bayesian approach to the Inverse
Problem - A Canonical Cortical mesh and Group Analyses
- Dynamic Causal Modelling (DCM)
47Functional vs Effective Connectivity
Correlations A B C 1 0.49 1 0.30 0.12 1
B and C correlated because of common input from
A, eg A time-series B 0.5 A e1 C 0.3
A e2
B
A
C
- Effective connectivity is model-dependent
- Real interest is changes in effective
connectivity induced by (experimental) inputs
48Basic DCM Approach
Design experimental inputs
Define likelihood model
Specify priors
Invert model
Make inferences
491. Neural Dynamics
David et al. (2006) NeuroImage
502. Observer Model
- Measurements assumed to reflect currents in
(large) pyramidal cells (x0)
B
A
- One option is a small number of equivalent
current dipoles (ECDs)
C
- Fix their locations, but allow orientations to
be estimated as 3 parameters (q)
y1
y2
Kiebel et al. (2006) NeuroImage
51EEG example MisMatch Negativity (MMN)
Standards (1kHz)
Deviants (2kHz)
HEOG
VEOG
MMN deviants standards
gtSeed 5 ECDs
Garrido et al. (2007) NeuroImage
52EEG example MisMatch Negativity (MMN)
IFG
IFG
IFG
-
STG
STG
STG
STG
STG
STG
A1
A1
A1
A1
A1
A1
input
input
input
Forward
Forward
Forward
Backward
Backward
Backward
Lateral
Lateral
Lateral
Garrido et al. (2007) NeuroImage
53EEG example MisMatch Negativity (MMN)
MisMatch reflects changes in forward and backward
connections Invalid top-down predictions fail
to suppress bottom-up prediction error?
Group-based posterior densities of connections in
FB model
IFG
0.93 (55)
1.41 (99)
STG
STG
1.74 (96)
4.50 (100)
2.41 (100)
5.40 (100)
A1
A1
input
Garrido et al. (2007) NeuroImage
54The End (Really)
55If want to try
- http//www.fil.ion.ucl.ac.uk/spm
- SPM5 Manual (/spm5/man/manual.pdf)
- http//www.fil.ion.ucl.ac.uk/spm/data/mmfaces.html
56Future Directions
Variational Bayes (VB), relaxing Gaussian
assumptions e.g, VB for ECD (Kiebel et al,
Neuroimage, 2007) Dynamic Causal Modelling
(DCM), for ECD or MSP Multi-level heirarchical
models (e.g, across subjects) Nonstationary
hyperparameters Proper Data Fusion (single
forward model from neural activity to both M/EEG
and fMRI)
57Optimal Regularisation
Single hyper-parameter for a coherent
(smoothness) constraint on sources (cf. LORETA)
Mattout Et Al (2006) Neuroimage, 753-767
58Where is an effect in time-frequency (2D)?
Kilner Et Al (2005) Neuroscience Letters 374,
174178
59Expectation Maximisation
How estimate J and ? simultaneously? Maximise
the free energy (F)
Y Data, n sensors x t time-samples ?
hyperparameter(s) q(J) any distribution over J
using EM algorithm, where E-step
Friston Et Al (2006) Neuroimage, 20, 220-234
and M-step is
- For Gaussian distributions, equivalent to ReML
objective function
In practice, this gives ReML estimates of ?,
which can then be used to give MAP estimates of J
(via Cj and Ce)
60Temporal Correlations
To handle temporally-extended solutions, first
assume temporal-spatial factorisation
In general, temporal correlation of signal
(sources) and noise (sensors) will differ, but
can project onto a temporal subspace (via S) such
that
Friston Et Al (2006) Human Brain Mapping,
27722735
V typically Gaussian autocorrelations
S typically an SVD into Nr temporal modes
Then turns out that EM can simply operate on
prewhitened data (covariance), where Y size n x
t
61Localising Power (eg induced)
Similarly, can extend over trials as well as
samples, such that given i1..N trials, Yi
Can be shown that expected energy for one trial
in a time-frequency window defined by W
and total energy (induced and evoked) across
i1..N trials becomes
Friston Et Al (2006) Human Brain Mapping,
27722735
62Localising Power (eg induced)
Friston Et Al (2006) Human Brain Mapping,
27722735
63Where is an effect in time-sensor space (3D)?
- SPM of F-statistic for EEG condition effect
across subjects - (NB MEG data requires sensor-level realignment,
e.g SSS)
Henson Et Al (2008) Neuroimage