An electrophysiological validation of stochastic DCM for fMRI The goal of Dynamic Causal Modelling (DCM) of neuroimaging data is to study experimentally induced changes in effective connectivity among brain regions. In this work, we assess the predictive validity of stochastic DCM of fMRI data (Daunizeau et al., 2012), in terms of its ability to explain changes in the frequency spectrum of concurrently acquired EEG signal. First, we use a neural field model to show how dense lateral connections induce a separation of time scales, whereby fast (and high spatial frequency) modes are enslaved by slow (low spatial frequency) modes. This slaving effect is such that the frequency spectrum of fast modes (which dominate EEG signals) is controlled by the amplitude of slow modes (which dominate fMRI signals). This means we expect the frequency modulation of EEG to follow temporal variations of slow neural states driving fMRI BOLD signal changes (as in Kilner et al., 2005). Second, we use conjoint empirical EEG-fMRI data – acquired in epilepsy patients – to demonstrate the electrophysiological underpinning of (spontaneous) neural fluctuations inferred from stochastic DCM for fMRI.
J. Daunizeau Brain and Spine Institute, Paris, France Wellcome Trust Centre for Neuroimaging, London, UK
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM for fMRI
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM: motivation
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
DCM for fMRI: introduction structural, functional and effective connectivity structural connectivity
functional connectivity
effective connectivity
O. Sporns 2007, Scholarpedia
•
structural connectivity = presence of axonal connections
•
functional connectivity = statistical dependencies between regional time series
•
effective connectivity = causal (directed) influences between neuronal populations ! connections are recruited in a context-dependent fashion
DCM for fMRI: introduction functional integration and the neural code
activation studies: functional segregation
u1
effective connectivity studies: functional integration
1
2
1
u1
u2
u1
2
u2
u1 X u2
“where did the experimental manipulation have a specific effect?”
“how did the experimental manipulation propagate through the network?”
DCM for fMRI: introduction neural states dynamics a24
b12
d24
gating effect
2 4
u2 modulatory effect
1 3 c1 u1
driving input
f f 2 f 2 f x2 x f ( x, u ) f x0 ,0 x u ux 2 ... x u xu x 2 0
nonlinear state equation: m n (i ) ( j) x A ui B x j D x Cu i 1 j 1 Stephan et al., 2009
DCM for fMRI: introduction the neuro-vascular coupling u t
m n x A ui B (i ) x j D ( j ) x Cu i 1 j 1
experimentally controlled stimulus
neural states dynamics
vasodilatory signal
s x s ( f 1)
f
s
s flow induction (rCBF)
f s
( h) { , , , , E0 } ( n) { A, B(i ) , C, D( j ) }
Balloon model changes in volume
v f v1/ v
( q, v )
( h ) {r0 , }
hemodynamic states dynamics
f
v
changes in dHb 1/ q f E ( f,E0 ) E q0 v q / v
q
S q V0 k1 1 q k2 1 k3 1 v S0 v
k1 4.30 E0TE k2 r0 E0TE k3 1
BOLD signal change observation
DCM for fMRI: introduction the variational Bayesian approach to model inversion
ln p y m ln p , y m S q DKL q ; p y, m q
free energy : functional of q
mean-field: approximate marginal posterior distributions:
q , q 1
2
,
p 1 ,2 y, m 2
p 1 or 2 y, m 1
q 1 or 2
DCM for fMRI: introduction example: audio-visual associative learning auditory cue
visual outcome
or
P(outcome|cue)
or
Put
response 0
200
400
600
800
2000
time (ms)
PMd
PPA
FFA
PPA
cue-dependent surprise
Put
FFA
PMd
cue-independent surprise Den Ouden et al., 2010
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM: motivation
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
Stochastic DCM for fMRI the effect of state noise on network dynamics
2-regions DCM structure +
-
+
state-space landscape of the normal rate of convergence x
-
1
2 +
-
-
E2
B
C
u1 t 1
x2
moving frame of reference v t E1 v f t
A
D
x t x1
Daunizeau et al., 2012
Stochastic DCM for fMRI mediated influence: canonical model comparison u2
Model comparison: evidence against the full model
1
LBFAB log 3 u3
A u2
0 -20 -40 -60 -200
ANOVA effects
2
1
3 u3
u u1
p y B, m
20
ANOVA fit
u u1
p y A, m
-100
0 LBFAB
100
200
1 2 1
1 2 2
residuals empirical distribution
2
0.4 0.3 0.2 0.1 0 -200
-100 0 100 ANOVA residuals
200
40 20 0 -20 -40
-60 network du DCM type
1 1 1
1 1 2
2 1 1
2 1 2
2 2 1
2 2 2
B Daunizeau et al., 2012
300
Stochastic DCM for fMRI
SPMresults: .\20070914MP_sDCM\analysis
350
Height threshold F = 11.230194 {p<0.05 (FWE)} Extent threshold k = 0 voxels
400 example: epileptogenic network 20 40
60 Design matrix
precuneus PFC thalamus
30 25 20 15 10 5 0
80
100
pC
pC
PF
PF
Th
Th
C: Th
pC
PF
pC
PF
Th
C: PC
pC
pC
Th
C: PF
PF
pC
PF
C: all
pC
PF
PF
Th
Th
Th
Th
B: fb
B: ff
B: ext
B: int
pC
pC
PF
D: fb
PF
null Th
Th
Log-evidence (relative) Log-evidence (r
Bayesian Model Selection
120 60 100 40
-0.72
80 20 60 0
-0.72 -0.02
pC 1 2 3 4 5 6 7 8 9 10 11 1213 14 15 1617 18 19 2021 22 23 2425 26 27 2829 30 31 3233 34 35 36
-0.12
Models
40
-0.01 -0.06
0.02
0.01
-0.05
20
PF
-0.01
-0.05
0.02
Th 0
1 2 3 4 5 6 7 8 9 10 11 1213 14 15 1617 18 19 2021 22 23 2425 26 27 2829 30 31 3233 34 35 36
-0.72
Models
del Posterior Probability Model Posterior Probability
1
null
null
dDCM
sDCM
Bayesian Model Selection 1
0.8
0.6 1
model 26 (sDCM, B: fb, D: fb)
A
0.5
Bayesian Model Selection 0
B 0.4 0.8
D
0.2 0.6
hemo
-0.5
-1 0 0.4
1 2 3 4 5 6 7 8 9 10 11 1213 14 15 1617 18 19 2021 22 23 2425 26 27 2829 30 31 3233 34 35 36
Models
A
B D
hemo
Stochastic DCM for fMRI estimating hidden states from fMRI BOLD signals simulated hidden-states time series
simulated observations vasodilatory signal blood inflow blood volume dHb content neural state
1
4
y
3
x
0.5
5
2
0
1 -0.5 0 -1
10
20
30
-1
40
10
20
time
30
40
time
full lagged correlation matrix
instantaneous states correlation matrix
lagged neural states correlation matrix
1
2
1
1 0.8
10
4
0.6
20
0.4
30
2
0.2
40
0
50
-0.2
60
-0.4
70
-0.6 -0.8
80 20
40
60
80
-1
0.5
6 8
3
0
10 12
4
-0.5
14 5
16 1
2
3
4
5
5
10
15
-1
Stochastic DCM are sDCM neural state estimates noise, or what? 3 consistent ROIs across patients (N=9)
network connectivity
PFC: structure of residuals Model fit: versus y 4
-0.60
pC
-0.60 -0.02
PF pC
0.01
excluded
2
PF
fitted
Th -0.03
-0.01
y
0.01
0
0.03
-2
Th -0.60
-4
-6
macrosopic sDCM “neural” states
-6
-4
-2
0
2
0.1
PF pC Th
0.05
xˆ ( n ) t
0 -0.05 -0.1 -0.15 0
4
50
100
150
200 250 time (scans, TR = 3s)
300
350
400
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM: motivation
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
Separation of time scales in neural fields preamble: Kilner’s heuristic fMRI BOLD signal: • Amplitude modulation of EEG frequency bands (inconsistent results)
Power spectrum
• Frequency modulation of EEG signals (Kilner et al., 2005)
EEG frequency
Separation of time scales in neural fields multiscale portrait of brain dynamics macro-scale
meso-scale Golgi
micro-scale
Nissl
EI
external granular layer external pyramidal layer
EP
II
mean-field firing rate
internal granular layer internal pyramidal layer
synaptic dynamics
Separation of time scales in neural fields from micro- to meso-scale j t : post-synaptic potential of j th neuron within its ensemble 1 N H j ' t H t p t d t N 1 j ' j S t mean-field firing rate
membrane depolarization (mV)
mean firing rate (Hz)
ensemble density
p(x)
S
H
mean membrane depolarization (mV)
Separation of time scales in neural fields from meso- to macro-scale
r, t G S r, t r 2 v 2t 2 G 0 1 1 G r, t t exp H t 4 2 2 v t
2
v 1
“dispersive” propagator (Bojak & Liley, 2010)
spatio-temporal dynamics of mean PSP:
2 2 2 1 r , t r , t 2 t t 0 v v 2 r , t G v S r , t 2 t 2 2
mean PSP neural field
propagated arrival rate of mean PSP
Separation of time scales in neural fields emergence of slow and fast modes
projection of mean PSP onto Laplacian (spatial) eigenmodes:
2 w( k ) r k w( k ) r : k 0 0
z (k )
(k ) (k ) z t w r r , t dr 1 D z ( k ) t w( k ) r r , t dr 2 D
0 0 1 z1( k ) v ( k ) (0) 2 z2 c z 1 k 0 z ( k ) O z 2 2 z3( k ) very slow time scale of fundamental mode 1 2 2 this happens if: v 0
G 0v cz S z1(0) 1 S z1(0) 2 (0)
: slow modes enslaves the dynamics of the fast modes
Separation of time scales in neural fields The effect of neuronal activation on EEG frequency spectrum EEG frequency spectrum (enslaved by large-scale dynamics of slow modes)
EEG centre frequency
v 2 1 2k i 1 i 2 v 2 1 k kc 1 2k i i 2 c z (0) i 2
2
d
d
centre EEG frequency
normalized EEG frequency spectrum normalized frequency power
centre frequency (Hz)
22
20 -1
10
18
2 16
-2
10
-3
10
14
12 0
16
32
48
64
80
frequency (rad/s)
96 112 128
10
20
30
40
50
z (0)
fundamental mode dV (mV)
10 0
10
20
30
fundamental mode dV (mV)
z
(0)
40
50
Separation of time scales in neural fields Heuristic hemodynamic correlate of the EEG
large-scale coupling of distal neural fields (sDCM for fMRI):
x ( n ) f ( n ) x ( n ) , u, ( n ) f ( n ) x ( n ) , u, ( n ) Ax ( n ) ui B (i ) x ( n ) Cu x (jn ) D ( j ) x ( n ) i
j
linear
gating
2 feedback of fast modes onto slow modes → neural noise : O z
fMRI: slow rate of energy dissipation
x( n ) z1(0) z1(0) z1(0) z3(0)
EEG: fast PSP, with centre frequency
1
' (0) 3
z
2
x(n) O x2
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM: motivation
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
Stochastic DCM validation derivation of the EEG frequency modulation 8
EEG setup of the recording session
neural field model: fast modes empirical EEG data: subject MP
frequency power (A.U.)
log-power (A.U.)
EEG log- frequency power (A.U.)
6
4
2
0
-2
-4
-6
-8 20
40 60 80 frequency (Hz)
100
120
frequency (Hz)
slow EEG frequency modulation
t
...to be compared with sDCM “neural” states estimates: GSW GSW
frequency (Hz)
time (scans)
H1 : xˆ ( n )
X 0 e
H0 : X 0 e
Stochastic DCM validation single channel comparison to sDCM “neural” states estimates best channel: structure of residuals
best channel: model fit
36
34
34
32 30
100
32 30
observed predicted 100
200
26
400
28
30
best channel: adjusted fit
36
20
40
60
80
100
confounds
best channel fit: adjusted EEG frequency modulation
best channel: adjusted correlation
1
0.5
0.5 adjusted
EEG centre frequency (Hz)
34
best channel: adjusted residuals
adjusted predicted
1
250
GSW
best channel fit: adjusted EEG frequency modulation
0 -0.5
0 -0.5
-1
-1
-1.5
-1.5
0
32
predicted
time
GSW
200
350
26 300
150
300
28
28
0
50
time (scans)
36
observed
EEG centre frequency (Hz)
best channel fit: EEG frequency modulation
26
matrix of confounds
best channel fit: EEG frequency modulation
100
200
300
time
400
-1.5
-1
-0.5
0
0.5
1
predicted
GSW
GSW GSW
GSW
Stochastic DCM validation subject-level and group-level Bayesian model comparison LBF log p H1 log p H 0
→ Bayes factor for each channel/subject:
best subject: BMC across channels
group distribution of log Bayes factors
Bayesian model comparison accross channels 40
250 200
log-Bayes factor
log- Bayes factor
30 20 10 0 -10 0
posterior on model frequencies
150 100 50 0
5
10
15
20
EEG channels
25
30
-50
1
2
3
4 5 6 subjects
Exceedance probability (N=9):
7
8
9
P f H1 0.5 0.99
→ sDCM neural states estimates predict concurrent EEG frequency modulation above and beyond (>100) confounds
Overview 1
Dynamic Causal Modelling (DCM): introduction
2
Stochastic DCM: motivation
3
Separation of time scales in neural fields
4
Predicting EEG frequency modulation from sDCM for fMRI
5
Conclusion
References Daunizeau et al., 2012: Stochastic Dynamic Causal Modelling of fMRI data: should we care about neural noise? Neuroimage,62: 464-481.
Den Ouden et al., 2010: Striatal prediction error modulates cortical coupling. J. Neurosci, 30: 3210-3219. Bojak et al., 2010: Axonal velocity distributions in neural field equations. PLoS Comp. Biol. 6: e1000653. Stephan et al., 2009: Bayesian model selection for group studies. Neuroimage, 46: 1004-1017. Daunizeau et al., 2009: Dynamic Causal Modelling of distributed electromagnetic responses. Neuroimage, 47: 590-601. Stephan et al., 2008: Nonlinear dynamic causal models for fMRI. Neuroimage, 42: 649-662.
Kilner et al., 2005: hemodynamic correlate of EEG: a heuristic. Neuroimage, 28: 280-286.
many thanks to:
Karl J. Friston (London, UK) Klaas E. Stephan (Zurich, Switzerland) Louis Lemieux (Chalfont, UK) Anna Vaudano (Roma, Italy) Dimitris Pinotsis (London, UK)