I ran a decoding analysis with stimuli of two kinds (i.e., social and non-social) from a fMRI experiment. I computed accuracy for each functional ROI I consider in the experiment (e.g., LOC from a previous meta-analysis).
I found on the tutorial (haxby experiment) on the nilearn website, a way to map the SVM weights.
My question is: Are these weights can be used in statistical analysis in order to found a spatial pattern (or something like this) that could be distinct from one group to another? Is this correct? and if not, what can we infer from these weights?
Thank you very much!
Yes, it’s possible, but not at all trivial - just looking for the biggest weights isn’t sufficient.
an example of some of the complications is at:
And here are some relevant papers:
Lee, et al. (2010) Effective functional mapping of fMRI data with support vector machines. Human Brain Mapping.
Haufe, et al. (2014) On the interpretation of weight vectors of linear models in multivariate neuroimaging. NeuroImage.
Schrouff, et al. (2018) Embedding Anatomical or Functional Knowledge… Neuroinformatics. And her PR4NI talk @ OHBM 2017 (https://www.pathlms.com/ohbm/courses/5158/sections/7784/video_presentations/75947).
Gaonkar, et al. (2015) https://doi.org/10.1016/j.media.2015.06.008
Thank you very much for your very helpful answer and for the papers!