We gathered fMRI data for a study on a heat pain task, employing a multi-echo sequence (TR = 1,000 ms, TE1 = 13.0 ms, TE2 = 31.26 ms, TE3 = 49.52 ms, Multi-band factor = 4). Our aim was to assess the predictive modeling performance of pain ratings, comparing single echo (the second echo), optimal combination, and Multi-Echo ICA (using the TEDANA tool).
During the GLM step, we regressed out the following:
- For single echo: linear trend + CSF + WM + 24 rp
- For optimal combination: linear trend + CSF + WM + 24 rp
However, we are uncertain about which components to include in ME-ICA. Currently, we have included CSF and WM in the ME-ICA GLM.
I understand that ME-ICA eliminates non-bold signals, but I am unclear about the specific components that are eliminated. If you have any insights on this, I would greatly appreciate your input.
Thanks a lot for your help
I would recommend including the components classified by tedana as “non-BOLD” (i.e., rejected). I would also recommend orthogonalizing your non-BOLD components (and possibly your other nuisance regressors) with respect to your BOLD components. Tedana has a page dedicated to this: Denoising Data with Components — tedana 23.0.2.dev14+g2cb05ab documentation.
Could you expand on this? The actual components will differ based on the data.
Are you asking which characteristics are used to reject components? If so, then tedana does have a page that describes the decision trees that are available: Included Decision Trees — tedana 23.0.2.dev14+g2cb05ab documentation. The different output files can be used to determine which exact node was used to reject each non-BOLD component. In 23.0.0+, the best files to look at are the status table and the tedana metrics file. In previous versions, you could look at just the tedana metrics file.