Dear Amy,
>I have followed your lab's work on effective connectivity, and I have a
quick question for you. We just completed a pilot study, in which we lowered
the blood glucose levels in diabetic patients and a control group, while
using fMRI imaging. Our focus was the hypothalamus, and whether there were
major differences in the timing and magnitude of activation between the
diabetic and control groups. There were differences in both categories, and
while analyzing the data, I noticed that a certain set of brain regions
consistently activated during hypoglycaemia, and I thought it might be
interesting to examine the effective connectivity of these regions during
hypoglycaemia, and possible differences between the two groups. Is this
something that would be appropriate to analyze using Dynamic Causal
Modelling? I am wondering if it is possible to measure differences in
connectivity that result from a physiological process rather than a
cognitive task using this method. If DCM is not appropriate, do you have any
ideas about other methods that might be applicable? Would SEM be more
>appropriate?
Yes, DCM could certainly be applied in this context. DCM is usually used
to assess changes in coupling induced by some experimental factor (these
changes are encoded by the bilinear coupling parameters - B). The
experimental factors
can be cognitive, sensory, physiological or anything else that may affect
coupling.
DCM is preferred over SEM, because SEM (i) does not model the mapping from
neuronal
to hemodynamic responses (ii) assumes the source of covariance is induced by
random innovations, as opposed to experimental effects, (iii) uses a rather
ad hoc estimation scheme based on covariances, as opposed to the data, that
precludes
realistic connections (i.e. reciprocal connections and loops).
I think the critical thing here is whether you used a primary factor to
evoke the
activation you referred to (e.g. some task or stimulus). If you did then your
design has three factors (task - within subject, Hypoglycaemia - within
subject and
diagnosis - between subject). To use DCM you would concatenate the
time-series from each
subject and perform a set of subject-specific DCMs, using task as a
perturbing input and
Hypoglycaemia as a bilinear input (i.e. allowing for blood sugar to affect
the sensitivity of
a region to its afferents). You would then take the estimates of coupling
strength
(A - represent experimentally-independent coupling strengths and B glucose
dependent-coupling)
and compare then using T-tests (or ANOVA) at the second (between-subject)
level.
If you have simply subjected subjects to changes in blood sugar, you would
use this as the
perturbing factor in the first-level (subject-specific) DCMs and take the A
parameters to
second-level (between subject) T tests. This may not be so easy, unless
you can specify the
glucose perturbation in terms of onsets and durations (as required by SPM
and the integration
scheme used by DCM).
If you wanted to use SEM, you could simply rely on endogenous variation in
hemodynamic responses and assume that these are a veridical reflection of
stochastic,
low-frequency neuronal responses. In this instance, you would use a stacked
model approach
to compare models that allowed for different connections between the two
groups (as assessed
with the Chi-squared statistic).
In both cases, you would have to specify the graph (DCM) (i.e. the nodes
and connections that
capture the observed responses). We unusually used the maxima of a
conventional SPM analysis
to identify regions responding to experimental manipulations.
With very best wishes,
Karl
|