[Eeglablist] general question on ICA and EEGLab

[Scott Makeig]

Peng -

See some comments below.   -Scott Makeig

>dear advanced users of eeglab:
>
>   I have a group of subjects' data, each have several conditions.
>   For each subjects, I imported their data from neuroscan format and
>then filt, epoch according each condition and then did baseline
>correction. I then rejected the bad trials for each condition, after
>which I did ICA analysis for each condition. At last I removed noise
>components (blink etc.).
>   The questions are
>   1 Am I right to did the above? (I think yes according to the
>manual, but need confirmation.)
>  
>
We normally decompose the concatenated data from all relevant conditions 
recorded during the same session for one subject -- assuming that the 
EEG source regions do not change between conditions. This appears to be 
accurate (at least to 1st order) - though if one condition were task 
performance, and the other sound sleep or anesthesia (anything with a 
real state change), it might not be as true!

>   2 is it possible to get weights for each componet in eeglab? There
>were too many components to analysis, I wish to deal with only the
>main ones.
>  
>
An important question is - What are 'the main ones' [components]? For 
this we are preparing a set of functions for clustering components 
across subjects and conditions, based on the component maps (and/or 
dipole locations), and activity measures (ERP, ERSP, ITC, etc.) in one 
or more conditions. We will discuss this software at the EEGLAB workshop 
in Porto, and hope to release it in the coming month (assuming 
sufficient documentation is ready).

>   3 Is there certain direct linkage between ICA components and neuro activities?
>  
>
In our experience and belief, 'good' ICA decompositions of 'good' data 
produce some-to-many 'good' independent components that each represent 
the activity of independent cortical EEG sources (for an example, see 
Onton et al., current issue of Neuroimage for an example, also available 
via sccn.ucsd.edu/publications.html).

However, this is certainly far from claiming that any ICA component, 
from any decomposition of any data, represents the activity of a 
physiological distinct EEG source!

>   4 If yes to 3, is it appropriate to average component between
>subjects? eg. By viewing the wave and distribution of componet A1 in
>sub1 and componet A2 in sub2, ... componet Ax in subx, I think they're
>same or very similar. Is it OK to average them together to form
>componet A. If it's philosophically OK, how to realize in eeglab
>technically?
>  
>
Averaging across subjects can be as (or more) legitimate for component 
waveforms compared to scalp channel waveforms, since there is no 
guarantee that the same mixture of EEG sources project to a given scalp 
site (e.g Cz, etc) in all subjects. However, averaging (or applying 
other statistical measures to) component activities is most useful when 
carried out across a cluster of components that are equivalent to each 
other by some measure and definition -- For example, one could well be 
interested in time-locking ERP or ERSP averaging across ' all mu sources 
whose equivalent dipole is located in or near right hand motor cortex'  
(see Makeig et al., PLOS 2004 and Makeig et al., Science 2002, also 
available via sccn.ucsd.edu/publications.html, for first examples of 
this approach).

The EEGLAB clustering software will have functions (plus a graphic user 
interface) for plotting component cluster mean maps and activities, both 
within one condition or across multiple conditions. A further step is 
to  obtain statistical difference measures for e.g. two conditions. We 
are developing functions to support this, a few of which should be 
available with the first release of the clustering routines.

>   5 If 4 is OK. Is it safe to compare componet A, B, C in condition I
>with compont A',B',C' in condition II? eg. latency prolonged, 
>amplitude decreazed etc.
>  
>
This depends on what you mean by 'safe'! Changes in activities of the 
same (spatial) sources across (task) conditions may well be of 
experimental interest. However, the possibility should be entertained, 
at least, that the cortical areas in which partial local field synchrony 
contributes to the far-field scalp EEG (i.e., the area comprising the 
EEG source) change slightly between conditions -- the nature of the 
boundaries of EEG source regions/patches are not yet observed or 
understood.  A UCSD student is now developing a detailed statistical 
method for asking whether and to what extent separate decompositions of 
data from different task conditions (collected in the same experimental 
session) return identical components. These are not easy questions, 
computationally or theoretically.

Scott Makeig