[pymvpa] multi-parameter mvpa?

[Yaroslav Halchenko]

just to make clear:  your goal is to predict psychological measure
(categorical or continuous?) based on 4 values available for each
location/voxel ( FA, AD, RD and MD ), right?

should then be easily possible and there are multiple ways to skin the
cow here -- e.g. 

* load series of those 4 volumes using the same mask image into 4 datasets,
  e.g. ds_pd = fmri_dataset(['file1.nii.gz', ...], ...)

* (tricky one) assign each one of them some unique .fa (e.g. call it domain, so
ds_pd.fa['domain'] = 0, ds_ar.fa['domain'] = 1, ...) , which is needed for
searchlight later on to tell it to get multiple domains at once

* hstack them all (ds = hstack((ds_pd, ds_ar, ...))),
  assign the phys measure to targets (ds.sa['targets'] = your_measure_per_each_of_those_volumes),
 
* assign chunks sample attribute (that is where there are options
  depending either you have groups of subjects there etc), e.g.

  ds.sa['chunks'] = np.arange(len(ds))

and then for the searchlight you would need to compose the neighborhood
specification which would not only get a Sphere in space but also across
those hstacked volumes, e.g. for quick GNB searchlight to try, smth like

        sl = GNBSearchlight(GNB(), OddEvenPartitioner(),
                            IndexQueryEngine(voxel_indices=Sphere(3),
                                             domain=Sphere(10)), # 10 just to get all domains at once
                            errorfx=mean_mismatch_error,
                            roi_ids=np.arange(ds_pd.nfeatures))
        map = sl(ds)

roi_ids here is to say that you want to "go" only through  the voxels on 1
volume since all other modalities (ar, rd, ...) would be picked up by that
Sphere(10).

.... I guess that would be the start (just typed here -- so could be having
typos and incorrect specs ;) proceed with caution)

mimiquing Event-related dataset could be another way I guess ;)

On Fri, 06 Apr 2012, Xiangzhen Kong wrote:

>    Hi, all
>    What I want to do now is to predict a psychlogical mearsure using
>    multi-parameters of a voxel in DTI data, such as FA, AD, RD and MD.
>    Is there a convenient way to do this?
>    I have tried to use the way just like in Event-related Data Analysis. But
>    when I added a node Searchlight, I was lost. Do you think I can do this
>    way?
>    If so, how can I do the 'multi-space' analysis with only 4 timepoints (I
>    means 4 DTI-parameters).
>     
>    B.T.W.��can someone talk sth. about the result of the Event-related Data
>    Analysis?
>     
>    Best,
>     

>    --------------------------------------------------------------------------

>    Xiangzhen Kong

> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa


-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic