[Eeglablist] Merging files for preprocessing and ICA

[Tarik S Bel-Bahar]

Hello Caroline, yes ICA is said to have a big stomach so it can detect
those spatial patterns! I would say the merging is ok in your case,
especially if you're getting good & similar ICs across both merged
datasets, with the exception of the big signal drop you've pointed out.
Some notes below, best wishes.

*There's plenty of published work that uses  short periods of time (~ 2 to
5 minutes) with hdEEG and ICA to get seemingly valid results  in top-tier
journals. Similarly, some researchers publish on source estimates from
16-channel EEG, whereas others would only use hdEEG. Just because something
gets published does not mean the methods/findings are valid...it's just
that in practice some rules are flexible or change depending on the
researchers, methods, and biases therein. That being said, it's better for
the field to adhere to best practices so findings can be comparable.

*the ideal number of samples for hdEEG is an ideal guideline, though in
practice  much shorter periods can be ICAed with reasonable results. My
understanding is that this approach works well for artifactual IC
detection, even with low channel numbers (16 or 32) and brief periods of
time (even for 64 and 128 channel data). And, if you're looking for the
neural ICs - if one's data has good examples of that kind of brain activity
(ie a relevant task) - then one should also be able to get valid ICs with
less channels and brief time periods. In your case, I would consider also
down-sampling to 64 channels and comparing resulting patterns.

*note there are few to no publications really doing comparisons on this
methodological topic.

*note there are also advancements and alternatives to traditional ICA (see
a range of source separation techniques used with real-time protocols such
as BCI, for example).

*artifacts from the same person, with the same cap/recording, should be
quite similar, it's good that you have confirmed this. If your thesis on
nap effects is correct, you should literally be able to see more relaxation
in the signal at it is being recorded, more relaxation on their faces and
body, and perhaps on a self-report scale.

*however, the big signal drop in the ICs between the two periods
(presumably across all ICs) is worrisome. Does the straight EEG look
different across the two ? Is this the same across multiple participants ?

*something may have changed in signal quality or artifacts between the two
recordings you're trying to join. Try a control case with rest but nothing
that could change the signal. Try also checking impedances across the two
recordings, and at different times between the two recordings. If you were
using a egi hydrocel, check whether the electrodes or nets were
significantly moved, rewet, or adjusted during that time.

*another check may be to record continuously across all tasks and all naps,
and see if similar changes occur in the signal.

*you may be safer off by doing separate ICAs for the periods, and then
finding matching ICs across them. See if the problems remain when you do it
this way.

*generally, recordings that are done in different sessions (with a new
putting on of the EEG cap/net) are better off with separate ICAs- per
session/subject.




On Wed, Nov 30, 2016 at 2:06 PM, Caroline Lustenberger <
lustenberger.caroline at gmail.com> wrote:

> Dear EEGLAB Team
>
>
>
> There is always the discussion about how many samples are needed for a
> good ICA and based on the tutorial pages from your lab it seems to be quite
> many. People that use high density EEG with 128 electrodes would likely
> need ~1000000 data points, which means more than 1 hour of EEG recording
> (considering we do downsample to 250 Hz). I doubt that all of us have
> recordings that are that long. At least I often have tasks that last for ~
> 10-20min. One idea that I thought of is to merge datasets of a person if
> she/he had multiple sessions on the same day with the same task (that is
> having the same net application). I would like to hear your opinion about
> this idea, whether you also do this approach and what I should consider
> when merging files.
>
>
>
> Here is a specific example: I tried this approach for a visual memory task
> during which my participants were seeing and rating pictures before a nap
> and after a nap. I merged the two datasets and after preprocessing (filter,
> clean_rawdata, interpolation, average referencing) I performed an ICA
> (tried both AMICA and RUNICA, results are very similar). The components
> seemed to be captured fine, however I clearly saw that many of the signals
> (e.g. muscle, but also physiological brain components) changed in
> activation after the transition of the merged files. More specific, the
> components were capturing the same type of activity (e.g. muscle,  eye
> blinks or alpha) but the amplitudes of the component clearly changed
> between the first and second data set. That is for instance one muscle
> component had strong activation in one dataset and was very low amplitude
> in the second part of the data. People are likely more relaxed after sleep,
> are maybe more focussed and I was wondering whether merging of datasets is
> not ideal because conditions are slightly different and I even expect that
> oscillatory activity relative to the stimuli will slightly differ before
> and after sleep for the same task. What is your impression? Would you
> rather process the files separately or merge them together to one?
>
>
>
> Thanks for your valuable input and best wishes,
>
> Caroline
>
>
>
> _______________________________________________
> 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
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20161201/196f65a3/attachment.html>