A robust model-based neural-machine interface across different loading weights applied at distal forearm

Abstract Musculoskeletal models (MMs) have recently been proposed to decode electromyography (EMG) signals for movement intent recognition. Since the robustness is critical to retain the performance of neural-machine interface (NMI) during daily activities and the loading weight change is one of the critical factors that would affect the performance of NMI, this study aimed to further investigate the robustness of a generic MM-based NMI across different loading conditions. Eight able-bodied (AB) individuals and one individual with a transradial amputation were recruited and tested while performing a real-time virtual wrist/hand posture matching task under different loading weights (AB subjects: 0 kg, 0.567 kg, and 1.134 kg; amputee subject: 0 kg and 0.567 kg) applied at the distal forearm. All tasks were achieved by both AB individuals and the individual with the transradial amputation. There was no significant difference among the real-time performance (completion time, the number of overshoots, and path efficiency) of AB individuals under different loading conditions. We calculated the average muscle activations of each muscle during the initial 0.5 s and last 0.5 s respectively for each target across all subjects and trials. The analysis of muscle activations showed that additional weights caused muscle co-contractions. However, the subjects can cope with the increased muscle co-activation level, modifying muscle activation patterns, and still complete tasks successfully. We obtained similar results from the individual with the transradial amputation. These results demonstrated the robustness of MM-based NMI across different loading conditions. The outcomes indicate the potential of the multi-user NMI toward practical applications.