9: Artifact Correction with Independent Component Analysis
Opening the book…
Learning Objectives
In this chapter, you will learn to:
Conceptualize artifact correction in terms of the overarching goal of your research and the specific problems that artifacts pose when you try to reach that goal
Decompose the EEG into a set of independent components (ICs), identify ICs that represent artifacts, and reconstruct the EEG without the artifactual ICs
Apply special preprocessing steps to obtain optimal ICs
Intelligently choose which ICs should be removed from your data
Combine artifact correction with artifact rejection to maximize data quality, eliminate confounds, and ensure that the sensory input is properly controlled
Evaluate the effectiveness of the combined artifact correction+rejection procedure
This chapter explains how to use independent component analysis (ICA) to correct certain kinds of artifacts (especially blinks and eye movements). ICA-based artifact correction is a real godsend for experiments in which artifact rejection would throw out too many trials. And it can improve the data quality for almost any experiment by allowing you to include most or all of the trials in your averaged ERPs.
However, ICA-based artifact correction massively changes your data. Every single data point is impacted. And if done improperly, ICA can make your data worse and lead to incorrect conclusions. One analogy I like is using a backburn to deal with a wildfire (i.e., starting a controlled fire to eliminate the fuel for the wildfire). If you’re not careful with a backburn, it can get out of control and burn what you were trying to save. Similarly, if you're not careful with ICA, it can get out of control and make your data worse rather than better. You really need to know what you’re doing with ICA to get the best results and avoid getting burned.
Before I wrote this chapter, I did a lot of reading to make sure I was up to date and that the strategies described in this chapter would reflect the current state of the art. I also spent a lot of time applying ICA to the ERP CORE data and carefully assessing the results. The Makeig group at UCSD are the world’s experts at ICA-based artifact correction, so most of the recommendations in this chapter are based on their recommendations. EEGLAB’s ICA documentation is an excellent resource, especially the videos created by Arnaud Delorme. The page of informal advice from Makoto Miyakoshi (called Makoto’s Preprocessing Pipeline) is also extremely useful. I recommend that you read these sources after reading the present chapter. I also recommend reading the general overview of artifact correction near the end of Chapter 6 of Luck (2014), along with the online supplement to that chapter.
As you read this chapter, keep in mind that the ultimate goal of artifact correction is the same as the ultimate goal of artifact detection, which is to accurately answer the scientific question that the experiment was designed to address. In addition, you should keep in mind the three main problems that we need to address in artifact rejection and correction: reduced statistical power as a result of increased noise, systematic confounds, and sensory input problems. Artifact correction will never be perfect, but it can still help you address these problems (especially combined with intelligent rejection of problematic artifacts that cannot be corrected with ICA). In the last section of this chapter, I describe an approach developed by my lab to evaluate whether the combination of artifact correction and rejection is actually effective (Zhang et al., 2024).