[Eeglablist] ICA as artefact correction method - dilemma

[David Groppe]

On Thu, Jul 15, 2010 at 4:24 AM, Kris Baetens <Kris.Baetens at vub.ac.be> wrote:
> Dear all,
>
> I have been messing for some time with filter/ICA issues. I would be very grateful if anybody could shed some light on the matter.
>
> In a number of experiments, I have used sentences as stimulus material. We collected ERP responses to the final word of the last of a series of sentences and are interested in N400 and P300-like effects. Participants were instructed by means of an icon to “do their blinking” as much as possible during short pauses of a few seconds that followed about 2 seconds after each final sentence. We have used a DC amplifier with an average recording reference.
>
> Regardless of whether I use FIR or IIR filters, the higher my high-pass filter cut-off, the more drift I get in the participant ERP averages following the final sentences. That is, if I use a 6th order two-way Butterworth filter with half-amplitude cut-off of 0.01 Hz, for example, there is no particular drift in the ERP following the critical end sentences, whereas a similar filter with a 0.3Hz cut-off results in drifts that go from 0 to 30 µV over the course of a one second in participant averages.
> These drifts are outspoken in the vertical EOG channel but in the frontal channels as well. Considering the fact that many trials are followed by eye blinks (+/-2 or three seconds after the time lock), it seems obvious that the drift is a result of the eye blinks and the filtering applied to them.
> However, the “normal” drift left in the trials (taking all channels into account) is much higher when I use a 0.01Hz high pass than when I use a 03Hz high pass, as one would expect.
>
> I'm wrestling a bit with the following dilemma:
> -I have seen that when I use an adequate high-pass filter (0.5 or 1Hz) I get a very nice decomposition of my data, enabling the precise removal of eye blink activity, jaw muscle activation etcetera.
> However, when using such filters, I get enormous drifts in the frontal channels, as explained above (and somehow, this doesn’t  attract too much ‘attention’ of the ICA algorithm, still enabling a proper decomposition). Also, I am concerned that using such filters in classical ERP research might cause some problems (cf. Prof. Luck’s book), especially when the ERP components of interest are rather big slow ones like the N400 and P300.
> -On the other hand, when using a high-pass filter in the range of 0.01 – 0.1Hz (as is recommended by many), the ICA algorithm fails to decompose the data well. I can still get rid of some substantial EOG activity, but no real proper correction.
>
> My questions are the following:
> -Given the fact that the ICA algorithm works well only when one uses high pass filters in the range of 0.5-1Hz, and that using such filters is most often advised against by people working in classical ERP research, is ICA really utilizable in classical ERP-grand-average-style research as a method of eye blink correction?

Hi Kris,
  I've found that with EEG data broken up into 1-2 second epochs, ICA
can do a decent job removing EEG artifacts without high-pass filtering
(if you have sufficient data).  ICA does appear to perform better
though when you dampen low frequencies.  A simple way to do this is to
epoch your data and then remove the mean of each epoch.  We found that
doing this tends to massively improve the reliability of ICA's
results.

See Figure 8 of:
Groppe, D.M., Makeig, S., &  Kutas, M. (2009) Identifying reliable
independent components via split-half comparisons. NeuroImage, 45
pp.1199-1211. (http://www.cogsci.ucsd.edu/~dgroppe/PUBLICATIONS/Groppe2009.pdf)

> -Is it generally a bad idea to instruct participants to do their blinking at a fixed moment that starts a few seconds after your time-lock stimulus?
> -What sort of distortions or invalid conclusions could possibly arise from using high-threshold high pass filtering (i.e., 0.5Hz 6th order Butterworth) when one applies it to all conditions, on a grand average ERP-level?

Luck nicely explains this in the book you mentioned.  The problem is
that high pass filters can induce oscillations that push effects far
forward or backward in time and can cause them to flip polarities.  If
you follow my recommendation and simply remove the mean of each epoch,
you won't be inducing any oscillatory activity, so general waveform
shape should be relatively preserved.
       hope this helps,
           -David

> -What sort of high-pass filter would you advise in general for DC recordings?
>
> Many thanks in any case,
>
> Kris Baetens
> Ph.D. fellow of the Research Foundation - Flanders (FWO)
> Dept. Experimental and Applied Psychology
> Faculty of Psychology and Educational Sciences
> Vrije Universiteit Brussel
> Pleinlaan 2, 1050 Elsene
> +32 2 629 23 31
>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
>



-- 
David Groppe, Ph.D.
dgroppe at cogsci.ucsd.edu
http://www.cogsci.ucsd.edu/~dgroppe/