8: Artifact Detection and Rejection

EEG recordings are often filled with large artifacts. In most areas of research, blinks are the most problematic. They’re large (often 200 µV), occur frequently (on over 50% of trials in many participants), and may differ systematically across groups or conditions, creating a significant confound if they aren’t properly addressed. In research with infants, small children, or people who are required to move around during the task, movement-related artifacts are also a major issue. In my own area of research, eye movements toward lateralized targets are the most significant artifact.

However, in most of the ERP papers I read that use artifact rejection, it doesn’t seem that much thought went into the strategy for dealing with artifacts. These papers typically use a very primitive algorithm for detecting trials with artifacts, and they use the same rejection threshold for all participants (even though artifacts differ quite a bit across individuals). This chapter is designed to help you conceptualize and implement artifact rejection in a more sophisticated manner, allowing you to minimize artifact-related confounds and maximize your data quality, thereby achieving the ultimate goal of providing an accurate answer to your scientific question.

Most ERP researchers use artifact correction instead of artifact rejection to deal with blinks. In artifact correction, a method such as independent component analysis is used to estimate and correct for the artifactual activity so that trials containing these artifacts do not need to be rejected. This means that more trials are available for averaging, which typically leads to cleaner averaged ERPs and greater statistical power.

However, most sophisticated ERP labs use a combination of artifact correction and artifact rejection. Artifact correction is mainly used to correct for eye blinks and eye movements and tends not to work well for some of the other common kinds of artifacts, such as movement artifacts. Rejection is therefore used for those artifacts. My lab has actually demonstrated that the combined use of artifact correction and artifact rejection is valuable across a broad range of ERP paradigms, both because it avoids confounds and because it improves statistical power (Zhang et al., 2024c).

In addition, ICA-based artifact correction requires reasonably long recording sessions, especially if you have a lot of channels. If you have short recordings and a lot of channels, you may need to reject trials with blinks rather than correcting them.

The bottom line is that you need to have a good understanding of artifact rejection even if you plan to use artifact correction. That is the goal of this chapter. The next chapter will discuss artifact correction and how to combine correction and rejection.

---