Classification of motor commands using a modified self-organising feature map

In this paper, a control system for an advanced prosthesis is proposed and has been investigated in two different biological systems: (1) the spinal withdrawal reflex system of a rat and (2) voluntary movements in two human males: one normal subject and one subject with a traumatic hand amputation. The small-animal system was used as a model system to test different processing methods for the prosthetic control system. The best methods were then validated in the human set-up. The recorded EMGs were classified using different ANN algorithms, and it was found that a modified self-organising feature map (SOFM) composed of a combination of a Kohonen network and the conscience mechanism algorithm (KNC) was superior in performance to the reference networks (e.g. multi-layer perceptrons) as regards training time, low memory consumption, and simplicity in finding optimal training parameters and architecture. The KNC network classified both experimental set-ups with high accuracy, including five movements for the animal set-up and seven for the human set-up.