Artifact Subspace Reconstruction (ASR)#

ASR estimates a reference covariance from relatively clean calibration data, detects windows whose component variance exceeds that reference, and reconstructs the affected subspace. It is intended for transient, high-variance artifacts in continuous EEG [1][2].

Usage#

from mne_denoise.asr import ASR

asr = ASR(sfreq=250.0, cutoff=20.0)
clean = asr.fit_transform(data)  # data: (n_channels, n_times)

For supported MNE objects, pass the object directly. fit calibrates the state, transform applies it without mutating the input, and fit_transform composes both operations.

Key points#

  • Calibration should represent the clean covariance of the processed channel type. calibration="auto" selects windows from robust RMS statistics; an explicit calibration mask or array can supply a trusted period.

  • cutoff is a multiplier in the calibrated component space. Lower values generally reconstruct more components, but its effect depends on calibration and processing settings.

  • method="standard" is the default. "riemannian_windowed" uses the windowed Riemannian backend; "riemannian" requires experimental=True [3].

  • AdaptiveASR updates calibration state between chunks; JugglerASR changes reference-sample selection [4][5].

  • GuidedASR adds artifact and preserve covariance guidance. Soft guided reconstruction is experimental and requires explicit experimental=True.

  • Inspect the fitted diagnostics and repaired spans. Artifact attenuation alone does not establish preservation of the signal of interest.

References#

On this page