Classification of combined motions in human joints through learning of individual motions based on muscle synergy theory

This paper proposes a novel method of pattern classification for user motions to create input signals for human-machine interfaces from electromyograms (EMGs) based on muscle synergy theory. The method can be adopted to represent non-trained combined motions (e.g., wrist flexion during hand grasping) using a recurrent neural network by combining synergy patterns of EMG signals preprocessed by the network. This approach allows combined motions (i.e., unlearned motions) to be classified through learning of individual motions (such as hand grasping and wrist flexion) only, meaning that the number of motions can be increased without increasing the number of learning samples or the learning time needed to control devices such as prosthetic hands. The effectiveness of the proposed method was demonstrated through motion classification tests and prosthetic hand control experiments with six subjects (including a forearm amputee). The results showed that 18 motions (12 combined and 6 single) could be classified sufficiently with learning for just 6 single motions (average rate: 89.2 ± 6.33%), and the amputee was able to control a prosthetic hand using single and combined motions at will.