Small models that read the body’s motor signals.

From the signal to a supervised decision, in four auditable steps.

Assistive devices for movement disorders depend on reading noisy biological signals correctly and on knowing when not to act. Nanomax builds that chain one measured step at a time, and each model must beat a transparent classical method on people it has never seen before it is adopted.

A real recording and its model output, with no clean reference.

This four-second excerpt comes from subject S013, held out of training. Large frontal transients and changes in C3 are visible in the recording and model output. Without an independent clean reference, these changes cannot establish selective artifact removal or preservation of neural activity. Switch electrodes and signals, and hover or tap the trace to read the values.

Results depend on the synthetic contamination and baseline.

Across 20,000 two-second windows from the five held-out people and 37 contamination conditions, EEG-Clear’s advantage concentrates on blinks, eye movements, electrode pops and muscle activity: artifacts that overlap brain frequencies. For mild slow drift, the classical high-pass filter remains the right tool, and the data show it.

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

The model’s gain varied by 1.0 dB across people; the filter’s by 3.1 dB.

We found our own model’s flaw, and fixed it.

For a motor programme, clean waveforms are not enough: the mu (8 to 13 Hz) and beta (13 to 30 Hz) rhythms over the motor cortex are what reveal movement intent. A downstream probe on real left-fist and right-fist trials showed that the first model, trained on waveform error alone, blurred those rhythms. It scored well on every waveform metric and still made moderately noisy trials less faithful than leaving them alone.

A second training stage, 22 minutes with a band-power term added to the objective, raised rhythm fidelity from 0.57 to 0.70. Under heavy contamination it restored features that were otherwise lost (0.70 against 0.52) and brought left-versus-right accuracy from 0.53 to 0.59, next to 0.60 on screened reference data. That finding now shapes every Nanomax model: spectral fidelity is trained for and reported explicitly, never assumed.

225 real-movement trials, five held-out people. Clean-data accuracy is 0.60. With 45 trials per person, accuracy differences are indicative only; fidelity carries the statistical weight.

A research method that others can audit.

The gait model still needs data and a laboratory.

No public dataset combines paediatric cerebral-palsy EEG, muscle activity and gait. Nanomax’s next step is to collect synchronised, consented recordings under ethics approval, train the 1,603-parameter gait-event model against its rule-based comparator on held-out participants, and test both on target hardware with the output simulated. That work needs laboratory access, clinical collaborators and appropriate approvals.

Read the numbers with their scope attached.

• All data are public healthy-adult recordings from one laboratory amplifier; the numbers do not cover children, other devices or anyone with cerebral palsy.
• Scored artifacts were synthetic, so the true signal was known; real blinks, as in Figure 1, can only be inspected, not scored.
• Five test people, one random seed. The held-out probe informed model development; an independent final evaluation is still needed.
• No Norysr hardware, participant or stimulation was involved, and no device is available for use.