[FieldTrip] strange observations using ICA

[Schoffelen, J.M. (Jan Mathijs)]

Dear Immo,

I understand your concern with respect to filtering, but replacing original recordings with linear combinations of all channels (which you essentially do with the ICA decomposition/backprojection) is also potentially harmful, and possible more harmful than careful filtering.

Anecdotally, if your artifact is a clear line in the spectrum, it is actually often better to notch it out (with an as narrow as possible filter) if you aim to estimate quantities like spectrally resolved Granger, using non-parametric factorization. If you don’t do this, the spectra typically look strange.

If you consider to go for more ‘straightforward’ frequency resolved methods, which rely on FFTs without a subsequent introduction of across-frequency bin dependencies of the local estimates (which is what happens with the spectral factorization), I wouldn’t be worried about the 130 Hz artifact (and only refrain from interpreting any result at that frequency).

Best wishes,

Jan-Mathijs

> On 10 Aug 2018, at 11:31, weberi at staff.uni-marburg.de wrote:
> 
> Dear Eelke,
> 
> thank you for the explanation. I am not sure I understand
> completely, but I get the gist of it. The reason why we don't use
> a band-stop filter is that we fear it may have a great influence on the subsequent
> analysis methods, especially regarding directionality measures or measures
> working in the time domain.
> 
> Thank you and best regards,
> Immo
> 
> Zitat von Eelke Spaak <e.spaak at donders.ru.nl>:
> 
>> Dear Immo,
>> 
>> While I don't have a clear-cut answer, I could imagine that something
>> like the following is going on. The ICA decomposition is never
>> perfect, and the isolation of individual components will get worse
>> with decreasing component power. So, as you say, removing the first
>> few components might indeed decrease the artifactual power, but at
>> some point removing components that have a clear 130 Hz peak might end
>> up instead *adding* 130 Hz power into channels for which the component
>> weight is nonzero (but which, in a perfect decomposition, should have
>> been zero). (I'm not sure I'm explaining this well enough; in any
>> case, the key is to keep in mind that ICA is simply a linear
>> projection of the data.)
>> 
>> Since you say that the artifact has very well-defined spectral
>> characteristics, might it not be simpler and more effective to use a
>> band-stop filter to take out 129-131 Hz (and similarly for the
>> harmonics)?
>> 
>> Hope that helps,
>> Best,
>> Eelke
>> 
>> On 9 August 2018 at 09:26,  <weberi at staff.uni-marburg.de> wrote:
>>> 
>>> Dear FieldTrip community,
>>> 
>>> I have a question concerning some strange observations when using ICA.
>>> In our lab we try to get rid of an EEG artefact induced by deep brain
>>> stimulation.
>>> For this purpose we compute an ICA of our EEG-channels and sort the
>>> components
>>> relative to their similarity (quantified using mutual information) with the
>>> raw artefact
>>> measured  directly from the stimulator. We then successively remove the
>>> components
>>> according to their mutual information and calculate the mean frequency
>>> spectrum of the
>>> back-transformed corrected data. The artefact has a very defined frequency
>>> distribution
>>> with peaks at multiples of 130 Hz.  With our approach, we would expect the
>>> artefactual
>>> peaks to successively become smaller. However, this is not the case. While
>>> the peaks
>>> ultimately become smaller after removal of several components, the
>>> artefactual power
>>> even rises again after the first couple of components. How can this be
>>> explained?
>>> 
>>> Any insights would be much appreciated,
>>> 
>>> Best regards
>>> Immo Weber
>>> 
>>> _______________________________________________
>>> fieldtrip mailing list
>>> https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>> https://doi.org/10.1371/journal.pcbi.1002202
>> _______________________________________________
>> fieldtrip mailing list
>> https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>> https://doi.org/10.1371/journal.pcbi.1002202
> 
> 
> 
> _______________________________________________
> fieldtrip mailing list
> https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> https://doi.org/10.1371/journal.pcbi.1002202