They say they want to make improve it so patients can choose letters and spells stuff. But with with just a binary choice that can be done with 5 bits. Of course they will need more than that in order to increase accuracy.
A few years ago I was at a demonstration for an interface like this. It worked by measuring the effect of having an expectation fulfilled, i.e. the patient is shown a number of shapes in sequence, each representing a different choice, and when the one the patient wants appears, this can be inferred from an EEG.
IIRC, they used five different shapes and repeated each choice at least three times for error correction, and two or three choices selected one keyboard key. The poor grad student demonstrating the procedure made quite a few typos, after which he had to select backspace, and the process was excruciatingly slow anyway, so it took a few minutes until he managed to write "Hello, welcome to ...". But definitely better than nothing if you are locked in.
