Evaluation of Hand Action Classification Performance Using Machine Learning Based on Signals from Two sEMG Electrodes

Abstract

Classification-based myoelectric control has attracted significant interest in recent years, leading to prosthetic hands with advanced functionality, such as multi-grip hands. Thus far, high classification accuracies have been achieved by increasing the number of surface electromyography (sEMG) electrodes or adding other sensing mechanisms. While many prescribed myoelectric hands still adopt two-electrode sEMG systems, detailed studies on signal processing and classification performance are still lacking. In this study, nine able-bodied participants were recruited to perform six typical hand actions, from which sEMG signals from two electrodes were acquired using a Delsys Trigno Research+ acquisition system. Signal processing and machine learning algorithms, specifically, linear discriminant analysis (LDA), k-nearest neighbors (KNN), and support vector machines (SVM), were used to study classification accuracies. Overall classification accuracy of 93 ± 2%, action-specific accuracy of 97 ± 2%, and F1-score of 87 ± 7% were achieved, which are comparable with those reported from multi-electrode systems. The highest accuracies were achieved using SVM algorithm compared to LDA and KNN algorithms. A logarithmic relationship between classification accuracy and number of features was revealed, which plateaued at five features. These comprehensive findings may potentially contribute to signal processing and machine learning strategies for commonly prescribed myoelectric hand systems with two sEMG electrodes to further improve functionality.

Keywords: classification; hand actions; machine learning; myoelectric prosthetics; sEMG; upper limb.

Figure 1

Study overview.

Figure 2

Block diagram of the measurement system illustrating steps involved in sEMG signal acquisition.

Figure 3

(a) sEMG electrodes placed on lateral and medial sides of a participant’s forearm. (b) A photo demonstrating hand and arm positions on a desk prior to data collection. (c) Illustrations of the six hand actions. (d) Time sequence of a hand action session.

Figure 4

Signal processing block diagram.

Figure 5

(a) Typical sEMG signals from WE action as a function of time, (b) segmented signals, (c) full-wave rectified signals, and (d) filtered signals, including exemplar identification of features of minimum and maximum voltages (V_min and V_max).

Figure 6

Flowchart outlining machine learning algorithms for (a) KNN, (b) SVM, and (c) LDA.

Figure 7

Mean and ±1 SD of sEMG signals for (a) HO, (b) WE, (c) WP, (d) HC, (e) WF, and (f) WS actions.

Figure 8

Example of confusion matrices obtained with SVM algorithm based on signals obtained from one participant using (a–c).

Figure 9

Mean profiles of (a) true positive (TP), (b) true negative (TN), (c) false positive (FP), and (d) false negative (FN) across all participants.

Figure 10

Mean profiles of (a) precision, (b) sensitivity, (c) action-specific accuracy, and (d) F1-score across all participants. Vertical lines indicate plateau onset values for number of features and horizontal lines indicate corresponding mean performance metrics.

Figure 11

Overall classification accuracy as a function of number of features averaged across all participants, where the corresponding values in the literature are indicated by the shaded band and the dashed line indicates accuracy at which plateauing occurs.

Figure 12

Comparison of overall classification accuracy across KNN, LDA, and SVM algorithms for 3, 6, 9, and 12 features.