Researchers at three institutions showed that artificial intelligence (AI) can improve how future visual prostheses, such as a bionic eye, send signals to the brain. A deep-learning model, which is a computer program that finds patterns in large amounts of data, designed electrical stimulation patterns for electrodes temporarily placed in the visual cortex of a blind participant. The visual cortex is the brain region that processes sight information. The model gave better control over how brain cells responded and helped predict what the person actually saw.
The work is a proof-of-concept study published in Neuron. It moves toward visual cortical prostheses, devices placed in the visual cortex that send electrical signals directly to the brain. These devices bypass damaged eyes or optic nerves and may one day help people whose blindness comes from stroke, disease, or injury while their visual cortex can still respond. The approach could be especially useful for those who had sight earlier in life and later lost it.
In tests at a hospital in Spain, a 27-year-old man with vision loss from brain injury received a temporary 96-channel electrode array in his visual cortex. When the electrodes delivered current, he saw phosphenes, which are spots or shapes of light that can resemble flashes or stars. The researchers trained the deep-learning model on recordings of the man’s brain activity after different stimulation settings, along with measurements of the brain’s resting state just before each test. The model then selected patterns most likely to produce a desired response.
How the model improved results
When those artificial-intelligence patterns were tested, they matched the targeted brain activity more accurately and needed less electrical current than other methods. The recorded brain activity predicted the participant’s perceptions better than the stimulation settings alone. The model also adjusted for the brain’s changing state, which may help future devices stay reliable over time. The findings show that an implant must learn and adapt to an individual brain rather than rely on fixed settings.