Robyn -

Very good project!
I'd be happy to have a look at relevant parts of your
dissertation (if/when convenient) or to answer any
questions that come up (as possible).

Re nonlinear ICA - Summation at electrodes is linear
(up to megaHz, Nunez estimated), so ICA is of use
to perform spatial separation to undo this linear mixing.

While nonlinear ICA is possible (with constraints and,
so far very few channels - see ICA99 and ICA00 meeting
books), you should think very clearly about what
nonlinearity you want ICA to model. If you know it
well enough (e.g. exp()) then linearizing the data (e.g. log())
should allow linear ICA to work well.

For epilepsy, the problem of propagation is important.
Say you find 10 spatial components that account for
some aspect of spikes in the data.

Then, by delaying the component activations appropriately,
you might try performing a second ICA on the 10 activation
time series (delayed relative to one another to overlap
the spike activities?) to find components of the spike 
trajectory implied by your delay choice. 

Else, you might take your 10 spike-producing components
and delay them in N successive 1-frame increments,
making an 10*N-times matrix that you could decompose.
Again, ICs of this matrix would be trajectory/movie components,
with the columns of winv() being weights. By back projecting,
undoing delays, and back projecting again, each such component
could be visualized as a movie (see eegmovie() code). 

The interesting thing would be to find out whether there
is just one such dominant movie component for a given
subject...

Good luck on your important and interesting project.

Scott

P.S. Incidentally, a real ICA (though not EEG) expert at CMU
is Mike Lewicki, a very clear thinker and nice guy as well.

P.S.S. Martin McKeown here worked on epilsepsy data some with ICA
(look for a reference on my web site or on the cnl web site 
if I haven't added it yet). 

P3S. May I use your question (anaonymously) for the tutorial?

P4S. I have added much to the toolbox tutorial since my last message.
        www.cnl.salk.edu/~scott/tutorial/



	Date: Fri, 26 Jan 2001 16:59:02 -0500
	From: Robyn Bates <rrbst10+@pitt.edu>
	Subject: Re: ICA/EEG
	To: Scott Makeig <scott@salk.edu>
	Cc: "Bates, Robyn Rebecca" <robyn@neuronet.pitt.edu>
	Organization: Neurological Surgery & Biomedical Engineering

	Dr Makeig,

	Sorry I didn't get back to you sooner -- I'm in throes of
	dissertation panic.

	Anyway, thanks for the tutorial information.  I've been
	using your ICA package to analyze 94-channel epileptic
	EEG data with the goal of isolating the most prominent five
	or six signal sources (if they exist).

	I was also wondering if there is a nonlinear extension to
	your MATLAB package, or how one might go about
	developing such an extension.

	Thanks for the attention!

	Robyn

	================================================
	Robyn Rebecca Bates
	Clinical Neurophysiology & Biomedical Engineering
	University of Pittsburgh
	robyn@neuronet.pitt.edu



