Nonlinear mappings between discrete and simultaneous motions to decrease training burden of simultaneous pattern recognition myoelectric control.

Abstract

Real-time simultaneous pattern recognition (PR) control of multiple degrees of freedom (DOF) has been demonstrated using a set of parallel linear discriminant analysis (LDA) classifiers trained with both discrete (1-DOF) and simultaneous (2-DOF) motion data. However, this training method presents a clinical challenge, requiring large amounts of data necessary to re-train the system. This study presents a parallel classifier training method that aims to reduce the training burden. Artificial neural networks (ANNs) were used to determine a nonlinear mapping between surface EMG features of 2-DOF motions and their 1-DOF motion components. The mapping was then used to transform experimentally collected features of 1-DOF motions into simulated features of 2-DOF motions. A set of parallel LDA classifiers were trained using the novel training method and two previously reported training methods. The training methods evaluated were (1) using experimentally collected 1-DOF data and ANN-simulated 2-DOF data, (2) using only experimentally collected 1-DOF data and (3) using experimentally collected 1- and 2-DOF data. Using the novel training method resulted in significantly lower classification error overall (p<;0.01) and in predicting 2-DOF motions (p<;0.01) compared to training with experimental 1-DOF data only. These findings demonstrate that using a set of ANNs to predict 2-DOF data from 1-DOF data can improve system performance when only discrete training data are available, thus reducing the training burden of simultaneous PR control.