Repeat ICA after an initial ICA

Hi Mike and community,

Hope you’re all doing great,
I’m analysing an EEG data set where my pipeline is basically import data, filter, remove non-stereotyped noise, run ICA and unselect “bad” components (I try to eliminate no more than 2 or 3 components), do epoching, and then do the usual averaging. I’m dealing with some very noisy data and I was wondering if it would be advisable to try a second ICA decomposition just after I epoch the data (since I’m using a tool [SASICA] to visualize how the component-related activity change in different epochs this is also nice because I can choose to delete a given trial instead of the whole component). What would you think?

Many thanks for your input,
José Luis

Have you tried the ASR plugin for EEGLAB?
It is excellent at removing/repairing artifacts in noisy data.
https://sccn.ucsd.edu/wiki/Artifact_Subspace_Reconstruction_(ASR)

Definitely run ICA again. This is an accepted approach: Clean the data, then ICA and remove bad components, then run ICA again on the cleaner data. Hopefully, the second ICA will better isolate the artifacts.

I’m not familiar with the subspace method John linked to below, but that’s also worth looking into.

Ultimately, the best approach is to try various cleaning methods on one dataset, then apply that protocol to all the other datasets. Data preprocessing tends to be idiosyncratic to each experiment, hardware, analysis goals, etc.

Thank you both!
I’ll check ASR