Motion and physiological artefacts in real-world signals

In everyday recording, the largest voltages an electrode picks up are usually not brain. Eye movement and blink (electro-oculographic) potentials, jaw, neck and facial muscle (electromyographic) activity, the heartbeat, sweat-driven slow skin-potential and impedance drifts, cable sway and triboelectric charging, footfall-locked motion, and mains pickup all sit on top of a cortical signal that is often smaller than any one of them. The problem is not that these exist — it exists in the lab too — but that in the field they are large, non-stationary, and frequently time-locked to the very behaviour being studied.

Removal is a spectrum with real trade-offs. Simple approaches — bandpass filtering, regression against an eye or motion reference, thresholding — are fast but blunt. Decomposition methods such as independent component analysis, and adaptive cleaning such as artifact subspace reconstruction, separate or reject contaminated components but need enough channels and can distort the neural signal if pushed too hard. The honest danger is a decoder that learns to read the artefact: a system that appears to classify intent may in fact be reading jaw clench or gaze, which is why artefact-controlled paradigms are essential.