Home » Resources for teachers, parents and researchers » Resources For Researchers » Automatic ICA Cleaning for Infant EEG Data

Automatic ICA Cleaning for Infant EEG Data

Looking at EEG recorded during naturalistic interactions, as we advocate here, is a good idea in principle but brings with it a lot of methodological problems! One of the biggest is when looking at artefacts to keep or reject. Manual artefact rejection is time consuming, but automatic ICA cleaning for infant EEG data comes with its own issues. For example, in traditional screen-based/ event-related tasks in which the child is passively exposed to a set of stimuli, artifacts are randomly distributed with respect to stimulation.

Topographical EEG maps example for automatic ICA cleaning for infant EEG

So removing sections containing artifact (e.g. times when the child’s getting fussy and not paying attention to the screen) can be beneficial – as it doesn’t affect the real signal. In naturalistic paradigms, though, removing whole sections of data is trickier because data segments contaminated by artifact often covary with cognitive/ attentional processes of interest. Specifically, in naturalistic paradigms, the ‘stimulation’ is often child-controlled (e.g., the child turning to the parent in a naturalistic interaction), and so artifacts are more likely to be time-locked to neural signals of interest; the removal of artifact is thus likely to also affect the analysis of neural signals. So we need approaches to artifact correction that remove artifactual signals from the EEG recording throughout the session, rather than removing whole segments of both signal and noise.

My PhD student Ira Marriott Haresign is working on this problem. As a first step, he has developed and tested some algorithms for automatic ICA cleaning for infant EEG in naturalistic data. (He didn’t do this from scratch – it was building on a previous paper from Irene Winkler and colleagues.)

ICA cleaning for infant data is inherently noisier – as more of the components contain a mix of neural and artifactual data, so it’s harder to classify them as either neural or artifactual. But we think we got them up to a standard where they’re useable.

You can download the algorithms from here.

One of the things that running the validation tests for this paper really rammed home to us though is that, as we know already but try to forget, no method of cleaning EEG data gets rid of all of the artifact completely. This is why we’re going on to look in more detail at how eye movements affect the EEG