[Eeglablist] ICA with bad electrodes
arno at ucsd.edu
Sat Nov 22 14:59:57 PST 2008
> a discussion arised in our lab (between biologists and engineers)
> having some (in the order of 10 out of 64 ) very bad channels
> producing a
> lot of different artifacts due two high impedance, bad skin contact or
> movement, the more correct and succesfull solution is:
> A: to eliminate such channnels (and eventually the controlateral
> symmetrical ones) from the dataset before performing ICA , thus
> the number of channels recorded or
> B: perfoming ICA on the all dataset hoping that ICA will segregate
> channels in separate components.
It depends on the level of noise and the number of electrode you have.
If the noise is relatively low, it would be better to use ICA. ICA
will most likely dedicate one component to this artifactual signal. If
the noise is too strong, this might make the decomposition unstable.
This is what I would recommend:
1) You have a lot of electrodes (>64), then you should just remove the
electrode as you can afford to loose some.
2) You have few electrodes and the noise is low: then I would first
try ICA and include such electrodes.
3) You have few electrodes and the noise is high but only transiently:
try removing selected portion of data where the signal is bad for all
4) You have few electrodes and the noise is high on one electrode all
the time: if you are using ICA to remove artifact only, then try ICA
to remove the bad signal on this channel. If you want to use ICA to
isolate specific brain rhythms, i would recommend to remove such bad
> A further question: theoretically is it better to have symmetrical
> distributed electrodes for example on the right left side of the
> brain or
> ICA works as well with unbalanced montages?
In theory, it is best to have homogeneously distributed electrodes so
source localization on cognitive ICA components will be more accurate.
ICA does not take into account the electrode localization. ICA itself
should work well with unbalanced montage.
Hope this helps,
More information about the eeglablist