[Eeglablist] Interpolation and ICA Update?

[Matthew Stief]

Well I have a dense electrode array (128) so I can afford to be somewhat
draconian.  However the bad electrodes do tend to be in groups so I worry
about the ICs and dipole fittings for components in the areas of these
gaps.  I have some participants where I think they must have shifted the
frontal electrodes somehow, or perhaps started to sweat a lot or some such,
because all of the electrodes toward their eyes start to get a lot of high
frequency noise (given that I have already run a low pass filter at 55hz)
halfway or toward the last third of the experiment.   So in some cases it
looks like I should remove almost all of the frontal electrodes (meaning
the first two rows or so).  Which may be fine for me since I'm interested
in the visual P1, but then again it may cost more than it buys me.  The
severity of this high frequency noise also tends to be in a gradient out
from a little clump, so for example I may have two or three very noisy
electrodes, and then the ones immediately above and below it are somewhat
less noisy, etc.  So I'm not sure where to cut it off, and could cut out a
little clump of five or six electrodes in some cases.  The high frequency
noise electrodes also generally seem to have some nicely stimulus-locked
activity underneath the noise so I don't know if I need to remove them at
all.  The electrodes that deviate more idiosyncratically and with larger
amplitude changes are more obvious to isolate out and tend to be on their
own and so are easier to deal with.

Regarding interpolation, I am not planning on backprojecting to channel
ERPs, though we were considering presenting our analysis both in ERP-space
and in component-space as a methodological demonstration of the utility of
ICA.  But even assuming we just work with components could you elaborate a
bit on the issue of needing equivalent scalp maps across ICs.  What exactly
do you mean by equivalent?  For example if I use scalp maps as one of the
variables to feed into the clustering process, does that mean they need to
be "equivalent"?

However, I would still love to have people comment on whether or not there
is ANY theoretical benefit from including the virtual channels in the ICA
to balance against the deficit from the introduction of nonlinearity.
Assuming you use PCA to reduce the dimensionality from the loss of
channels, which also introduces nonlinearity, then how bad for the ICA
decomposition is the nonlinearity introduced by the interpolation and is
there any reason to think that given that deficit there is also a benefit
from having the "extra" data from the virtual channels?  I understand that
the archived discussions seem to conclude interpolation afterwards or not
at all is defensible as long as bad channels are removed, but has anyone
ever actually tested it to compare the two approaches?  Is there a paper I
could be directed to?

I would also be grateful if anyone could help me understand why the
polarity of IC scalp maps is arbitrary.  I honestly am confused by that
point, and in particular am unsure of what it implies for calculating the
mean amplitude in some latency window for a cluster ERP if the ICs that go
into that cluster have arbitrary polarities.

Grateful as always.

-Matthew


On Sun, Feb 19, 2012 at 2:34 PM, Tarik S Bel-Bahar
<tarikbelbahar at gmail.com>wrote:

> Leave questionable electrodes out before ica.
> If you have a sparse electrodes array, dont be draconian. some of my best
> decompositions come with discounting up to 40% sometimes, although I prefer
> not to remove more than 25% or so.
>
> Interpolate if you must after ICA.
> After decomposition, Normal ic analyses should not require interpolation,
> unless you need equivalent scalp maps across ICs, or you are backprojecting
> to scalp/erp land.
>
> You could run a second Ica yourself, after interpolation, to see what
> happens to extant IC s
> in a new, second decomposition.
>
> Please try  be sure to feed your final conclusions back to the list.
>
-- 
_________________________________________________________________
Matthew Stief
Human Development | Sex & Gender Lab | Cornell University
http://www.human.cornell.edu/HD/sexgender


Heterosexuality isn't normal, it's just common.
-Dorothy Parker
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120227/3d779094/attachment.html>