The EEGLAB News #15

Question: We've recently begun adding the use of ICLABEL to our pre-processing pipeline. What we've found is that it labels only a minority of components as brain activity, with the majority being labeled as "other." If we reject the components labeled as non-brain activity, we end up rejecting the majority of the components, and this seems wrong. Can anyone who has been using ICLABEL comment on how to select components for removal? Thanks in advance, Michael


Scott Makeig. ICLABEL compares component properties in your data (e.g., component scalp projection maps, power spectra) with datasets in its large training data (mainly datasets from our 20-yr history of applying ICA decomposition to EEG data at SCCN). ICA decomposition can be negatively affected by several factors: too little data, abundant non-brain noise in the data, etc. - and by the conditions under which it was recorded (e.g., Were participants moving? Were the electrodes securely placed? etc.).

So the first thing I would suggest you look at is whether your data preprocessing and data rejection process was adequate for the data. Next, I would suggest you see how much of the data is accounted for by the labeled Brain components plus the non-brain components of known origin (e.g., Eye Movement components). Many times components rated as 'Other' by ICLabel account for quite little of the data (e.g., single-channel ICs) - forming an (as it were) ICA 'noise subspace'.

If you perform PCA decomposition on your dataset and look at the values of the resulting eigenvalue spectrum, you will typically find that a large proportion of EEG data 'lives' in relatively few dimensions - ICA decompositions typically find a relatively small set of ICs that account for most of this subspace such that each basis element (Independent Component) is as temporally distinct from the others as possible -- and is thereby typically functionally distinct from others, And spatially distinct from others. 


David Gilbert. A related question I have concerns the fact that ICLabel returns percent probability of an IC of being EEG, heart, EMG, other, etc. Can you refer me to how these percentages were calculated. I believe that there were a number of experts, but were they looking at epoch-by-epoch portions of the data?
Thanks in advance.


Scott Makeig. The ICA 'experts' viewed IC properties windows like this one in entering their 'expert' opinions to give ICLabel training a coherent beginning.

Follow the ICLabel tutorial on making property-guided IC-classifying decisions here.


Makoto Miyakoshi. Lisa pointed me to this post.

See this paper: Automated preprocessing and phase-amplitude coupling analysis of scalp EEG discriminates infantile spasms from controls during wakefulness (Miyakoshi et al, 2021). In Figure 4, it shows Brain 53%, Muscle 12%, Eye 9%,
Channel Noise <1%, Line noise <1%, Heart < 1%, Other 24% for awake state. During the sleep state, Brain 74%, Other 23%, everything else < 2%. This is for the case of 19 channels.

See also this Wikipedia article. It shows Brain 52%, Muscle 30%, Eye 6%, Heart 2%, Other 11%. This is for the case of 38 channels.

I have several other unpublished datasets with 64 channels (the LEMON datasets; Babayan et al., 2019. Scientific Data). They also showed Brain class around 55%. Although this cross-number-of-channel test is not official, my impression is that the EEG data with standard quality seem to show 50-55% of Brain class rate regardless of the number of channels.

I also calculate percent variance of the Brain etc. classes, but these are not published. I will next time.