Hi Jinseok,

Yes, ICA is a difficult topic.

First you can try to reject clear artifacts in your data to obtain a better ICA decomposition.

Then you should worry about which components to reject, which seems to be very subjective.

Mike Cohen has some interesting YouTube tutorials about this topic, where he mentions two important things:

1) if there's doubt, there is not doubt

2) it's better to leave a little noise than to remove a real signal

Some components are a contamination of EEG signal and noise, I would look into the ERP waveforms of the components and see whether this is clearly stimulus based or completely random.

If it's this latter, you would consider to remove this component.

Also keep in mind that components are ordered based on how many they contribute to the total signal, so removing the last component although it is very noisy, would not have that much effect in the signal.

All the best,

Cara

Ultimately how you reject ICs depends on research question and quality of data. I have often used a minimal approach to ICA, rejecting blinks, and eye movements and very clear instances of electrical noise, and complementing with some more general rejection of large amplitude epochs. That is a common approach, and suitable if you are not very familiar with IC. In some cases, with problematic data, I have tested using stronger criteria using IC-label such as rejecting all artifacts labeled with more than 80% confidence. It is very hard to evaluate such a criteria, I would avoid that kind of strategy if I'm wasn't very familiar with the type of data. With IC-label only ICs with very clear alpha is labeled 'brain' in my experience, and although that is a good positive indicator of a brain source, using this criteria only would reject too much data. IC-label is good help to learn about ICs but I wouldn't use it without other criteria as well. So, start with a minimal approach (that match the competence of your team in identifying ICs), and as you learn more you can perhaps use more radical IC-rejection criteria when needed.

Criteria to reject a persons data completely could be based on having good data in a sufficient number of epochs. For an N400 incongruence paradigm it is often enough with 30-40 epochs per condition, for other components and paradigms you often need more. Look at the specific literature for guidelines.

Dear Jinseok Eo,

Probably you might find useful the following thresholds for ICLabel. Those are specifically designed in preprocessing pipelines for artifact rejection (since ICLabel can be also used as a "marker" for brain components):

1) Automagic approach:

https://github.com/methlabUZH/automagic/wiki

'ICLabelParams', struct('brainTher', [], ... 'muscleTher', [0.8 1], ... 'eyeTher', [0.8 1], ... 'heartTher', [0.8 1], ... 'lineNoiseTher', [0.8 1], ... 'channelNoiseTher', [0.8 1], ... 'otherTher', [], ... 'includeSelected', 0, ... 'high', struct('freq', 2.0, 'order', [])),...

2) from_bids2stats - Pernet et al. approach:

https://www.frontiersin.org/articles/10.3389/fnins.2020.610388/full

% Run ICA and flag artifactual components using IClabelfor s=1:size(EEG,2) EEG(s) = pop_runica(EEG(s), 'icatype','runica','concatcond','on','options',{'pca',EEG(s).nbchan-1}); EEG(s) = pop_iclabel(EEG(s),'default'); EEG(s) = pop_icflag(EEG(s),[NaN NaN;0.8 1;0.8 1;NaN NaN;NaN NaN;NaN NaN;NaN NaN]); EEG(s) = pop_subcomp(EEG(s), find(EEG(s).reject.gcompreject), 0);end

Going to your questions:

1) As you have seen, both are conservative pipelines that try to reject artifactual components (at least eye and muscle components) using a threshold >= 80% according to the given ICLabel classification %.

Contrary to other answers, in my experience with ICLabel in ERP data (bimodal oddball), brain components don't always have the expected prominent alpha peak at 10Hz (which can be easily observed in ICLabel with resting-state eyes closed). A good alternative to evaluate if a component contributes to the ERP you are recording is to calculate the Percent Variance Accounted for (PVAF) using the envtopo() function or the GUI.

2) Regarding subject exclusion, I strongly recommend the Automagic approach, since it gives objective values (with plots and metrics) that facilitate the exclusion of any recording you have. You can find the details in the Github since their paper was not updated when they included ICLabel (but the documentation of Automagic - ICLabel integration, and quality control is in their github).

Hope this delayed answer can help you for further works.

Best,

A