[pymvpa] classification based on individual parameter estimates from FSL

Nick Oosterhof nikolaas.oosterhof at unitn.it
Tue Aug 5 10:01:27 UTC 2014


On Aug 4, 2014, at 12:44 PM, David Soto <d.soto.b at gmail.com> wrote:

> for some reason the maptonifti works fine with the output of the searchlight (eg img=map2nifti(sl_res) gives the same as  img=map2nifti(dataset=ds, data=sl_res).
> 
> I think this may be the header info was incorporated in the fmri_dataset call by adding the mask and the add_fa
> ds = fmri_dataset(samples=os.path.join(datapath1, 'predbothsi.nii.gz'),
>                   targets=attr.targets, chunks=attr.chunks,
>                   mask=os.path.join(datapath1, 'mask.nii.gz'),#based on FEAT analyses
>                   add_fa={'unmbral_glm': os.path.join(datapath1, 'mask.nii.gz')}) .

I don't agree, it seems the nifti header comes directory from the 'samples' input ('predbothsi.nii.gz').
(The mask and add_fa values are not used to set the nifti header)

> 
> What is intriguing is that the output of the FEAT GLM gives a robust univariate signal in visual cortex for the contrast a vs. b in task 1 and for a vs. b in task 2.
> Yet I tried with different searchlight radius and only get near chance classification
> from task1 to task 2.....I guess this could simply mean the patterns of responses in visual cortex are very different across task contexts, yet the signal associated with activation level is picked up by the GLM?

It is possible that the response magnitude differs across tasks. If in task A responses to the two conditions are, say, 2 and 4 (arbitrary units), and in task B they are 20 and 22, then you would expect chance classification accuracy.

z-scoring or de-meaning the data (for each task separately) may address this. 

> if so would it be reasonable to investigate this further for instance by deriving similarity matrices across individual 
> parameter estimates for task 1 and task 2?

Usually similarity matrices are based on correlations, which involves subtracts the mean response. It may work, but this analysis is qualitatively different from SVM classification. The latter can detect information content if, for each condition separately, the response is the same for all features (but differs across conditions); the former cannot.



More information about the Pkg-ExpPsy-PyMVPA mailing list