[pymvpa] Fast event-related data and PyMVPA

[Yaroslav Halchenko]

well -- there are so many factors which could effect the results here --
from choice of parameters (C for SVM, how many best-anova features you
take), choices of how do you aggregate sensitivities (maxofabs might
be of the noisest ones) and going back to the design (# of trials,
proper order balancing) etc and indeed what do you put as features...
for the last factor you might like to see

http://www.sciencedirect.com/science/article/pii/S1053811911010081
Deconvolving BOLD activation in event-related designs for multivoxel pattern classification analyses
doi:10.1016/j.neuroimage.2011.08.076

for the choice of parameters, if number of trials permits -- nested
cross-validation to select the best parameters.

On Thu, 06 Oct 2011, Kimberly Zhou wrote:

>    Hi Everyone,
>    I searched for this in the list archives but didn't see anything
>    regarding this so here goes...
>    Has there been any success using PyMVPA with fast event-related data?
>    Looking at the sensitivity map in AFNI (mapped back into original space
>    with map2nifti), the high sensitivity areas look completely random.
>    They seem to mainly be on the edges of grey matter/in spaces. I wonder
>    if this is because, with fast event-related data, there is just too
>    little information for the classifier, or is there something wrong with
>    what I have been doing?
>    We used 3dDeconvolve's -stim_times_IM option to estimate the amplitude
>    of the response to each individual stimulus, plus this in PyMVPA:
>    from mvpa.suite import *
>    datasetfile = 'allstim.nii.gz'
>    attrfile = 'attr.txt'
>    attr = SampleAttributes(attrfile)
>    ds = fmri_dataset(datasetfile, targets = attr.targets, chunks =
>    attr.chunks, mask ='resampledmask.nii.gz')
>    zscore(ds, chunks_attr='chunks')
>    #no detrending...done in afni
>    clf = LinearCSVMC()
>    fsel = SensitivityBasedFeatureSelection(OneWayAnova(),
>    FractionTailSelector(0.08, mode='select', tail='upper'))
>    fclf = FeatureSelectionClassifier(clf, fsel)
>    sclf = SplitClassifier(fclf, enable_ca=['stats'])
>    cv_sensana = sclf.get_sensitivity_analyzer()
>    sens = cv_sensana(ds)
>    sens_comb = sens.get_mapped(maxofabs_sample())
>    map2nifti(avgds, sens_comb).to_filename('test.nii.gz')
>    Anybody have any ideas or info...?
>    Thanks,
>    Kim

> _______________________________________________
> 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