[Eeglablist] line noise removal issues

[David Groppe]

Hi Tim,
  You might want to try the ICA algorithm Second Order Blind
Identification (SOBI).  It learns spatial filters from autocorrelation
matrices so it might be better than infomax ICA at extracting things like
60 Hz that have a regular, clearly-defined temporal structure.  The
following paper provides some evidence that SOBI is effective for
extracting 60 Hz:

Tang et al. (2005)
Validation of SOBI components from high-density EEG
Neuroimage, 25(2), 539-553

The function sobi.m implements SOBI and should come with EEGLAB.
    -David

dgroppe at cogsci.ucsd.edu

On Mon, 10 Dec 2007, Tim Mullen wrote:

> Dear EEGLAB users,
> 
> I have a question regarding the use of ICA for line noise removal.
> 
> I have some electrocorticographic (ECoG) data with a strong 60Hz line noise
> artifact as well as a 180Hz harmonic (only odd harmonics seem to be present,
> probably due to symmetrical clipping).
> 
> I am applying frequency-domain granger causality to this data, but have run
> into some serious problems with the presence of this line noise. Oddly
> enough, the line noise dominates as a *directional *effect in the granger
> causality (unless there is an apparent temporal delay between channels at 60
> Hz, a peak at 60 Hz should only be present in the instantaneous causality).
> This is likely because the phase at 60Hz appears to differ between channels.
> The strength of the directional effect at 60 and 180Hz is so strong that it
> dominates any other interesting nearby features, making it impossible to
> analyze causal interactions within a wide range of frequencies of interest.
> 
> The noise band is far too wide for notch filtering to be considered a
> suitable solution.  I have then tried extended infomax ICA (as implemented
> in EEGLAB's *runica* function), to isolate the subgaussian noise components.
> I have attempted this both in automatically estimating the number of
> sub-gaussian sources and also fixing the number of subgaussian sources to 1,
> 2, etc. None of these approaches have been successful.  ICA appears to
> converge properly and the covariance matrix of the estimated components is
> the identity matrix (it's at least second-order independent).
> 
> It is possible that the tanh function used to model the subgaussian source
> distributions is unsuitable for this line noise source. Has anyone used or
> implemented any other families of distributions to calculate the score
> function for ICA?
> 
> Des anyone have any recommendations on how to remove this line noise, either
> via source separation or other techniques?  In particular, if anyone has
> developed a plugin for EEGLAB or their own code for automatic line noise
> removal, that would be optimal.
> 
> 
> Thanks much for your input!
> 
> Regards,
> Tim
> 

--