Myoelectric hand prostheses use surface electromyography (sEMG). Although various studies about effective machine learning technique have been conducted, there are few studies regarding reducing the number of sEMG channels. In this study, we investigated how the number and location of channels affect the robustness of the classification accuracy in hand prosthesis.

Total 14 healthy subjects (12 male and 2 female, median age 26 years) were recruited in the study. We placed six sEMG channels targeting forearm muscles (targeting method); extensor carpi radialis (ECR), extensor digitorum communis (EDC), extensor carpi ulnaris (ECU), flexor carpi radialis (FCR), flexor digitorum superficialis (FDS), and flexor carpi ulnaris (FCU). Then, the participants were instructed to perform 14 hand movements as follows; thumb flexion/extension, index finger flexion/extension, middle finger flexion/extension, ring finger flexion/extension, little finger flexion/extension, hand open/close, and wrist flexion/extension. We recorded sEMG signal during each movement and classified the signal using support vector machine algorithm. Auto- regressive coefficients, discrete wavelet transform, root mean square, and time domain feature sets were used. Also, we compared the classification accuracy with the untargeting method, which is placing six uniformly spaced sEMG electrodes just below the elbow.

No statistical differences were observed between targeting and untargeting method when using five or six channels. However, there were statistical differences between those methods when using less than five channels (P-value<0.05). In targeting method, no statistical differences were shown between three to five sEMG channels, while a significant decline of accuracy was observed when reducing channels in untargeting method (Fig. 1). Also, ECR muscle was found to be the most informative (Fig. 2).

We found out that placing the electrodes on targeting muscles can classify EMG pattern more accurately, especially when reducing the number of channels. Also, we investigate optimal channel subset for sEMG classification.