Force classification using surface electromyography from various object lengths and wrist postures

Pattern recognition using myoelectric control of upper-limb prosthetic devices is essential to restore control of several degrees of freedom. Although much development has been relevant, the prediction of force level in finger movements is scanty. In this study, we propose the surface electromyography (sEMG) to predict the force level of the thumb-index pinch. Ten non-amputee subjects are asked to do five force levels with three different wrist positions and five object lengths. The sEMG data are recorded from three muscle regions (12 channels) of the right forearm. Twelve traditional time-domain features are extracted from collected sEMG signal. The sequential forward floating selection (SFFS) method is investigated to find the optimal set of muscles and features for force prediction. Performances from seven linear and nonlinear classifiers are compared. The results show that k-nearest neighbor and neural network outperform other classifiers with the accuracy of about 99% and 97%, respectively. The accuracy from the set of muscle groups and features selected by SFFS method is slightly better than that from the set of baseline (all of channels and features). The frequently selected muscles are from the hand region. However, the combination of lower and upper muscles also performs well, which is useful for the prosthetic design in a hand and wrist disarticulation amputee and a transradial amputee in the future.