Automated optimal coordination of multiple-degree-of-freedom musculoskeletal actions in feedforward neuroprostheses

Musculoskeletal joint actions are planned and coordinated by the nervous system and carried out by typically redundant muscles generating moments at multiple joints. Injuries to the spinal cord (SCI) damage some of the nerves that convey motor commands from the brain to the muscles, resulting in muscle paralysis. Motor neuroprosthesis can restore some lost function by delivering electrical impulses to the muscles via electrodes connected to a stimulator. A command signal maps these impulses to desired movement kinetics/kinematics using feedforward control laws, which rely on expert human observation and intervention. These maps are typically implemented for limited operational modes, which causes current neuroprostheses to not provide independent control of coupled degrees of freedom, limiting the amount of function restored. We examined the automated creation of these input/output maps from clinically-measured input-output data, and evaluated the use of artificial neural network-based inverse controllers to decouple multiple degrees of freedom. We tested this approach in computer simulations with a biomechanical model of a thumb and in experiments by stimulating the extensor pollicis longus, abductor pollicis brevis, and adductor pollicis of ten able-bodied individuals and one SCI patient. Muscle stimulation generated time-varying and coupled thumb-tip forces along two degrees of freedom (i.e., extension/flexion, abduction/adduction). We trained a neural network-based forward model (system model) to create a mathematical representation of the experimentally measured input-output data properties, capable of reducing time variability in the responses. The system model did not eliminate mechanical redundancy, but allowed us to choose new, non-redundant input-output data to optimize muscle cocontraction with a criterion that minimized the summed squared muscle activations. We trained an inverse model feedforward controller with the optimized data set. RMS errors smaller than 2N and significant correlations (R2>0.7, p=0) in seven isometric force generation tests showed the ability of the controller to independently control both degrees of freedom. The experimental results show that we can construct feedforward controllers, which accurately invert the input/output properties of nonlinear, multi-degree of freedom, redundant neuromuscular systems. Accurate isometric force control directly confirms accurate muscle input control, which implies that this technique can be extended to kinematic control if kinematic variables are used as the measured output variables in the system model training. The demonstrated technique has the advantage of generality since it does not rely on a-priori knowledge of the underlying structure of the system being controlled.