High-performance handwriting BCI

The high-performance handwriting BCI decodes attempted handwriting — the neural activity a paralyzed user produces when trying to write letters by hand — into text, rather than decoding point-and-click cursor movements onto an on-screen keyboard. A recurrent neural network converts the motor-cortex activity accompanying each attempted pen stroke into character probabilities, and a language model cleans up the output. In its landmark demonstration this reached typing rates around ninety characters per minute with raw character accuracy in the mid-nineties percent, rising above ninety-nine percent with autocorrect — roughly the speed of able-bodied smartphone typing and far faster than prior cursor-based spellers for this population.

The conceptual insight is that temporally rich, variable movements are easier to decode than simple straight reaches. Different letters trace distinct, complex spatiotemporal trajectories, so their neural correlates are more separable than the near-identical activity for reaching to nearby keyboard keys. This reframed the design goal from making movements simple and stereotyped to choosing movements whose neural signatures are maximally distinguishable — a principle that generalizes beyond handwriting.