[Eeglablist] ICA Problem

[Arnaud Delorme]

Dear Habib,

I agree with Makoto that you should not reduce dimensions.
However, you will need to make sure that you have at least 14^2 datapoints (since this is the number of value in the ICA matrix). In practice it is best to have 20 times 14^2 (it is a rule of thumb based on our day to day experience - not based on actual rigorous comparisons).

For finding out which solution are stable, running ICA multiple times and computing correlation is a solution. For correlating scalp topographies, use the "corrmap" or the "mapcorr" EEGLAB functions (and plugin). In practice, it would be better to bootstrap the data then perform correlation of ICA solutions rather than use the same data every time.

When using FASTICA, use the symmetric approach, ('approach', 'symm') which is not incremental. It is better by all means compared to the default (this is the opinion of Aapo Hyvarinen who is the developper of FASTICA). In any case, know that, despite its name, FASTICA is not faster than other ICA algorithms and in our comparison of many ICA algorithms, FASTICA is not in the top performer group.

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0030135

Best,

Arno

On 10 Dec 2012, at 11:41, Makoto Miyakoshi wrote:

> Dear Habib,
> 
> Without any experience with FASTICA toolbox, let me say:
> 1. you should not reduce your data dimension (unless your data length is super short).
> 2. sounds like FASTICA output is not sorted by variance, nor is the polarity taken care of (but this is the original nature of ICA and nothing is wrong with it).
> 
> Why don't you try EEGLAB's ICA. You can run it from GUI. It takes care of sorting after decomposition, and also polarity reversal. It also detects sub-Gaussian components such as 50/60 Hz line artifact.
> 
> Makoto
> 
> 2012/12/6 Habib Paracha <ra_lums at hotmail.co.uk>
> Hi,
>  
> I am performing fast ICA on a 14 Channel EEG data. Reducing the dimension of data from 14 to 5 components. When I perform the ICA I get a different result every time I run it. I am using the FASTICA toolbox available for matlab.
>  
> The Question is how will I be able to figure out which is the most suitable component of all the extracted ICs. I am trying to extract the P300 response from the data. I get the P300 data component randomly in the 1st, 2nd, 3rd or 4th Component. I have to visually figure out which is the best extracted IC and then rest of the code.
> Sometimes I even get the P300 signal but it is inverted(instead of the positive peak ICA outputs a negative peak data).
>  
> Someone kindy tell me a technique to automatically detect the best component and run the rest of the code using this component.
> Also a way to cater the wrong results of ICA.
>  
>  
> Regards,
>  
> Habib Paracha
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
> 
> 
> 
> -- 
> Makoto Miyakoshi
> JSPS Postdoctral Fellow for Research Abroad
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20121212/3a3bf2af/attachment.html>