[ENH] add transformation free model

[Remi-Gau]

related to #11

uses stats model to implement 3 different noising strategies on resting state data

inspired by nilearn's : https://nilearn.github.io/stable/modules/generated/nilearn.interfaces.fmriprep.load_confounds_strategy.html

TODO:

• decide on sensible filter cut of

Could possibly still use transformation to create other scrubbing regressors, but probably best to keep this simple

[Remi-Gau]

Reading the data descriptor paper, found some GLM description for the task data, that is also in there.

Could be interesting for FSL users who use FLOBS

Example voxelwise GLM analysis.

To illustrate an example of the simplest kinds of analyses that could be performed on our data, we conducted a basic GLM analysis. After the minimal preprocessing from fMRIPrep, we applied the following additional preprocessing steps. First, the functional data was spatially smoothed with a Gaussian kernel (FWHM = 8 mm), the data were de-trended, and framewise censoring was applied to volumes with framewise displacement >0.5 mm. To model the effect of stimulation, we did not use a canonical hemodynamic response function (HRF) and instead estimated the HRF by convolving a boxcar function denoting the electrical stimulation (50–90 ms, depending on the actual duration of the stimulation) with an optimized basis set function generated using FLOBS (FSL Linear Optimal Basis Sets)39. This optimization helped account for any non-canonical HRF shapes caused by the stimulation.

A mass univariate analysis was performed for one subject using a linear regression model. The analysis was performed in subject space. Statistical parametric maps were thresholded at an uncorrected voxelwise p < 0.001. The following confounds were used in the regression model: 6 motion parameters, one global signal regressor, anatomical noise components (aCompCor, n = 6).

[Remi-Gau]

Reading the data descriptor paper, found some GLM description for the task data, that is also in there.

This is also an interesting case because the events.tsv in there have no trial_type: only onset and duration

[adelavega]

Nice, looks good to me. I would say let's not close #11 though. I'd like to add an example that is transformation free and also event related.

[Remi-Gau]

@effigies @adelavega
This should be good to go AFAICT

[Remi-Gau]

Added some "comments": I believe we should make extensive / intensive use of them in the model zoo.

[effigies]

You have ds002799 added twice as submodules. The model names have hyphens instead of underscores. I might suggest the following:

• model-denoise_desc-compcor_smdl.json
• model-denoise_desc-scrubbing_smdl.json
• model-denoise_desc-simple_smdl.json

Finally, your code editor seems to be omitting or removing trailing newlines, which is counter to standard practice (see https://thoughtbot.com/blog/no-newline-at-end-of-file). Would you be willing to update your config?