Missing physiological data

[HDhanis]

Hi there,

This is not exactly a technical question but I thought I would give it a go anyway.

I have been using the PhysIO for a while now in multiple fMRI projects, however, now I am going to start working with a delicate patient population for which quite some data is already acquired. Unfortunately, due to complications with having these patients in the scanner, physiological data is missing for around 10% of them.

The main question is : Have you encountered this, and do you have any advice on how to deal with it?

Some thoughts: I am reticent about changing the entire preprocessing pipeline to do denoising with ICA versus retroicor, however ICA would be doable in all patients including the ones with missing data. Would it be acceptable to do ICA based correction in the participants that don't have physio, and then include a covariate in the follow-up analyses that differentiates the strategies?

Really a bit at a loss as none of these solutions seems optimal, and discarding 10% of the patients would be really sad.

Best wishes,
Herberto Dhanis

[mrikasper]

Dear Herberto,

Thank you for being a loyal user of the PhysIO Toolbox! Indeed, what you ask is a very common scenario, because the quality of the physiological recordings, or its availability varies between subjects in most studies.

I understand that many researchers decide against having subject-specific analysis pipelines, because they are concerned about biasing results. The golden rule is typically standardization, to the extent that some studies don't even adjust slice geometry to the actual position of the head in the individual, just to keep the physical parameters the same.

My perspective on all these approaches is that in the end, we would like to perform statistical inference at the group level. For that, it is beneficial to reduce the variance in the data already at the single-subject level. You would sacrifice a lot of sensitivity by not using the optimal denoising model for the individual (be it recording-based or data-driven like ICA), and if you don't try to denoise every participant, you will have unequal residual variance between your subjects, which might be a concern for certain statistical approaches (although summary t-statistics based on the contrast images is surprisingly robust).

So, in summary, I think your approach of using a data-driven method for the participants that don't have physiological recordings seems plausible.

1. Within PhysIO, you can use the aCompCor method directly (Behzadi et al., 2007), which is using principal components of white matter/CSF regions as nuisance regressors. This way, you can use a near-identical pipeline of processing, where you can assess explained variance directly in the GLM, e.g., using F-contrasts over the nuisance regressors (which is a bit more cumbersome to assess in the ICA framework, I think, because it's more of a preprocessing step and you don't see task correlations etc. that you remove).
2. In any case, it makes sense to check the residual images (in SPM, Res.nii in your GLM folder) for the GLM, to see what residual error variance is not explained by your regressors of interest and nuisance regressors combined. This should be similar across participants.
3. Including the covariate for the denoising strategy in the group analysis is definitely a good idea (but you might also just see the indirect effects of why participants don't have physiological recordings, e.g., because setup time took longer due to them being more uncomfortable in the scanner in general, which might impact their task performance or motion patterns).

I hope this helps - let me know if you find some interesting differences between the denoising strategies!

All the best,
Lars