It has long been thought that neuroprosthetic devices
might be useful for restoration of motor functions in
patients afflicted with paralyzing neurological diseases.
In particular, neuronal activity recorded in the motor
might be useful as a source of "motor command" information
to drive a neuroprosthetic device. At its simplest level,
such a device could be conceived as a robotic arm capable
of being controlled from a patient's motor cortex to the
extent that it could reproduce some normal limb functions
of the arm. Further developments would include the ability
to allow a patient to control his own muscles through
functional neuromuscular stimulation (FNS). Such techniques
might ultimately make it possible to electronically create
posture and locomotion under the direct control of the
brain. Finally, it might be possible to create a sensory
neuroprosthesis capable of carrying cutaneous and proprioceptive
feedback from the limbs, allowing ongoing movement to
be more accurately controlled. Beyond this, the development
of such neuroprosthetic technologies would provide new
concepts and tools for the study of information processing
in the cerebral cortex, which is of immense scientific
importance.
To determine whether this basic idea is feasible, we
developed the following experimental paradigm: up to 46
neurons were simultaneously recorded in the forelimb motor
cortex, ventrolateral thalamus and/or cerebellum of rats
initially trained to obtain water by pressing a lever
to control movement of a robot arm. Neuronal population
activity peaked just before onset of lever movement, but
this peak encoded the trajectory of movement over the
next 3-500 ms.
In order to use this multi-neuronal signal to control
a robot arm it was first necessary to "decode" the information
present in the neuronal population. In particular, we
needed to define a mathematical transformation capable
of taking the signals from up to 46 separate neural recordings
and using them to create a single analog signal appropriate
for controlling the position of the robot arm.
First, principal components analysis (PCA) was used to
extract major sources of signal in this multichannel data
set, while removing noise. To appropriately transform
the time domain of this signal we utilized multivariate
statistical techniques, or neural networks to yield output
functions that closely matched the timing, magnitude,
and direction of forelimb movements. After training, the
control of the robot arm was switched from the lever press
to the neuronal population. Animals with at least 25 recorded
neurons successfully used the first principal component
of the multi-neuron population signal to move the robot
to the water source and return it to their mouths.
These results showed that the animals were to use the
neural population signal as a surrogate for limb movement
to move the robot arm and obtain reward, even though its
tendency to peak in the premovement period resulted in
the delivery of water reward before the onset of movement.
This therefore provided a test of whether the animal could
alter its operant behavior and/or its neural activity
to take advantage of the changed situation. Typically,
during the first few minutes of neurorobotic mode, the
animals would continue for several trials their welltrained
behaviors of pressing the lever down to the original threshold
position for obtaining the water reward. In fact, this
was necessary considering that the peak amplitude of pre-movement
neural activity was still proportional to the ultimate
magnitude of the movement.
Over continued trials, however, the normally high correlation
between the neural signal and the movement magnitude tended
to disappear such that the neural signal became independent
of the movement. The correlation usually became insignificant
within about 10-20 trials (over about five minutes). After
that, it would often go back to full or partial movement,
but exhibited many more movements where it only reached
to contact the lever but not to press it. Even so, the
cortical signal reached a much higher level than it would
have when performing similar movements during previous
training in the neurorobotic mode. Thus the neural coding
necessary to obtain the reward did not require that the
associated reaching movement be an operant for the previously
learned behavior, and instead, the neural signal itself
could be rapidly replaced as the operant.
Overall, these findings have demonstrated the feasibility
of using real time neuronal population activity to control
external motion devices, and also suggest techniques for
extracting multiple motor codes from these populations.
Such techniques for large scale chronic recording of brain
neurons, and for transformation of these recordings into
appropriate output signals, could someday be used by paralyzed
patients to control external movement devices or even
their own muscles through functional electrical stimulation.
