Effect of dynamic change of arm position on myoelectric pattern recognition

Robustness of pattern recognition is a key point for myoelectric prosthesis developed in laboratory to be applied successfully in real life. This study investigated the issue about effect of dynamic change of arm position on myoelectric pattern recognition. Six kinds of surface electromyography (sEMG) features were used and compared in detail. The classifier was trained in static (S) condition (i.e. the arm position was stable) and dynamic (D) condition (i.e. the arm position was changed) separately, and then it was tested in S-condition and D-condition respectively. The performances were analyzed among the four groups: S (training)-S (testing), S-D, D-S, and D-D. We found that the dynamic change of arm position had significant impact on sEMG pattern recognition for all features (TDS: average intra-set error 4.72% to inter-set error 17.94%; AR4: 6.66% to 21.93%; AR6: 5.21% to 20.23%; CA6: 5.26% to 20.77%; AR6+RMS: 4.34% to 19.58%; TDS+AR6+RMS: 2.79% to 17.41%). There was no significant difference on robustness among these features (p = 0.2789). The performance of S-D and that of D-S were significantly different (p <; 0.0001). A method that combined training data from multi-experiment set was proposed here and the classified error decreased evidently (from 13.75% to 4.18% with TDS+AR6+RMS). In this paper we also pointed out that the training data were indispensable to contain experiment of dynamic change of arm position.