Simultaneous estimation of 2-DOF wrist movements based on constrained non-negative matrix factorization and Hadamard product
Published on Feb 1, 2020in Biomedical Signal Processing and Control2.943
· DOI :10.1016/J.BSPC.2019.101729
Abstract Decoding hand/wrist movements of multiple degrees of freedom (multi-DOF) plays a significant role in the simultaneous and proportional control of dexterous prostheses. In this paper, we propose a new algorithm, CNMF-HP, that integrates the constrained non-negative matrix factorization (NMF) and Hadamard product, for estimating 2-DOF wrist movements (DOF-1, flexion/extension; DOF-2, adduction/abduction) from myoelectric signals. We suppose that, by supplementing L2 norm regular term and Hadamard product to the objective function, the performance of NMF can be improved. We evaluate our method through both offline (cross-validation) and online (target-tracking) experiments, on metrics such as regression accuracy (ASNR and R2) and those from 2D Fitts’ law (completion rate, path efficiency, throughput, etc.). Our results show that, compared with the existing methods (classic NMF; NMF with sparse constraint, SCNMF; NMF with Hadamard product, NMF-HP), our method CNMF-HP shows superior both in the offline cross-validation (metrics: average ASNR, R2) and online experiments (metric: throughput).