Hi everyone,
I was wondering if there were any guidelines or recommendations for which additional regressors to include when running XCP-D on resting state data while including components from multi-echo denoising using tedana.
In an older version of XCP-D, you could point to a --nuisance_regressors strategy, such as acompcor, and also to a --custom_confounds file with excluded ICA components from tedana (orthogonalized with --tedort). Now, it appears that all confounds need to be included in the custom setup under the --nuisance_regressors flag, so I have been revisiting what to include.
After looking around these forums and the literature, it’s not clear to me which regressors to include, whether tedana components only, or other options from the fMRIPrep confounds such as movement, aCompCor, or even bandpass filtering. There are a few posts where @roeysc used custom confounds from tedana as well as the acompcor regression strategy, and some discussion here that tedana + the 36P option is overkill, but I’m not sure if there is a general consensus of what works.
My understanding of multi-echo denoising is that it can detect non-BOLD artifacts such as movement and scanner drift, but may leave behind BOLD-like artifacts. So, there may be an argument for additional regressors past those created by tedana. For instance, this paper suggests that ME-ICA + aCompCor is more effective than ME-ICA alone at 7 T (but less aggressive than ICA-AROMA).
However, I’m concerned about the degrees of freedom left in the data after regression. The paper above used scans with 1350 volumes, while I have data from a 9-minute, 300 volume scan. Given that tedana often outputs 20-30 nuisance components for regression, even a bandpass filter and 6 basic movement regressors are going to leave very few degrees of freedom.
So, could anyone suggest a set of regressors that has either worked well for them in the past, or has a solid theoretical basis, for post-processing multi-echo resting state data? Do you think the data needs additional filtering, movement regression, or anatomy-based regression, or is multi-echo ICA denoising alone sufficient? Thanks!
Sincerely,
Keith G. Jones