Simultaneous and proportional force estimation for multifunction myoelectric prostheses using mirrored bilateral training.

This study presents a novel method for associating features of the surface electromyogram (EMG) recorded from one upper limb to the force produced by the contralateral limb. Bilateral-mirrored contractions from ten able-bodied subjects were recorded along with isometric forces in multiple degrees of freedom (DOF) from the right wrist. An artificial neural network was trained to provide force estimation. Combinations of processing parameters were evaluated and an estimation algorithm allowing high accuracy from relatively short signal epochs (100 ms) was proposed. The estimation performance when using surface EMG from the contralateral limb was 0.90 ± 0.02 for the able-bodied subjects. In comparison, the estimation performance for one subject with congenital malformation of the left forearm was 0.72 which, albeit lower than for able-bodied subjects, is still comparable to or better than previously reported results. The proposed method requires only the measured forces from one limb, such as in the case of unilateral amputees and has thus the potential to be used in clinical settings for intuitive, simultaneous control of multiple DOFs in myoelectric prostheses.

Enhanced EMG signal processing for simultaneous and proportional myoelectric control.

A new signal processing scheme is presented for extracting neural control information from the multi-channel surface electromyographic signal (sEMG). The extracted information can be used to proportionally control a multi-degree of freedom (DOF) prosthesis. Four time-domain (TD) features were extracted from the multi-channel sEMG during a series of anisotonic, isometric wrist contractions, which involved simultaneous activations of the three DOF of the wrist. The forces produced at the three wrist DOFs during these contractions were also collected using a customized force sensor. The extracted features and the recorded force signals, as input/target pairs, were then used to train a multilayer perceptron (MLP) neural network. A five-fold cross-validation training/testing method was applied. The resulting performance is a significant improvement over a previously proposed sEMG processing method for the proportional, multi-DOF myoelectric control task.