dynamic causal modelling (dcm): theory demis hassabis & hanneke den ouden thanks to klaas enno...

28
Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging Neuroscience Institute of Neurology University College London

Upload: jade-horton

Post on 29-Dec-2015

221 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Dynamic Causal Modelling (DCM): Theory

Demis Hassabis & Hanneke den Ouden

Thanks to

Klaas Enno Stephan

Functional Imaging LabWellcome Dept. of Imaging NeuroscienceInstitute of NeurologyUniversity College London

Page 2: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Overview

• Classical approaches to functional & effective connectivity

• Generic concepts of system analysis

• DCM for fMRI:– Neural dynamics and hemodynamics– Bayesian parameter estimation

• Interpretation of parameters– Statistical inference– Bayesian model selection

Page 3: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

System analyses in functional neuroimaging

System analyses in functional neuroimaging

Functional integrationAnalyses of Analyses of inter-regional effectsinter-regional effects: : what are the interactions what are the interactions between the elements of a given between the elements of a given neuronal system?neuronal system?

Functional integrationAnalyses of Analyses of inter-regional effectsinter-regional effects: : what are the interactions what are the interactions between the elements of a given between the elements of a given neuronal system?neuronal system?

Functional connectivity= the temporal correlation = the temporal correlation betweenbetween spatially remote spatially remote neurophysiologicalneurophysiological events events

Functional connectivity= the temporal correlation = the temporal correlation betweenbetween spatially remote spatially remote neurophysiologicalneurophysiological events events

Effective connectivity= the influence that the elements = the influence that the elements

of a neuronal system exert over of a neuronal system exert over anotheranother

Effective connectivity= the influence that the elements = the influence that the elements

of a neuronal system exert over of a neuronal system exert over anotheranother

Functional specialisationAnalyses of Analyses of regionally specific regionally specific effectseffects: which areas constitute a : which areas constitute a neuronal system?neuronal system?

Functional specialisationAnalyses of Analyses of regionally specific regionally specific effectseffects: which areas constitute a : which areas constitute a neuronal system?neuronal system?

MECHANISM-FREE MECHANISTIC

Page 4: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Models of effective connectivity

• Structural Equation Modelling (SEM)

• Psycho-physiological interactions (PPI)

• Multivariate autoregressive models (MAR)& Granger causality techniques

• Kalman filtering

• Volterra series

• Dynamic Causal Modelling (DCM)Friston et al., NeuroImage 2003

Page 5: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Overview

• Classical approaches to functional & effective connectivity

• Generic concepts of system analysis

• DCM for fMRI:– Neural dynamics and hemodynamics– Bayesian parameter estimation

• Interpretation of parameters– Statistical inference– Bayesian model selection

Page 6: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Models of effective connectivity = system models.But what precisely is a system?

• System = set of elements which interact in a spatially and temporally specific fashion.

• System dynamics = change of state vector in time

• Causal effects in the system:– interactions between elements– external inputs u

• System parameters :specify the nature of the interactions

• general state equation for non-autonomous systems

)(

)(

)(1

tz

tz

tz

n

),,...(

),,...(

1

1111

nnn

n

n uzzf

uzzf

z

z

),,( uzFz

overall system staterepresented by state variables

nz

z

zdt

dz

1

change ofstate vectorin time

Page 7: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Example: linear dynamic system

LGleft

LGright

RVF LVF

FGright

FGleft

LG = lingual gyrusFG = fusiform gyrus

Visual input in the - left (LVF) - right (RVF)visual field.z1 z2

z4z3

u2 u1

state changes

effectiveconnectivity

externalinputs

systemstate

inputparameters

2

1

12

21

4

3

2

1

444342

343331

242221

131211

4

3

2

1

0

0

0

0

0

0

0

0

0

0

u

u

c

c

z

z

z

z

aaa

aaa

aaa

aaa

z

z

z

z

CuAzz

},{ CA

Page 8: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Extension:

bilinear dynamic system

LGleft

LGright

RVF LVF

FGright

FGleft

z1 z2

z4z3

u2 u1

CONTEXTu3

CuzBuAzm

j

jj

)(1

3

2

112

21

4

3

2

1

334

312

3

444342

343331

242221

131211

4

3

2

1

0

0

0

0

0

0

0

0

0

0

0000

000

0000

000

0

0

0

0

u

u

uc

c

z

z

z

z

b

b

u

aaa

aaa

aaa

aaa

z

z

z

z

Page 9: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Bilinear state equation in DCM

state changes

intrinsicconnectivity

m externalinputs

systemstate

direct inputs

CuzBuAzm

j

jj

)(1

mnmn

m

n

m

j jnn

jn

jn

j

j

nnn

n

n u

u

cc

cc

z

z

bb

bb

u

aa

aa

z

z

1

1

1111

11

111

1

1111

modulation ofconnectivity

Page 10: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Overview

• Classical approaches to functional & effective connectivity

• Generic concepts of system analysis

• DCM for fMRI:– Neural dynamics and hemodynamics– Bayesian parameter estimation

• Interpretation of parameters– Statistical inference– Bayesian model selection

Page 11: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

DCM for fMRI: the basic idea

• Using a bilinear state equation, a cognitive system is modelled at its underlying neuronal level (which is not directly accessible for fMRI).

• The modelled neuronal dynamics (z) is transformed into area-specific BOLD signals (y) by a hemodynamic forward model (λ).

λ

z

y

The aim of DCM is to estimate parameters at the neuronal level such that the modelled BOLD signals are maximally similar to the experimentally measured BOLD signals.

Page 12: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

BOLDy

y

y

hemodynamicmodel

Inputu(t)

activityz2(t)

activityz1(t)

activityz3(t)

effective connectivity

direct inputs

modulation ofconnectivity

The bilinear model CuzBuAz jj )(

c1 b23

a12

neuronalstates

λ

z

y

integration

Neural state equation ),,( nuzFz Conceptual overview

Friston et al. 2003,NeuroImage

u

z

u

FC

z

z

uuz

FB

z

z

z

FA

jj

j

2

Page 13: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

-

Z2

stimuliu1

contextu2

Z1

+

+

-

-

-+

u1

Z1

u2

Z2

2

112

22

2

112

21

12

2

1

2

00

0

0

0

12

u

uczuz

a

a

z

z

CuzBuAzz

bb

2

112

22

2

112

21

12

2

1

2

00

0

0

0

12

u

uczuz

a

a

z

z

CuzBuAzz

bb

Example: generated neural data

u1

u2

z2

z1

Page 14: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

sf

tionflow induc

s

v

f

v

q q/vvf,Efqτ /α

dHbchanges in

1)( /αvfvτ

volumechanges in

1

f

q

)1( fγszs

ry signalvasodilato

},,,,{ h},,,,{ h

important for model fitting, but of no interest for statistical inference

signal BOLD

qvty ,)(

The hemodynamic “Balloon” model

)(tz

activity • 5 hemodynamic parameters:

• Empirically determineda priori distributions.

• Computed separately for each area (like the neural parameters).

Page 15: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

LGleft

LGright

RVF LVF

FGright

FGleft

Example: modelled BOLD signal

Underlying model(modulatory inputs not shown)

LG = lingual gyrus Visual input in the FG = fusiform gyrus - left (LVF)

- right (RVF)visual field.

blue: observed BOLD signalred: modelled BOLD signal (DCM)

left LG

right LG

Page 16: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Overview

• Classical approaches to functional & effective connectivity

• Generic concepts of system analysis

• DCM for fMRI:– Neural dynamics and hemodynamics– Bayesian parameter estimation

• Interpretation of parameters– Statistical inference– Bayesian model selection

Page 17: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Bayesian rule in DCM

)()|()|( pypyp )()|()|( pypyp posterior likelihood ∙ prior

Bayes Theorem• Likelihood derived from error

and confounds (eg. drift)

• Priors – empirical (haemodynamic parameters) and non-empirical (eg. shrinkage priors, temporal scaling)

• Posterior probability for each effect calculated and probability that it exceeds a set threshold expressed as a percentage

Page 18: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

sf (rCBF)induction -flow

s

v

f

stimulus function u

modelled BOLD response

vq q/vvf,Efqτ /α1)(

dHbin changes

/αvfvτ 1

in volume changes

f

q

)1(

signalry vasodilatodependent -activity

fγszs

s

)(xy )(xy eXuhy ),(

observation model

hidden states},,,,{ qvfszx

state equation),,( uxFx

parameters

},{

},...,{

},,,,{1

nh

mn

h

CBBA

• Combining the neural and hemodynamic states gives the complete forward model.

• An observation model includes measurement error e and confounds X (e.g. drift).

• Bayesian parameter estimation: minimise difference between data and model

• Result:Gaussian a posteriori parameter distributions, characterised by mean ηθ|y and covariance Cθ|y.

Parameter estimation in DCM

ηθ|y

neural stateequation

CuzBuAz jj )( CuzBuAz j

j )(

Page 19: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Overview

• Classical approaches to functional & effective connectivity

• Generic concepts of system analysis

• DCM for fMRI:– Neural dynamics and hemodynamics– Bayesian parameter estimation

• Interpretation of parameters– Statistical inference– Bayesian model selection

Page 20: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

DCM parameters:interpretation & inference

Hypothesis: modulation by context > 0

–How can we make inference about effects represented by these parameters

–At a single subject level?

–At a group level?

– How do we select between different models?

- DCM gives gaussian posterior densities of parameters (intrinsic connectivity, effective connectivity and inputs)

LGleft

LGright

RVF LVF

FGright

FGleft

z1 z2

z4z3

u2 u1

CONTEXTu3

Page 21: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

• Assumption: posterior distribution of the parameters is gaussian

• Use of the cumulative normal distribution to test the probability by which a certain parameter (or contrast of parameters cT ηθ|y) is above a chosen threshold γ:

• γ can be chosen as zero ("does the effect exist?") or as a function of the expected half life τ of the neural process: γ = ln 2 / τ

Bayesian single-subject analysis

ηθ|y

ηθ|y

Probability

Page 22: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Group analysis

• In analogy to “random effects” analyses in SPM, 2nd level analyses can be applied to DCM parameters:

Separate fitting of identical models for each subject

Separate fitting of identical models for each subject

Selection of bilinear parameters of interestSelection of bilinear

parameters of interest

one-sample t-test:

parameter > 0 ?

one-sample t-test:

parameter > 0 ?

paired t-test: parameter 1 > parameter 2 ?

paired t-test: parameter 1 > parameter 2 ?

rmANOVA: e.g. in case of

multiple sessions per subject

rmANOVA: e.g. in case of

multiple sessions per subject

Page 23: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Model comparison and selection

Given competing hypotheses on structure & functional mechanisms of a system, which model is the best?

For which model i does p(y|mi) become maximal?

Which model represents thebest balance between model fit and model complexity?

Pitt & Miyung (2002), TICS

Page 24: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Bayesian Model Selection

Bayes theorem:

Model evidence:

The log model evidence can be represented as:

Bayes factor:

)|(

)|(),|(),|(

myp

mpmypmyp

dmpmypmyp )|(),|()|(

)(

)()|(log

mcomplexity

maccuracymyp

)|(

)|(

jmyp

imypBij

Penny et al. 2004, NeuroImage

Page 25: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

The DCM cycle

Design a study thatallows to investigatethat system

Extraction of time seriesfrom SPMs

Parameter estimationfor all DCMs considered

Bayesian modelselection of optimal DCM

Statistical test on

parameters of optimal

model

Hypothesis abouta neural system

Definition of DCMs as systemmodels

Data acquisition

Page 26: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Inference about DCM parameters: Bayesian fixed-effects group analysis

Because the likelihood distributions from different subjects are independent, one can combine their posterior densities by using the posterior of one subject as the prior for the next:

)|()...|()|(),...,|(

...

)|()|(

)()|()|(),|(

)()|( )|(

111

12

1221

11

ypypypyyp

ypyp

pypypyyp

pypyp

NNN

)|()...|()|(),...,|(

...

)|()|(

)()|()|(),|(

)()|( )|(

111

12

1221

11

ypypypyyp

ypyp

pypypyyp

pypyp

NNN

1,...,|

1|

1|,...,|

1

1|

1,...,|

11

1

NiiN

iN

yy

N

iyyyy

N

iyyy

CC

CC

1,...,|

1|

1|,...,|

1

1|

1,...,|

11

1

NiiN

iN

yy

N

iyyyy

N

iyyy

CC

CC

Under Gaussian assumptions this is easy to compute:

groupposterior covariance

individualposterior covariances

groupposterior mean

individual posterior covariances and means

See:spm_dcm_average.m Neumann & Lohmann, NeuroImage 2003

Page 27: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

Approximations to model evidence

Laplace approximation:

Akaike information criterion (AIC):

Bayesian information criterion (BIC):

pmaccuracymyAIC )()|(

Penny et al. 2004, NeuroImage

yppypT

py

eT

e hh

mcomplexitymaccuracyF

||1

|

1

log2

1log

2

1)()(

2

1

))(())((2

1log

2

1

)()(

CCθθCθθ

θyCθyC

yppypT

py

eT

e hh

mcomplexitymaccuracyF

||1

|

1

log2

1log

2

1)()(

2

1

))(())((2

1log

2

1

)()(

CCθθCθθ

θyCθyC

SNp

maccuracymyBIC log2

)()|(

Unfortunately, the complexity term depends on the prior density, which is determined individually for each model to ensure stability. Therefore, we need other approximations to the model evidence.

Page 28: Dynamic Causal Modelling (DCM): Theory Demis Hassabis & Hanneke den Ouden Thanks to Klaas Enno Stephan Functional Imaging Lab Wellcome Dept. of Imaging

DCM parameters = rate constants

azdt

dz )exp()( 0 atztz

The coupling parameter a thus describes the speed ofthe exponential growth/decay:

)exp(

5.0)(

0

0

az

zz

Integration of a first order linear differential equation gives an exponential function:

/2lna

05.0 z

a/2ln

The coupling parameter a is inversely proportional to the half life of z(t):