A deep-learning algorithm that can analyze the brain activity of a patient with tetraplegia—the paralysis of the arms and legs—is reported in a paper published online this week in Nature Medicine. This algorithm has been employed to deliver electrical stimulation to the patient’s forearm muscles, thereby restoring functional movements to the previously paralyzed limb.

The quality of life of individuals with chronic paralysis can be improved using brain–computer interfaces. These can provide a gateway between the central nervous system circuits that govern movement and assistive devices, such as computer cursors or robotic devices. Recently, brain–computer interfaces have been used to bypass spinal cord damage, restoring function in a paralyzed limb through direct muscle stimulation. While promising, obstacles to taking this approach into real-world use remain. These include the need for accurate and quick responses, capacity to serve multiple functionalities and efficient day-to-day recalibration as required.

Over a two-year period, Michael Schwemmer and colleagues collected cortical brain activity recordings of a patient with tetraplegia performing ‘imagined’ arm and hand movements. The brain activity was collected through an invasive, chronically-implanted microelectrode array situated in the patient’s motor cortex. These microelectrodes directly sampled neuronal activity with high spatiotemporal resolution. From this large dataset, they used a deep-learning approach to develop a brain–computer interface decoder that provides accurate, rapid and sustained performance and learns new functionalities, all with little need for retraining. The authors show that the decoder can be used to control an electrical stimulation device, reanimating the patient’s paralyzed forearm in real-time.

The authors note that although the example patient can use the decoder to grasp and manipulate objects, confirmation will be required so that this approach can work with other patients and over longer periods of real-world use. They conclude that future work should investigate whether a similar functioning decoder can be generated from training data obtained through real-world use rather than in controlled, laboratory conditions.