[Eeglablist] Running ICA with 128 channels

[Tarik S Bel-Bahar]

>
>
> Hello Heather, some notes below that should help you. Best wishes.
> ​there were two typos below which I corrected :)​
>
>
> *******************BEGIN NOTES FOR HEATHER
>
> 0. Generally speaking if one does not meet ICA's constraints (enough time
> points relative to total amount of channels) then it is totally possible
> that the solutions are not stable, or do not "settle" accurately on ICs.
> Note it's also possible that the artifacts you are interested in are not
> well-represented enough, and "regular/stereotyped" enough for ICA to
> accurately detect them.
>
> 1. Make sure that the visual cleaning
> ​d​
> oes no
> ​t​
> involve removal of eyeblinks and stereotyped artifacts (such as EMG). ICA
> needs clear information in the data in order to "retrieve" good ICs.
>
> 2. Although some some researchers have published ICA results for
> high-density systems with short time periods, it's probably best to reduce
> some of the information in you data if you're worried about the ICA
> results, and then compare. Reducing can involve dropping a set of channels
> (e.g., 32 or 64) and also downsampling (e.g, cut the sampling rate in half).
> However, I and others have had success at least pulling out artifactual
> ICs from ICA of high-density data with relatively short time periods (e.g.,
> 2 to 5 minutes). There are many articles with 32 and 64 channel data with
> short time periods that use ICA for cleaning at least.
>
> 3. Not all ICA solutions for every subject are stable or great. There are
> usually some percentage that are a bit weird. In the case of multiple eye
> blink ICs, I usually remove them if they are within the first 25 ICs, even
> if there is more than one. I don't remove them when they seem thoroughly
> mixed with other scalp patterns, suggesting that those ICs have blink +
> other patterns within them.
>
> 4. You may want to try ASR or PREP (both eeglab plugins) that may get your
> data cleaner before visual inspection and ICA. You may also want to try PCA
> to reduce total number of components (which may "merge" some of you similar
> components).
>
> 5. You may want to only remove the truly worst periods, run ICA, then
> apply ICA to the continous data that has received no cleaning, then remove
> clearly artifactual ICs, and THEN
> do the visual cleaning. This should
> ​increase
>  total time left in your conditions.
>
> 6. You may also benefit from the below if you haven't had a chance to yet:
> a. Luca's IC classification training site (just google eeglablist luca IC
> classification)
> b. Makoto's general recommendations for ICA pipelines
>
> 7. If you keep having issues with the processes you raised, I recommend
> sending more details and pictures of ICs at least.
>
> 8. try one or two subjects with several different cleaning and ICA paths,
> and compare across them. This might help you settle on best-possible
> ​pipeline
>  for your data.
>
> 9. Review publications that use iCA with 128 channel data and relatively
> short time periods for further  ideas.
>
>
> *******************END NOTES FOR HEATHER
>
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170605/ba34406c/attachment.html>