questions on ICA to remove EOG artifacts from MEG data

Elisabeth May ElisabethSusanne.May at UNI-DUESSELDORF.DE
Fri Mar 19 10:01:16 CET 2010

Dear Masaki, dear Laurence,

thank you a lot for your really quick and extensive replies to my long
list of questions!

Masaki, you are right with your first point: so far I am not completely
sure that my plotting scripts correctly sepparates the two gradiometer
types. I do get the same results with my plotting function and
ft_topoplotIC. However, I use my own layout files for horizontal and
vertical gradiometers in both functions so there might be something
wrong with the layout files. I'll check those again to be sure. I'll
also try to scale gradiometers and magnetometers and do ICA on all
sensors together.

Laurence, thank you for the plots, they will give me a better idea of
what to look for! I only want to use ICA to remove artifacts from my
data so I guess I might try the approach you suggested, too.

I'll let you know as soon as I make any progress with my data!

Thanks again and best wishes,

> Dear Elisabeth,
> I'm not very experienced with using ICA nor with using fieldtrip, but
> I do have some experience of obtaining reliable eyeblink spatial
> topographies from Neuromag 306 data (using SPM), so I thought I'd chip
> in with some advice...
> On 18 Mar 2010, at 08:51, Elisabeth May wrote:
>> Dear Fieldtrip users,
>> I am trying to use independent component analysis
>> (ft_componentanalysis) to remove eye artifacts from my MEG data. The
>> dataset was recorded with the Neuromag 306 system and contains two
>> conditions with 40 trials each (one where a painful laser stimulation
>> was applied to the subject's right hand and one where the same
>> stimulation was applied to the left hand). So far, I am trying to do
>> the ICA just for the gradiometers on preprocessed data (after removal
>> of bad channels) and then plotting the topographies of the components
>> to identify and then reject components representing the eye artifacts
>> (as suggested in the example script on the Fieldtrip webpage). I have
>> used different settings with the default option cfg.method = 'runica'
>> (no PCA before ICA, different numbers of principal components to be
>> retained (cfg.runica.pca), both conditions sepparately and
>> together...). However, I have never gotten component topographies
>> that I can clearly identify as components representing eye artifacts
>> (showing two sources right above the eyes), although there are almost
>> 200 eye artifacts in the 80 trials of this data set (i.e. there was
>> usually more than one EOG artifact in every single trial). So I am
>> still wondering about a few questions on the exact settings for the
>> ICA in my case and if I am missing something really obvious:
>>   * Is it correct to do the ICA on both horizontal and vertical
>>     gradiometers together and then plot them sepparately to identify
>>     artifacts?
> Yes, this sounds sensible - there should be covariance across
> horizontal and vertical gradiometers during an eyeblink, which should
> help you in isolating a component, but of course they will have a very
> different spatial topography, which means you would want to plot them
> separately. I have attached a snapshot of an eyeblink component that I
> identified, in vertical and horizontal gradiometers (top left and top
> right), magnetometers (bottom left), and 60 EEG sensors (bottom
> right). Note that I didn't obtain this component with ICA, but with
> PCA of an 'average eyeblink' (see below).
>>   * Should I do the PCA before the ICA and if so, how many principal
>>     components should I retain?
>>   * Should I do the ICA on both conditions together or sepparately
>>     (since I expect different sources to be activated in both
>>     conditions, e.g. one of the two primary somatsensory cortices
>>     activated according to stimulation of the right or left hand)?
>>   * Is it accurate enough to use complete trials for the component
>>     analysis or do I have to extract and use just the parts containing
>>     EOG artifacts?
> In my (limited) experience, I have had good success identifying
> eyeblink components with a slightly simpler method than ICA. I epoch
> the data with respect to well-identified eyeblinks (detected using
> EOG), then create an 'average eyeblink' using these epochs, so that I
> know what the timecourse of an eyeblink should look on average look
> like. I then just run a simple PCA on this 'average eyeblink', and
> normally the eyeblink is well captured by the first (and sometimes
> second) spatial component of this PCA. I then use this PCA map just as
> you might use an ICA map, i.e. regress it out of the raw (original)
> data. As a sanity check, you can correlate the timecourse of this
> component against the EOG channel, and this should correlate quite
> nicely. You can also try recomputing the average eyeblink after you
> have regressed the component out of the data, in order to see that it
> has been effective in removing most of the artefact.
> I find this a bit simpler than ICA, and maybe comes with a few
> advantages. First, it doesn't use all the data to derive the
> components, only the data epoched around EOG-detected eyeblinks, so it
> is quite a bit faster on big datasets. Second, by creating an average
> eyeblink, it removes any variability in the data driven by other
> boring stuff (like the brain :) ), so it should pretty much always be
> the component which explains the most variance in the 'average'
> eyeblink (i.e. the first principal component). Finally, it is
> important to note that although the method requires you to detect
> /some/ eyeblinks using the EOG channel, it is still robust against you
> failing to detect /all/ eyeblinks, as the regression is done on the
> raw data, rather than just on the epochs where you have already
> detected the blinks.
> However, there are some caveats; the first being that you assume that
> all eyeblinks look similar (in my experience they do), and perhaps
> more importantly that the component doesn't change (spatially) during
> the course of the experiment - with Neuromag data, you might be able
> to get around this to some extent by using Maxfilter for movement
> compensation. I think these caveats both also apply to ICA, but I'd be
> interested to hear other people's thoughts (particularly those who
> have more experience with ICA). Finally, I guess if you are interested
> in isolating interesting (brain) signals as well as boring (noise)
> signals simultaneously, then ICA is the way to go. I use this method
> just to remove eyeblinks before I do subsequent processing.
>>   * And finally, are the component topographies usually sufficient to
>>     identify eye artifacts?
> I find with this method they are very easy to identify, and I imagine
> they would be reasonably straightforward in ICA also? Again, perhaps
> someone with more ICA experience can comment...
>> Any help or advise on this would be greatly appreciated! Thanks in
>> advance,
>> Elisabeth
> Hope this helps,
> Laurence
> ===========================================
> Laurence Hunt, DPhil Student
>  Centre for Functional MRI of the Brain (FMRIB),
> University of Oxford
> lhunt at <mailto:lhunt at>
> Phone: (+44)1865-(2)22738
> ===========================================
> ----------------------------------
> 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.

Dipl.-Psych. Elisabeth May

Universitätsklinikum Düsseldorf
Institut für Klinische Neurowissenschaften und Medizinische Psychologie
Universitätsstr. 1
40225 Düsseldorf

Tel: +49 211 81-18075

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 and

More information about the fieldtrip mailing list