Evaluating denoising

Artifact attenuation and preservation of the signal of interest are separate questions. A lower peak, PSD, variance, or residual amplitude is not by itself evidence that the desired neural signal was preserved.

Use controls suited to the experiment, such as held-out data, simulated mixtures, unaffected events or channels, surrogate data, or a forward model when appropriate. Compare both artifact attenuation and changes to the signal of interest.

The public mne_denoise.qa module provides spectral and data-change metrics. mne_denoise.quantify_overcorrection() compares a fitted sensor operator with a lead field using amplitude, correlation, relative-error, and goodness-of-fit definitions owned by this package. The individual APIs document formulas, units, and sentinel values.

Useful evaluation categories include:

• target-artifact attenuation;

• broadband or signal-of-interest change; and

• forward-model distortion when a source model is available.

Choose and report the frequency bands, windows, references, and preprocessing used for each comparison.

See the Evaluation and QA API for the complete metric reference.