For those with advanced ALS (amyotrophic lateral sclerosis) or certain brainstem strokes, they can be what is referred to as “locked in”. They are mostly paralyzed. A tip of the basilar artery brainstem stroke, for example, can leave one only able to move their eyes. Many with ALS, long before they get to this point, lose the ability to speak because of bulbar weakness – weakness in the tongue and muscles in their throat.
The resulting inability to communicate can be devastating. Even simple needs, such as asking someone to scratch their nose or reposition them in bed, can be laborious. There are various methods to improve communication. For those with any remaining hand control, there are communication devices. For those with only eye control, there are eye-tracking systems that can allow them to spell out their messages. Stephen Hawking famously used such a device, with the computer voice being recognized as “his” voice. These techniques are extremely useful, but also can be slow and laborious.
The ultimate goal for treating the inability to communicate because of some form of paralysis, while the mind is still relatively spared, is the brain-machine interface (BMI). A sufficiently robust and trained interface could theoretically allow for patients to simply think what they want to say while the computer “reads” their thoughts and converts them to text or to speech.
This technology goes back to 1970 with the first animal research. In 1998, the first invasive BMI was performed on a human with implanted electrodes on the brain’s surface. This allowed for the control of a cursor on a computer screen. The technology has incrementally advanced since then, including three main components: the electrodes that provide for the interface, the computer hardware that runs the algorithms, and the computer software that interprets the signals.
A recent demonstration shows where the technology currently is. Researchers at UC Davis Health have implanted a BMI on a patient with ALS who has lost the ability to speak. With that interface, they were able to train the patient and the software to produce real-time computer speech based on the patient’s intentions with 97% accuracy. This is an amazing accomplishment and could be life-changing for those with similar disabilities.
This advance is also, in my opinion, more than incremental. I have been following this technology closely for the last couple of decades, and this achievement is about 15-20 years ahead of where I thought we would be based on my assessment from about 5 years ago. What has changed? Simply put – AI.
The latest crop of artificial intelligence applications or machine learning is very good at finding subtle patterns, such as the patterns in electrical signals from a patient’s brain. By adding such machine learning into the mix, BMI technology is able to be trained much faster and produce results much more quickly. The result, in this case, is real-time BMI-based communication. This would simply not be possible today without AI.
It is important to note that this is a single case, and there are some limitations, so let’s get into some details. First, the electrodes were placed in the brain tissue itself, meaning they required an invasive surgery. They used “256 intracortical electrodes” in four arrays on the motor cortex. So they were not recording his language cortex, but rather the muscles he intended to move in order to speak. So when he tries to speak (but cannot due to muscle paralysis) the software interprets the signals, determines the words he is trying to say, and then speaks them. Because there were recordings of his voice from prior to the onset of ALS the computer was able to speak in his voice. Here are the results in detail:
On the first day of use (25 days after surgery), the neuroprosthesis achieved 99.6% accuracy with a 50-word vocabulary. Calibration of the neuroprosthesis required 30 minutes of cortical recordings while the participant attempted to speak, followed by subsequent processing. On the second day, after 1.4 additional hours of system training, the neuroprosthesis achieved 90.2% accuracy using a 125,000-word vocabulary. With further training data, the neuroprosthesis sustained 97.5% accuracy over a period of 8.4 months after surgical implantation, and the participant used it to communicate in self-paced conversations at a rate of approximately 32 words per minute for more than 248 cumulative hours.
This is an incredibly rapid training period for a BMI. Also, 32 words per minute with 97.5% accuracy is essentially normal conversation (a bit on the slow side, but perfectly reasonable). And of course there are likely to be incremental advances from here.
The main limitation is the intracortical electrodes. These have a lifespan of 1-4 years. At some point within this time period their function will degrade until they are no longer functional. The two main limitations are slight movements of the electrodes in their relationship to the cortical tissue, and the formation of scar tissue that electrically isolates the electrodes.
For BMI technology there is currently a trade-off between fidelity on the one hand and longevity and invasiveness on the other. In this study they opted for maximal fidelity. In other words, the more intimately the electrodes interface with the brain, the better the signals you get, but the more invasive the procedure and the shorter the lifespan. Intracortical (embedded in brain tissue) electrodes give the best signals but are the most invasive, followed by brain surface, then inside the skull but not touching brain tissue, followed by scalp electrodes. You could also have a semi-invasive procedure of drilling holes in the skull beneath the scalp electrodes, or even place the electrodes through the skull but not touching the brain.
One intriguing technology is the stentrodes – which are placed inside veins within the skull, similar to stents used in cardiology. But these stents are electrodes, which can get close to brain tissue within the skull, but are not touching the brain.
Researcher are also working on nanowire electrodes, which go inside brain tissue, but the hope is they will be so small they will not cause scar tissue to form. They are also developing soft electrodes that can move with the brain, which can solve both technical problems – keeping a consistent relationship to brain tissue and not provoking inflammation and scar tissue.
Electrode technology is now the main limiting factor in BMI applications. As this latest study demonstrates, the computer hardware and AI-enabled software are highly functional. With high-quality enough signals they can quickly learn and then rapidly interpret those signals to essentially read the minds of patients in near real time.
Two potential paths forward include improving the software further so that it can make do with progressively lower quality (but less invasive) signals. If we can develop a system that functions nearly as well with scalp electrodes (not sure if this is even possible, but if it is) then this will make the technology vastly more accessible. Or, we need to make soft electrodes that are essentially permanent. Or we need some combination of the two – improve the AI so it can get by with progressively less invasive electrodes. Personally, I think the stentrodes may be the best compromise, but we’ll see.
In short, brain-machine interface technology is an exciting new area of medicine that has recently benefited from a huge functionality boost due to recent developments in AI. It is an example of what we can do when we invest in basic science and biomedical research.
