SCIENCE NOTES 2002 ¦ University of California, Santa Cruz Science Communication Program

KRISHNA SHENOY HAS MADE A CAREER out of eavesdropping. But he listens to thoughts, not voices.

In a world where brain meets silicon, computers do the work of injured neurons — even reading thoughts.

Shenoy, who heads a new lab at Stanford University, is one of several researchers across the country who are linking brains to computers as part of the growing field of neural prosthetics. He is devising ways to tap into a monkey's brain and read where the animal plans to reach its arm. He can route these signals to a computer icon that moves for the monkey.

By bypassing the need for the brain and arm to "talk" through the usual neural connections, this technology could eventually help people with spinal cord injuries to type, pick up a fork, or turn a page just by thinking about it. Electrodes set in the brain will talk to robots or stimulate distant muscles.

Shenoy is well aware that overblown research claims have raised the hopes of paraplegics in the past, only to fizzle. But he believes his monkey experiments are leading to practical results. "We have a view toward human patient tests. We've initiated those conversations with neurosurgeons here at Stanford," he says. "We have to have a bigger picture, an ambitious goal, or we're frittering away our time."

Researchers in neural prosthetics build devices that make up for lost neural activity. In the healthy body, the brain communicates with the limbs via the spinal cord. Messages zip along as electric pulses through end-to-end nerve cells, moving from the brain to the spinal cord and from the spinal cord to the limbs. Any break in this line of communication stops the message cold, usually permanently. Using sophisticated new electronic devices, researchers hope to bypass such breaks.

One approach to treating spinal cord injury, for example, is to build a neural prosthetic that mimics the work of the spinal cord. Three steps are involved in building such a device: plucking neural signals from the brain, making sense of them, and carrying out the intention encoded in the signals.

"The biggest bottleneck has been getting neural signals out of brain correctly," says Daniella Meeker, a graduate student who collaborated with Shenoy when he was a post-doc at the California Institute of Technology, before he moved to Stanford. Each electrode listens to a single nerve, and there's no wiggle room. If the electrode moves even 50 microns (the size of a pinhead) away from the neuron it’s recording, it will lose communication.

Unfortunately for a scientist trying to place an electrode, the brain is a bit wiggly. The pliable brain moves slightly relative to the skull, threatening to move the target neuron out of earshot of the electrode, which is fixed in the bone of the skull. This loose connection between brain and electrode may be the limiting factor for using neural prosthetics in humans, Shenoy says.

Moreover, electrodes get gummed up with sticky fluids after a while, insulating them from local signals. The electrodes that Shenoy and colleagues plant in a monkey's brain have limited lifetimes. Improving the robustness and longevity of the electrodes also will be critical to transferring this technology into humans.

Nevertheless, these challenges haven't prevented researchers from achieving some startling successes in laboratory monkeys. Shenoy's research team at Caltech, led by Richard Andersen, trained a rhesus monkey to touch the right or left side of a computer screen in response to an on-screen flash of light. All the while, the scientists snooped into the monkey's brain, recording neural pulses. Using this code, a computer read "right" or "left" from the monkey's brain activity and flashed an arm icon on the corresponding side of the screen. The crafty monkey soon realized it didn't have to lift a finger to get its reward, a sip of juice; it just had to think about moving. The thought alone was enough to get the virtual arm to do the work and earn the reward.

"They preferred using the icon to play these video games we provided them instead of using their real arm," says Meeker. She and others were surprised that it was so natural for the monkeys to quit moving their arms.

The researchers bring the monkeys to a dark, isolated room where there is no background interference. It's so quiet in these chambers that you can almost hear yourself think. And that's exactly what Shenoy is trying to do–hear the monkey's thoughts.

What exactly does a thought sound like? "If you're listening to it, it is sort of like a buzzing, and the buzzing increases or decreases its frequency," describes neuroscience expert Andrew Schwartz, who does related work at the Neurosciences Institute, a private foundation in San Diego. The raw language of neural pulses is better suited to a computer's ear than a human’s.

But even for a computer, reading these buzzing thoughts is tricky. "We don't know the language of the brain. We're tourists with only a visitor's guide book," Shenoy says. "The brain is magic. How do wet squishy neural cells compute? It's just fascinating."

Information is contained in the speed and intervals at which neuron cells fire their electric pulses. By monitoring that process for a while, researchers can correlate the nerve-firing rate of a nerve cell with a monkey's actual movement.

"We listen in during the normal behavior, and we make our little map. For example, 100 spikes per second means right, 10 spikes per second means left," says Shenoy. Thereafter, they can predict movements from the rate of cell firings in the recorded neuron.

Once the monkey's intention has been read, that intention must be acted out. In Shenoy's experiment, the researchers simply flashed an arm icon to the correct side of the screen. Eventually, researchers aim to move a real arm through muscle stimulation or to move a robotic arm.

If you ask Shenoy when this technology will be available in humans, one answer he gives is "two years ago." Though there are no systems that move arms, researchers Philip Kennedy of Georgia Tech and Roy Bakay of Emory University have implanted electrodes in humans with amyotrophic lateral sclerosis ("Lou Gehrig's disease") or strokes in their brain stems. These patients can't move a muscle but are cognitively alert. The implants allow them to move an icon over a virtual keyboard and slowly tap out messages, simply by thinking. This is the first example of a human brain communicating directly with a computer.

Kennedy and Bakay implanted two glass cones, each about the size of a tip of a ballpoint pen, into the brain of Johnny Ray, a 53-year old brain-stem stroke victim who is completely paralyzed. His brain functions perfectly but the signals don't get anywhere. With special chemicals, Kennedy and Bakay induced neurons in the motor cortex—which controls movement—to grow into the glass cones, ensuring that the electrodes would stay in place. Ray was told to think about moving his finger. A circuit routed this signal to an icon on the screen instead of into his arm. After practicing, Ray eventually learned to will the cursor to move right or left and up or down. The brain signals act as a computer mouse. They move the cursor across the screen and select pre-scripted phrases, such as "See you later. Nice talking with you," or "I'm thirsty."

Beyond these initial human tests, the field of neural prosthetics is embroiled in many controversies. One major quandary is where to place electrodes within the still-mysterious brain.

Shenoy's group placed electrodes deep in the brain, in an area called the "parietal reach region." This area of the brain first specifies where to you want to go, and precedes any formal plan for how to get there. It's the place where thoughts are born.

"This is the highest level, the most abstract plan of how you want to move your arm," Shenoy says.

Most other researchers place electrodes in the motor cortex of the brain, which is the last place thoughts visit before they exit the brain for the spinal cord. But tapping into the planning region of the brain has advantages, Shenoy believes. Whereas motor neurons coordinate movement along a pathway, planning neurons simply tell where and when the arm should go next— an easier set of instructions to read and transfer. If the neuron just specifies a target, then scientists should be able to engineer a robotic solution of how to get there, without having to read tons of neurons. Recording electrical impulses from a few neurons is technologically simpler and surgically less invasive, and thus may be more feasible to do in humans in the near future, Shenoy says.

Planning neurons may also be less susceptible to the changes that may take place in motor neurons after paralysis, when the muscles they control become inactive. "We're going to a deeper, more isolated, more central part of the brain, farther from the sites of potential injury," Shenoy says. "It may well be that, since the motor cortex is closer to the periphery, if you have a spinal cord injury the motor cortex reorganizes and the parietal reach region remains intact."

But not everyone agrees with this theory. "I think most of the data are against them," Schwartz says. "My point of view is even if it [the motor cortex] does reorganize you can train the individual to reorganize it again to the way you want it to work. In my mind it's not such an issue."

Shenoy's experiment involved only one neuron, but he says this was just a proof of concept. He plans to expand to reading from several neurons, using electrode arrays. "It could be that if we then go listen to a second neuron or a third or a fourth or even 100 neurons all at the same time, then we can do a very good job of predicting where the monkey wants to reach— not just left versus right, but up versus down, and near versus far," Shenoy says.

Indeed, there are distinct advantages to reading more than one neuron. John Donoghue, a top neuroscientist at Brown University, says that it is crucial to read from populations of neurons. "How we're coming to understand the brain is like trying to understand one instrument at a time in a symphony," Donoghue says. "Certain things arise from interactions, such as harmony, that can't be heard one at a time."

Donoghue and his collaborators at Brown look at groups of 6 to 25 cells in the motor cortex using multi-electrode arrays. They read out specific motor plans, three-dimensional pathways with direction and speed, not just binary movements. "Our lab is interested in turning thoughts into behaviors," he says.

In Donoghue's experiment, a monkey plays a video game, rather like ping-pong, where it has to capture an on-screen target by moving a mouse with its hands. It doesn't take the monkey long to master the game. After the monkey has played for a few minutes, the scientists disconnect the mouse from the computer and switch from mouse control to brain control, unbeknownst to the monkey. Instantaneously, the monkey controls the video game from its brain.

"What's coming out of the brain is some kind of code that mathematical filters can decipher in minutes," Donoghue says.

Donoghue was surprised the monkeys could do it so well. Eventually, one monkey even realized he didn't need to move the mouse, and he quit moving his hand altogether.

Based on these findings, Donoghue says he could reconstruct how a person was scribbling on a paper just from recording his brain activities. "Once you have that signal, you can control any kind of device that you can imagine," he says.

The system performs better when the team reads more nerve cells, he says. However, it's a trade-off. Breaking into the brain is one of the biggest obstacles to this type of technology. The more electrodes in the brain, the greater the chance of infection— a particular danger once the procedure is moved outside the controlled environment of a lab.

Says Donoghue, "If you had simply paralyzed one leg, would you do this [in order to walk normally again]? I’d say we’re not sufficiently comfortable with this technology to recommend it in this case."

"The holy grail in these communities would be to have a totally non-invasive way of reading out the brain and what you want to do," Shenoy says. "We're not there, but we're at least getting much closer to the invasive way of doing what we've been discussing."

There are procedures that involve cutting into a part of the body other than the brain, and these might be better for people who are only partially paralyzed. For example, scientists have sent signals from a working shoulder to a non-working hand through external electrodes, letting the shoulder take on some duties of the injured spinal cord.

He cites functional Magnetic Resonance Imaging (fMRI), which remotely images brain activity by measuring blood flow changes. However, like normal MRI, the machine takes a huge room. Even if you could miniaturize the technology to a pinhead, the resolution is not good— you're not able to say, "That neuron just fired one spike," Shenoy says.

The history of practical successes in the field of neural prosthetics is rather short. The two biggest success stories involve reading signals into the brain instead of reading them out.

The cochlear implant, a commercially available device, restores hearing to some deaf people, was the first real interface between the brain and an external, man-made device. The implant takes over for damaged cochlea organs, which normally turn sound waves into electric pulses that stimulate nerve cells in the brain. A receiver under the ear receives digitized sound from a microphone and converts these signals to electric pulses. The pulses trigger microelectrodes in the cochlea, which then spark the brain neurons

Another electronic device, made by MedTronics, Inc., prevents tremors in Parkinson's patients by writing signals into the brain and disrupting neural circuits.

In fact, for neural prosthetics to be truly useful, they must be able to both read signals out of the brain and read them in. "The typewriter is helpful for a paraplegic, but from a longer-range scientific view, we want to be able to do much more than this," Shenoy says.

For example, just picking up a glass is a complicated coordinated process between the brain and the fingers. Grip too hard and you might break the glass. Grip too lightly and you’ll drop it. The prosthetic either has to be intelligent enough to gauge how to react, or it has to be able to talk back and forth with the brain.

Ultimately, the neural prosthetic should be able to learn to work with the brain. "The brain is going to change. Therefore, our algorithms and our electronics have to keep up with, if not encourage, the brain’s behavior. That way, the whole system improves itself, just like a child learning to catch a ball," Shenoy says.

"Eventually, you want to have the computer system intelligent enough that it fine tunes itself, sort of like modern cars giving themselves tune-ups. "

Listening to the brain is going to satisfy Shenoy for only so long. Eventually, he wants to have a conversation.