Intracortical brain–computer interfaces

Spinal cord injury, brainstem stroke and ALS can disconnect the brain from the body and leave the motor cortex still producing the activity of movements that never arrive. An intracortical brain–computer interface reads that activity at the source: fine electrodes in motor cortex, and a decoder that turns their spike counts into the velocity of a cursor, the reach of a robot arm or, lately, words. The array is micromachining given to neuroscience. The decoder rests on a finding about motor cortex given to engineering: a movement’s direction is carried by a population of broadly tuned cells, not by any one of them.

A hundred needles from a block of silicon

The Utah array was built to stimulate. Richard Normann’s group at the University of Utah wanted a visual prosthesis that would evoke spots of light from visual cortex, and in 1991 Campbell, Jones, Huber, Horch and Normann described how to make the interface: 100 platinum-tipped needles, each 1.5 mm long, standing on a 4.2 mm square of silicon. Diodes isolate the needles from each other: aluminium thermomigrated through an n-type block leaves trails of p+ silicon, so between any two trails sit two opposing pn junctions, one of them reverse-biased whichever way the voltage points, and the trails are then machined into needles.

A version with 96 of its 100 needles wired became the sensor of the BrainGate trial, a microelectrode array left in people for years. In 2006 Hochberg and colleagues reported MN, tetraplegic three years after a spinal cord injury, moving a cursor with it, opening simulated email and working a television. In 2012 S3, unable to move her arms or speak after a brainstem stroke, lifted a bottle of coffee to her mouth with a robot arm and drank, on 4 of 6 attempts, using an array implanted more than five years before. In 2023 four arrays decoded the attempted speech of a woman with ALS at 62 words a minute, at a 24% word error rate on a 125,000-word vocabulary (Willett and colleagues); in 2024 a man with ALS reached 97.5% word accuracy with a similar system and used it in conversation for more than 248 hours (Card and colleagues).

None of this was first of its kind. Philip Kennedy’s neurotrophic electrode gave a locked-in patient on/off control in 1998, and monkeys were steering cursors with motor cortex by 2002. What the Utah array brought to people was channel count.

The population code for movement

The decoder goes back to 1982. Recording in the arm area of monkey motor cortex during reaches in eight directions, Apostolos Georgopoulos and colleagues found three quarters of the active cells tuned to direction, most of them as a cosine: the firing rate $D$ is highest when the movement’s direction $\theta$ matches the cell’s preferred direction $\theta_0$ and falls off as the two diverge, with $b_0$ and $c_1$ fitted for each cell.

$$D = b_0 + c_1 \cos(\theta - \theta_0)$$

That is broad tuning, and one cell says little about direction, but preferred directions differ from cell to cell. In 1986 Georgopoulos, Schwartz and Kettner had each cell vote along its preferred direction, weighted by its change in firing, and the vector sum — the population vector — pointed where the arm went. To an engineer that is population coding as sensor fusion: broadly tuned, noisy sensors whose pooled estimate sharpens with the square root of their number, as long as their errors are independent.

The decoders kept the model. The 2012 robot-arm study fitted each channel’s rate as a baseline plus its preferred-direction vector projected onto the intended direction — the cosine in vector form — and those vectors, stacked, are the observation matrix of the filter that drove the arm. Reading a population also excused two things: the decoder need not know which neuron fired, only how often the signal dipped below about −4.5 times its RMS noise, and 13 to 50 channels were enough.

What came back was evidence about human cortex. MN’s cells were still modulated by intended hand movement three years after the injury, and S3’s drove the robot nearly fifteen years after her stroke. Willett’s team put two arrays in ventral premotor cortex (area 6v) and two in area 44, part of Broca’s area. A 3.2 mm square of 6v carried an intermixed code for jaw, larynx, lips and tongue, and still represented phonemes by how they are articulated; area 44 carried little information about speech production at all, in line with recent work questioning its traditional role. The population code for movement was borrowed from monkeys, and in people it proved still readable years after paralysis.

An estimator with a person in the loop

Stripped of physiology, the decoder is a state estimator. The state is intended velocity; the measurements are spike counts on each channel every 20–100 ms, a rate code read by counting; the model says velocity changes smoothly and the counts depend linearly on it, with Gaussian noise on both. That is the textbook case for a Kalman filter, as Wu, Black and colleagues set out for motor cortex in 2006: predict the velocity each bin, then correct it by the gap between measured and predicted counts, with a gain set by how noisy the counts are against how fast intent changes. In people, decoding velocity this way beat the linear position decoder MN had used (Kim and colleagues, 2008).

The user closes a feedback loop around the filter, watching the cursor and correcting it; as Gilja and colleagues put it, the prosthesis is a new physical plant, with dynamics unlike the arm’s. So their ReFIT Kalman filter (2012) is refitted on closed-loop data, with every decoded velocity rotated to point at the target, as if the user always meant to head straight there, and zeroed on arrival. In monkeys it halved the time to acquire targets.

Then the plant drifts. Within single sessions, Perge and colleagues found, 84% of units in three BrainGate participants changed firing rate significantly, by about half their mean rate on average and mostly for physiological reasons, enough to bias the cursor in 56% of assessments. The answer is an outer loop that re-estimates the decoder while it runs, which is adaptive control proper. Jarosiewicz and colleagues (2015) tracked baselines during pauses, subtracted a running velocity bias and refitted from ordinary typing, inferring afterwards which key the user had been heading for; typing held up for hours, and decayed with the methods switched off. The brain adapts too: Taylor, Tillery and Schwartz saw cells change their tuning once they were driving a cursor. Two adaptive systems, each inside the other’s plant.

The speech decoders swap the Kalman filter for a recurrent neural network and a language model, which is machine learning and beyond this site, but the drift remains: Willett’s participant recorded about 40 minutes of new sentences a day to retrain her decoder.

What it costs

A Utah array goes in through a craniotomy — 5 by 5 cm for the four arrays in Card’s participant — and, so far, reports out through a connector in the scalp. The array is passive: every needle has its own wire, in a gold bundle to a titanium pedestal screwed to the skull, and a cable runs from the pedestal to amplifiers that sample each electrode at 30 kHz and 16 bits — 46 Mbit/s per 96-channel array. It is the Neuropixels wiring problem in its original form.

The connector is where trouble collects. About half of the 68 device-related adverse events among BrainGate’s 14 participants from 2004 to 2021 involved the skin around the incision or pedestal, and there were no intracranial infections. In 78 arrays in 27 monkeys, Barrese and colleagues found most failures within a year, and the commonest kind, abrupt and mechanical, mostly at the connector.

In BrainGate, wireless has so far moved the problem rather than removed it: Simeral and colleagues’ 2021 transmitter, 20 kS/s at 12 bits per electrode and radiating well under a milliwatt, sits on the pedestal. Closing the skin means living on a battery or on power sent through it, as a cochlear implant does, and on either budget it pays to send what the decoder uses: 96 counts every 20 ms is 4,800 numbers a second, against 2.88 million raw samples.

The array also wears out. In Barrese’s series, spike amplitude and working channels declined slowly enough to project complete signal loss at about eight years, with failing insulation the main suspect, and 9 of the 62 arrays that failed were encapsulated by the meninges and pushed out of the cortex. S3’s array gave spikes on 41 of 96 electrodes at 1,000 days, fewer and smaller by her sixth year.

And the evidence is thin in the plainest sense. BrainGate, by its own account the largest and longest-running trial of an implanted brain–computer interface, had implanted 14 people by 2021 and counted 19 participants by 2024; each speech result here comes from one person. Its safety report notes that 33 person-years of observation gives less than a 1% chance of detecting even a doubling in the rate of intracranial infection.