A denoiser measured on people it never saw.

Before any two recordings can be compared, the artifacts have to come out the same way for everyone.

The programme's long-term interest is comparing whole-body recordings across people, devices and eventually clinical groups. Any such comparison needs a cleaning step that works on people it was not trained on, does not depend on the electrode layout, and does not erase the rhythms the comparison is about. That is a testable question with public data, so it was tested first, while the gait side of the project waits for a dataset it can legitimately use.

The study was also a chance to apply the project's evidence rules to a real result: person-level splits, recorded manifests and hashes, classical baselines, per-person reporting, a downstream check, and an immutable run bundle.

Seven steps, all of them reproducible from a command list.

SNR improvement on synthetic contamination: +14.42 dB versus +7.14 dB.

On 20,000 clean windows corrupted with the training mixture (input signal-to-noise ratio about −5 dB), the band-pass filter recovered +7.1 dB and EEG-Clear +14.4 dB. The relative error fell from 1.27 (filter) to 0.44, and the correlation with the true signal rose from 0.68 to 0.86. The advantage is concentrated on blinks, eye movements, electrode pops and in-band muscle activity, the artifacts that share their frequencies with the brain signal; it disappears, as it should, for mild slow drift, where a high-pass filter is already the right tool.

Means of per-person means over five held-out people; the interval is a 2,000-sample subject bootstrap.

Clean input passed through with a correlation of 0.991 to the target, against 0.985 for the filter. On 1,500 real frontal windows containing genuine blinks of several hundred microvolts, for which no clean reference exists, the median mean-square difference between band-pass and model outputs was 79 percent of band-pass-output power. This difference-power ratio does not measure selective artifact removal or preserved neural activity. Inference runs at 2,500 to 4,800 two-second windows per second on an eight-thread laptop CPU.

Waveform accuracy did not, by itself, preserve the movement rhythms.

A downstream probe took cue-locked epochs from the real left-fist and right-fist runs of each held-out person, computed mu (8 to 13 Hz) and beta (13 to 30 Hz) band power over seventeen sensorimotor electrodes, and asked two questions: how well do those features agree with the clean epochs, and can a classifier trained on clean epochs still tell left from right?

The first-stage model, trained on squared error alone, scored superbly on every waveform metric and yet made the band-power features of moderately contaminated epochs less faithful than leaving them alone (correlation 0.57 against 0.83 for the untouched epochs). Squared error rewards hedging: when the network cannot tell rhythm from artifact it shrinks both, which costs little signal-to-noise ratio but shifts every band-power feature by about the amount that separates one condition from another.

A second stage that added a band-power term to the loss raised that fidelity to 0.70, restored features that severe contamination had destroyed (0.70 against 0.52 for doing nothing), brought decoding accuracy from 0.53 back to 0.59 against a clean 0.60, and improved every other metric as well. The remaining gap at moderate contamination means that windows a screen already accepts as clean are best left alone: the rule that follows is to clean only what fails a screen, and to measure spectral fidelity explicitly whenever a cleaning step precedes a comparison across people.

Read the numbers with their boundary attached.

• The artifacts were synthetic; real ones were only inspected visually, and an artifact of an unmodelled shape may be handled worse.
• Every recording came from one laboratory amplifier and healthy adults. Transfer to other devices, children or anyone with cerebral palsy has not been validated.
• Five test people give wide intervals; one random seed was used; the clean-window rule was set after inspecting three training people.
• Classification accuracy in the probe was near chance even on clean epochs, so only its fidelity measure carried statistical weight.
• Seeing one channel at a time forgoes the spatial information that multi-channel methods use; cross-device performance remains untested.
• It is not a Norysr device result, it did not use Norysr hardware or participants, and it changes nothing about the gait programme's stage gates.

Run records and data attribution.

Data: Gerwin Schalk (2009), EEG Motor Movement/Imagery Dataset v1.0.0, PhysioNet, DOI 10.13026/C28G6P. Used under the Open Data Commons Attribution License v1.0. Displayed traces are excerpts; model and filter outputs are derived from the recordings. Original publication: Schalk, McFarland, Hinterberger, Birbaumer and Wolpaw (2004), BCI2000: A General-Purpose Brain-Computer Interface (BCI) System, IEEE TBME 51(6), 1034–1043. Platform citation: Pollard, Moody, Lehman, Gow, Fernandes, Xie, Johnson, Mark and Heldt (2026), PhysioNet as a global platform for biomedical research, Nature Health 1, 792–795, DOI 10.1038/s44360-026-00096-z.