[Eeglablist] Help regarding AMICA

[Naeem, Jasim]

Hello all,

My team and I are trying to use multi model AMICA via EEGLab to characterize changes in brain states during mobile EEG imaging task. We do not yet know how many brain states to look for, so we have been performing multiple runs with different model counts (2, 4, 6, 8). At higher model counts we find that some models are overfitting to bad data, denoted in the AMICA model probability plots as areas with positive log-likelihood values. Apologies if the following image does not load properly:

[cid:image003.png at 01D9A372.2A5BF890]

Looking at the data at those times where the log-likelihood is positive in the component space shows suppression of about half of the components in the data. We believe that this is the result of using ASR to clean the data. Additionally, calculating the sliding rank of the data shows that the data is rank deficient throughout, about half of what it was pre-cleaning.

My team and I have two questions-

1: Without knowing how many different states we expect, how would we decide what number of models to use for AMICA? Is there a way to quantify the usefulness or validity of a model?

2: If a higher model count is necessary, how do we prevent models from over fitting to bad data? We have a few potential ideas for this, including downselecting the number of channels before running AMICA, setting a lower number of sources to look for in AMICA, and using ICA to remove problematic sources from the data before AMICA, but we don't know what the best way to utilize a higher number of models while maintaining the integrity of both the data and the results is.

We're open to any suggestions and any help would be appreciated.

Thank you,

Jasim Naeem