Suggestions on how to set up GLM after fmriprep

Hi everyone,

I’m looking for help in order to set up my design matrix for first level analysis.
I’m using the fmriprep very last version 20.2.2.

The steps I would like to implement are below. I would like to start from the -preproc_bold.nii.gz 4D signal (normalized in the standard space MNI152NLin2009cAsym):

  1. regressing out the noisy IC calculated with ICA-AROMA from -preproc_bold.nii.gz as fmriprep provides a list of the noisy IC and the mixing matrix (using fsl_regfilt). Then, I’m planning to use this denoised signal as input for FEAT.
    I’m trying to avoid the use of the AROMA output (-smoothAROMAnonaggr_bold) and perform the denoising myself because: i) I have 1.5 mm isotropic fmri data and the output of AROMA is resampled to 2mm; ii) according to some previous topics on this website there is a debate on recalculating or not some confounds after AROMA.
    So, is this a reasonable step for denoising? I’m wondering if it is correct to use noisy IC estimated on a different signal (ICA-AROMA applies spatial filtering for example) to the -preproc-bold.nii.gz signal.

  2. highpass filtering and smoothing the data by switching on the built-in option of the Feat interface

  3. selecting the nuisance confounds for the design matrix. In order to do this step I’m looking at the barplot in the report which show the correlation between the estimated confounds and the global signal. My barplots show the highest correlation between global signal and: wm, csf, tcompcor, rot_x, rot_y, rot_z, trans_x, trans_y, trans_z (movements confounds show smaller correlation values). So, I’m planning to add all of them in the design matrix.

However, if I want to use tcompcor I need to add the cosine confounds which applies the highpass filtering and I thus need to switch off the temporal highpass filtering of Feat. Is this correct? If so, I need to add all the cosine confounds and let’s say the number of tempcor confounds which explains the 50% of the variance. Is this again a reasonable choice?

Finally, it makes sense to regress out the movement confounds after the denoising in step 1? This step should remove the movement-related IC components, so I’m not 100% sure to add movement confounds in the design matrix.

Thank you in advance for any help and suggestions.