ICA and artifact rejection

Markus Bauer m.bauer at FIL.ION.UCL.AC.UK
Wed May 28 14:57:48 CEST 2008


Hi Floris
> 1) ICA can be pretty slow, and it has been suggested in the Fieldtrip
> mailing list to use PCA and estimate only the 50 components that
> explain most variance. However, when I do this to e.g. identify
> ECG-related artifacts, I run into trouble when trying to decompose the
> ECG-locked data into the components, using the previously estimated 50
> components (in order to calculate the coherence between the components
> and the ECG signal). The ECG-locked data is 275 channels X number of
> time points, so it cannot be simply decomposed into 50 components.
> Does anyone know how to go about this?

I'm not quite sure whether I understand the problem correctly. But
obviously, if you have done a PCA first on the data, and do an ICA those
principal components, in order to get back to the original data you have
to multiply the unmixing matrix obtained from the ICA with that of the
PCA. So if you only kept 50 principal components and pruned the others,
your PC unmixing matrix will have 275 rows and 50 columns (ICA usually
then 50 columns and rows) and in order to get back to your raw data you
have to multiply the mixing matrices (i.e. pseudoinverse of the mixing
matrix) of ICA and PCA in the correct order. I can provide you some code
snippet if that is indeed the problem...

> 2) For EOG I suppose I could do like for ECG, and calculate the
> coherence with VEOG in order to select components that are artifacts.
> But this is not so easy for head muscle-related artifacts (jaw and
> neck movements, etc.), since we typically don't measure neck and jaw
> muscle EMG. Apart from looking at spatial topography, is there a
> robust criterion that one could use for detection of EMG artifact
> components?  Or would it be better to use 'classical' artifact
> rejection routines for these kinds of artifacts? And would ICA also be
> suitable to remove things like jumps and head movements in the MEG, or
> is it better/safer to just throw away those kind of data segments?
I would take out any short-lived non stationary (or non-frequently
occuring) phenomena such as short muscle twitches and jumps in any case.
For the more tonic muscle phenomena (like the neck and jaw muscles)  ICA
can do a decent job, but if subjects make movements the topography can
each time be quite different - if a subject makes too many movements the
whole idea of ICA as spatial filters is pointless cause the MEG sensor
array won't follow the movement

One problem with muscle artefacts is that there can be these huge
variations in high-frequency power (muscle tonus primarily in neck and
jaw muscle) over several blocks of trials. These can be detrimental for
identifying gamma-oscillations and also for finding a decent rejection
criteria for short-muscle artefacts (movements etc). In order to
compensate for that we've written in Berlin some own muscle artefact
routines that basically apply kind of a low-pass filter on the amplitude
envelope of the high-frequency signal (that you use to separate short
lived muscle artefacts from the ones with longer time course). You can
use that to throw away the movements and leave the more tonic effects in
(which may be in half of your trials or so)  - to either take care of
them by individual trial baseline correction or removing them by ICA.
I can give you the scripts - but they use a radically different reading
routine. I will also hand them over to Robert, they are written a bit
differently than the usual fieldtrip stuff, but maybe he'll like them anyway

best
Markus

PS:
Personally, I became sceptical about ICA, cause it's so subjhecttive
whether a compoentn is now an artefact component or not, whether to take
it out or not. Also, often one and the "same" phenomenon are distributed
acroiss components others you would expect to find you don't find. I
didn't like that much what I saw

----------------------------------
The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/fcdonders/fieldtrip.



More information about the fieldtrip mailing list