An AI-assisted neural bypass helped a paralysed man feed himself and feel touch

Keith Thomas could bring a cup to his mouth, feed himself and feel contact with objects using an experimental neural bypass. The July 2026 Nature Medicine report followed one man with complete tetraplegia after a spinal-cord injury. Its central achievement was combining useful hand movement with information travelling back to the brain about touch.[3]

Thomas injured his neck in a diving accident in 2020. Before implantation, the Feinstein Institutes team mapped the brain regions associated with his attempted movements and imagined sensations. During parts of the March 2023 operation, he was awake so that his descriptions of sensations could help guide electrode placement.[4]

The hand needs a command and a return signal

An earlier study by Chad Bouton and colleagues, published in 2016, had connected decoded brain activity to electrical stimulation of a participant’s forearm muscles. Machine learning enabled several wrist and hand movements and the manipulation of objects. This established a route from movement intention to the person’s own muscles.[5]

The newer system adds a return route. Two implanted arrays record movement-related brain activity; three arrays sit in the sensory region. Pressure sensors in the hand support trigger stimulation of the sensory cortex, allowing contact at the hand to become a perceived sensation.[2]

Muscle stimulation alone did not produce enough grip. The researchers added a powered orthosis: a support that pulls artificial tendons to bend the index and middle fingers. A long short-term memory neural network interprets attempted hand opening, closing and rest, letting Thomas command the support.[1]

The movement decoder remained useful for more than five months without repeated retraining. Keeping a familiar command reliable over time matters because a daily activity should not require the person to relearn the interface at every session.[3]

AI also steadies the grip

Directly linking the decoded command to the orthosis produced inconsistent squeezing. A reinforcement-learning controller, trained in simulated hand–object interactions, regulates force after Thomas initiates grasping. This separates his decision to take hold from the rapid adjustments needed to maintain the grip.[1]

In the hollow-eggshell task, average session success was 87% with the controller and 27% without it, across five sessions per condition. Those sessions contained 97 controller-on and 85 controller-off trials. Thomas could not see his hand or the shell, although a force display remained visible. The test used one object type and a predefined force range.[1]

Some gains continued after stimulation stopped

The rehabilitation programme also pursued lasting change. Spinal stimulation and activity training improved elbow strength before the AI hand interface was added. These gains should therefore be credited to that rehabilitation stage. Combining stronger arm movement with the assisted grip made self-feeding possible.[3]

For sensation, the team developed cortical mirroring. It recorded patterns associated with imagined touch, then delivered related stimulation to the sensory cortex while also stimulating the spinal cord and skin. After approximately 25 weeks targeting the right wrist, Thomas could detect touch in an area that had been insensate.[2]

The paper documented wrist-sensation gains persisting for more than two months after stimulation ended. The equipment still required trained personnel, so these demonstrations do not describe an independently operated home product.[3]

Thomas chooses to pick something up; the support produces a grip; the controller steadies it; sensation supplies information about contact. This sequence turns a decoded brain signal into a purposeful interaction with the physical world.

Report an error or suggest a correction