ICA with averaged data

Christian Hesse c.hesse at FCDONDERS.RU.NL
Tue Mar 13 09:45:19 CET 2007

Hi Niki,

I fully concur with Stefan's advice and add a brief (technical)
explanation for why doing ICA on time-locked averaged data is
unlikely to give you a good separation.

One of the most common application of ICA to EEG/MEG data in
cognitive neuroscience seems to be to remove (what the users assume/
judge to be) artifact (i.e. non-brain or non-task related brain)
activity from the data, and then to proceed with conventional sensor
level analysis on the "clean" data. ICA assumes that the underlying
source signals (component waveforms) in the data are statistically
independent and have non-gaussian marginal distributions. If either
of these assumptions are violated, then the sources are not uniquely
identifiable w.r.t. the ICA search criteria, meaning that ICA will
tend to give "funny" decompositions.

Now consider that (time-locked) averaging essentially involves
summing your data over trials at each time point, and that the
distribution of the sum of several random variables with any
distribution (including non-gaussian) approaches a gaussian
distribution in the limit. From this it is clear that time-locked
averaging tends to make all non-time locked activity in your data
(i.e. the stuff you want to remove, and need ICA to identify) appear
MORE gaussian and therefore LESS identifiable from an ICA
perspective. As a consequence, ICA cannot model the artifact activity
sufficiently well to be able to separate it from your brain activity
of interest.

Thus, if the aim of using ICA is to remove non-time locked artifacts
(i.e. asynchronous oscillatory activity or intermittent transients)
from your data, then following the reasoning above it is advisable to
decompose the raw data (making sure to include examples of the
artifact you wish to remove). Then either remove the artifacts and
project back to the sensor space to do your averaging, or just do the
averaging on the components themselves.

Hope this helps,

On 9 Mar 2007, at 13:11, Stefan Debener wrote:

> Hi Niki,
> It does, in my opinion, not make much sense to average first and
> then run ICA, you unlikely obtain a robust decomposition this way
> and may not be able to remove the stimulator artefact. ICA needs
> quite a few training data to find a decent decomposition, and this
> is not provided by ERPs, even though some commercial software
> packages may allow you to apply ICA to ERPs/ERFs. Only if (!) all
> your data are recorded such that there was no substantial head
> movement (that is, the position between sensors and sources did not
> change), you may get a decent artefact IC (or several ICs) from
> combined data (there is a few more ifs involved, please see
> previous ICA discussions here and in the EEGLAB mailing list). YI'd
> suggest using the pca option and decomposse, say, about 50
> components, to start with, and then try to optimize the result by
> playing with the pca option and by running ICAs for separate
> recording blocks.
> Best,
> Stefan
> Niki Ray wrote:
>> Hi,
>>   I'm very new to fieldtrip. I have three questions. I have
>> collected MEG
>> data from patients with a deep brain stimulator implanted. I'd
>> like to do an
>> ICA to reject the stimulator related artifacts. Sometimes the
>> stimulator is
>> on and sometimes off. So my first question is, do i combine all
>> the data
>> together and then perform the ICA, or do i do different ICA
>> analyses for
>> each condition(on and off stimulation), and hence reject different
>> components for each condition?
>>  One possibility for deciding wich of the artifacts were related
>> to the
>> stimulator was to use the example matlab script you provide for
>> rejecting
>> ECG components. Some of the channels (about 13, i'm using a ctf
>> 275 system)
>> have a lot of very distinct artifact that is obviously from the
>> stimulator,
>> so i wanted to use one of these channels as the "ECG" channel. Is
>> this
>> sensible, or not??
>>   Finally, I've been getting conflicting advice about whether it
>> is okay to
>> first average data, and then perform an ICA. What do you suggest
>> is the
>> right thing to do?
>> Many thanks!
>> Niki

Christian Hesse, PhD, MIEEE

F.C. Donders Centre for Cognitive Neuroimaging
P.O. Box 9101
NL-6500 HB Nijmegen
The Netherlands

Tel.: +31 (0)24 36 68293
Fax: +31 (0)24 36 10989

Email: c.hesse at fcdonders.ru.nl
Web: www.fcdonders.ru.nl

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20070313/8823476a/attachment-0002.html>

More information about the fieldtrip mailing list