Controlling confounding variables in nested cross-validation (nilearn)

burgerks · February 3, 2020, 5:27pm

Hi all,
We are considering collapsing across a couple task-based fMRI samples that assess the same construct with slightly different stimuli and on different magnets. Specifiically, they present different tastes (kool-aide, milkshake) with the same control taste for contrast, but also with different paradigm designs. So, I’m trying to figure out a way to control for confounds in our nested CV approach. I’ve come across a few articles re: controlling for confounds in CV. Hoping to apply this to nested CV. n > 780 scans total.

any thoughts?
thanks
Kyle

GaelVaroquaux · February 3, 2020, 5:58pm

We have just submitted a manuscript on this. I hope that it will come out soon.

In the mean time:

Some code for the manuscript
https://github.com/darya-chyzhyk/confound_prediction
A short paper on the topic:
https://ieeexplore.ieee.org/abstract/document/8423961

I hope this helps.

burgerks · February 3, 2020, 6:17pm

Thanks so much. Looking forward to it