View source for FBIRN:FBIRNStatsMon2006

questions:

# How does modeling like Dynamic Causal modeling etc. relate to Barry's methods? Ans: There are different sources of variability that have to be modeled in different ways.
## difference in tasks
## subject to subject differences
## block to block (run to run) differences
## trial to trial variability. They've tried to model each of these in a realistic way; eg, subject to subject differences as differences in weights in region to region connectivity, by subject.
# Can you do statistical tests on the model? Ans: In simulation approach can run all the different "subjects" and get mean/variance as in actual experimental data.
# Would it help to bring in MEG? Ans. Absolutely.
# Would it be better to work with non-human primates and monitor via electrophysiology and fMRI in monkeys? Ans: Yes, and they are trying to go that way.
# Can we find a better statistic to use to talk about fMRI activation? E.g., the thresholded map, or the percent signal change--would some other measure be better, like the integrated activity over a few mm of brain region? Ans: One reason they had to do their own experiments to collect fMRI data was to know the data that the published results wouldn't tell them. To model the patterns, they need the mean and the variance, not a convolution of the two in a z score for example.
# Prediction of level of BOLD response from the model averages over total activity but doesn't include HRF? Ans: Yes, they convolve it with a simple HRF (poisson function) but it can be convolved with any choice. The values shown for the % difference were from the sampled, hemo-dynmically convolved summed activation over a region. At the moment since they aren't too worried about details like the undershoot a simple HRF will do.
# Pulsatility of the brain and the vessels as a source of correlation among regions: Do we need to be concerned about this? Ans: Yes, and that is a problem. Don't know the source of a lot of the fMRI or PET noise. The functional connectivity *could* be an artifact of the scanner but it changes from group to group (in his experience, in PET) in a way that is consistent with what is known about the disease. And the model needs to simulate the connectivity from the imaging data...
# Have you looked at all in the abnormalities in Sz? Ans: no, don't think so. Once he tried to simulate the hypofrontality via a frontotemporal disconnection. He did get something like that but then had effects in other regions that didn't match up. Lesion data and degenerative disease findings still have to be incorporated into the model. For all the disorders, it's important to build in some plasticity, to allow the simulated brain to try to adapt to the problem.
# In the model he builds, are the dynamics are all from the integrators? Ans: yes. The neurons are all leaky integrator models. (Has tried it with spiking neuron models too)
# Parametrization of the neurons--how does changing the connectivity weights, variance parameters, etc. affect behavior on an individual basis? Ans: Nope, has been focussing on the group average and group variability, rather than individuals.