A man paralyzed from the neck down imagined writing with a pen. The implants read his writing thoughts and displayed them on screen. (Image Credit: Stanford University)


Stanford University researchers have developed a brain-computer interface capable of interpreting a human’s writing thoughts into text. A paralyzed man with this implant typed out approximately 90 characters per minute with 94.1% accuracy by imagining they were writing on paper. The researchers put two 4x4 mm implants in the paralyzed individual’s premotor cortex, where intentions to perform movements occur. Capturing these intentions could produce a clear signal compared to capturing movements, which is complex (multiple muscles cause movement) and dependant on context (hand is relative to writing on the page). Afterward, the researchers told the participant to imagine writing on a page and recorded the neural activity while he performed the task. 


Overall, there were around 200 electrodes in the participant’s premotor cortex. The researchers conducted a principal component analysis of the neural recordings, which showed more differences as the participant thought of different letters. The recordings were then converted into a two-dimensional plot, revealing that single characters clustered together. Meanwhile, characters like p and b or h, n, and r formed clusters close to each other. The participant was also instructed to use punctuation like a comma and question mark. He used a > to signify a space and a tilde as a period.


The researchers managed to decipher the character with 94.1% accuracy, but a slow analysis was performed on the system after recording the neural data. Real-time performance was achieved by training a neural network to predict the probability of a signal matching a letter.


Generally, the system performed very well even though it worked with 242 sentences’ worth of characters. It only took 0.5 seconds for a character to appear on screen after the thought process. The participant managed to generate around 90 characters per minute, higher than the previous typing-implant record of 25 characters per minute. It only produced a 5% error rate, and installing autocorrect could reduce the error rate to 1%.


All the tests used prepared sentences. However, after the system’s validation, the participant was instructed to type free-form responses to questions. The speed decreased to 75 characters a minute while errors increased to 2%. The team says this isn’t a “complete, clinically viable system.” The alphabet used in this system doesn’t have digits, capital letters, and most punctuation. Additionally, the implants’ behavior adjusts over time due to scar tissue build-up or small transitions relative to the readable neurons. This meant the system had to be recalibrated once a week to ensure an adequate error rate. 



Have a story tip? Message me at: http://twitter.com/Cabe_Atwell