ICA-S³M: switching state-space model guided automatic EEG artifact removal from independent components

Area of Science:

• Neuroscience
• Signal Processing
• Biomedical Engineering

Background:

• Electromyographic (EMG) artifact poses a significant challenge in scalp electroencephalography (EEG) due to spectral overlap with neural signals.
• Muscle contractions can cause widespread artifact distribution across independent components (ICs), a phenomenon known as EMG smearing.

Purpose of the Study:

• To introduce ICA-S3M, a novel two-stage pipeline for robust EMG artifact removal from EEG recordings.
• To evaluate the performance of ICA-S3M against a convolutional neural network (CNN) in diverse EEG datasets.

Main Methods:

• The ICA-S3M pipeline first decomposes EEG data using independent component analysis (ICA).
• A switching state-space model is applied to each IC, separating neural oscillations (damped oscillators) from broadband artifacts (AR(2) process).
• The method generates explicit artifact probabilities without requiring training data.

Main Results:

• ICA-S3M significantly improved signal-to-noise ratio (SNR) by 12.4 dB in simulations, outperforming a CNN (1.2 dB).
• On a semi-synthetic dataset, ICA-S3M surpassed the CNN in 38 of 42 scenarios across multiple metrics and scalp regions.
• The CNN exhibited "alpha hallucination," introducing spurious alpha peaks, a failure mode avoided by ICA-S3M's adaptive approach.

Conclusions:

• ICA-S3M provides an interpretable, principled, and training-free alternative for challenging EEG artifact removal.
• The method's ability to adapt to individual recording spectral content prevents common failure modes seen in template-based approaches.
• ICA-S3M offers a significant advancement in EEG signal processing for accurate neural activity analysis.