update: artifact detection

[Robert Oostenveld]

Dear Georges,

On 3-feb-2006, at 9:41, georges Otte wrote:
> I saw some paper from a group in Leuven eliminating EMG artefact based
> on the canonical component analysis (another Blind Source method).
> Google it with CCA- EEG. It worked very fast. I think that ANT
> compagny
> is looking into porting this into their standard software. The
> method is
> based on the difference in aotocorellation between EEG and EMG
> signals.

We have implemented that algorithm in Fieldtrip, in the
componentanalysis function (cfg.method=cca). I am not sure whether it
is compleetly in the ftp release version of fieldtrip.

We did try it out on our MEG data. It performed rather poor on our
MEG data, the obvious EMG artifacts in the lower temporal channels
were not very well separates into seperate components to be removed
from the data. I suspect the reason for it performing rather poor on
the MEG data is that it is based on 1-sample lag autocorrelation. The
seperation (linear unmixing) using autocorrelation seperates the
signal into high frequency components and low frequency components.
But our MEG data has a much higher sampling rate (usually 1200Hz),
and contains noise components at a much higher frequency than the EMG
artifacts.

It may work for clinical EEG data for which the Leuven group
develloped the algorithm, since that is sampled at typically 250 Hz,
which means that the EMG does correspond to the high frequency part
of the data. It may be possible to tweak the algorithm (e.g. use N-
lag autocorrelation) and to use differently preprocessed (filtered
and downsampled) data to improve the performance, but we have not
pursued that. Furthermore, we have not tested it on EEG data, so I
cannot judge how it performs there.

best
Robert