Extended ICA Removes Artifacts from Electroencephalographic Recordings

Tzyy-Ping Jung, Colin Humphries, Te-Won Lee, Scott Makeig, Martin J. McKeown, Vicente Iragui, Terrence J. SejnowskiNeurIPS 1997

Overview

This paper introduces an "Extended ICA" (Independent Component Analysis) framework designed to clean Electroencephalographic (EEG) data by treating it as a blind source separation problem. Unlike traditional methods that subtract noise using reference channels, this approach statistically separates brain activity from artifacts like eye blinks, muscle noise, and power line interference. By doing so, it preserves vital neural signals that were previously thought to be lost or irreparably corrupted.

The Challenge of Noisy EEG Data

EEG recordings are notoriously susceptible to noise. Because scalp electrodes pick up any electrical activity in their vicinity, the tiny voltages generated by the brain are often drowned out by much larger signals. These include eye movements (EOG), heartbeats (EKG), muscle tension (EMG), and even the 60Hz electromagnetic hum from building power lines.

Historically, researchers dealt with these artifacts in two primary ways:

  1. Rejection: Throwing away segments of data where an artifact (like a blink) occurred. This approach is highly wasteful and often impossible in clinical populations—such as children or patients with movement disorders—who cannot remain perfectly still.
  2. Regression: Measuring the noise (e.g., using an EOG channel for eyes) and subtracting a portion of it from the EEG. The fatal flaw here is that EOG channels also pick up underlying brain activity. By subtracting the EOG signal, researchers inadvertently subtract the very brain signals they are trying to study. Furthermore, regression requires a "clean" reference channel for every type of noise, which simply does not exist for muscle or line noise.

The Extended ICA Method

The authors propose using Independent Component Analysis (ICA), specifically an "Extended" version, to solve this issue.

This approach can be compared to the "Cocktail Party Problem": if you have several microphones in a room with multiple people talking, ICA is the mathematical tool that separates the mixed audio recordings back into individual voices. In the context of EEG, the "voices" are the brain, the eyes, the heart, and the muscles.

What makes it "Extended"? Standard ICA algorithms (like the original Infomax) were primarily designed to separate "super-Gaussian" signals—signals with occasional large spikes, such as speech or eye blinks. However, some artifacts, like 60Hz line noise, are "sub-Gaussian" (more uniform and periodic). The Extended ICA algorithm utilizes a learning rule that automatically switches based on the kurtosis (a statistical measure of "peakedness") of the signal. This allows the algorithm to isolate both types of noise simultaneously.

Key Insights and Ideas

  • Source Separation vs. Noise Subtraction: A core insight of this work is shifting from noise subtraction to source unmixing. Once an "eye blink source" is separated into its own independent component, researchers can simply set that specific component to zero and reconstruct a "clean" EEG without losing underlying brain data.
  • No Reference Needed: Because ICA relies on the statistical independence of sources, it can identify and isolate muscle noise or line noise without needing a dedicated reference sensor.
  • Revealing Hidden Signals: The authors found that after removing eye artifacts using ICA, they could clearly observe 8Hz alpha wave activity in the frontal channels that was completely invisible in the raw data. Traditional regression-based methods failed to recover this signal.
  • Bidirectional Contamination: The paper highlights that EOG electrodes are "contaminated" by the brain just as much as brain electrodes are contaminated by the eyes. ICA effectively addresses this bidirectional "leakage."
Comparison of ICA and regression for artifact removal, alongside spectral density analysis.

Key Results and Multi-Artifact Removal

The study evaluated the algorithm using two primary EEG datasets (one with 20 scalp electrodes plus 2 EOG placements, and another with 19 EEG channels and no EOG reference). Decompositions were performed on 10-second epochs.

In head-to-head comparisons with multiple-lag regression, ICA successfully recovered neural signals that regression accidentally deleted. More importantly, the algorithm successfully isolated and removed multiple distinct artifacts simultaneously:

  • Eye blinks and saccades: Isolated to specific frontal components.
  • Muscle spiking: Separated cleanly, even when the power spectra of the muscle and brain activity overlapped significantly.
  • Line noise (60Hz): Effectively concentrated into a single sub-Gaussian component.
  • Cardiac (EKG) signals: Widespread contamination was successfully reduced to one independent component.
ICA successfully isolates multiple artifacts including blinks, muscle noise, line noise, and EKG into independent components.

Remarkably, the training required only about 6.6 minutes on a standard workstation of the time (a DEC 2100A processor), making it computationally feasible for routine research.

Limitations and Future Work

While powerful, the authors acknowledge a few key constraints of the ICA approach:

  • Data Requirements: As a statistical method, ICA requires a sufficient amount of data (time points) and a sufficient number of channels to produce meaningful separations. Future work is needed to determine the exact minimum thresholds required to reliably remove specific types of artifacts.
  • Stationarity Assumption: ICA assumes that the spatial distribution of the sources remains fixed over the recording period. If an electrode moves or the physical source of the noise changes position significantly, the model's accuracy may degrade.

Summary updated

BibTeX
@inproceedings{jung1997extended,
  author = {Tzyy-Ping Jung and Colin Humphries and Te-Won Lee and Scott Makeig and Martin J. McKeown and Vicente Iragui and Terrence J. Sejnowski},
  title = {{Extended ICA Removes Artifacts from Electroencephalographic Recordings}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure