Using GLMdenoise for n-fold cross-validation decoding

[Alireza Karami]

Hello

My experiment has six runs, and I want to do a 6-fold cross-validation decoding analysis. Each condition is repeated twice in each run.

As far as I know, GLMdenoise needs at least two runs for denoising. One solution that comes to my mind to have six independent runs is to denoise every two runs together (e.g., splits = {[1 2] [3 4] [5 6]}), but by using the condition-splits strategy suggested here [1], I can split different trials of the same experimental condition within a run into separate conditions. As there are two trials of the same experimental condition in each run, I will end up having six independent runs, and I can do a 6-fold cross-validation decoding analysis. Am I correct?

[1] A critical assessment of data quality and venous effects in sub-millimeter fMRI; Kendrick Kay, Keith W. Jamison, Luc Vizioli, Ruyuan Zhang, Eshed Margalit, Kamil Ugurbil

Best regards,

Alireza

[Kendrick Kay]

> My experiment has six runs, and I want to do a 6-fold cross-validation decoding analysis. Each condition is repeated twice in each run.
> As far as I know, GLMdenoise needs at least two runs for denoising. One solution that comes to my mind to have six independent runs is to denoise every two runs together (e.g., splits = {[1 2] [3 4] [5 6]}), but by using the condition-splits strategy suggested here [1], I can split different trials of the same experimental condition within a run into separate conditions. As there are two trials of the same experimental condition in each run, I will end up having six independent runs, and I can do a 6-fold cross-validation decoding analysis. Am I correct?
>
> [1] A critical assessment of data quality and venous effects in sub-millimeter fMRI; Kendrick Kay, Keith W. Jamison, Luc Vizioli, Ruyuan Zhang, Eshed Margalit, Kamil Ugurbil
>

Hello,

Sorry for the delay.

Regarding the condition-split idea, what you said is not quite the original idea. If you "split" each condition into two and if you analyze all 6 runs together, you will ultimately just get 2 betas for each condition in total. (This is because the condition-split idea treats the two split conditions as "recurring" across the 6 runs.) That's probably not what you want, since there not much you can do with only 2 betas (trials).

What you could do is something like what you said:

1. Group together runs 1 and 2, and runs 3 and 4, and runs 5 and 6.

2. For each group of runs, perform a condition-split for each condition. This will produce two different betas for each condition.

3. Then, what you will have is two betas from the first group, two betas from the second group, and two betas from the third group. That means a total of 6 betas (per condition) using all your data. Then I guess you do 6-fold cross-validation based on those 6 betas.


It's a little clunky, but might be reasonable. You'll have to think about it.

Alternatively, the GLMsingle approach will try to just deliver a separate beta estimate for all 12 presentations of each condition during your whole experiment. That seems to simplify things, and maybe you could proceed down that path.


Kendrick