This article from MIT’s technology review discusses a paper from a team at Stanford University titled “A high-performance speech neuroprosthesis” which was published as a preprint on bioRxiv (21 January 2023). The paper relates to translating the thoughts of a patient into text on a screen. These so-called brain-computer interfaces are particularly useful for sufferers of paralysis providing them the ability to communicate without the need to physically speak.

The team claim to have achieved new state-of-the-art performance at 62 words per minute decoded speech, “which is 3.4 times faster than the prior record” and “begins to approach the speed of natural conversation (160 words per minute)“. The paper further claims a 9.1% word error rate on a 50 word vocabulary, and an astonishingly low 23.8% word error rate on a 150,000 word vocabulary.

This state-of-the-art performance was achieved using two machine learning models, one to translate the brain’s electrical signals into predicted phonemes, the other to output a given word or sentence based on these predictions. The article discusses the integration of natural language models in newer methods of brain-computer interface, and points to the rise of larger and more complex natural language models, as potentially driving further developments in the field.

Given how useful brain-computer interfaces are likely to be for patients, it is only natural to expect the developers to pursue patent protection. Applicants and attorneys familiar with the European Patent Office’s (EPO’s) approach may wonder how friendly the EPO’s current approach is to patent developments in this area, given the EPO’s position on Natural Language Processing (NLP). According to the EPO’s Guidelines for Examination “classifying text documents solely in respect of their textual content is however not regarded to be per se a technical purpose but a linguistic one“. Additionally, the EPO often seeks to distinguish “cognitive” data from technical data, and the EPO’s Guidelines for Examination state that “cognitive data…are those data whose content and meaning are only relevant to human users and do not contribute to producing a technical effect.” Given that the data in question relates to textual content/language and cognitive data, one may be forgiven for wondering if brain-computer interfaces are patentable at the EPO.

However, developments in this area is very likely to be patentable, and there are a number of granted European patents in this field. In the Enlarged Board of Appeal’s decision in G1/19, the ruling pointed to ways in which a technical effect can be achieved, including input or outputs that have direct links with physical reality. In the case of brain-computer interfaces, the measured electrical brain signals are clearly inputs with a direct link to physical reality and these inputs are processed to convert the electrical signals into a real world output.

As always in the field of machine learning, the key is to ensure that your patent application is drafted by an attorney with an expert understanding of the legal and technical issues, so that the pitfalls can be avoided and you give yourself the best chance of success.

More philosophically, the issues highlighted above do perhaps beg the question of whether the EPO’s approach to natural language processing is appropriate. If words themselves can be accurately predicted based on a real world technical input, such as an electrical signal, this arguably implies that words and language are not merely “cognitive” and “non-technical” as the EPO currently hold.