Dear community,
I’m analyzing a continuous task fMRI dataset with two goals in mind:
(1) connectivity analysis
(2) GLM analysis
The task lasts 15 minutes and has no trial structure: subjects continuously use an MR-safe mouse to move blocks on the screen. Data were preprocessed with fMRIPrep.
(1) Connectivity analysis
I’d like to calculate connectivity matrices and compare them across different phases of the continuous task (e.g., first vs. second half), as well as compare against resting-state connectivity. I’m using XCP-D to denoise the data (even though it’s task data).
Here, I think it makes sense to use censoring (e.g., FD > 0.3 mm). My concern is that if censoring is not applied, band-pass filtering could allow motion artifacts to “bleed” into nearby TRs. My understanding is that XCP-D handles this by interpolating across censored TRs before filtering and then re-censoring, which should help protect the surrounding volumes. Does that make sense?
(2) GLM analysis
For the GLM analysis, I’m leaning not to censor volumes. Instead, I plan to include one-hot regressors for the identified outlier TRs. This way, motion outliers are modeled out but the temporal structure of the data remains intact.
My question:
Does this approach (censoring for connectivity, regressors for GLM) align with best practices for continuous task fMRI? Of course, I plan to test robustness by comparing connectivity results with and without censoring.
Many thanks!
Roey
P.S.
I also realize band-pass filtering might be problematic if I expect slow drifts across the task (e.g., if subjects gradually shift strategies during the 15 minutes).
Related discussions: