[Eeglablist] Using ICA for labelling artifacts

[Scott Makeig]

Velu -

I am not much in favor of your proposed approach, as 4-channel ICA
decomposition of full-blown(!) EEG data must attempt to separate all
sources into only 4 component processes(!), meaning each component will
likely still represent a substantial mixture of sources. In addition,
ICLabel operates on component scalp maps (and activations), which would be
highly inaccurate when extrapolated from only 4 channel values.

Much better would be to decompose the whole (many-channel) data with ICA
and then, perhaps, experiment with using the EEGLAB tool:  Tools / Reject
data using ICA / All methods.
I am personally interested in using the data rejection feature of AMICA. By
default, it finds 'unlikely' frames and omits them from its further
learning iterations (recommended). To find what data points it rejected,
look in the EEGLAB structure for a variable lT or llT.

Jason also wrote a function that computes the likelihood of each data frame
under the AMICA decomposition. See EEGLAB > Tools > Post-AMICA tools. I
would try using a fairly narrow smoothing width (0.5-s ??), capturing the
output, and setting a rejection threshold. At the recent MoBI meeting,
someone in Klas Gramann's lab said they found that this approach is as good
as any more complicated procedure they tested (I don't have any reference
to this yet, though) - technically, I suppose, rejecting data on AMICA data
likelihood is likely about as complex/advanced a procedure as has yet been
attempted.

Scott

On Tue, Jul 19, 2022 at 3:07 PM Velu Prabhakar Kumaravel via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Hello,
>
> Does anyone have some thoughts to share on this?
>
> Thanks,
>
> Velu
>
> On Thu, 14 Jul 2022 at 12:00, Velu Prabhakar Kumaravel <
> velu.kumaravel at unitn.it> wrote:
>
> > Dear EEGLABers,
> >
> > As we know, it is rare to find data labeled for artifacts which makes it
> > hard to develop or validate artifacts detection algorithms. My colleagues
> > and I wonder if we could use ICA followed by ICLabel to label the
> segments
> > of data as "Artifact" or "Clean". The steps are as follows:
> >
> > 0) To facilitate reliable ICA decomposition for short windows of data,
> > only 4 channels are kept (known apriori)
> > 1) Bandpass filter (1-45 Hz)
> > 2) Segment data into 4s non-overlapping windows (sampling rate = 256 Hz,
> > 1024 samples)
> > 3) Perform ICA
> > 4) Classify components using ICLabel
> > 5) If there exists at least 1 eye or muscle Component with probability >
> > 0.95, the given window of data is "Artifact". Otherwise, "Clean".
> >
> > Do we see any potential problems in the approach? From our preliminary
> > analysis, more than 50% of windows are labeled for Artifact, hence the
> > concern.
> >
> > I appreciate your feedback.
> >
> > Thanks and regards,
> >
> > Velu Prabhakar Kumaravel, Ph.D. Student
> > Center for Mind/Brain Sciences,
> > University of Trento, Italy
> >
> _______________________________________________
> 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
>


-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott