[Eeglablist] 2qs: ICA on ERP noise removal

[Arnaud Delorme]

>
>
>1. is there a quantified crietia to determining what components are
>noise? if there isn't, do I need to visually go through all the 128
>components? It is not only very time consuming but also not objective
>because some components are very hard to tell whether they are noise or
>not.
>
This is true. We are working on some tools to automatically select 
artifactual components (I guess this is what you call noise?). Right 
now, you can either select component visually or write your own tools. 
You may use the PCA option of runica to reduce the number of dimension 
(i.e. output components) prior to running ICA.

>2. which kind of data should I apply ICA: single trial, averaged data,
>or continuing data? I think using average data makes no sense in theory.
>And continuing data is not actually continuing in time, because it is
>only a series of connected potentials, and they are not always
>successive. for example, in our experiment, we pause when we see the
>brainwave is a mess. what's more, we usually group the trials by
>conditions (stimuli) before analysis. in that way, trials in the same
>group are very unlikely to be successive in time.
>
Single-trial or continuous data is fine. If you run ICA on continuous 
data, you might first want to high-pass the data at 0.5Hz to remove DC 
currents. The data does not have to be made of successive data points 
since ICA shuffles data points at each step of the training algorithm 
(time has no meaning for ICA: for instance continuous EEG from 64 
electrodes only represents a collection of data points in 64 dimension 
for ICA).

Best

Arno

-- 

*Arnaud Delorme, Ph.D.*
Computational Neurobiology Lab, Salk Institute
10010 North Torrey Pines Road
La Jolla, CA 92037 USA

*Tel* : /(+1)-858-458-1927 ext 15/
*Fax* : /(+1)-858-458-1847/
*Web page *: www.sccn.ucsd.edu/~arno <http://www.sccn.ucsd.edu/%7Earno>