Artifact subspace reconstruction (ASR)

An automatic, online method for removing high-amplitude transient artifacts. ASR learns a covariance statistic from a clean calibration segment, then slides a short window over the data; within each window it performs a PCA and flags principal components whose variance exceeds a threshold (set as k standard deviations above the calibration statistics). Flagged subspace directions are reconstructed from the retained, calibration-consistent subspace, so bursts of eye, muscle or motion artifact are attenuated while the rest of the data passes largely untouched.

ASR is popular in mobile and real-time EEG because it is unsupervised, causal and fast, and it handles the large sporadic transients that ICA (a stationary global model) handles poorly. Its behavior hinges on the cutoff parameter k: too aggressive and it removes real high-variance neural events, too lax and artifacts survive; recommended ranges (roughly 10 to 100, commonly near 20) are heuristic, and validation on the specific paradigm is essential. It also needs a genuinely clean calibration segment, which is not always available.