العربية

ICA & artifact cleaning

EEGLAB's strongest tool for removing blinks, muscle and heart artifacts without discarding data, and the hardest beginner skill: reading components.

7 lessons: 5 core, 1 recommended, 1 optional

After this station you can:

  • Run ICA on a 1-Hz-filtered copy and copy the weights to your data
  • Classify components with ICLabel and check its decision yourself
  • Remove artifact components and verify the signal improved

Lessons

  1. ICA, simply explainedCore12 min
  2. How much data ICA needsRecommended10 min
  3. Running ICACore15 min
  4. 1-Hz filtering for ICA & copying weightsCore12 min
  5. ICLabel & removing componentsCore15 min
  6. Reading artifact componentsCore15 min
  7. AMICA & PicardOptional12 min

Test yourself

4 questions. Your score stays with you and isn't sent anywhere.

1. You ran ICA on your data and have not removed any component yet. What changed in the data?
Show the answer

The correct answer: Nothing: ICA gives you another way to see the same data as independent sources, and removal is a separate decision

ICA computes weights that unmix the channels into components, each with a scalp map and a time course, and removes no channel and no part of the recording. Removing a component is a decision you make after review, in the ICLabel lesson.

Review the lesson: ICA, simply explained

2. Put the ICA cleaning steps in the order of this site's path:
  1. Copy the weights to the main dataset (0.1 Hz)
  2. Remove the artifact components from the main dataset
  3. Create a copy with a 1 Hz high-pass filter
  4. Classify the components with ICLabel and review them by eye
  5. Run runica on the 1 Hz copy
Show the answer

The correct answer: 1. Create a copy with a 1 Hz high-pass filter, 2. Run runica on the 1 Hz copy, 3. Copy the weights to the main dataset (0.1 Hz), 4. Classify the components with ICLabel and review them by eye, 5. Remove the artifact components from the main dataset

ICA is computed on a 1 Hz copy because slow drift confuses it. Then the weights move to the main dataset that goes on to the ERP. Classification and review come next, and removal is always from the main dataset, not from the 1 Hz copy.

Review the lesson: Running ICA

3. Why is it valid to copy ICA weights from the 1 Hz copy to the main dataset filtered at 0.1 Hz?
Show the answer

The correct answer: Because the weights are a fixed way to combine channels (a spatial filter), and both copies have the same channels and the same mixing in the head; only their frequencies differ

ICA does not learn a signal; it learns how to combine the channels. That is why the weights move between two copies with the same channels in the same order. A 1 Hz filter does shrink P3, which is why the analysis continues on the main dataset, not on the ICA copy.

Review the lesson: 1-Hz filtering for ICA & copying weights

4. Which statements about ICLabel's decisions are true? (Select all that apply)
Show the answer

The correct answer: ICLabel's probabilities are an estimate from a trained classifier and can be wrong, so you review every component before removing it, A component whose top class is Other is not removed unless you are sure it is an artifact, A component at 55% Eye and 40% Brain needs your review before any decision

ICLabel is an assistant, not a judge. Removal is from the main dataset at 0.1 Hz, because the 1 Hz copy is only for computing the weights. Component order changes between runs, so find a component by its shape, not its number.

Review the lesson: ICLabel & removing components