EOG regression correction

The classical subtractive method for ocular artifact. One records dedicated vertical and horizontal EOG channels, estimates by linear regression the propagation coefficients (transfer weights) that describe how much EOG appears at each EEG electrode, then subtracts the scaled EOG from every channel. The Gratton, Coles and Donchin procedure computes these weights on blink and saccade segments after removing the stimulus-locked average, so that genuine ERP variance is not regressed away.

Regression is fast, causal and works online, which keeps it in use for real-time BCI. Its weakness is bidirectional contamination: the EOG electrodes also pick up frontal brain activity, so subtracting them removes some real EEG (overcorrection), and a single global weight cannot capture the differing topographies of blinks versus saccades. Multiple regression and separate blink and saccade models mitigate but do not eliminate this; component-based methods such as ICA are generally preferred offline because they do not require sacrificing part of the signal.