Seeking Assistance with GLMsingle Model Anomalies

[Yujie-cog]

I am currently employing GLMsingle with default parameters in Python for rapid event design analysis in fMRI research. My dataset consists of 9-11 runs per subject, with a TR of 1.5 seconds. Each trial in my study involves a 0.5-second stimulus presentation following a 0.5-second fixation period and a jitter ISI.

However, I've encountered an issue. After GLMsingle model fitting, I observed anomalous results manifesting outside the brain region, which has left me perplexed about the potential causes.

For context, I have preprocessed the fMRI data using SPM, including steps like Realign, Coregister, Segment, Normalize, and Smooth. I did not perform Slice Timing as I found it challenging to apply to multi-band sequences. I have attached my code, example of design matrix .

Your guidance and insights would be greatly appreciated. Thank you, Liu Yujie ! GLMsingle_code_for_issue.zip

[kendrickkay]

Hi, a few thoughts:

• What kinds of ISIs do you have? If your trials come too rapidly, it will be challenging to separate out adjacent trials' responses
• Given the figures you show, it's not clear to me whether there is any strong signals. Are you sure your design matrix is 100% correct?
• However, it could be that your experiment just generates very weak evoked responses, so it might be the case that nothing is "wrong".
• Have you looked at the figures written to disk? These figures that GLMsingle outputs can be very informative.
• Regarding your question about outside the brain stuff, I don't see anything particularly concerning. See this: https://glmsingle.readthedocs.io/en/latest/wiki.html#why-do-the-betas-look-crazy-outside-the-brain

[Yujie-cog]

Thanks for your response!!

1. ISI in my data is 2-5s. TR = 1.5s. In every sample trial, the flow is : fixation(500ms) - stimulus(500ms) - ISI(jitter), there's another response stage in catch trials(they are coded as a condition). The onsets of data correspond with the start of fixation in trial instead of stimulus. One of my condition is Instructions (they are presented for 4 s). I am also kind of unsure about how to choose 'stimdur' for this data and design ?

2. About the possibility that experiment just generates very weak evoked responses, I know another person used this data on some trial-based decoding and RSA, so I think it shouldn't be this case .

3. I didn't find any mistake in the design matrix up to now x), but I will double check all of them.

Here's one example design of a run, I set about last 10s as 0 in analysis.

[kendrickkay]

OK, what you said seems fine. I don't necessarily see any problems. Do you think there is a problem?

Regarding stimdur, you can set it to 0.5 or 1 (second). It won't make a big difference, as the shape of the HRF predicted for 0.5 or 1 second of stimulation is very similar.

[iancharest]

A close look at your design matrix suggests that you did not provide "single trial" predictors, which is how GLMsingle expects designs to be input.

[kendrickkay]

@iancharest - can you elaborate on what you mean?

[iancharest]

ok, so looking at the design matrix above, I can see that the trials are concatenated with multiple blips per condition in a column. GLMsingle is optimised as the name says, for estimating single trial betas (which could in the end be averaged across conditions if this is what one wants.) here is an example trial based design matrix:

The following example is perhaps useful to then average the single trial betas and convert to t-volumes/maps.

https://github.com/cvnlab/GLMsingle/blob/main/examples/example5_average_single_trial_betas_t_stat.ipynb

[kendrickkay]

Ian, yes, but remember that GLMsingle wants a design matrix with the onsets associated with a given condition to be indicated in a single column (and not separated into multiple columns).