Brain-computer interfaces are typically systems which measure neural activity and convert it into artificial output. These systems have shown great potential for assisted movement in patients with motor impairments. The interfaces typically work by directing the patient to think about making a movement and allowing the system to repeatedly record the neural activity associated with that movement. Through supervised learning methods, the interface learns to associate specific activity patterns with intended movements. However, because the baseline firing of neurons can shift from day-to-day based on a variety of factors, the interface must be retrained at the start of each session, a process which can take up to two hours.
A collaborative team from Battelle and the Ohio State University Wexner Medical Center has released new findings in Nature Medicine demonstrating a novel method using a deep neural network that drastically reduces the training time required for their brain-computer interface. Their method involves initially training the system using the traditional supervised approach, and then allowing an unsupervised neural network to monitor the performance of the system without explicit labels for the intended motions of the user. David Friedenberg, senior author on the study, explains that, after the initial training setup, “You only need a couple of examples to say, ‘this is the basic pattern I’m looking for,’ and then the rest of the data is helping to counteract those baselines that might be shifting around.” The hybrid decoder was able to sustain its performance of correctly predicting motion signals with 90 percent accuracy in one subject for over a year without explicit retraining.
The group believes that their new system addresses many of the desires of potential end-users for brain-computer interfaces, which include high accuracy, minimal daily setup, and fast response times. The team is currently looking at ways of reducing the size of the system so that it can be taken outside of the lab and tested in a natural environment.
Here’s a Bloomberg video report about the research in which you can see how one patient is able to do things he thought he’d never do again:
Study in Nature Medicine: Meeting brain-computer interface user performance expectations using a deep neural network decoding framework…
Via: Battelle…