Adaptive neural network controller for an upper extremity neuroprosthesis

The long term goal of this project is to develop an adaptive neural network controller for an upper extremity neuroprosthesis targeted for people with C5/C6 spinal cord injury (SCI). The challenge is to determine how to simultaneously stimulate different paralyzed muscles based on the EMG activity of muscles under retained voluntary control. The controller extracts the movement intention from the recorded EMG signals and generates the appropriate stimulation levels to activate the paralyzed muscles. To test the feasibility of this controller, different arm movements were recorded from able bodied subjects. Using a musculoskeletal model of the arm, inverse simulations provided muscle activation patterns corresponding to these movements. The model was modified to reflect C5/C6 SCI and the optimization criteria were varied to reflect different nervous system motor control strategies. Activation patterns were then used to train a time-delayed neural network to predict paralyzed muscle activations from voluntary muscle activations. Forward simulations were performed to obtain predicted movements and use the kinematic errors to design an adaptive strategy to account for disturbances and changes in the system.