Hello all,
I wanted to know if it was possible using TDT to decode whether a subject is from group A or from group B (I searched through tdt tags, not sure I was exhaustive enough, but I didn’t find this question, if someone does, please feel free to forward the thread 
).
I have two groups of participants, of 19 subjects each. They went under the same experiment in fMRI, with 5 conditions, but consider it was 2 for simplicity.
I know how to decode between these 2 conditions for each, but I wanted to know if it was possible to do the opposite?
If yes, I don’t know what would be the best method.
For example, should I train the decoder on maps of decoding between these 2 conditions for each participant, telling him if they are from group A or B, then test it on these maps? Or should I train him 4 different conditions, i.e. groupAcond1, groupAcond2, groupBcond1, groupBcond2?
Or something else?
Thanks all for your answers!
Best
Fabien
Hi Fabien,
It really depends on your goal. Are you interested in the difference between conditions and want to know how that is affected by group? Then I’d probably run decoding of conditions and see where you find information in the brain (or check your ROIs) and then see how that is different between groups.
Or are you interested in whether the same classifier generalizes between participants? Then perhaps in MNI space where you can somewhat make the simplifying assumption that one persons voxel k is comparable to another persons voxel k, you could train the classifier between both conditions on all but one participant (I would probably not do this at the trial level but at the mean beta level in each subject) and then you would test this on the left out subject. Then you could also train it on one group and see how well it generalises to the same group as compared to the other group.
Or are you interested in decoding the effect between groups? Then I’d probably just compute the difference between both conditions and feed that difference as the relevant contrast in the between group analysis. I’d use the difference in mean betas, train on 18 participants on either side, and test on a left-out pair of participants. The stats will just be a little tricky since there are not a lot of independent data points. Maybe you can do something with multiple trials per participant.
In any case, coming back to the question for TDT usage: check the template called …nobetas. That allows you to manually specify what data you want to use for training and testing.
Hope that helps!
Martin
Dear Martin,
Thanks for your answer.
I’m not sure I understand it all well, but I think what I want to test is something else: based on a classifier I would have trained on (? beta, contrasts, I don’t know?), I would like to be able with this classifier to decode based on what I have trained it on, to classify a participant into the 2 different groups, am I clear?
Best
Fabien
Hi Fabien,
Ok, great! This was my third suggestion. Hope that helps.
You may still want to use the template called something with “nobetas” and then manually add the paths to all relevant file names. I would probably suggest to use the contrast between pairs of conditions.
Best,
Martin