More PNC
Opened this issue · 4 comments
Also would you happen to know whether it makes sense to use rapidtide on non-lfo regressors (RETROICOR cardiac and respiratory signals)? Do you have an idea, whether rapidtide yields better results than standard methods (slice-wise delays according to slice-timing)?
It's not really optimized for that - those signals are aliased, and while you can use it to remove aliased signals, I'm not sure if it's better than RETROICORR. If your TR is short enough that they aren't aliased, then just using a spectral filter is probably better.
For cardiac signals though, there is a method that was inspired by happy (the companion app in the package that extracts cardiac signals from multi band fmri) called WHOCARES. That's specifically designed to remove cardiac noise.
Happy does have a cardiac removal mode, but it's sort of half baked at the moment.
Let me restate my question: In your publication https://doi.org/10.1016/j.neuroimage.2012.01.140 (e. g. fig 3) you split the NIRS signal into different frequency bands (lfo, resp, card) and use rapidtide (/RIPTiDe) on each of them if I understand it correctly. I was curious if one could similiarly apply rapidtide to RETROICOR regressors and whether you already tried something like this. Or am I missing something?
First of all, rapidtide can certainly do that if you want (you can select the cardiac and respiratory bands instead of the default LFO band). In practice, since nowadays I'm doing primarily data driven analyses using only the fMRI data (no external recordings), I haven't really done that lately.
RETROICORR regressors use both sin and cosine components of the Taylor expansion of the periodic cardiac and respiratory signals, so the proper linear combination of the regressors can account for any phase shift within the period. So I suppose you could use rapidtide on one or the other, and halve the number of degrees of freedom lost by regressing them out, but I'm not sure you gain that much. I've also experimented with fitting the timeshift and regressing out the under sampled, aliased cardiac and respiratory time courses. It does work, but not as well as I'd hoped (which could definitely be due to boneheaded coding on my part - I'm not entirely convinced I did it properly). That said, you should probably look into WHOCARES, by Colenbier et al. (https://doi.org/10.1088/1741-2552/ac8bff). They did it properly, using entirely fMRI derived information (from happy, also part of the rapidtide package).
Thank you for your answer and your prolonged help. My thesis has by now shifted away from rapidtide though but just one last comment:
I was also curious about larger time shifts than the period.
To get it to work one has to disable the antialiasing filter and manually set the oversampling frequency as the regressor frequency to ensure rapidtide doesn't downsample the regressor prior to shifting.