Brain-computer interfaces (BCIs) seek to bridge neuroscience and engineered systems, allowing neural engineers to record electrical activity in the brain, analyze it to infer what the individual is trying to do, and use it to control devices like prosthetics. Aside from furthering our understanding of how the brain works, extracting information about intended physical motion could be used to restore movement in individuals with disabilities. Still, it’s challenging to develop a system capable of recording small electrical signals from large numbers of the billions of neurons in our brains over long periods of time without breaking down. Large numbers of tiny electrodes (electrical sensors) can be made from silicon- the same material used in computer chips. However, these devices tend to shift over time, in some cases in as little as a day, changing which neurons are being recorded by which electrodes, causing “instabilities” in the recorded data.
BCIs have to be calibrated to each user initially to accurately capture what their brain is trying to accomplish. Recalibrating the device once instabilities appear can be difficult, time-consuming, and often frustrating. Some engineers have focused on developing electrodes from different, more flexible materials, while a group of scientists from Carnegie Mellon University and the University of Pittsburgh recently published their new computational algorithm, which accounts for the presence of instabilities. The new algorithm outperforms the current standard, a method called supervised self-recalibration, and can function continuously over days without needing recalibration.
The authors hypothesized that, while the identity and location of individual neurons may change over time, there is likely some underlying pattern of activity that could be used to “realign” the system without needing recalibration. They can use these base patterns like true north on a compass to make adjustments over time. They developed a computer program to identify such patterns of activity and tested it in two monkeys using a BCI to control a cursor on a screen. The algorithm was then trained to predict the cursor motion intended by the monkeys’ thoughts. The authors then simulated large instabilities expected to be found in clinical BCIs that disrupted the recorded data. Their algorithm was able to automatically correct performance in these cases back to near-perfect accuracy within a couple of minutes and could work for at least five days without needing recalibration. When comparing their algorithm to an existing supervised recalibration test, they found that, in addition to being faster, it was more successful at reducing the error of the cursor movement, especially during times where the animal may not have been focused on the task. This previous standard self-recalibration method forced the algorithm to “re-learn” the animal’s intent once an instability caused inaccuracy in cursor movement, essentially starting from scratch. In clinical systems, this would require the user to stop whatever they were doing and spend approximately ten minutes re-calibrating the machine, often with the help of a technician. This makes such an algorithm unrealistic for any long-term use outside of a hospital, while independent at-home use is the ultimate goal of BCIs in restoring function. With the new algorithm, no long recalibrations were required. The user didn’t even have to be overly focused on the task for the device to function properly again after an instability was introduced, while in supervised recalibration, the algorithm needs to know the user’s intended movement to recalibrate correctly.
This work is preliminary – only two monkeys were tested with the algorithm and it has yet to be used on humans. Additionally, the task used for training was relatively simple compared to the complex motions that an individual would need in a fully restored prosthetic device. Even so, the mathematical foundation of this algorithm suggests that it should be easily applied to more complicated problems, and the excellent performance over extended periods of time is promising. While we may not be seeing fully bionic arms and legs any time soon, neural engineering is now one step closer to making “science fiction” less fictional. It also provides some insight into how underlying patterns in the brain may contain information that we don’t yet fully understand, and with the introduction of these smart algorithms, we may not even need to for effective BCIs.
Managing Correspondent: Andrew T. Sullivan
Press Articles: “Stabilizing brain-computer interfaces,” ScienceDaily
“Mind over body: The search for stronger brain-computer interfaces,” Neurosciencenews.com
Original Journal Article: “Stabilization of a brain-computer interface via the alignment of low-dimensional spaces of neural activity,” Nature Biomedical Engineering
Image Credit: Pixabay