Controlling an advanced neuroprosthetic hand via a myoelectric interface requires the ability to record multiple, independent muscle action potentials, infer intended movement, and drive the prosthesis appropriately. As proof-of-concept, we implanted nine implantable myoelectric sensors (IMES) in the hand muscles of the forearm of a rhesus monkey, and wirelessly recorded myoelectric activity while the monkey was performing an individual finger movement task with a manipulandum. We developed an off-line estimator of each finger’s joint position (flexed state, extended state, or neither flexed nor extended) through multivariable linear regression of the muscle action potential firing rate, and through systematically varying the design of the estimator. In the training data set, the changes in finger state predicted by the myoelectric decode algorithm provided an excellent fit to the actual state as sensed by the manipulandum. The same algorithm was also able to predict finger state when applied to a different testing data set (~80% correct on a moment-by-moment basis). Subsequently, the algorithm was used to drive a virtual prosthetic hand in real time. Our decode and control strategy differ conceptually from a pattern recognition nominal classifier approach, in which myoelectric signals are used to determine which of a finite set of classes (e.g., power grip) is intended. Because we are independently predicting the desired position of each finger on a continuous basis, our strategy is not limited to inferring a finite set of possibilities, and can allow changes and corrections to the movement while the movement is occurring.

Support by DARPA BAA 05-26.