GLMsingle across sessions and trial durations

[AnnekeGoblirsch]

Dear GLMsingle community,

We want to use GLMsingle for a thorough analysis of our real-time fMRI Neurofeedback experiment but are a bit unsure how to best apply it.

Our paradigm includes localizer runs (to identify emotion-related brain patterns) followed by neurofeedback runs using the same conditions but longer trials. We collected two sessions the first focuses on localizers and task introduction, while the second emphasizes neurofeedback training. We now want to perform an offline analysis to evaluate the accuracy of our real-time results and improve the paradigm.

We are unsure what to do about trial duration since our trial structures differ:
Localizer: 15 s trials, 15 s baseline (3 runs in the first session and one in the second)
Neurofeedback: 24 s trials, 19.5 s baseline (one run in the first session and 3 runs in the second)

Our current plan is to analyze localizer and neurofeedback runs separately, pooling across sessions and including a session indicator. This would mean running GLMsingle once for all localizer runs and once for all neurofeedback runs.

Does this seem like a reasonable approach?
Additionally, for comparisons across sessions and tasks, would you recommend any normalization steps (e.g., mean-centering betas within runs), or alternative strategies?

Thank you very much for your help!

[kendrickkay]

Our current plan is to analyze localizer and neurofeedback runs separately, pooling across sessions and including a session indicator. This would mean running GLMsingle once for all localizer runs and once for all neurofeedback runs. Does this seem like a reasonable approach?

Yes, this definitely seems reasonable. However, I guess I am wondering about the approach. From the point of view of GLMsingle, it is attempting to boil down the fMRI data into a single "trial-wise response amplitude" (i.e. beta). In your experiment, are you fine with that? Are there any potentially interesting dynamics within your very long trials (15 s, 24 s) that you care about?

Additionally, for comparisons across sessions and tasks, would you recommend any normalization steps (e.g., mean-centering betas within runs), or alternative strategies?

For sure, there are many factors that can change across scan sessions, some of interest and many presumably not of interest. This requires a deeper dive into what the goals/assumptions are for your experiment. Another critical issue is the interpretation of betas (response estimates) across the different trial durations. Ultimately, for the two cases (15 s vs 24 s), GLMsingle internally constructs the predictor (convolution of selected HRF with the specified duration) and normalizes the resulting predictor to peak at 1. Then, betas can be interpreted as scale factors on this (and then converted to units of percent signal change). But, this might be apples-and-oranges in the sense that the timecourse for the 15 s case is sort of fundamentally different for the 24 s. See attached figure. 

[AnnekeGoblirsch]

Thank you for your reply.

You raise a good point about potential dynamics within trials. We do expect that there may be meaningful within-trial variation given the long durations. However, our current research question focuses on the stability of spatial patterns across preprocessing approaches and between the localizer and neurofeedback tasks. Since our real-time analysis is based on RSA, we want to compare those real-time results to the same analyses performed on more thoroughly preprocessed data. After these analyses we will also look into the processes happening within each trial but then using other more suitable methods.

Regarding the normalization to 1 that the tool does internally, we don't think it is a problem for us as our focus is on spatial patterns rather than absolute amplitude.

For the RSA, our plan is to reconstruct t-statistics from the beta estimates after ridge regression (so reconstructing the GLM and getting the residual variance). This would be important for us since t-statistics add a step of pattern normalization and should therefore make the RSA more stable.
Our question now is would this reconstruction of t-values be a valid approach?

Again thank you very much for your help!

[kendrickkay]

Re: reconstructing t-statistics.... that's a thorny issue. First, the use of ridge regression (which is a departure from standard least squares) implies that the usual notion/calculation of t-statistics is not really applicable. Second, GLMsingle unfortunately does not make it easy to "reconstruct the GLM / residual variance stuff", so I'm not sure I would try to do that (or at least, it's quite tricky). If your goal is to simply get units that are a bit more "stable" or more indicative of which voxels are reliable, perhaps you could lean on the typeD R^2 and/or the FRACvalue values. Alternatively, if it is possible, you could quantify how reliable the single-trial beta estimates are (e.g. across trials that are comparable) and use that to weight your voxels, or something like that.