Automatic bad-channel detection

Individual electrodes fail in characteristic ways — a broken lead or dried gel gives flat or saturated (railing) signals, high impedance gives excess noise, a lifted electrode gives drift — and one bad channel can corrupt average referencing, ICA and spatial filters for the whole montage. Automated detection flags channels by deviation statistics: abnormal variance, low correlation with neighbours, high-frequency noise, or poor predictability from surrounding electrodes via a RANSAC and spherical-spline interpolation criterion, as in the PREP pipeline.

Flagged channels are typically interpolated from their neighbours rather than dropped, so the montage stays complete for downstream spatial methods, though interpolation invents no new information and cannot recover a source seen only by the lost electrode. Distinguishing a globally bad channel from one that is merely artifact-contaminated on some epochs matters: the former should be interpolated, the latter handled by epoch-level cleaning.