Effects of Decomposition Parameters and Estimator Type on Pseudo-online Motor Unit Based Wrist Joint Angle Prediction

Abstract

The decomposition of HD-EMG into motor unit (MU) discharge timings permit a detailed window into the motoneuronal manifestation of motor intent. Recently, the feasibility of MU-driven wrist joint angle estimation was preliminarily demonstrated although the influences of certain parameter selections have yet to be fully investigated. Here, a decomposition algorithm was used to predict wrist joint kinematics over three DoFs in a pseudo-online manner. Three separate estimator types were tested and the effects of two key parameters on their prediction accuracies were studied: the decomposition extension factor and process window length. Pre-recorded EMG from four able-bodied subjects was decomposed in a simulated real-time manner as to permit parameter scanning, with the tested estimators being linear regression (LR), linear discriminant analysis (LDA), and LDA with LR for proportionality control (LDA-LR). Results showed the best performing combination of parameters were an extension factor of 8 with window length of 50 ms which allowed the LDA-LR estimator to yield an average R2 value of 0.86 ± 0.05. Under the most computationally demanding set of parameters, the median processing time of the algorithm on a desktop computer was 47 ms which was within the update rate of the proposed system. Such results also indicate that parameters optimal for online control applications deviate from those ideal for offline physiological studies.