When discontinuity artefacts are detected using ICA decomposition, the identified 'bad' independent component (IC) might contain a mixture of true artefact and data from neural sources.Throwing away that IC would then reduce the amount of 'true' EEG: this is quite obvious when using ADJUST toolbox to detect and remove bad ICs based on discontinuities, because running ICA again after such cleaning results in far fewer ICs than data channels.
Has anyone published estimates of how much other signals are carried along with discontinuity ICs? What would be a more clever way to deal with these artefacts?
To answer the gist of your question, it probably depends on how correlated your neural components are with your artefactual ones (as long as you accept the assumptions of ICA). For instance, if your stimulus onset tends to evoke a blink, evoked potentials might well be decomposed into the same component.
Thank you Matt and Matthew. To clarify, when referring to true EEG I don't mean only neural sources, but rather anything that is not the artefact in question.
Matthew: I would guess that discontinuities are highly uncorrelated with EEG.
The goal is: discos happen quickly and not very often, and everywhere else in time the activation vector for an IC that has been identified as capturing discontinuity will probably be non-zero: this data should be preserved. In order not to do this manually, what options are there?
To make it clear, you are talking about spatial discontinuities right? I would also guess that they're not temporally correlated with genuine EEG, so any associated genuine EEG should be essentially low-amplitude random noise and not likely to bias your results in any particular direction.
I generally go with what Matt suggested, run ICA once, eliminate the epochs in which the component is expressed (if reasonably possible), then run again. It's been a couple of years since I needed to do that, but I couldn't find a satisfactory automated solution. There have probably been improvements though.
Thanks Matthew. I'm strongly motivated to minimise epoch deletion, because I'm running complex protocols which can't have massive numbers of trials per condition without inducing serious fatigue effects. And I've been searching for methods to clean, rather than reject, all types of artefact. Preferably automated, because a) I'm lazy and b) I think my subjective judgements are too prone to drift.
There might be some ways to automate the suggestion by Matt by detecting IC activations outside confidence intervals, i.e. bad epochs. There might also be ways to avoid deleting those epochs, using changepoint smoothing based on Gaussian processes.
On the other hand, there might be a good chance of never getting published with such esoteric methods!
In that case you might want to look at artefact subspace reconstruction (ASR) as well. Sometimes you may have to just bite the bullet and do things by visual inspection. There might be some drift, but you can hopefully reduce that through practice.