Systems and methods of training and using a neural decoder

A method includes implanting a neural probe configured to record and transmit bioelectrical signals into the brain of a subject. One or more external stimuli having an expected response are provided to the subject. In response to the one or more external stimuli, a recorded bioelectrical signal from the neural probe is received at a computing device in communication with the neural probe. The bioelectrical signal is mapped, at the computing device, to the expected response. The steps of providing, receiving, and mapping are repeated a plurality of times to generate a training dataset of mapped bioelectrical signals. The method further includes, at the computing device, training a neural decoder algorithm based on the training dataset of mapped bioelectrical signals and without external subject feedback.

The abstract as published by the patent office, from WO2026096357A1. It describes the invention; it is not a statement of what is protected.

• Interface and electrodes

  Electrodes and arrays

• Signal processing

  Signal processing

• Decoding

  Neural decoding

What the filing appears to be about, read by published rules from its title, abstract and classification codes (A61B5/00, A61B5/291, A61B5/293, A61B5/31, A61B5/316). The claims define what a patent protects, and they are not read here.

Interface and electrodes

Everything downstream works with what the contacts pick up. Their number, size and distance from the neurons set the ceiling on the signal. The problem behind it: more contacts, closer to neurons, that still record after years in tissue that reacts to them.

Signal processing

A decoder is only as good as the features it is given, and recordings drift and pick up noise from movement, muscle and stimulation. The problem behind it: stable features from signals that change from day to day, computed fast enough to act on.

Decoding

This is the step that turns brain activity into a command or a word. Its accuracy and speed are what a user experiences. The problem behind it: accuracy that holds over months without frequent recalibration, for more than one user and one task.

Why the layer matters to the field, not a judgement of this filing. A patent family shows that protection was sought. It does not show that the technology was built, is used in a product, or works in people. The whole stack

One invention, published 1 time in 1 office. Each publication opens in Espacenet.

• International (PCT)a route to several countries, not a countryWO2026096357A17 May 2026

1. 28 Oct 2024First filingEarliest priority, at United States. The family is counted in this year.
2. 27 Oct 2025ApplicationFiled at International (PCT).
3. 7 May 2026First publicationWO2026096357A1, International (PCT).

Dates as the offices published them. Records of lapses, withdrawals and changes of owner have not been read for this family, and whether a patent is in force today is not assessed.

Organisation

Axoft · 3 families identified

• Bibliographic data, family members and office records: European Patent Office, Open Patent Services (the DOCDB and INPADOC databases), read 11 Oct 2026.
• The record itself, with its claims and description: WO2026096357A1 in Espacenet
• Counted as a BCI patent family: uses the field's wording; classified A61B5/293 (invasive electrodes for brain recording); filed by AXOFT, INC., an organisation already in the graph; describes neural hardware or methods without naming the field. How families are found and judged