[FieldTrip] coherence/connectivity measures after applying ICA

[Tessa van Leeuwen]

Dear Olga, Thank you very much for your reply, this is very useful information! Because my trials are short I do not have sufficient data to let ICA compute all 275 possible components from my 275 channels - I use component reduction and estimate 80 components. This could then very well be the reason why coherence is altered after ICA. It makes sense that the interpolation would destroy any original phase relationships that are there. Thanks! Best wishes, Tessa ----- Original Message -----
> From: "Olga Vladimirovna Sysoeva" <sysoevao at psychiatry.wustl.edu>
> To: "Email discussion list for the FieldTrip project"
> <fieldtrip at donders.ru.nl>
> Sent: Wednesday, March 28, 2012 4:01:22 AM
> Subject: Re: [FieldTrip] coherence/connectivity measures after
> applying ICA
> Hi, Tessa,
> I do not have much experience for myself with coherence, but I'm very
> interested in the question you raised. Here is what I've seen in
> R.Thatcher's tutorial on Adulteration of Phase Relations when using
> Independent Components Analysis/Blind Identification and other
> Regression Methods to “Correct for Artifact
> http://www.appliedneuroscience.com/Tutorial%20on%20ICA%20Phase%20Adulteration.pdf
> .
> "Let us consider the rejection of eye movement artifact (EOG) using
> the ICA method which is the mathematical basis for Blind Source
> analysis. The ICA method decomposes a time series into a set of
> globally independent time series. Mathematically ICA can invert the
> equation and exactly reproduce the original time series based on the
> mathematical components. In this case the originally measured time or
> phase differences between each electrode are preserved and the
> cross-spectrum is unaltered.
> However, everything changes when the ICA or PCA or Blind Source, etc.
> methods are used to remove artifact by reconstructing 19 channels
> based on a smaller set of components, e.g., 17 or 18 channels and then
> deleting or omitting the EOG “components” to reconstruct 19 channels.
> This is because a regression process (interpolation) is used to
> minimize the deviation of the original 19 channel time series from the
> 17 or 18 “components” by which the remainder is used as weights to
> produce a second time series that has altered all of the phase
> relations in the original time series. The regression uses the average
> of the entire time series and therefore “smoothes” time or phase at
> each time point. Averaging a time series and computing deviations
> (whether information min-max, equimax, or simple linear regression)
> and then using the result to weight each time point by averages
> thereby smoothes each data point’s phase relations at each moment of
> time, distorting coherence, and destroys the time and phase relations
> present in the original time series."
> Hope it can help...
> Best Regards,
> Olga
> Olga Sysoeva,
> Research Associate, PhD
> Washington University School of Medicine
> Campus Box 8134
> 660 South Euclid Ave
> Saint Louis, MO 63110-9909
> From: fieldtrip-bounces at donders.ru.nl
> [fieldtrip-bounces at donders.ru.nl] on behalf of Rodolphe Nenert
> [batrod at gmail.com]
> Sent: Tuesday, March 27, 2012 7:20 PM
> To: Email discussion list for the FieldTrip project
> Subject: Re: [FieldTrip] coherence/connectivity measures after
> applying ICA
> Dear Tessa,
> ok this is more clear now.
> ICA is a wonderful tool which can easily become your worst enemy in
> data analyzing.
> By removing yourself certains epochs with blinks before doing an ICA,
> its is possible that you remove relevant data that can be used by your
> ICA algorithm to better separate components. Certain people would
> argue that even filtering your data before ICA can be a bad idea.
> Therefore, when you apply the ICA on your "cleaned" data, there is
> more chance that the component that will reflect the most what is left
> from blinks will also contains more relevant data.
> This data removal could potentially decrease your data variability
> therefore increase your coherence.
> Also, the number of epochs you keep after your cleaning (which should
> be different from ICA compared to visual removal) could potentially
> explain such a difference.
> Hope this helps,
> Rodolphe
> On Tue, Mar 27, 2012 at 4:34 PM, Tessa van Leeuwen <
> tessa.vanleeuwen at fcdonders.ru.nl > wrote:
> > Dear Rodolphe,
> > Thank you for your response, I tried to clarify below.
> > > 1) After your first "cleaning", do you test coherence on
> > > particular
> > > components or your entire data minus the "EOG" component?
> > I tested the coherence on the entire dataset after removal of the
> > EOG
> > component.
> > > 2) When you say that you redo ICA on cleaned data (of course, an
> > > ICA
> > > analysis made on a result of a previous ICA analysis with
> > > components
> > > removed is a bad idea), do you remove another component or do you
> > > test
> > > your coherence on particular components?
> > Sorry, I could have been more clear about this. I initially compared
> > two versions of the same dataset: one in which all trials containing
> > a
> > blink were removed from the dataset after manual inspection; another
> > version in which only trials with a blink during the actual stimulus
> > period were removed manually and the rest of the trials, including
> > trials with blinks outside the stimulus window, were 'cleaned' with
> > ICA, i.e. the EOG component was removed. Eye-balling coherence in
> > occipital channels, this was increased in the ICA-cleaned version.
> > To check whether the increase in coherence could have been explained
> > by an increased number of retained trials in the
> > ICA-cleaned-version,
> > I also applied the unmixing matrix obtained from the ICA to the
> > manually cleaned version of the data, i.e., in which only smaller
> > eye-movements or unidentified EOG artifacts would have been
> > remaining
> > after manual inspection. Now of course I could have been
> > over-removing
> > non-existent noise in this case, but also here the same difference
> > in
> > coherence appeared, now with the same number of trials in both
> > condition. When I look at the time-frequency representations from
> > both
> > versions, these look highly similar, only small intensity
> > differences
> > can be seen.
> > I only used a small amount of data to test this and these results
> > might therefore not be completely reliable. But I know other people
> > have experienced similar problems with altered coherence and I was
> > wondering whether any effect of ICA-preprocessing on
> > coherence/connectivity measures was generally known on the list and
> > in
> > the literature. Perhaps the removal of common noise with ICA can
> > already explain the differences?
> > Best wishes,
> > Tessa
> > > Rodolphe
> > > On Tue, Mar 27, 2012 at 11:27 AM, Tessa van Leeuwen <
> > > tessa.vanleeuwen at fcdonders.ru.nl > wrote:
> > > > Dear Fieldtrip experts,
> > > > I have noticed enhanced coherence (sensor level) in my data
> > > > after
> > > > applying ICA during preprocessing, removing only 1 EOG
> > > > component.
> > > > Of
> > > > course the (mainly quantitatively) enhanced coherence could be
> > > > due
> > > > to
> > > > the removal of (artifact induced) noise from the data. But this
> > > > increase also occured when applying ICA to previously cleaned
> > > > data,
> > > > implying changes induced by ICA somehow affect coherence.
> > > > One of the aims of our project is to compute
> > > > coherence/connectivity
> > > > measures at the source level. Since connectivity measures are
> > > > often
> > > > difficult to interpret as they are, I would like to ask whether
> > > > anyone
> > > > has experience with connectivity analyses after preprocessing
> > > > that
> > > > involved ICA. Are people aware of possible influences of ICA on
> > > > connectivity measures and is there a way to deal with this? Or
> > > > would
> > > > it be advisable NOT to use ICA when later looking at
> > > > coherence/connectivity at the source level?
> > > > We initially aim to compare across conditions (data that have
> > > > been
> > > > preprocessed together and from which the same ICA component has
> > > > been
> > > > removed). But we also have different experimental groups for
> > > > which
> > > > we
> > > > would like to qualitatively compare active networks during our
> > > > task.
> > > > Thank you in advance for any input, it is highly appreciated.
> > > > Best wishes,
> > > > Tessa
> > > > ---
> > > > Tessa van Leeuwen, PhD
> > > > postdoctoral researcher
> > > > Department of Neurophysiology
> > > > Max Planck Institute for Brain Research
> > > > _______________________________________________
> > > > fieldtrip mailing list
> > > > fieldtrip at donders.ru.nl
> > > > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> > > _______________________________________________
> > > fieldtrip mailing list
> > > fieldtrip at donders.ru.nl
> > > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> > _______________________________________________
> > fieldtrip mailing list
> > fieldtrip at donders.ru.nl
> > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> The materials in this message are private and may contain Protected
> Healthcare Information or other information of a sensitive nature. If
> you are not the intended recipient, be advised that any unauthorized
> use, disclosure, copying or the taking of any action in reliance on
> the contents of this information is strictly prohibited. If you have
> received this email in error, please immediately notify the sender via
> telephone or return mail.
> The materials in this message are private and may contain Protected
> Healthcare Information or other information of a sensitive nature. If
> you are not the intended recipient, be advised that any unauthorized
> use, disclosure, copying or the taking of any action in reliance on
> the contents of this information is strictly prohibited. If you have
> received this email in error, please immediately notify the sender via
> telephone or return mail.
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20120328/985298d4/attachment-0002.html>