[Eeglablist] Should Average Reference Include Eye Channels?

Mon Feb 27 05:41:07 PST 2012

Thank you very much Joe.   I've copied and pasted critical sentences from
your message to organize my response.

"I definitely wouldn't drop the EOG channels until after you've dealt with
the EOG artifacts in whatever manner."

My experimental paradigm is trying to use increased P1 amplitude as a
measure of covert attention capture, so during stimulus presentation I do
not want any blinks or saccades present at all.  I am primarily using the
eye channels to identify blinks and saccades while the stimuli are present
on the screen and removing those epochs.  My trials are very very short and
I have 1000 of them so I can afford to lose them.  I am going to be doing
my data analysis in component space and so won't be using ICA for artifact
removal.

So my goal is simply to clean the data as effectively as I can to prepare
it for a single optimal ICA decomposition.  I was originally using the mean
of the two mastoids as a reference, but I found that if I switch to the
average reference the data appears much cleaner.  While I inspect the data
for blinks and saccades, remove bad epochs, and remove bad channels I just
want to have the cleanest data to look at as possible.  So in that context
and for that purpose, it seems I should just use the scalp electrodes for
the average reference?

Since my target component is occipital and I am not using ICA to remove eye
artifacts it seems like it is simplest and defensible to just remove the
eye channels from the data processing stream once blinks and saccades have
been identified and removed.  Does that make sense?

"That depends on how high quality the signal from the EOG channels is. In
this context, "noise" means electrode noise due to poor contact with the
skin and due to physical displacement of the electrode as opposed to
unwanted but real electrical fields from either eye movements or from
background EEG."

Well thank you for that distinction, that makes a lot of sense and hadn't
occurred to me.  It is very possible that the facial electrodes could have
been less stable than the scalp electrodes, as the scalp electrodes are
plugged in to the cap and the facial electrodes are secured with a sticker,
and tend to have more "tug" on them from the way we have the cords set up.
I have no idea.  Obviously they are visibly more noisy, but that of course
could be soley because of the real electrical activity in their face.

"I can't see any way that operations like epoching and baseline correction
can cause problems for average reference.

My understanding is that it theoretically would cause problems for the ICA
decomposition.  I am basing this on this post that Arnaud made a while
ago...

http://sccn.ucsd.edu/pipermail/eeglablist/2008/002539.html

Specifically he said:

"*>Do I need to reference the data before epoching or can I do so **after?<
* If you do not remove the baseline after epoching, it is equivalent.
However, if you do, it is not. It only valid to rerefence non-epoched data.
When you remove the baseline, you introduce some offset in each data epoch
and for each channel. If there is an EEG/ERP source that project linearly
to all channels, the linear relationship will be lost after baseline
removal. We usually do not subtract baseline or use long baselines (1
second). Simulations have showed that ICA is unstable if the epochs
baseline is too short (and baseline is removed of course). It is described
in more detail in David Groppe's conference papers."

As a matter of fact, since I am interested soley in isolating the P1 I have
decided to filter at 1hz based on the logic that this should attenuate the
P1 minimally.  Unfortunately there is still drift present in some of my
data even with this aggressive filter, so I have also pursued the alterate
strategy suggested by David Groppe of taking the mean of the entire epoch
and subtracting it to serve as a kind of additional high pass filter.
This does indeed seem to get rid of the remaining drift, and from what I
gather should minimally affect the quality of the ICA decomposition.  I am
planning on removing the prestimulus average after ICA has been done.

"Not sure what you mean by baseline correcting and not epoching though, but
agree that it sounds problematic."

This is a misunderstanding from my vague language I think, I was just
referring to the idea that it is the removal of the baseline that is
non-optimal for the ICA, not epoching itself.  I meant to confirm that it
was a possibility to epoch but NOT remove the baseline, then run ICA, then
remove the baseline.

"5) If you rereference only the 128 channels and then add in the 4 EOG
channels then the data will be badly distorted. "

I don't believe that I did this but I'm not sure.  What do you mean by add
in the eye channels?  I didn't remove them, I simply indicated that they be
excluded from the calculation of the average reference.  Is that what you
mean?  Or do you mean specifically that it will be problematic for some
later process like ICA?  Because if that's the case then it shouldn't
matter since I am not planning on including them after I reject epochs
containing blinks and saccades (again because I am not using ICA to detect
and remove eye artifacts).  For that purpose what I did didn't seem to
change things too much.

Thanks for all the help.

-Matthew Stief

On Mon, Feb 20, 2012 at 7:24 PM, Joseph Dien <jdien07 at mac.com> wrote:

>
> 1) One question is whether you're re-referencing before or after dealing
> with the eye blinks etc.  I definitely wouldn't drop the EOG channels until
> after you've dealt with the EOG artifacts in whatever manner.  Keeping the
> EOG channels in there during the EOG artifact correction/rejection process
> can only help.  I normally use ICA to remove eye blinks and deal with other
> types of artifacts and bad channels before going to average reference to
> avoid spreading the artifacts around to the other channels.  The high
> amplitude of the EOG artifacts in the EOG channels can only help.  Think of
> it as being especially high signal-to-noise (where signal is the blink
> activity that you're trying to characterize and then remove).
>
> 2) Once you've dealt with the major EOG artifacts, then it's less clear
> whether to keep them or not.  I generally leave them in since with a
> high-density montage the distinction between an EOG channel and a regular
> channel is fairly arbitrary (assuming they are all referenced to a common
> reference channel(s), as you indicated yours are.  I do usually leave the
> EOG channels out during source analysis since they are more susceptible to
> signal leakage out the optic tract, which distorts source analysis
> algorithms that don't account for the presence of the optic tract.
>
> 3) Regarding the average reference itself, keep in mind that some of the
> signal from the EOG activity is going to end up in the other channels
> anyway (which is why we record EOG channels in the first place) so as far
> as the average reference goes, the more completely you are characterizing
> the fields the better.  The flip side is whether you might be keeping more
> noise than signal.  That depends on how high quality the signal from the
> EOG channels is. In this context, "noise" means electrode noise due to poor
> contact with the skin and due to physical displacement of the electrode as
> opposed to unwanted but real electrical fields from either eye movements or
> from background EEG.  The latter is also ending up in the other channels so
> characterizing that real electrical activity is good.
>
> 4) I can't see any way that operations like epoching and baseline
> correction can cause problems for average reference.  Average reference is
> computed on a timepoint by timepoint basis.  It makes no difference what is
> happening at other timepoints.  In fact, informally, I've found that you
> get better results from data that are both epoched and baseline corrected
> prior to performing ICA for eyeblink correction (this is how my EP Toolkit
> does it).  This seems to be the case because it helps deal with slow high
> amplitude variations in the EEG data that can otherwise swamp out the EEG
> and EOG activity.  Whether this is an issue likely depends on the filter
> settings and the nature of the equipment (I observed this with an EGI
> system with a digital .1 high-pass setting).  Been meaning to write up my
> automated artifact correction algorithm up.  Hope to do so soon.  If you'd
> like to try it out, it's at:
> http://sourceforge.net/projects/erppcatoolkit/
>
> Not sure what you mean by baseline correcting and not epoching though, but
> agree that it sounds problematic.
>
> 5) If you rereference only the 128 channels and then add in the 4 EOG
> channels then the data will be badly distorted.  Don't do it!
>
> Cheers!
>
> Joe
>
>
> --
_________________________________________________________________
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/bdcc0733/attachment.html>