[Eeglablist] Can I use independent components derived from filtered data for correction of unfiltered data?

[Scott Makeig]

Niko -

In this case you'd be decomposing   your    whole_data = highpass_data + 
lowpass_data
as follows

     wts*sphe*(hipass_data + lowpass_data) =  wts*sph*hipass_data + 
wts*sph*lowpass_data;

 The first term on the right above would be separated into its 
independent components (if you trained ICA on the highpass data). The 
second term would be separated into components without any guarantee at 
all of independence.

You could subtract components of the first term  (i.e. of the 
highpass_data) from the whole data. though this would fail to separate 
out the lowpass portions of the eye artifacts.

You could run a separate decomposition on the lowpass data, subtract 
from it any components you wish, and then add together the 
artifact-adjusted highpass and lowpass data portions. This might or 
might not be effective  - we have no experience since we do not try to 
work with the lowpass (< 1 Hz) phenomena in our data -- but is possibly 
worth a try.

Scott


Niko Busch wrote:

> Dear EEGLAB users,
>
> is it reasonable to compute ICA components on a highpass filtered (0.5 
> Hz) dataset, and then use the solution to correct eyeblinks in the 
> unfiltered dataset?
>
> I recently conducted a recognition memory experiment. The epochs are 
> long, there are a lot of eyeblinks, and, unfortunately, a lot of slow 
> electrode drifts which are probably sweat artifacts (beware of 
> recording experiments in July!). I would like to use ICA to correct 
> the eye artifacts, but I am afraid that the slow drifts would trouble 
> the ICA solution. On the other hand, I do not want to highpass filter 
> the data, either, since this also eliminates the slow potential 
> effects I am interested in. Therefore, could I run ICA on the highpass 
> filtered dataset, then transfer the weights to the unfiltered dataset, 
> and finally use them for eyeblink correction? Or would that distort my 
> data in some way?
>
> Thank you and best regards,
> Niko.
>