[Eeglablist] filters, ICA and erp

Tue Oct 11 09:40:24 PDT 2011

Alex,

Thanks much for providing those references on EEG/EMG source separation.  I
would welcome to this list messages (from anyone) about any papers bearing
on the use of EEGLAB and/or its tools.

I would like to comment on one sentence in the 2011 paper in which you
justify your procedure of reducing your 128-channel data to a 64-dimension
principal subspace by saying, "*Preliminary visual inspection of the 128
components extracted from the native electrode array indicated overﬁtting,
evidenced by the ﬁssion of artifactual signals across components (e.g.,
ocular and cardiac artifacts loaded onto an excessive number of components;
see also Viola et al., 2009)*."

The concept of overfitting is drawn from neural networks in which
overfitting produces (e.g. classification) models whose performance do not
generalize to new data, and more generally from models of the variance of a
data set, for which it is hoped that the non-interesting ('noise') portion
of the data may be orthogonal to the data phenomena of interest.

In information-based signal processing using ICA, the concept is rather
different. Here, there need be no fixed concept of the data as consisting of
signal plus noise, and no assumption that the 'noise' (component) subspace
(however that may be defined) is orthogonal to the rest of the data.
Clearly, measured data always does have some level of additive and
multiplicative variability from non-intrinsic causes (sensor noise,
temperature fluctuations, etc.). However,  for (enough) well-recorded EEG
from a subject in a stable state (more on this below), this
noise-variability in the data is relatively small, certainly *far* smaller
than considered by e.g. the implicit ERP model (data  =  ERP + 'noise')
 that has colored thinking about EEG in the minds of most researchers for
decades.

Recent thoughts in the signal processing community about ICA and its
generalizations consider the case of independent subspaces. Imagine a
(sleepy?) subject is blinking slowly during the experiment. Infomax ICA uses
a static spatially static source assumption; but assume that in this case
the artifact produced by the slow blinks moves through its time course.
 Then ICA must find a subspace of components with nearly the same
(eyeblink-like) scalp maps and overlapping time courses -- something like
overlapping movie 'frames,'  each explaining a portion of the blink time
course. (In my talks, I use a slide, also linked to the EEGLAB main page,
showing an example of this).  This is not an example of overfitting, but
rather of fitting a complex phenomenon into an independent subspace.

A recent exciting result from Jason Palmer is that under some (reasonably
assured) assumptions, ICA will separate such a subspace into a component
subspace independent of the other component activities. Jason finds slightly
(or more) dependent subspaces in ICA decompositions by attempting to block
diagonalize the pairwise mutual information matrix (I hope he can publish an
EEGLAB tool for this soon).

Rather than attempting to eliminate such subspaces by PCA dimension
reduction, it is important to study their details, since they may shed light
on the microstructure of cortical (or muscle) activity. To find and separate
such subspace details, however, requires more dimensions -- e.g., at least
as many (suitably positioned) channels as one is able to record.

Since in general EEG 'noise' subspaces cannot be assumed to be spatially
orthogonal to the brain and non-brain sources of (varying degrees of
present) interest, reducing the data to a principal subspace using PCA
guarantees that some portion of the noise subspace will be spread (i.e.
projected) into the remaining dimensions. Therefore, to me PCA dimension
reduction should only be a procedure of last resort when the number of time
points available is too small for complete ICA decomposition -- and even
then, I believe better procedures than PCA are likely possible.

My comments above do not address the question of whether infomax ICA as
usually applied to EEG data equally well separates component process with
different frequency characteristics, etc. There is much more to be learned
about EEG/EMG decomposition, and I am pleased to hear of your and others'
careful explorations.

One point not much addressed in your 2011 paper is the scalp projection
pattern of muscle EMG, which has been known on basic principles, and is also
routinely observed by us (using 256 channels with coverage below the inion)
to be that of an equivalent dipole at the end of the muscle (e.g., the
muscle/tendon junction, which for head/neck muscles is near the tendon
attachment to the skull) -- and pointed in the direction of the muscle.

Collecting more channels allows ICA to separate more (scalp/neck muscle)
sources, but this may be equally an effect of having a wider montage
including the muscle (and its equiv. dipole) polarity reversals.  It should
be interesting to compare ICA decompositions of two 128-channel subsets of
our 256 electrodes, one with all electrodes located above or at the level of
the nasion/inion, and another containing still lower electrodes --- which
will separate more muscle components? If you or anyone is interested in
this, please contact me.

It also should be of interest for you to compare your results to date using
(extended) infomax ICA with resulting on the same data obtained using AMICA,
which our comparisons (so far for EEG sources, mainly) show to be still more
successful at separating physiologically distinct sources.

Scott Makeig

On Mon, Oct 10, 2011 at 12:56 PM, Alexander J. Shackman
<shackman at wisc.edu>wrote:

> I'm looking forward to reading Arno's paper!
>
> In the meantime, some may find our group's experiences using ICA to remove
> EMG from the EEG useful:
>
> http://psyphz.psych.wisc.edu/~shackman/mcmenamin_shackman_ni2011.pdf
> and
>
> http://psyphz.psych.wisc.edu/~shackman/mcmenamin_shackman_davidson_ni2010.pdf
>
> Cheers,
> Alex
>
>
> On Fri, Oct 7, 2011 at 9:43 PM, Scott Makeig <smakeig at gmail.com> wrote:
>
>> Does EMG in EEG data have to do with movements?
>>
>> Muscle tension is what generates EMG, not muscle movements per se. Two
>> opposing muscles, when tensed together, may not produce movement, but do
>> increase stiffness of the affected joint. Moving the body is only one of the
>> two main things muscles do -- the other is to stiffen the body, e.g., to
>> hold the head upright, but also for defensive and/or offensive purposes
>> (actual or imagined). Thus I believe there is useful cognitive and emotional
>> information in the configuration and strengths of neck muscle activations in
>> situations when the head is not moving.
>>
>> Scott Makeig
>>
>> p.s. Arno is first author of a paper, long in preparation and now in final
>> revision,  on differences between ICA algorithms applied to EEG data. We
>> will share it with the list when finished-- I believe it
>> has useful objective info on this subject.
>>
>> On Fri, Oct 7, 2011 at 5:58 AM, Sara Graziadio <
>> sara.graziadio at newcastle.ac.uk> wrote:
>>
>>> Hello Alonso,
>>> I agree that the subjects are not always moving and so the ICA does not
>>> work perfectly, but this is still an issue as the cortical activity is
>>> intermixed with the emg and if you remove that IC you are cleaning the data
>>> on one side but you are removing some cortical activity on the other side.
>>> And you don't know anything about where that cortical activity is coming
>>> from, what it is and so on. If you study the cortico-muscular coherence for
>>> example, removing channels with emgs will reduce your coherence. Almost in
>>> all the ICs I obtain with fastica (that should be the best algorithm or at
>>> least one of the best) I have also some cortical activity, you can see it if
>>> you look at the full signals in time, sometimes it is also obvious from the
>>> PSD of the IC. I think this is a very important limit of ICA that make me
>>> dubious when I read in the articles that ICA was used for artefact
>>> rejection. To which extend it is usually not clear. Probably more
>>> information should be provided in th!
>>>  e methods about this but nobody does it as far as I know. Perhaps
>>> somebody with more experience than me could comment on this.
>>>
>>> The 0.5Hz limit is something my colleagues and I found and I think it is
>>> the common experience for everybody, probably it was also previously
>>> discussed in this forum. I am not sure if there are any references for this.
>>> I don't think that the computer crashes because of your filter though,
>>> but I am not the best person to reply to you on this.
>>>
>>> I am not sure this could help you though!
>>>
>>> Best
>>>
>>> Sara
>>>
>>>
>>> >-----Original Message-----
>>> >From: Alonso Valerdi, Luz M [mailto:lmalon at essex.ac.uk]
>>> >Sent: 07 October 2011 12:44
>>> >To: Sara Graziadio; 'David Groppe'; 'japalmer29 at gmail.com'
>>> >Cc: eeglablist at sccn.ucsd.edu
>>> >Subject: RE: [Eeglablist] filters, ICA and erp
>>> >
>>> >Hello Sara,
>>> >
>>> >I've been following your questions and replies and I have undergone the
>>> >same experience that you are describing. The EMG seems to be mixed with
>>> >another component, but over the time I started to believe that maybe the
>>> >EMG doesn't come up over all the trial because the subject didn't move
>>> all the
>>> >time. Are you with me?
>>> >
>>> >On the other hand, I'd like to ask you why you posted previously that
>>> the
>>> >high-pass filter below 0.5Hz is not suitable for ICA processing? Do you
>>> have
>>> >any reference to recommend me about this issue? The point is that I'm
>>> >filtering my data between 0.1 - 40Hz, but the problem is that certain
>>> datasets
>>> >are stuck during the ICA processing and my computer crashes after
>>> several
>>> >hours. I wonder if it is because of the filtering. Do you have any
>>> comment?
>>> >
>>> >I'll really appreciate it!
>>> >Cheers
>>> >Luz
>>> >
>>> >-----Original Message-----
>>> >From: eeglablist-bounces at sccn.ucsd.edu [mailto:eeglablist-
>>> >bounces at sccn.ucsd.edu] On Behalf Of Sara Graziadio
>>> >Sent: 06 October 2011 10:51
>>> >To: 'David Groppe'; 'japalmer29 at gmail.com'
>>> >Cc: eeglablist at sccn.ucsd.edu
>>> >Subject: Re: [Eeglablist] filters, ICA and erp
>>> >
>>> >Hello,
>>> >Thanks for your suggestion.
>>> >
>>> >As I was planning to do also a PSD analysis on the data I guess that to
>>> remove
>>> >the mean is not the best method if it works as a non-selective high pass
>>> filter,
>>> >am I right?
>>> >
>>> >I am applying the PCA before applying the ICA to reduce the number of
>>> >components. How the data rank would be modified in this case?
>>> >I have to admit that it never happened to me that the muscle artefact is
>>> put in
>>> >a single source with the ICA. Usually it spreads on half of the
>>> components, is
>>> >this only my experience?
>>> >
>>> >Thanks again
>>> >
>>> >Best wishes
>>> >
>>> >Sara
>>> >
>>> >
>>> >>-----Original Message-----
>>> >>From: David Groppe [mailto:david.m.groppe at gmail.com]
>>> >>Sent: 05 October 2011 23:10
>>> >>To: Sara Graziadio
>>> >>Cc: eeglablist at sccn.ucsd.edu
>>> >>Subject: Re: [Eeglablist] filters, ICA and erp
>>> >>
>>> >>Hi Sara,
>>> >>   I found that a good way to improve the performance of ICA for ERP
>>> >>analysis is to
>>> >>1) Epoch your data into one or two second chunks time locked to the
>>> >>event of interest
>>> >>2) Remove the mean of each epoch at each channel
>>> >>3) Run ICA to remove artifacts
>>> >>4) Use a standard pre-event time window to baseline your data
>>> >>5) Compute your ERPs
>>> >>
>>> >>Removing the mean of each epoch acts as a crude high-pass filter.
>>> >>It's not nearly as selective as a "true" high pass filter but it
>>> >>doesn't distort the ERP waveforms as much either.  Moreover we've
>>> >>found that the procedure described above massively improves the
>>> >>reliability of ICA when compared to standard ERP prestimulus
>>> >>baselines:
>>> >>
>>> >>Groppe, D.M., Makeig, S., & Kutas, M. (2009) Identifying reliable
>>> >>independent components via split-half comparisons. NeuroImage, 45
>>> >>pp.1199-1211.
>>> >>
>>> >>Hope this helps,
>>> >>       -David
>>> >>
>>> >>
>>> >>
>>> >>On Wed, Oct 5, 2011 at 10:46 AM, Sara Graziadio
>>> >><sara.graziadio at newcastle.ac.uk> wrote:
>>> >>> Hello,
>>> >>> I would like just a suggestion about some data cleaning/analysis I am
>>> doing.
>>> >I
>>> >>am doing an ERP analysis and I want to clean my data first with the
>>> ICA. In
>>> >>theory, though, I should not use an high-pass cutoff higher than 0.1 Hz
>>> to not
>>> >>reduce the erp amplitude. On the other side the ICA does not work well
>>> if
>>> >the
>>> >>high-pass cutoff is lower than 0.5 Hz...what is then the best method to
>>> >apply?
>>> >>Has anybody tested how robust the ica is with a 0.1Hz filter?
>>> >>> I have also another question: I am doing the analysis on 94
>>> electrodes
>>> >>referenced to Fz. I planned to average reference the data but actually
>>> there
>>> >is
>>> >>quite a large spread of noise on all the electrodes with this method
>>> >(muscular
>>> >>artefacts for example from the temporal electrodes). But actually
>>> almost all
>>> >>the papers are using the average reference so I was surprised, am I the
>>> only
>>> >>one having this problem of noise? Would not be better just to keep the
>>> Fz
>>> >>reference and then perhaps to average the erps for every different
>>> cortical
>>> >>area and do the analysis on these averaged erps?
>>> >>>
>>> >>> Thank you very much
>>> >>>
>>> >>> Best wishes
>>> >>>
>>> >>> Sara Graziadio
>>> >>> Research Associate
>>> >>> Newcastle University
>>> >>>
>>> >>>
>>> >>>
>>> >>> _______________________________________________
>>> >>> 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
>>> >>>
>>> >>
>>> >>
>>> >>
>>> >>--
>>> >>David Groppe, Ph.D.
>>> >>Postdoctoral Researcher
>>> >>North Shore LIJ Health System
>>> >>New Hyde Park, New York
>>> >>http://www.cogsci.ucsd.edu/~dgroppe/
>>> >
>>> >_______________________________________________
>>> >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
>>>
>>> _______________________________________________
>>> 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
>>>
>>
>>
>>
>> --
>> 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
>>
>> _______________________________________________
>> 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
>>
>>
>
>
> --
> Alexander J. Shackman, Ph.D.
> HealthEmotions Research Institute | Lane Neuroimaging Laboratory
> Wisconsin Psychiatric Institute & Clinics
> University of Wisconsin-Madison
> 6001 Research Park Boulevard
> Madison, Wisconsin 53719
>
> Telephone: +1 (608) 358-5025
> Fax: +1 (608) 265-2875
> Email: shackman at wisc.edu
> http://psyphz.psych.wisc.edu/~shackman
>

-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20111011/e736b73d/attachment.html>