ICA based artifact correction and phase-locking

Stefan Debener s.debener at UKE.UNI-HAMBURG.DE
Fri Feb 23 15:25:44 CET 2007


Dear Markus,

While Christian is certainly correct, I do have a more optimistic
interpretation of ICA for artefact removal. In order to tell with some
scientific "objectivity" whether any artefact removal improves your data
or not, you should, in my opinion, always look at the amount of artefact
removal AND the change in signal quality at the same time. For both
artefact and signal you should have some a priori assumptions in mind
what actually your signals and artefacts look like (you implicitely have
this in mind by excluding some epochs and keeping others, or, more
generally, for every single EEG signal processing step). If you keep
this in mind, you will find that ICA sometimes, even though reducing the
amount of  artefact, also reduces the SNR. For a nice demo that this can
happen, you may look up Debener et al., 2007, Neuroimage, 34, 587-97.
However, in my experience, this happens only if you massively violate
ICA assumptions. If you compare ICA artefact removal with any other
artefact processing techniques (Gratton & Coles, epoch rejection, etc.)
or uncorrected data, you will very likely find that ICA, if applied
correctly, returns substantially better results. For an example where we
could analyse ALL(!) recorded single trials from EEG data recorded
inside a 3T MRI see: J Neurosci, 25, 11730-37. Marcel Bastiaansen from
the FCD and others used the same ICA approach but with different tasks
and components, and were similarly successful...In sum, from a practical
standpoint, ICA can massivley improve your SNR and this in turn allows
you to do things with your EEG data that otherwise seem impossible.

You wrote that you observed channel saturation in your data. GIGO? ICA
is not very good in the very high and very low frequency domain
(personal opinion) and won't deal well with DC recorded data. In fact, a
few drifting channels can easily spoil a decompositon....this and
further issues are (hopefully) discussed in the EEGLAB tutorial, which
also explains how to pre-process your data in order to obtain reasonable
decompositions.

Best,
Stefan


Markus Werkle-Bergner wrote:
> Dear Christian,
>
>
>> Why do you remove these epochs? Do the signals saturate or is there
>> excessive electrode(-cap) or subject movement in these?
>>
> Yes, that's correct. I remove epochs before any further analysis, if I
> find saturated channels or excessive subject movements.
>
>>
>>> This means, my trials are a mixture of clearly contaminated as well
>>> as clearly uncontaminated trials. Because my main goal is to
>>> identify the eye and muscle artifacts, would it improve the
>>> detection of components reflecting artifacts if I do the ICA
>>> decomposition only on pre-identified artifactual trials? (My idea is
>>> that then the components reflecting artifacts would catch less
>>> 'true' brain activity. But perhaps I have a missconception here ...)
>>
>> The accuracy of the ICA decomposition relies fundamentally on the
>> accuracy of the mixing matrix estimate A from which you form the
>> de-mixing matrix (or vice versa). So if only some columns of A (e.g.
>> artifact source sensor projections) are accurately estimated while
>> the sensor projection estimates of brain activity are inaccurate
>> (individually or as a subspace), then the de-mixing the data using W
>> = inv(A) will not be fully successful in removing artifact activity
>> from components reflecting the brain activity (subspace). In order to
>> accurately estimate the sensor projections of ALL sources, the data
>> needs to comprise example activity from ALL sources, i.e. BOTH
>> artifact activity and brain activity.
>>
>
>
> Thanks for this clarification.
>
>
>> So since you need to estimate your brain activity (subspace) as
>> accurately as your artifact sources, you should use as much of the
>> data, i.e. contaminated and clean segments, as possible.
>>
>> How many EEG channels do you have, by the way? I need to know in
>> order to continue with the 'slightly longer answer'
>>
>
> I recorded my data from 60 Ag/Ag-Cl electrodes embedded in an elastic
> cap (with BrainAmp amplifiers) positioned according to the extended
> 10/20 system. Additionally, I record EOG with two electrodes placed at
> the outher canthi and one electrode placed below the left eye. All
> channels are referenced to the right mastoid during recording, while
> the left mastoid electrode is recorded as an additional active channel
> (off-line re-refrenced to the mean of both mastoids).
>
>
> Best regards,
> Markus
>
>
>
>
>
>> Regards,
>> Christian
>>
>>
>>
>>>
>>> Thank you very much for your help.
>>>
>>> Best regards,
>>> Markus
>>>
>>> Christian Hesse schrieb:
>>>> Hi Markus,
>>>>
>>>> the short answer is: if (and only if, big if by the way) ICA
>>>> correctly separates artifacts from brain activity, and you
>>>> correctly identify the artifact components, removal of the artifact
>>>> components from your data does not affect the time-frequency
>>>> properties (including phase) of the other components (i.e. the rest
>>>> of the data) if you project back to the sensor level. This is
>>>> because ICA assumes a linear (instantaneous) mixing of
>>>> statistically independent signals.
>>>>
>>>> I would just try it out for the  time being - if you get funny
>>>> results (or things too good to be true), please get back in touch
>>>> as there is also a slightly longer answer.
>>>>
>>>> Hope this helps,
>>>> Christian
>>>>
>>>>
>>>> On 23 Feb 2007, at 12:24, Markus Werkle-Bergner wrote:
>>>>
>>>>> Dear all,
>>>>>
>>>>> in my studies, I'm investigating early preceptual binding (visual)
>>>>> across the lifespan (i.e., I have data form children, younger and
>>>>> older
>>>>> adults) with EEG measures. My main interest concerns changes in
>>>>> gamma-power and measures of phase-synchronization in the gamma
>>>>> frequency range(e.g., phase-locking index, n:m (theta:gamma) phase
>>>>> synchronization).
>>>>>
>>>>> Currently  I use a 'semi-automatic' procedure for artifact
>>>>> rejection, i.e., I use thresholding in the time-domain (min/max in
>>>>> segment -/+ 100µV)to 'suggest' contaminated epochs. After that I
>>>>> visually inspect the data again for eye-blink and muscle activity,
>>>>> and completely reject the contaminated epochs.
>>>>>
>>>>> The problem with this procedure is that, especially in the older
>>>>> adults group, for many subjects only too few trials remain in the
>>>>> final sample.
>>>>>
>>>>> Therefore, I thought I could use ICA for artifact correction
>>>>> (instead of complete rejection). After identification of the
>>>>> components that reflect muscle activity (and also other
>>>>> artifacts), I thought to recombine the remaining ICs and perform
>>>>> my analyses (power, PLI, n:m synchronization) on the recombined
>>>>> (cleaned data).
>>>>>
>>>>> Now my question(s): Is there any experience whether removing
>>>>> certain ICs
>>>>> may change the phase spectrum, i.e. may this approach induce some
>>>>> systematic bias? If there is a systematic bias, are different
>>>>> frequency
>>>>> bands affected differentialy? Could anyone give me some references on
>>>>> these issues?
>>>>>
>>>>> Any comments are very much appreciated.
>>>>>
>>>>> Best regards,
>>>>> Markus
>>>>>
>>>>> --
>>>>> **************************************************************
>>>>> Markus Werkle-Bergner, Dipl. Psych.
>>>>> Predoctoral Research Fellow
>>>>>
>>>>> Center for Lifespan Psychology
>>>>> Max Planck Institute for Human Development
>>>>> Lentzeallee 94, Room 211, D-14195 Berlin, Germany.
>>>>> Phone: +49(0)30-82406-447       Fax: +49(0)30-8249939
>>>>> **************************************************************
>>>>>
>>>>
>>>> ----------------------------------------------------------------------
>>>> 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 <mailto:c.hesse at fcdonders.ru.nl>
>>>> Web: www.fcdonders.ru.nl <http://www.fcdonders.ru.nl>
>>>> ----------------------------------------------------------------------
>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>> --
>>> **************************************************************
>>> Markus Werkle-Bergner, Dipl. Psych.
>>> Predoctoral Research Fellow
>>>
>>> Center for Lifespan Psychology
>>> Max Planck Institute for Human Development
>>> Lentzeallee 94, Room 211, D-14195 Berlin, Germany.
>>> Phone: +49(0)30-82406-447       Fax: +49(0)30-8249939
>>> **************************************************************
>>>
>>
>> ----------------------------------------------------------------------
>> 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 <mailto:c.hesse at fcdonders.ru.nl>
>> Web: www.fcdonders.ru.nl <http://www.fcdonders.ru.nl>
>> ----------------------------------------------------------------------
>>
>>
>>
>>
>
>



More information about the fieldtrip mailing list