fMRI

Movement tasks and data analysis and artifacts

CAUTION - EXTERNAL EMAIL:Do not click links or open attachments unless you recognize the sender.

Hi Paul,

Thank you for your swift and very detailed reply!

I will need some time to dig deeper into the details of the RMANOVA. Your remarks and your papers will definitely help me with that.

Unfortunately, I don’t have access to one of your papers, which I would be very interested in: the 2004 Neuroimage publication (A method for removal of global effects from fMRI time series). Would it be possible for you to share a pdf of this paper with me?

I found it interesting that you mentioned your other study with foot movement:

Another set of data where we moved people’s feet (currently under review) showed big (non-neural %’s) signal drops at the time of foot movement, even with motion correction etc. These were clearly movement-related, but could be indirect – BP and muscle blood flow changes due to lifting the leg, for example.

As it happened, we did a test measurement just last week for a new paradigm which involves hand movements in amputees. The subject in the scanner was a HC and was moving his hands (opening – closing (fist)) in a block design, while watching them in a mirror. The fMRI sequence was a high resolution multiband sequence with a MB factor of 4.

At rest he rested his hands at his side and lifted them during the opening and closing blocks (we need to work on the instructions to prevent bulk motion in a better way). This led to huge artefacts (signal drops) in the BOLD images, which are probably only partially directly related to head movement.

There are at least two other mechanisms by which the artefacts might be (partially) caused. (I am certain that you know well about them, but just did not mention them).

We would think that lifting the hands changes the homogeneity of the magnetic field in a major way, so that the pre-scan shim is not good any more. That might also happen with movement of the feet which probably cannot be done in complete isolation from the rest of the body.

Also, there are spin-history effects, which cannot be corrected by motion correction, i.e. alignment of the images.

Here are some slices of one TR from this series after motion correction. You can still see the artefacts very clearly. It looks like every 4^th slice is affected in a rather similar way (due to the fact that every fourth slice was recorded at the same time.). Some images have lower signal only in parts of the brain (e.g. 46 anterior vs 48 posterior). The artefact is in fact pretty similar to the uncorrected image.

Open

In the next TR the artefacts had a completely different distribution across slices.

Probably, for this study we will have to consider other ways to mitigate the effects of the artefacts, like ICA to clean up the data (FSL-FIX or FSL-AROMA), or censoring of individual outlier data-points (‘scrubbing’). The later strategy might not be feasible as most of the artefacts would happen during the TRs which we would be interested in. A further complication is that movements can have effects lasting well longer than the movement as has been shown for resting state fMRI.

In some of our earlier studies with facial motion I tried to choose block-length of the movement in a way to make the expected motion artefact rather independent of the expected hemodynamic response (i.e. independence of the box-car predictor and the HRF-convolved box-car, under the assumption that the motion artefact is only concurrent with the movement). But while this might help in a traditional fMRI analysis, it would probably not help much with your RMANOVA approach.

Again, thanks for your kind reply,

Best,

Carsten

From: Macey, Paul [mailto:PMacey@sonnet.ucla.edu]
Sent: Samstag, 30. April 2022 03:53
To: Schmidt-Samoa, Carsten
Subject: RE: Pal et al. PLoS One 2021

Hi Carsten,

Thanks for your interest, and I agree the figures are complex and the mixture of lateralizing, anterior-posterior and relative signal changes is confusing!

I think you are correct about mistakes in the last three figures – repeating what you found, the legends are accurate, the titles in the figure images (bold text at top) are accurate, but the figure images are mismatched - the image in now in Fig. 5 should be Fig 3, the image now in Fig 3 should be Fig 4, and the image now in fig 4 should be Fig 5. I will contact the journal.

Re you other questions, I’ve responded below.

Best wishes,

Paul

Paul M. Macey

Professor

Director of Evaluation

UCLA School of Nursing

424-234-3244

From: Schmidt-Samoa, Carsten <carsten.schmidt-samoa@med.uni-goettingen.de>
Sent: Friday, April 29, 2022 7:51 AM
To: pmacey@ucla.edu
Subject: Pal et al. PLoS One 2021

CAUTION - EXTERNAL EMAIL:Do not click links or open attachments unless you recognize the sender.

Dear Dr. Macay,

In our lab (Department of Cognitive Neurology, University Medicine Göttingen, Germany, head: Dr. M. Wilke) we have started to look at the central autonomic network and autonomous challenges during functional imaging. A colleague pointed me to your publication in PLoS One 2021 (‘Insular response …’), which I have read with great interest.

I have some points I am not sure about, and I would like to ask you if you could help me to get them right.

I was puzzled by the figures 3-5, because the captions of the figures do not agree with the figure legends and the paper text. I guess something went wrong in the editing process.

I think I figured it out, but it would certainly be good if you could check and – if I am not mistaken -  fix the issue for future readers.

Figure legends seem to be ok, but to me it looks like printed figure 5 should be figure 3 (lateralization) because it contains PLG which is not in the other two figures where it is used as a reference region. Likewise fig 3 should be fig 4 and fig 4 should be fig 5 (assuming the caption is ok).

I could not completely figure out: was there a formal statistical test for sex differences (ASG left-right balance negative in male vs positive in female)?

No – we simply present the results separately. We do have an earlier paper just looking at sex differences in healthy people

https://pubmed.ncbi.nlm.nih.gov/28435658/

How many time-points (‘during and after the challenge’) were included in the RMANOVA? From the tables in the Harvard Dataverse I gathered it is 38 ((16s+60s)/2s).

Yes, there is one timepoint per 2 sec. Note that we average the three handgrip task & recovery, so we consider (for example) 8 time-points during the handgrip, where each time-point is the average of 3 handgrip tasks.

A main effect of time means that the average BOLD signal (% change from the baseline) -  pooled across all participants (OSA and Controls) at each time-point - is not completely stable over time. Is this correct?

That is correct – it is saying that is a change from baseline, accounting for the other variables

But this does not tell you when in the time-series the deviation(s) occurred?

For this, we use the within-group timepoint-specific tests. We only do these tests if there is an overall time effect; if there is, we do post-hoc tests are each timepoint – but only within group, since the groups are a priori considered different.

We did publish the method here:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4995690/

For this you did t-tests per time-point, as in the excel files in Dataverse. Why is the StdErr the same in all data-points but the first (‘0’)? Is the first one special? (In the plots it looks like the period before the challenge (including the timepoint 0) has the same low error-bars and the later time-points might have the same higher error-bars). Is this related to the normalization? Why does not every time-point have its own StdErr?

Those sheets come from the processed data from SAS. This may help, but let me explain perhaps more simply

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4995690/

RMANOVA as implemented in SAS has different options for considering covariance of each measure (covariance structure – see attached). The most liberal is to allow the SE to vary at each timepoint for each group; this is the CSH option. However I used the CS option, which assumes similar covariance, hence similar SE (for timepoints with the same number of measures. Since the baseline is all lumped into 1 time-point, it has more measures than others and hence a lower SE). While CSH option is appealing, it exponentially increases computational time and in practice makes no difference in sensitivity, so after testing a few times I stopped using it.

All 0’s: this is because we analyze the change from baseline (since fMRI is relative). So, the baseline mean is calculated, the subsequence points are % change from that mean, and the baseline is set to 0.

The ‘increased anterior insular’ activity (in ASG) (-> discussion) is with respect to the PLG activity which is mainly negative in the beginning of the challenge? Is there increased activity without this normalization?

You’re correct, we can look at the intrinsic signals in Fig 2, and see that “increased” may be better read as “increased with respect to PLG”, or “less decreased with respect to PLG”. In this case, the wording in the discussion is perhaps too simplistic.

Some of the significant time-points (male right ALG vs. PLG: 2s, 4s) seem to be very early in the challenge (given the slow hrf). Do you think these can be effected by neural activity (instead of vasculature or motion)?

I have also wondered about this neural vs vascular/motion effect. I have done several other challenges where there are early effects, and it’s probably true that early signal is not always neural. However, motion per se should be mostly corrected in preprocessing. That leaves vascular responses – but in theory they should be global so the detrending should remove their effects. I suspect there are non-linear effects not corrected by the preprocessing – and in some cases we have observed anticipation effects, meaning the fMRI changes BEFORE the challenge! Another set of data where we moved people’s feet (currently under review) showed big (non-neural %’s) signal drops at the time of foot movement, even with motion correction etc. These were clearly movement-related, but could be indirect – BP and muscle blood flow changes due to lifting the leg, for example.

Anyway, here are studies where I looked at the actual global signal….

https://pubmed.ncbi.nlm.nih.gov/25166862/

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4148259/

Finally, here is a simple brute force method I use for remove global effects –like I say, I don’t think effects are linear, or even uniform over the brain; this method would remove global signal, but (for example) vascular effects are not occurring at the same time (meaning flow takes time to go to some parts of the brain, an issue in ASL processing), so vascular effects are uneven and basically impossible to truly remove – especially since the BOLD signal IS a vascular effect.

https://pubmed.ncbi.nlm.nih.gov/15110027/

I think the take-home is that we should be careful how we interpret fMRI. Of course, you are working with leading experts in the field!! 😊

Thank you very much in advance,

Yours sincerely,

Carsten Schmidt-Samoa

UCLA HEALTH SCIENCES IMPORTANT WARNING: This email (and any attachments) is only intended for the use of the person or entity to which it is addressed, and may contain information that is privileged and confidential. You, the recipient, are obligated to maintain it in a safe, secure and confidential manner. Unauthorized redisclosure or failure to maintain confidentiality may subject you to federal and state penalties. If you are not the intended recipient, please immediately notify us by return email, and delete this message from your computer.