Myoelectric digit action decoding with multi-label, multi-class classification: an offline analysis

Abstract

The ultimate goal of machine learning-based myoelectric control is simultaneous and independent control of multiple degrees of freedom (DOFs), including wrist and digit artificial joints. For prosthetic finger control, regression-based methods are typically used to reconstruct position/velocity trajectories from surface electromyogram (EMG) signals. Although such methods have produced highly-accurate results in offline analyses, their success in real-time prosthesis control settings has been rather limited. In this work, we propose action decoding, a paradigm-shifting approach for independent, multi-digit movement intent decoding based on multi-label, multi-class classification. At each moment in time, our algorithm classifies movement action for each available DOF into one of three categories: open, close, or stall (i.e., no movement). Despite using a classifier as the decoder, arbitrary hand postures are possible with our approach. We analyse a public dataset previously recorded and published by us, comprising measurements from 10 able-bodied and two transradial amputee participants. We demonstrate the feasibility of using our proposed action decoding paradigm to predict movement action for all five digits as well as rotation of the thumb. We perform a systematic offline analysis by investigating the effect of various algorithmic parameters on decoding performance, such as feature selection and choice of classification algorithm and multi-output strategy. The outcomes of the offline analysis presented in this study will be used to inform the real-time implementation of our algorithm. In the future, we will further evaluate its efficacy with real-time control experiments involving upper-limb amputees.3/11 5/11

Discrete action control for prosthetic digits

“…We extracted two EMG features from each channel, namely, waveform length and log-variance [5]. We based our selection on previous findings showing that these features are effective both for multi-output regression and multilabel classification [14,23]. We used the same EMG processing pipeline for the two control schemes presented in the following section.…”

“…We set the action step for the "open" and "close" commands such that a single-DOF movement from the bottom to top position, or vice-versa, would require 1.5 s. This translated in using an action step of 0.043. Based on previous findings [23], we trained six independent linear discriminant analysis (LDA) classifiers, one for each available DOF.…”

“…position) control in a robotic hand tele-operation task with an instrumented data glove [22]. In addition, we have performed a systematic offline investigation of multilabel classification for digit action decoding [23]. Here, we provide for the first time a real-time implementation of the algorithm for myoelectric digit control.…”

sEMG-based Regression of Hand Kinematics with Temporal Convolutional Networks on a Low-Power Edge Microcontroller

“…Alternative approaches to hand kinematics focus on the velocity of hand joint movements. For instance, velocity was targeted in [13], adopting a hybrid classification-regression setup which thresholds speed into 3 levels, thus still limiting the prediction to discrete classes. A completely orthogonal approach for sEMG regression focuses on hand dynamics instead of kinematics.…”