BCIwiki (bciwiki.com) — An adaptive version of Artifact Subspace Reconstruction (ASR) that tracks EEG structure with an exponential moving average (EMA) cut blink artifacts by 53.1% in 24-channel recordings from 10 subjects, compared with 24.0% for Original ASR and 26.7% for a memory-limited variant, according to a preprint posted to arXiv on September 28, 2026. The work has not been peer reviewed.
Removing artifacts from electroencephalogram (EEG) signals remains a critical challenge for deploying brain-computer interfaces (BCIs) in real-world settings, the authors wrote, because signal statistics are non-stationary and calibration data are limited. ASR suppresses artifacts automatically based on variance, but relies on a static covariance model derived from calibration, which limits stability when a subject's signals drift and makes it sensitive to mis-specified parameters.
The proposed method combines short-horizon covariance estimation with slow recursive assimilation, creating two adaptation timescales: it responds to evolving EEG structure while staying stable against transient artifacts. The authors say the design also yields smoother behavior across rejection thresholds, reducing sensitivity to hyperparameter choice in continuous use.
The recordings came from 10 subjects across cognitive states, all with prominent blink artifacts. EMA-ASR's stronger suppression came at the cost of increased reconstruction and spectral deviation, while the memory-limited variant performed comparably to Original ASR. The authors describe the approach as lightweight and deployment-oriented, with an explicit trade-off between suppression strength and signal preservation.
The paper is by Vishnu KN and Cota Navin Gupta; code is available on GitHub.