Forearm Orientation and Contraction Force Independent Method for EMG-Based Myoelectric Prosthetic Hand

An accurate and robust pattern recognition system is crucial for advanced electromyography-based upper limb prosthetics. Variations in the user's forearm orientation and contraction force varies the electromyography features for the same movement. It degrades the performance of electromyography-based movement classification. Few attempts reported to increase the classification accuracy under such dynamic factors include utilizing multi-stage classifiers, multi-modal sensory data, high-dimensional data, or set of power spectral descriptors. This work aims to overcome limitations in the existing methods through a novel feature set. The combined effect of forearm orientations and contraction force levels on the classification of six hand movements are considered in this study. In this paper, a novel feature set extracted from electromyography's wavelet bispectrum, obtained as the combination of continuous wavelet transform and bispectrum, is proposed. The proposed method remains independent of the multiple dynamic factors due to its invariant and uniqueness properties. Linear discriminant analysis classifier with the proposed wavelet bispectrum feature set showed superiority (classification accuracy=86.43 +/- 3.67%) over conventional time-domain feature set (classification accuracy=75.92 +/- 5.41%) when trained with data from single orientation and force level. When trained with data from appropriate variations (three orientations and medium force), the accuracy with the proposed method increased to 90.35 +/- 2.01%. Further, by training with data from all three orientations and force levels, this method achieved an accuracy of 96.11% +/- 1.34%. The statistical significance of the higher classification accuracy achieved with the proposed feature set compared to other conventional feature sets is proven (repeated measure one-way ANOVA, p < 0.05).