Spinal cord injury impairs neural pathways between the brain and limbs, but spares both the motor cortex and muscles. Recent studies have shown that quadriplegic patients could volitionally modulate activity of neurons in hand area of motor cortex, even several years after paralysis6
, and that monkeys could use cortical activity to control a robotic arm to acquire targets4
and feed themselves5
. These and other brain-machine interface studies used sophisticated algorithms to decode task-related activity of neural populations and calculate requisite control parameters for external devices4–6,8–10
. An alternate strategy to restore limb function is to directly connect cortical cell activity to control stimulation of a patient’s paralyzed muscles (). Here we show that monkeys can learn to use direct artificial connections from arbitrary motor cortex cells to grade stimulation delivered to multiple muscles and restore goal-directed movement to a paralyzed arm.
Brain-controlled functional electrical stimulation (FES) of muscle
In previous biofeedback studies monkeys rapidly learned to control the discharge rates of newly isolated neurons in motor cortex to obtain rewards14,15
. We used a similar operant conditioning paradigm for single neurons in hand and wrist area of motor cortex of two monkeys (see methods and supplementary information
). We tested volitional control of cell activity by displaying smoothed discharge rate as cursor position on a monitor and rewarding the monkeys for maintaining activity within randomly presented high- or low-rate targets. The directional tuning of most cells was also characterized in an isometric 2-dimensional wrist tracking task. However, our experiment employed all sufficiently well-isolated cells encountered, with no selection bias for possible association to movement or directional tuning.
Monkeys demonstrated volitional control of the discharge rates of nearly all cells tested within the first 10-minute practice session. Although cell activity controlled the cursor directly, monkeys often continued to produce wrist torques during these initial sessions (Figure S1
). We then blocked peripheral nerves innervating the wrist muscles with a local anesthetic (see methods). Despite loss of motor function and sensory feedback from the innervated forearm, monkeys continued to control the cursor with cell activity for 45 of 46 cells after the nerve block. Figure S1
shows the loss of flexor and extensor torques following injections of local anesthetic, while the monkey continued to volitionally control the cell activity. The nerve block was confirmed by the monkey’s inability to perform the 2-dimensional torque tracking task.
We then converted cell activity into proportional stimuli delivered to paralyzed muscles. The cursor was now controlled by wrist torque, and the monkey was rewarded for maintaining FES-evoked torque within peripheral and center (i.e., zero-torque) targets for 0.5 – 1.0 s. To allow the monkey to grade contraction force, stimulation current was made linearly proportional to cell rate when the cell discharged above a threshold.
The example in shows a monkey modulating cell activity to control FES and generate appropriate torques via paralyzed wrist extensor muscles. The monkey learned to increase cell activity to activate the stimulator and acquire the extensor targets, and to maintain activity below the stimulation threshold to relax the muscle and acquire the center targets. Both monkeys were able to control muscle FES during nerve block and acquire torque targets with 44 of the 45 cells tested (5 cells from monkey I and 39 from monkey L).
For each cell the monkeys’ control improved with practice, as evidenced by more rapid acquisition of targets and fewer errors. Monkeys began using cell activity to control the stimulator almost immediately, and improved substantially during the relatively brief practice sessions with each cell (mean duration 66 min). To quantify improvement we compared performance during the initial two minutes of practice and during the two-minute period with the highest performance, typically just before task difficulty was increased to probe the limits of FES control. The rate of target acquisition with FES control was over three times greater during peak performance (14.1 ± 5.3 torque targets acquired /minute; mean ± SD) compared to the beginning of practice (4.0 ± 4.3 targets/min; p < 0.001; Figure S2
). Peak target acquisition rates during brain-controlled FES were similar to those seen when cell activity controlled the cursor directly before nerve block (13.2 ± 5.5 targets/min; p = 0.66).
With continued practice monkeys also learned to control the torque more precisely with cell activity, making fewer target acquisition errors and more often acquiring targets on the first attempt. A target acquisition error was defined as triggering the stimulator to acquire the peripheral target when the center target was displayed. Monkeys made target errors on only 0.8 ± 5.1% of targets during peak performance for each cell compared to 20.7 ± 28.9% of targets at the beginning of practice (p < 0.001; Figure S3
). They also made 81% fewer failed attempts to acquire the target during peak performance (0.10 ± 0.31 failed attempts per target) compared to the beginning of practice (0.52 ± 0.93; p <0.001).
To test whether FES could also be controlled by decreases in cell activity, we set stimulation current to be inversely proportional to cell rate below a threshold for 11 cells. Monkey L learned to control stimulation with this inverse relation just as well as with a positive relation between cell rate and stimulus current (38 cells, some tested in both groups; p > 0.46), acquiring 13.4 ± 3.9 targets per minute and making no errors during peak performance.
The activity of a single cell could also be used to control stimulation of antagonist muscle groups and restore bi-directional movements. shows an example of one cell that controlled stimulation of flexor muscles with high discharge rates and extensor muscles with low rates. The monkey learned to control cell activity and grade contraction force to rapidly satisfy targets at five different torque levels. The nerve blocks remained very effective, as evidenced by negligible torques produced in either direction when the stimulators were turned off during target presentation (). Seven cells tested with such bi-directional control performed similarly to cells that controlled only one muscle group, although target acquisition rates were marginally slower (9.8 ± 3.7 targets/min; p = 0.06).
Brain-controlled FES of multiple muscles restores graded torque in two directions
The assumptions underlying common neural decoding schemes would predict that monkeys should be able to control FES torque better with cells that are strongly related to wrist movements than with unrelated cells. To investigate this, we documented cell activity during a 2-dimensional wrist tracking task before the nerve block, and calculated the directional tuning for each cell (). The magnitude of directional tuning did correlate significantly with the monkeys’ ability to bring the cursor into the optimally placed targets with cell activity during the initial 10-minute practice period (r2 = 0.33, p < 0.001; ). However, cell tuning was not a good predictor of the peak target acquisition rates during subsequent brain-controlled FES (r2 = 0.03, p = 0.33; ). For example, with the untuned cell on the left in the monkey acquired 18.5 targets per minute. The tuned (n = 9) and untuned (n = 29) cells showed no difference in any measure of FES control (target acquisition rates, errors, or failed attempts; p > 0.51).
Cell directional tuning is unrelated to FES control
Extending the strategy of direct neural control to more complex movements will require additional control signals. As a first step toward this goal, we tested a monkey’s ability to simultaneously control two cell-muscle pairs. shows monkey L using high discharge rates of one cell to control FES of flexor muscles and high rates of a second cell to control extensor muscles. The monkey learned to independently modulate the activity of five cell pairs in order to control antagonist muscles and rapidly acquire bidirectional torque targets at rates similar to single cells (11.6 ± 3.8 targets/min, p = 0.32).
These findings have several implications for future approaches to neuroprosthetic control. In contrast to the conventional strategy of deriving control signals from the combined activity of a neural population4–6,8–10
, it may prove efficacious to maintain separate signal pathways from cells to muscles. Using direct channels from single cells to specific muscles may provide the brain with more distinguishable outcomes of the cell activity16
and allow innate motor learning mechanisms to help optimize control of the new connections. The brain’s ability to adapt to novel but consistent sensorimotor contingencies has been amply documented17,18
, and motor cortex can adapt rapidly to learn new motor skills19,20
. Motor circuitry can compensate for drastic changes in connectivity, such as surgically cross-connected nerves controlling wrist flexor and extensor muscles21
, or targeted reinnervation for control of prosthetic limbs22
Our finding that monkeys could learn to use virtually any motor cortex cell to control muscle stimulation, regardless of the cell’s original relation to wrist movement (), suggests another advantage of directly tapping single cell activity. Strategies based on decoding the activity of neural ensembles to obtain movement parameters or muscle activity depend on finding cells that modulate sufficiently with the output variables during actual or imagined movements4–6,8–10
. Instead, arbitrary cells available on recording arrays could be brought under volitional control using biofeedback, substantially expanding the source of control signals for brain-machine interfaces. This and previous biofeedback studies14,15
have shown that even cells with no discernable relation to muscles can be volitionally modulated after brief practice sessions. Issues concerning the use of individual cells and neural populations for prosthetic control are further discussed in the supplement
The degree of FES control demonstrated here was limited by the relatively brief training time provided by the transient nerve block. Implanted electronic circuitry will enable adaptive learning over much longer times and under more varied conditions1
. For example, the autonomous ‘Neurochip’ system can discriminate single cell activity and deliver stimulation through days of free behaviour23,24
. In several preliminary FES sessions, we confirmed that this system would allow a monkey to trigger stimulation of a paralyzed muscle with cell activity and acquire torque targets (Figure S4
). Such autonomous low-power circuits could permit subjects to practice continuously with an artificial connection from brain to muscles or the spinal cord25–27
, without requiring complex decoding algorithms or robotic arms. Further development of such direct-control strategies may lead to implantable devices that could help restore volitional movements to individuals living with paralysis.