Background
Brain Machine Interfaces (BMIs) are an emerging technology whose purpose is to allow amputees and spinal cord injury patients to control a prosthetic limb using signals from the brain. One particular area of BMI research uses signals from ensembles of individual neurons that are recorded using densely packed arrays of cortical electrodes. Because such systems record directly from individual neurons, this class of BMIs is expected to be able to achieve the most dexterous prosthesis control.
The large numbers of densely packed recording electrodes, combined with the generally poor quality of the neural signals and the signal processing requirements for estimating neural intent, create a set of unique design challenges for the data acquisition system. Presently, progress towards implantable, wireless BMI instrumentation is hindered by the absence of a definition of the minimum allowable data acquisition system and the lack of a means for rapidly testing new prototype systems. To this end, a novel model is being developed in our lab that aims to replicate the interaction between an adaptive motor control loop in the brain, spike-trains generated in real-time by cortical neurons, a BMI recording system, and a prosthetic limb. The system will be comprised of both hardware and software components, and will use real-time closed-loop visual feedback to adaptively train synaptic weights in the virtual brain. This model will make it possible to investigate how (and to what extent) the brain is able to adapt its behavior in response to performance deficits in each of the BMI instrumentation subsystems. The model will also facilitate the establishment of minimum performance criteria for each of the instrumentation subsystems below which the brain is incapable of adapting its behavior to adequately control the prosthesis. Finally, the proposed system will allow for rapid evaluation of prototype BMI instrumentation.
In the initial phase of this work, the basic adaptive closed-loop robotic system will be developed. The system will be capable of smoothly guiding a multiple degree-of-freedom robotic arm through a trajectory in three-space using a virtual cerebellum trained with visual feedback to correct for prosthesis dynamics. Subsequently, an instrumentation testbed will be developed that is capable of implementing entire families of BMI data acquisition subsystems and systematically manipulating their parameters. The system will handle up to 100 channels and will collect performance statistics that quantify how and where information is lost or altered in the data pathway. Finally, a virtual motor cortex will be developed that translates the desired prosthesis motor angles into real-time spike trains. Virtual neurons will code for multiple parameters simultaneously including confounding factors such as end effector position and velocity, as well as random drift.
Figure 1. Block diagram of the BMI model under development. The controller/cerebellum complex create signals for the two robot arm motors. These signals are then rate-encoded by the neurons of the motor cortex and communicated to the robot arm via the BMI instrumentation. A visual feedback system determines the error between the desired and actual robot arm position and uses this error to train the cerebellum. The system is first being developed in two dimensions and will be subsequently upgraded to 3D.
