Choose a method

mne-denoise includes methods that rely on different kinds of information in the recording. Some use clean calibration periods, some spatial or spectral structure, some reference channels or repeated trials, and others a forward model.

No denoising method is appropriate for every recording. Match the method to the artifact, the available structure, and the signal you need to preserve.

The sections below are intended as a starting point. The individual method pages describe the assumptions and behavior in more detail.

Transient high-amplitude artifacts

ASR

Detect and reconstruct windows whose component variance departs from a clean reference. Main assumption/input: a relatively clean calibration period or calibration data. See the ASR guide; adaptive ASR variants remain part of this family.

Sensor-specific noise

SNS

Reconstruct sensors from spatially redundant neighboring sensors. Main assumption/input: noise is specific to individual sensors while the signal of interest is spatially shared. See the SNS guide.

SOUND

Estimate channel-specific noise from forward-model geometry and reconstruct a cleaned sensor-space signal. Main assumption/input: a compatible EEG montage or an explicit lead field. See the SOUND guide.

Power-line contamination

Spectrum interpolation

Replace amplitudes around target line-noise frequencies using neighboring spectral bins. Main assumption/input: a known line frequency and enough neighboring frequency bins. See the spectrum interpolation guide.

ZapLine

Use line-locked DSS components to suppress power-line noise and harmonics. Main assumption/input: a line frequency and data with suitable spectral resolution. See the ZapLine guide.

Reproducible or structured components

DSS

Extract, retain, or subtract components defined by a user-supplied bias. Main assumption/input: a baseline covariance and a reproducibility, spectral, temporal, or lagged-trial bias. See the DSS guide, which also covers TimeShiftDSS and the other DSS variants.

SSA

Decompose each channel in a delay-coordinate space and reconstruct selected components. Main assumption/input: an embedding window and a frequency or local-subspace selection rule. See the SSA guide.

CCA and reference-informed cleaning

BSS-CCA

Use canonical correlation with a lagged copy of the signal to identify components for cleaning. Main assumption/input: a lag choice and a rule for selecting the canonical-correlation end to remove. See the BSS-CCA guide.

iCanClean

Remove shared variance between primary channels and a physical or derived reference. Main assumption/input: matched observations and a suitable reference block. See the iCanClean guide.

TMS-evoked muscle artifact

SSP-SIR

Suppress a fitted artifact subspace and reconstruct the signal with source-informed geometry. Main assumption/input: a TMS-evoked artifact window and a forward model or compatible EEG montage. See the SSP-SIR guide.

Compact comparison

Method	Main target	Main structure used	Additional requirement
ASR	Transient high-variance artifacts	Clean calibration covariance	Calibration period/data
SNS	Sensor-specific noise	Spatial sensor redundancy	None beyond channel data
SOUND	Sensor-specific noise	Forward-model geometry	Montage or forward model
Spectrum interpolation	Power-line contamination	Neighboring spectral amplitudes	Line frequency and sampling rate
ZapLine	Power-line noise and harmonics	Line-locked DSS components	Line frequency and DSS settings
DSS	Reproducible or structured components	Bias covariance relative to baseline	Bias or segment definition
SSA	Structured channel-wise components	Delay-coordinate decomposition	Embedding and selection settings
BSS-CCA	Components separated by lagged CCA	Lagged temporal correlation	Lag and component rule
iCanClean	Shared reference variance	CCA with reference channels	Reference block or pseudo-reference
SSP-SIR	TMS-evoked muscle artifact	Artifact subspace and lead field	Forward model or EEG montage

Evaluate the result

Whichever method you choose, evaluate both artifact attenuation and preservation of the signal of interest. See the Evaluating denoising guide.