[Eeglablist] effects of re-lowering electrode impedance during recordings

[Tim Mullen]

Hi Michael,

If impedances change dramatically over a session, this could be a potential
concern for ICA, since this can certainly change the statistics of the data
-- particularly the variance/scale.

However, bear in mind that the gradual change in impedance due to gel
drying over a typical recording session is an equally significant form of
non-stationarity insofar as Infomax ICA is concerned -- though ICA seems
fairly robust to this. I don't have much personal experience with recording
such long sessions that gel needs to be reapplied, but if reapplication of
gel only serves to restore impedances to their approximate original values
from the start of the session, and the gel is reapplied frequently enough
so that you don't have a massive range of impedances over the entire
recording session, my opinion is that -- although stationary assumptions
are being violated -- ICA will be sufficiently robust to this and you will
be able to recover suitable components (others with a different experience
may wish to chime in).

However, bear in mind that since a change in impedance may produce (among
other things) a change in scale of the channel data ("X" part of X=AS), and
since Infomax ICA will return a single unmixing matrix for all blocks
(fixed "A"), this change in impedance can translate to a change in scale
for some IC activations ("S" part of X=AS). So you might see prominent
changes in amplitude of IC activations for trials following block
boundaries. This will be apparent if you plot an ERPImage of the epoched
ICA activations for a given component. However, since ICA activations have
arbitrary units of measurement, one possibility is to temporally
z-normalize the activations for each trial independently to adjust for
this. Note, however, that temporal normalization may have other
implications for interpretability of ERPs, ERSPs, etc and for certain forms
of signal processing, which you'll want to consider.

I would also caution that you take extra care to ensure that electrodes do
not move during the re-application of gel as this could dramatically impair
ICA.

As an alternative to Infomax ICA (runica) you might also consider using
AMICA with the number of models chosen to match the number of blocks
(assuming gel is reapplied after each block). If the data distributions are
significantly different between blocks, AMICA may be able to learn an
optimal ICA model for each block. Ergo, problem solved.

Insofar as MVAR model fitting, the non-stationarity in mean and variance
can be mitigated by normalization of the data (typical preprocessing step)
and the use of an adaptive modeling approach (sliding-window or kalman
filter).

Tim


On Fri, Jan 3, 2014 at 6:08 AM, Michael Schubert <mschuber at mail.upb.de>wrote:

> Hi Tim and others,
>
>
>
> if I record several blocks of EEG, I might need to re-lower impedances
> with extra gel due to drying etc. in-between. When I then concatenate
> blocks as raw signals to do my processing, can that be bad in terms of
> signal stationarity and thus affect ICA decomposition and MVAR model
> fitting?
>
>
>
> Best,
>
> Michael
>



-- 
---------  αντίληψη -----------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20140104/c061d89a/attachment.html>