[Eeglablist] epoch baseline removal experience?
Jason Palmer
japalmer29 at gmail.com
Tue May 15 15:11:28 PDT 2012
Hi Yuan-fang,
Some notes in response to your questions:
1) Removing major artifacts, such a large spikes in the data, before
filtering can be preferable since filtering can "spread" the artifact out
over "good" data, requiring more data to be rejected after filtering. When
you remove major artifacts, and "boundary" event replaces the removed data.
Filtering is only applied to continuous data segments, not across
boundaries.
2) The main considerations regarding epoched vs. continuous data are:
a) amount of data given to ICA, and b) stationarity of data given to ICA.
ICA expects the data to be stationary, i.e. the same statistical model is
generating all time points. If you have enough data after epoching, then
epoched data may be preferable since it will increase stationary as you
mention. However, you may want to epoch on different events to produce
different datasets, with the same ICA decomposition for all conditions. More
data generally gives a better ICA decomposition, assuming all the data is
similar statistically.
3) So the idea is to run ICA on the continuous, cleaned, filtered data,
then epoch. Filtering will removed slow drifts in the data, and make the
channels zero mean. There is obviously no baseline until you epoch the data.
It is possible if you only care about one condition and want to run ICA on
epoched data to remove the baseline first before running ICA. Results may
differ slightly (particularly when using multiple ICA models) so you may
want to compare results on your data. In my view, removing the baseline
before running ICA on epochs is not "bad practice", and shouldn't affect the
estimation of e.g. theta and alpha components, or higher frequency
components. The main issue is whether the original (high-pass) filtering
works and produces epochs with "flat" zero mean baseline periods, or whether
the filtering does not work sufficiently well, and your epoch baselines are
not always zero mean. You might also try "detrending" the epochs before ICA
if they are particularly recalcitrant.
4) Longer epochs are preferable because they yield more data for ICA
(assuming stationarity holds.) However, if you are epoching before ICA, you
don't want to give ICA overlapping epochs, since it will then have
duplicates of some data skewing the statistical model. So the epoch start
time should not be before the stop time of the previous epoch, and the stop
time should not be after the start time of the next epoch.
5) Dipolarity, i.e. residual variance (R.V.) of dipole fit, of a
component map is one indication of component quality. If the R.V. is low,
then the component is physiologically localized and concisely accounted for
potentially by a single patch of cortex. A good decomposition should have
many dipolar components. Mutual information can be used to assess how
independent a component is from other components. And the "Quality" of
spectral peaks of independent component activation is also informative. E.g.
theta and alpha components should have stronger theta or alpha peaks than
the raw channel data, and should account for most of the theta or alpha
activity in the data in one or a few components. ERP averages should also be
"cleaner" and "simpler" in the independent components than in the raw
channel data.
Best,
Jason
From: eeglablist-bounces at sccn.ucsd.edu
[mailto:eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Yuan-fang Chou
Sent: Sunday, May 13, 2012 6:41 AM
To: Matthew Stief
Cc: smakeig at ucsd.edu; eeglablist at sccn.ucsd.edu; mmiyakoshi at ucsd.edu
Subject: Re: [Eeglablist] epoch baseline removal experience?
Dear Makeig,
I still feel quite confused about your post.Please forgive me for my
ignorance as a newbie in ICA.
My questions can be outlined as follows:
1)Why should we first remove major artifacts and then filter the data?Can we
invert these two procedures?
2)ICA should be conducted in epoched data or continuous data?If both are
ok,which is better?For continuous data often contains large amounts of
artifacts,which happens during the interval of each trial in experiment,I
think it may be better to do ICA on epoched data.
3)Why should baseline removal be done after ICA?I really don't understand
the reason under this practice.
4)Why longer epochs are more enjoyable for ICA?
5)Are there some indicators which can used to identify if the result of ICA
are good enough to make inference?
Sorry for the long question list and wish for your reply.
2012/5/13 Matthew Stief <ms2272 at cornell.edu>
Hi Scott,
Thanks for this. If you're going to baseline-zero epochs after ICA, then
what's the point of baselining the whole dataset before epoching? Just to
have an additional kind of high pass filter? You're saying that doing this
AND a ~1Hz high-pass filter would be better for the ICA than just doing the
high-pass filter, right? I thought that the advantage of doing the
whole-epoch baseline (and thus also i assume this whole dataset baseline
removal), was that it ameliorated problems of low frequency drift for the
ICA without suffering from the attenuation of large later components caused
by an aggressive high pass filter. So I was thinking of it as an alternative
to high pass filtering, not an addition to it. In my current data processing
strategy I've gone for not baseline removing before ICA at all, and just
relying on an aggressive 2 Hz high-pass filter (all I care about is the P1),
and then doing a baseline removal for epochs after the ICA. But you're
saying doing this big baseline removal and a high pass produces superior
results, right?
Also, I wasn't sure from your e-mail whether you thought the whole dataset
baseline removal should occur before or after filtering. I've been doing
major artifact removal after filtering because it makes bad patches easier
to see, but i'd be happy to do it this way if it creates a better ICA
decomposition to do this kind of total baseline removal.
Thank you!
-Matthew
On Fri, May 11, 2012 at 11:09 PM, Scott Makeig <smakeig at gmail.com> wrote:
Even whole-epoch baseline removal is not ideal. It is better to
zero-baseline the data after major artifact-period removal but before
epoching (and, typically, high-pass filtering above ~1 Hz). Only then
extract epochs for ICA decomposition (IF you do not want to decompose the
continuous data -- our more typical procedure). After ICA decomposition,
data epochs can be individually baseline-zeroed without affecting the ICA
account of them.
Scott
On Fri, May 11, 2012 at 12:31 PM, Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
wrote:
Dear Ida and Scott,
> As I understood, the purpose of Baseline Removal is for me/us to have
> better insight when event in observed epoch happened, so the value around
> corresponding marker is expected to be zero. Right?
That sounds right, although I may not understand you perfectly.
ERP show up usually after the event (unless it is expectation-related
nature), so it makes sense to set the baseline period before stimulus
onset during which brain activity is supposed to be neutral, and
whatever ERP can be compared against it.
> I have one more question regarding this - does it matter if I Remove
> Baseline for example (-1000ms to 0ms) if I have epoch that is longer (-4
> secs to 4 secs)? I read in Q&A list Arno's answer regarding similar
question
> where he said that ICA can be unstable if the epochs baseline is too
short,
> so he suggests longer baselines (i.e 1 sec).
Although I don't know what Arno meant in that specific context, I
guess he was probably referring to the finding reported by Groppe,
Makeig, and Kutas (2009). In the paper, the authors reports
whole-epoch baseline produced better ICA results compared to short
pre-stimulus baseline. Therefore, for ICA purpose, it's even better to
use an entire epoch for a baseline. The authors says 'It is not clear
what causes this difference.' in the paper (pp.1208), though I heard
Scott say a brief explanation. What do you think, Scott?
Makoto
2012/5/10 ida miokovic <ida.miokovic at gmail.com>:
> Dear Makoto,
>
> thank you for your answer, it cleared the doubts in my head regarding this
> =). As I understood, the purpose of Baseline Removal is for me/us to have
> better insight when event in observed epoch happened, so the value around
> corresponding marker is expected to be zero. Right?
>
> I have one more question regarding this - does it matter if I Remove
> Baseline for example (-1000ms to 0ms) if I have epoch that is longer (-4
> secs to 4 secs)? I read in Q&A list Arno's answer regarding similar
question
> where he said that ICA can be unstable if the epochs baseline is too
short,
> so he suggests longer baselines (i.e 1 sec).
>
> Thanks,
>
> Ida
>
>
> On Thu, May 10, 2012 at 9:45 PM, Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> wrote:
>>
>> Dear Ida,
>>
>> The consequence would be that you may not have near-zero potential
>> at/around time zero (and this time zero which should be an onset of
>> whatever event). Usually people want to reset their data to zero
>> microvolt at/around time zero, so they subtract mean of short time
>> period immediately before it (for example, -200 ms to 0 ms as a
>> baseline period). Am I answering to your question? If not, let me
>> know.
>>
>> Makoto
>>
>> 2012/5/10 ida miokovic <ida.miokovic at gmail.com>:
>> > Hello everyone,
>> >
>> > Since I do not have experience in eeg signal processing, I am asking
you
>> > for
>> > the opinion regarding epoch baseline removal (a window for this pops up
>> > after I do the data epoching). Epochs I am extracting are quite long:
-4
>> > secs before and 4 secs after Marker of my interest.
>> >
>> > Why is following suggested in tutorial:
>> >
>> > "Using the mean value in the pre-stimulus period (the pop_rmbase()
>> > default)
>> > is effective for many datasets, if the goal of the analysis is to
define
>> > transformations that occur in the data following the time-locking
>> > events."
>> >
>> > What are the consequences if I leave the fields in pop up window (Epoch
>> > Baseline Removal) empty and therefore have the whole epoch used as a
>> > baseline?
>> >
>> > Thank you in advance,
>> >
>> > All the best,
>> >
>> > Ida
>> >
>> >
>> >
>> >
>> >
>> >
>> > _______________________________________________
>> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> > To unsubscribe, send an empty email to
>> > eeglablist-unsubscribe at sccn.ucsd.edu
>> > For digest mode, send an email with the subject "set digest mime" to
>> > eeglablist-request at sccn.ucsd.edu
>>
>>
>>
>> --
>> Makoto Miyakoshi
>> JSPS Postdoctral Fellow for Research Abroad
>> Swartz Center for Computational Neuroscience
>> Institute for Neural Computation, University of California San Diego
>
>
--
Makoto Miyakoshi
JSPS Postdoctral Fellow for Research Abroad
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
--
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation; Prof. of
Neurosciences (Adj.), University of California San Diego, La Jolla CA
92093-0559, http://sccn.ucsd.edu/~scott <http://sccn.ucsd.edu/%7Escott>
_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
For digest mode, send an email with the subject "set digest mime" to
eeglablist-request at sccn.ucsd.edu
--
_________________________________________________________________
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
_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
For digest mode, send an email with the subject "set digest mime" to
eeglablist-request at sccn.ucsd.edu
--
Yuan-Fang Chao
School of Psychology
SouthWest University
Beibei,Chongqing,China
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120515/206054f7/attachment.html>
More information about the eeglablist
mailing list