DCM demo

André Bastos and Martin Dietz

Wellcome Trust Centre for Neuroimaging

pseudo-random auditory sequence

80% standard tones – 500 Hz

20% deviant tones – 550 Hz

time

standards deviants

Mismatch negativity (MMN) paradigm and hypothesis

time (ms)

Paradigm:

amp

litu

de

(μV

)

0

200

-

+

0

Deviant ERPStandard ERP

Hypothesis:MMN is caused by recurrent dynamics enabled by backward connections

Garrido et al., Neuroimage 2007

Mismatch Negativity scalp topography of ERPs

time (ms)

sens

ors

sens

ors

standard

deviant

Deviant ERPStandard ERP

time (ms)

amp

litu

de

(μV

)

0

200

-

+

0

Hypothesis:

MMN is caused by recurrent dynamics enabled by backward connections

100 2000

The generative model

Source dynamics f

states x

parameters θ

Input u

Evoked response

data y

),( xgy

Observation model g

x.

f (x,u,)

David et al., Neuroimage 2006; Kiebel et al., Neuroimage 2006

DCM specification

A1 A1

STG

input

STG

IFGseveral plausible models…

modulation of effective connectivity

Forward - F

Backward - B

Both - FB

1 2

3 4

5

Garrido et al., Neuroimage 2007

Opitz et al., 2002

lSTG rSTG

rIFG

Deviant responseStandard response

time (ms)

amp

litu

de

(μV

)

0

200

-

+

0

What set of areas and interconnections caused the MMN?

A1 A1

STG STG

ForwardBackward

Lateral

STG

input

A1 A1

STG STG

ForwardBackward

Lateral

input

A1 A1

STG

ForwardBackward

Lateral

input

Forward - F Backward - BForward and

Backward - FB

STG

IFGIFGIFG

DCM specification of different models

modulation of effective connectivityGarrido et al., Neuroimage 2007

Analysis steps

0. Have a HYPOTHESIS!

1. Preprocessing and SVD decomposition

2. Model specification: specify cortical areas and inter-areal connections for various competing models that you think might explain your data

3. Model inversion: find the parameters that minimize differences between observed measurements and model predictions for each of the competing models

4. Bayesian model comparison: make a statistical inferences about which model best describes the data