ICA on 275ch MEG... removing artifacts

[Suresh Muthukumaraswamy]

Thanks Stefan and Michael for your help with this!
All the best,
   Suresh

Suresh Muthukumaraswamy, PhD
CUBRIC
Cardiff University
Park Place
Cardiff, CF10 3AT
United Kingdom
email: sdmuthu at cardiff.ac.uk
Phone: +44 (0)29 2087 0353 

>>> Stefan Debener <s.debener at UKE.UNI-HAMBURG.DE> 03/27/09 8:54 pm >>>
Hi Suresh,

>   Typically how many components do people normally estimate and then how
> many of these would normally get rejected as containing eye artefact? 270
> components is alot to look through! I see one can limit the number of
> components the function can return....
>   
Be careful with assuming a fixed number of components to explain eye 
blinks. In EEG, it largely (but not solely!) depends on the number of 
channels/components; so, while most 32 channel decompositions return 1 
eye blink component, this cannot always be expected, and in 128 channel 
data it can be anything between, roughly, 1 and 6. The problem with ICA 
of course is that we have no clue about the number of sources 
contributing to the data, so over/underfitting seems the rule rather 
than the exception. Therefore (and for other reasons), one should 
evaluate the quality of the decomposition, which usually includes 
evaluating the time course information as well as the reliability of a 
decomposition.

With regard to spatial stationarity, as mentioned by Michael already, it 
seems important to recognize that many artefacts violate this assumption 
to some more or less relevant extent. Eye blinks, for instance, seem to 
be caused primarily by the movement of the eye lids (not so much of the 
eye balls; there is a wonderful paper published on this in Clin 
Neurophysiol, but I keep forgetting the reference, sorry). So, it is a 
moving source which often can be 'perfectly' modelled with ICA. In my 
experience, the same holds for lateral eye movements, which also can be 
nicely modelled with ICA: these components show a very robust 
topography, that is, you need not much data to find such ICs quite 
quickly and reliably, and they show up in nearly all datasets. Thus, ICA 
seems to tolerate to some extent stationarity violations. My speculation 
is that this is related to the relatively poor spatial sampling of these 
artefacts in conventional EEG/MEG recordings: imagine decomposing from 
100 channels that covered the whole face: I would not be surprised if 
those recordings would be more sensitive to stationarity violations 
contributed by these artefacts.

A case where the violation of stationarity typically messes up ICA is 
for inside scanner EEG recordings. Here, it is my experience that ICA 
cannot deal well with the ballistocardiogram, which contributes a very 
complex, dynamically moving signal to the EEG (but see plenty other 
published papers proposing just the opposite). The reason for this 
inconsistency seems that the ballistocardiogram scales with the scanner 
B0 field: in the case of 1.5T, ICA sometimes still pulls out a more or 
less reasonable decomposition, but at 3T, or even 7T, it performs 
poorly, under otherwise identical conditions (Debener et al., 2008, Int 
J Psychophysiol).
>  Do people normally reject solely on topography or do they do a frequency
> analysis of the component time-course or perhaps other things?
>   
Depends. As it happens, there is a paper in press (Viola et al., Clin 
Neurophysiology, will be available online in a few days) evaluating the 
use of topographical information for eye blink and eye movement 
component identification and clustering (across datasets). This works 
surprisingly well, from 30 to 128 channel EEG recordings. Temporal 
correlations of component activations with aux channels have also been 
used for component identification (e.g., Srivastava et al., 2005, 
Neurimage) but I have mixed experience with this approach and would not 
recommend it. However, it really depends on the type of artefact you are 
after: there are cases where topographical information is not of much 
help, while temporal information gives a detailed picture of which 
component reflects artefact and which does not (e.g., Sandmann et al.,in 
press in Brain; or Ohla et al., in press in Brain Topography; both 
available online).

> Moreoever, I was also wondering if anyone had carried out any kind of
> systematic comparison between this approach for MEG compared to 
> traditional
> EEG approaches to this problem (e.g. projecting out the EOG channels)?
>   
Interesting you mention this - that's one of the issues we are very much 
interested in.


Hope this helps,

Stefan

----------------------------------
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/neuroimaging/fieldtrip.

----------------------------------
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/neuroimaging/fieldtrip.