<p style="margin-bottom: 0cm;"><font><font size="2">Jordi Costa Faidella wrote "</font></font>Is it correct to
perform an ICA on a dataset in which some of the channels have been
interpolated?"<br>
</p>
<p style="margin-bottom: 0cm;"><font><font size="2">This is an interesting question and we considered both orders
(each order has some advantages and disadvantages) for the FASTER
method. Ultimately, we decided to run interpolation first followed ICA.
Here was our rationale:<br>
</font></font></p>
<p style="margin-bottom: 0cm;"><font><font size="2">As the EEGLAB manual recommends – “ICA works best when given
a large amount of basically similar and mostly clean data.” (see p.59).
Therefore, an ICA on a dataset in which some channels are noisy
(perhaps with a lot of non-stereotypic data due to a problem with the
electrode) may decrease the quality of the ICA (i.e., dissimilar
activations are mixed into the ICs).</font></font></p>
<p style="margin-bottom: 0cm;"><font><font size="2">On the other hand, interpolating before ICA raises a couple
of issues 1) it reduces the dimensionality of the data and 2)
introduces some non-linearity into the data (if the interpolation
method was not linear), which is detrimental to the ICA solution. We
dealt with Issue 1in FASTER by restricting the maximum number of ICs to
correspond with the reduced rank of the data after interpolation.</font></font></p>
<p style="margin-bottom: 0cm;"><font><font size="2">The choice then was between reducing the quality of the ICA
by introducing noisy channels or reducing the quality of the ICA by the
non-linearity introduced due to spherical interpolation. </font></font><font><font size="2">Although ICA assumes
linearity, there is almost certainly some non-linearity in the signals
recorded at the scalp, and the non-linearity introduced by spherical
interpolation is likely only a small contributer to the overall
non-linearity. </font></font><font><font size="2">In any case, based on pilot testing we found that when the
ICA was done with noisy channels included (i.e., not interpolated out)
the resulting components were less useful than when the data were
cleaner (i.e., the channels were interpolated). As an aside, testing
algorithms on real data proved much more informative than testing on
the simulated data, perhaps due to the inclusion of non-stereotypic
artefacts in the real data.</font></font> <br>
</p>
<p style="margin-bottom: 0cm;"><font><font size="2">That said, we are certainly open to persuasion on this issue </font></font><font><font size="2">and/or
suggestions about how to quantify which order is better</font></font><font><font size="2">. Also, might
there be situations in which one order is superior to the other,
perhaps depending on the maximum number of ICs that can be generated? <br>
</font></font></p>
<p style="margin-bottom: 0cm;"><font><font size="2">If there is demand, we can also configure FASTER so that the
user can select the order of the processing steps. Email me directly <a href="mailto:robert.whelan@tcd.ie" target="_blank">robert.whelan@tcd.ie</a> or <a href="mailto:whelanrob@gmail.com">whelanrob@gmail.com</a>
if this is something that people might want or with any other
suggestions.<br>
</font></font></p>
<font><font size="2"><br>
Best Regards, <br>
<br>
Rob & Hugh</font></font><br clear="all"><br>-- <br>Robert Whelan, PhD<br>Senior Research Scientist<br><br>Trinity Centre for Bioengineering<br>Trinity College Dublin<br><br>Department of Neurology<br>St. Vincent's University Hospital<br>
Elm Park, Dublin 4<br><br>webpage: <a href="http://www.mee.tcd.ie/~neuraleng/People/Robert">http://www.mee.tcd.ie/~neuraleng/People/Robert</a><br>