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

Background Electromyographic (EMG) artifact is among the most challenging contaminants in scalp EEG: its broadband spectrum overlaps directly with neural activity, and sustained muscle contractions distribute across many independent components (ICs) rather than segregating cleanly—an effect we term EMG smearing. Methods We propose ICA-S 3 M, a two-stage pipeline that first decomposes the recording via independent component analysis (ICA), then applies a switching state-space model to each retained IC. This model separates neural oscillations, modeled as damped oscillators fitted to the recording's own spectral content, from broadband artifact, modeled as an autoregressive (AR(2)) process. The method produces an explicit artifact probability at every time point and requires no training data. We evaluated ICA-S 3 M against a CNN from EEGdenoiseNet in three settings: simulated EEG with known ground truth; a semi-synthetic real-EEG dataset constructed from 29 participants through 11,025 expert IC–trial classifications, in which a trained rater classified every IC on a per-trial basis across interleaved rest and facial-movement segments; and a naturalistic recording of musical improvisation. Results In simulation, ICA-S 3 M reduced RRMSE from 0.83 to 0.23 and improved SNR by 12.4 dB, versus 1.2 dB for the CNN. On the semi-synthetic dataset, ICA-S 3 M significantly outperformed the CNN in 38 of 42 tested scenarios spanning seven metrics, two time periods, and three scalp regions, while leaving clean segments essentially untouched. Critically, the CNN exhibited an “alpha hallucination” failure mode, injecting spurious alpha-band peaks where none exist in the ground truth; we reproduced this with an alpha-free control simulation and observed it again on real scalp recordings. Discussion ICA-S 3 M avoids this failure by adapting to each recording's own oscillatory structure rather than imposing learned spectral templates. The method offers a principled, interpretable, and training-free alternative for EMG artifact removal in challenging EEG recordings.