[Eeglablist] Splitting the data for ICA

[Jan Karsten]

Dear Scott and Makoto,

Thank you for your answers, they give me some ideas to test out and discuss with my colleagues (& share them, if meaningful). But as you said, I guess only a quantitative evaluation can give an answer regarding best practices.

Running the ICA on the bandpass filtered data (0.05 - 5 Hz) does not give any helpful results (even with a very low rank) for my data; the decomposition was not able to adequately isolate obvious artefacts (eyeblinks, cable artefacts...).

Regards,
Jan

-----Ursprüngliche Nachricht-----
Von: eeglablist <eeglablist-bounces at sccn.ucsd.edu> Im Auftrag von Makoto Miyakoshi via eeglablist
Gesendet: 29 August 2023 23:32
An: eeglablist at sccn.ucsd.edu
Betreff: Re: [Eeglablist] Splitting the data for ICA

Dear Jan,

> Do you think that the low-frequency data within the ICA sources is
problematic or can it be ignored? What are your suggestions on how to best deal with such data?

Here, you are asking a qualitative question. Often, this kind of question is answered by seeing quantitative evaluation.

> I am concerned, that this low frequency content will be randomly
distributed across the ICA components when applied to the second dataset (in my case with a band pass filter of 0.05 - 5 Hz).

It is NOT random.
You can think of the problem as follows.
Let's say you apply ICA 2 Hz, then copy the ICA weight matrix to the broadband data that contains 1-2 Hz signals.
What you can do is that you band-pass filter the data to 1-2 Hz first. This tells you the exact data ICA missed. You can study their non-stationary scalp topography changes etc.
Now, you apply the > 2Hz ICA to this 1-2 Hz band-pass filtered data. This tells you what will be added when you apply > 2 Hz ICA to your broadband data.

If you like, you can also run ICA on your 1-2 Hz data (I recommend you specify very low rank decomposition because such as narrow band-pass filtered data could be severely rank reduced; see my ICA-rank paper from the link below. The obtained ICA may show more correlation to your broadband ICA results rather than > 2Hz ICA results, IF the 1-2 Hz signal has dominant power over the entire freq spectrum (there is 1/f PSD curve, so lower the frequency, higher the power in general).

https://urldefense.com/v3/__https://www.frontiersin.org/articles/10.3389/frsip.2023.1064138/full__;!!Mih3wA!FTktodiwtkv_Agi2VLBbfwOviUhXg5O8pVlMiA1I0MAQxJqds0JJt441KG9uPc-s2qpW3JM4wLxchFs0UArw40XAVX8$ 

To conclude, your question can be answered by looking at data processed with a band-pass filter and ICA. After seeing these results, you can made a decision. Maybe the results are very obvious to you to see its meaning, maybe not.
See also this Wiki article.

https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline#What_happens_to_the_.3C_1_Hz_data_if_ICA_is_calculated_on_.3E_1_Hz_data_and_applied_to_0.1_Hz_data.3F_.2805.2F18.2F2022_Updated.29

Makoto


On Tue, Aug 29, 2023 at 12:07 PM Jan Karsten via eeglablist < eeglablist at sccn.ucsd.edu> wrote:

> Dear list members,
>
> I am currently working on a preprocessing procedure for EEG data with 
> the aim to analyse the movement-cortical potential. My main purpose 
> for the preprocessing  is to reduce the amount of artefact 
> contamination (especially eye blinks) using ICA. Since we are 
> interested in the MRCP, I am following the suggestion to split the 
> data before the ICA to apply a high-pass filter (1 Hz) to the data I 
> am running the ICA algorithm on and attach the solution to the second dataset I am using for the analysis.
>
>
> Now my question is:
>
> Because the ICA is trained with data that does not contain low 
> frequency content (due to the high pass filter at 1 Hz), I am 
> concerned, that this low frequency content will be randomly 
> distributed across the ICA components when applied to the second 
> dataset (in my case with a band pass filter of 0.05 - 5 Hz). Hence, in 
> the worst case, an ICA component that contains mostly eye blinks will 
> also contains more valuable information than usual in the low 
> frequency domain. Do you think that the low-frequency data within the 
> ICA sources is problematic or can it be ignored? What are your suggestions on how to best deal with such data?
>
>
>
> I attached a picture to the mail for clarification (I hope it can be 
> seen through the list), showing an ICA component containing eye blinks 
> acquired from the split datasets.
>
> Thanks in advance for your help and I am looking forward for your 
> replies,
>
> Jan
> _______________________________________________
> 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