I'm working toward an ICA analysis on healthy controls and AD patients for the first time and thus I am very interested in making sure I appropriately deal with motion confounds.

I've been looking through past message board correspondence on this topic and I can't seem to find anything too conclusive wrt just when/where to include the fsl_motion_outliers output as a confound EV (specifically using the Power (2012) methodology).

It seems like the most recent discussion about this topic suggested doing so at the single subject level, PRIOR to doing any group ICA/dual-regression steps (i.e. doing a single-subject melodic run and including the output in the post-stats model/contrast). Is this correct?

Some of the messages also discuss interest in developing a melodic-based classifier to automatically pick out/remove motion-based components. This is yet to be incorporated in FSL, correct?

What are other folks doing to address this issue?
--