ICLabel & removing components

What does this show?

Why do I need this step?

Judging every component by eye takes time and experience. ICLabel is a plugin that ships with EEGLAB. It was trained on thousands of components labeled by hand, and it gives each component a probability for seven classes.

ICLabel is an assistant, not a judge. You review its suggestion, then remove the artifact components from the main dataset.

How do I do it in EEGLAB?

Start with the main dataset loaded and holding ICA weights, as in the Starting point.

First: classify the components

1. Open Tools › then Classify components using ICLabel › then Label components.
2. In Select which icversion of ICLabel to use:, keep Default (recommended) and press OK.
3. The component scalp maps appear, each with its most likely class and probability above it.
4. Click any component to open its panels: scalp map, spectrum, time series, and the probability of each class.

Second: review by eye

For each component you plan to remove, check the three signs:

The component panels you open by clicking a component come from the Viewprops extension, which ships inside ICLabel. You can open them for all components at once from Tools › then Classify components using ICLabel › then View extended component properties; the window title starts with View components properties - pop_viewprops(). The Viewprops files are in the folder <EEGLAB>/plugins/ICLabel/viewprops. The menu adds this folder to the MATLAB path by itself, but a script must add it with an addpath line before pop_viewprops, as this lesson’s script does.

Third: flag artifact components

A flag is a mark on the component, not a removal. Removal is the next step.

1. Open Tools › then Classify components using ICLabel › then Flag components as artifacts.
2. The Select range for flagging component for rejection window shows a Min and a Max for each class. Defaults: Muscle from 0.9 to 1, Eye from 0.9 to 1, the rest empty.
3. Press OK. The command window prints how many components were flagged.

Fourth: remove the components and compare

1. Open Tools › then Remove components from data.
2. A window lists the flagged component numbers and asks whether to remove them. Press Manual rej. to see and edit the list, not Yes, because Yes removes them without review. Ignore the STUDY note in that window for now; it is for multi-participant analysis in station 8.
3. The List of component(s) to remove from data field holds the flagged components. Edit it based on your review: add a component you are sure is an artifact, or delete a number you doubt. Then press OK.
4. A confirmation window appears. Press Plot single trials: black is before removal, red is after.
5. If the blinks are gone and the rest of the signal is unchanged, press Accept.
6. Name the new dataset D1 main clean and leave Overwrite it in memory (set=yes; unset=create a new dataset) unchecked. Save it with File › then Save current dataset as as D1_main_clean.set; the power spectrum lesson can start from it.

Fifth: apply the postponed low-pass

Now that the data are clean, apply the 30 Hz low-pass from the filtering lesson: Tools › then Filter the data › then Basic FIR filter (new, default), and type 30 in Higher edge of the frequency pass band (Hz). Name the resulting dataset D1 main clean lp30, then save it with File › then Save current dataset as. The epoching lesson starts from it.

How do I write it as code?

% Classify all components with the default ICLabel classifier
EEG = pop_iclabel(EEG, 'default');

% One row per component, one column per class
classes = EEG.etc.ic_classification.ICLabel.classes;
probs   = EEG.etc.ic_classification.ICLabel.classifications;
disp(classes);
disp(round(probs(1:10, :) * 100));

% Flag components above 90% Muscle or Eye
% Row order: Brain, Muscle, Eye, Heart, Line Noise, Channel Noise, Other
thresh = [NaN NaN; 0.9 1; 0.9 1; NaN NaN; NaN NaN; NaN NaN; NaN NaN];
EEG = pop_icflag(EEG, thresh);
artifactComps = find(EEG.reject.gcompreject);

% Review the flagged components
% pop_viewprops is in the ICLabel plugin folder; the menu adds it to the path, a script must add it
addpath(fullfile(fileparts(which('iclabel')), 'viewprops'));
if ~isempty(artifactComps)
    pop_viewprops(EEG, 0, artifactComps);
end

% Remove them from the main dataset
if ~isempty(artifactComps)
    EEG = pop_subcomp(EEG, artifactComps, 0);
end
EEG.setname = 'D1 main clean';
EEG = pop_saveset(EEG, 'filename', 'D1_main_clean.set', 'filepath', pwd);

% Low-pass at 30 Hz, postponed until after ICA cleaning
EEG = pop_eegfiltnew(EEG, 'hicutoff', 30);
EEG.setname = 'D1 main clean lp30';
EEG = pop_saveset(EEG, 'filename', 'D1_main_clean_lp30.set', 'filepath', pwd);

If no component is flagged, the script skips the review and the removal. If your review changes the list, edit artifactComps by hand before the removal block.

Which value should I choose?

In older papers you will see other tools for classifying components, such as MARA, ADJUST, and SASICA. These are older tools, and ICLabel is the default today.

How do I check it worked?

1. Compare the signal. Open Plot › then Channel data (scroll) on D1 main, then on D1 main clean. Blink peaks should be gone from the frontal channels, and the rest of the signal should be almost identical.
2. Count the components. In the command window, type size(EEG.icawinv, 2). It should be the previous number minus the components you removed.
3. Open the component maps with Plot › then Component maps › then In 2-D on D1 main clean. The blink component should no longer be among them.

What mistakes do people make?

• Trusting ICLabel blindly. It is an automatic classifier and it is sometimes wrong. Review every component with the three signs before removing it.
• Removing everything that is not Brain. A component whose top class is Other may carry part of the brain activity. If you are not sure it is an artifact, keep it.
• Removing components from the ICA copy instead of the main dataset. The 1 Hz copy is only for computing ICA. Remove components from the main dataset filtered at 0.1 Hz.
• Copying component numbers. Component order changes between runs. Component 3 in the official tutorial, or in this lesson’s picture, may not be the blink in your data.
• An old ICLabel version. Check its version in File › then Manage EEGLAB extensions, and update it if it is older than 1.6. EEGLAB 2026.1 ships with version 1.7.

How do I write it in my paper?

A template: edit the data, settings and versions to match what you ran.

Components were classified with ICLabel (version 1.7; Pion-Tonachini et al., 2019), and components with a probability above 0.9 for the Eye or Muscle class were reviewed visually and removed ([N] components). Data were then low-pass filtered (passband edge 30 Hz, cutoff frequency [33.64] Hz at −6 dB, filter length [59] points; pop_eegfiltnew, firfilt plugin 2.8).

Going further

• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html (6. Reject artifacts > d. Indep. Comp. Analysis: Automated detection of artifactual ICA components)
• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html (Subtracting ICA components from data)
• ICLabel Tutorial: https://labeling.ucsd.edu/tutorial/about (ICLabel Tutorial)
• ICA applied to EEG part 9: Automatically detecting ICA component classes: https://www.youtube.com/watch?v=Gi-0KiV_Az8 (Full video)