[FieldTrip] a question about preprocessing

Stephen Whitmarsh stephen.whitmarsh at gmail.com
Sat Apr 16 14:41:50 CEST 2022

Dear Masoud,

ICA can certainly help you with removing artefacts from eye movements and
eye blinks, and perhaps ECG artefacts. Please see the following
documentation for explanation and use case:

Before you do ICA you can reference (I assume to a common reference),
especially if your original reference you recorded with is bad. Otherwise,
I don't think it would matter much. But definitely don't do any "clinical"
or "sleep" referencing that reference against different electrodes (i.e.
nothing but either a single, linked or average reference).

In general, I would not remove too much slow activity (hp filter) yet, but
"give it to the ICA" to better estimate EOG components (which include slow
activity). Same for downsampling, I would not do it (too much) beforehand,
as you want to be able to unmix EOG/ECG artefacts. It all depends on the
compromise between time/memory, as downsampling can drastically reduce
time/memory use for ICA. But this is all relative to the quality of your
data: ICA does a better job when it has good data to work with, and if you
can remove a lot of slow drifts and e.g. 50/60Hz noise, it will be able to
unmix the data in useful EOG/ECG components. A high-pass of 1Hz sounds
reasonable to me.

Similarly, if you have a very noisy channel and not that many electrodes,
removing it beforehand might help the ICA to unmix the artefacts of
interest and not "spend components" on the noisy electrode. In fact, I
would not use ICA to remove noisy channels, as that might introduce some
uncontrolled changes in the rest of your channels, as components in my
experience rarely pick up only one noisy electrode (always check the
topographies of the components), and even if does so, you wasted a
component on it.

In fact, ICA for artefact detection works best if you have sufficient
electrodes, so it can unmix the data in enough (and detailed enough)
components. You will only get n components, with n being the number of
electrodes. If you have say only 8 electrodes, you might want to reject
trials with e.g. blinks instead, rather than relying on ICA and remove
several 1/8th of your data. If you have separate EOG channels it will allow
you to better identify EOG components. Always check your component time
courses and topographies, and be confident that the unmixing worked well,
and the component represents a clear EOG/ECG component before you remove
it. In contrast, rejecting trials will always be 100% successful in
removing the artefact :).

ICA is not a sure success, but it can be very impressive. In the end, it
depends on the details of your data and experiment, so some empirical tests
could be the most informative.  The questions you ask are good ones, and at
the end of the day you might just want to try out different ways and see
what works best for you. My experience is mainly with MEG, and even there
we used to try out what works best before deciding what to do for the rest
of the study. One important consideration is whether it works for all
datasets in your study. E.g. I wouldn't turn things around for a single
subject with a bad reference, but I would also like to try to treat every
dataset the same.

I hope this helps,

P.s I now see you preprocess your data in EEGlab first - before analysing
in FieldTrip? Please note that the code for ICA is shared between EEGlab
and FieldTrip. My answer does not consider what is more practical when
having to navigate between them.

On Sat, 16 Apr 2022 at 13:27, masoud via fieldtrip <fieldtrip at science.ru.nl>

> Dear fieldtrip community
> I hope this email finds you well. I am writing to you with the hope that
> you might be able to help me to find a proper answer for a question which I
> have been dealing with.
> I am a master student in Biomedical Engineering and I've been using the
> EEGlab toolbox to preprocess the acquired EEG data for my master thesis.
> My question is about EEG preprocessing orders, more specifically when is
> more suitable to apply ICA  to my data. In my current setting, first, I
> downsample data and use a hp filter (cut-off frequency is 1 Hz), then I
> interpolate the suspected bad channels, but I don't know if I should use
> ICA and then use re-referencing or vice versa?
> Although I know there is no definite answer for this question, it would be
> highly appreciated if you kindly share your opinion and experience as an
> expert in this field  with me, and let me know if my current preprocessing
> setting is appropriate and, specifically, should I use re-referencing after
> ICA or before that?
> Thank you in advance for your consideration. I am looking forward to
> hearing from you.
> Best regards,
> Masoud
> _______________________________________________
> fieldtrip mailing list
> https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> https://doi.org/10.1371/journal.pcbi.1002202
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20220416/62512e26/attachment.htm>

More information about the fieldtrip mailing list