[Eeglablist] Flexible preprocessing (ICA then epoching and vice versa)

[Scott Makeig]

Velu -

A paper published to test methods of 'bad data' rejection using ICA
decomposition showed that simply turning on the default data rejection
parameters used by AMICA decomposition performed as well (by their
measures) as using an explicit pipeline. In addition, it has the advantage
of separating the eyeblinks (and other non-brain source activities) from
the rest of the component data. You can then select what sources in the
data you want to study, and which you do not. E.g., you can back-project
and sum all components but those you deem outside the realm of your
research interest to constitute a 'cleaned' dataset.

Naturally, if you do not also remove all (else,?some portion?) of the
AMICA-ignored data points, you cannot guarantee that the ICA decomposition
will accurately reflect the sources active in the rejected data (e.g.,
non-stereotyped noise episodes, etc.).

I prefer in most cases not to back-project and sum components of interest,
but to study their individual time courses, both in relation to exp. events
and in relation to each other.

An exception is the case in which two or more ICs span a common 'dependent
subspace' in which ICA cannot make their time courses quite independent
(while separating their joint activity from other independent components
and component subspaces if any). The Post-AMICA Tools plug-in has a
function to detect 'dependent subspaces' in an AMICA decomposition of your
data.

Scott Makeig

On Tue, May 7, 2024 at 3:07 PM Zaeem Hadi via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Dear Velu,
> Many thanks for your suggestion.
> By high threshold, I meant the "max acceptable 0.5 sec window SD" option
> in EEGLAB. By default, it is set to 20. I have tried much higher values for
> it (ranging from 20-50) however it doesn't do well in many of the high
> amplitude or high-frequency noise periods in my data but keeps clipping the
> eye-blinks. I also have task related data and I am interested in
> time-frequency measures with a 4 second epoch (-1 to 3s), which is why I
> avoided ASR rejection as it can result in too much rejection within the
> trials. I was concerned that it might impact the time-frequency dynamics.
> I have resorted to manually interpolating the noisy time periods and then
> ICA as I wanted more control of the artifact removal process.
> Anyways, my main question was still about ICA then epoching or epoching
> then ICA, are there any expected differences due to the choice of either?
> (apart from the reordering of components due to variance differences, which
> is expected as epoch data is lesser than continuous data)
> I know both are possible choices of a pipeline, but could this be flexible
> within the same dataset? Can I do ICA-->epoching in some participants and
> epoching-->ICA in other participants within the same dataset,
>
> Best wishes,Zaeem   On Tuesday, May 7, 2024 at 09:51:06 AM GMT+1, Velu
> Prabhakar Kumaravel <velu.kumaravel at unitn.it> wrote:
>
>  Hi Zaeem,
> 1) When you say high threshold, do you mean a higher value for the ASR
> parameter? In ASR, the strict cleaning occurs at a lower ASR cut-off
> parameter (e.g., 3). But, from our experience on newborn EEG data
> (characteristic of high noise levels with non-stereotypical artifacts), a
> strict threshold of 3 removes an excessive amount of neural information.
> Moreover, we have repeatedly observed (both in adults and newborns) that
> ASR Rejection is better than ASR Correction. (see figure 7 in this
> manuscript:
> https://urldefense.com/v3/__https://www.sciencedirect.com/science/article/pii/S1878929322000123?via*3Dihub*sec0190__;JSM!!Mih3wA!GdxOq_kFP5WazJaxA_Jg655ZzjFnuk9RxuNskUzgJA_J-z8ehjf52KezVMGD5kdyZ16cQHUaiYAunlSq4m7l0CQ$
> )
> 2) It was not clear from your email whether or not you considered
> integrating ASR and ICA for your pipeline. In this case, you could use a
> relaxed ASR threshold (e.g., 20) and the residual eye-related artifacts can
> be removed by ICA + ICLabel.
> Hope this helps.
> Best,
> Velu Prabhakar Kumaravel, PhD
> On Mon, 6 May 2024 at 19:02, Zaeem Hadi via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
> Hi,
> I was wondering if it would be reasonable to keep the preprocessing
> pipeline a bit flexible as below in different participants of the same
> dataset. Particularly the order of ICA and epoching.
>
> In some subjects, continuous data is much cleaner and could be used for
> ICA with minimal rejection/interpolation. Whereas in some subjects the
> continuous data is very noisy and I was considering epoching first (which
> takes away most of the noise) and then doing ICA on epoched data in those
> subjects. I tried ASR but it keeps picking up eye-related data (blinks, eye
> movements) even with a very high threshold which I would like to keep for
> ICA and for later removal from data, so I would prefer to avoid it.
> In each case, I am doing re-referencing and epoch baseline correction
> after ICA as recommended.
> Is there an appropriate reference for justification for using this
> approach?
> Kind Regards,
> Zaeem
> _______________________________________________
> 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
>
> _______________________________________________
> 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



-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott