ICA-based artifact removal

Independent component analysis unmixes the multichannel recording into maximally statistically-independent components, each with a fixed scalp topography and its own time course. Because eye, muscle, heart and single-channel artifacts are generated by sources spatially and statistically distinct from brain rhythms, they tend to concentrate into a few identifiable components; zeroing those components and back-projecting the rest reconstructs cleaned data without discarding whole epochs — the key advantage over rejection.

The method assumes linear, stationary mixing with at least as many well-conditioned channels as sources, and it needs enough data to converge; results depend on high-pass filtering the training data (near 1 Hz) and on the algorithm (Infomax, FastICA, AMICA). Component selection was historically manual and subjective; automated classifiers (ICLabel, ADJUST, MARA) now label components as brain, eye, muscle, heart, line or channel with reasonable accuracy. ICA cannot separate sources that mix nonlinearly or that are fewer-than-full-rank after average referencing, and aggressive removal can distort the brain signal it leaves behind.