Reliability of individual finger force prediction from electromyographic signals recorded from surface arrays:...

Objective. High-density surface electromyography (HD-sEMG) has shown promise as a means of providing intuitive user control of assistive devices. Using motor unit (MU) firing activities obtained from HD-sEMG, neural drive-based approaches enable continuous fingertip force prediction. MU decomposition, however, requires intensive computation and frequent calibration, which are not practical to perform daily. This study aimed to evaluate predictive model performance for continuous user control of finger forces across multiple sessions with repeated donning and doffing of the HD-sEMG arrays between each testing session and investigated the impact of simulated electrode channel loss.Approach. Eight neurotypical participants completed three sessions on separate days. In each session, index and middle fingertip forces and HD-sEMG signals from finger flexor and extensor muscles were recorded simultaneously during isometric uni- and bi-directional finger force tasks. sEMG signals were decomposed using independent component analysis to extract MU firing activities. Force prediction performances using only agonist muscles (single array) were compared to those combining agonist-antagonist (both electrode arrays) activity. Cross-day reliability was assessed by applying models derived from the first session to subsequent sessions. Robustness under electrode loss conditions was further evaluated by randomly removing selected EMG channels from data collected during later sessions.Main results. Models using both electrode arrays outperformed those using a single array across all sessions for flexion tasks (p< 0.05) but not for extension tasks. Applying models developed from session 1 data to subsequent sessions' sEMG signals did not significantly increase average root mean square error, which rose by 2.6 ± 2.1% MVF (maximum voluntary force) in session 2 and 3 ± 2.6% MVF in session 3 compared to session 1. Simulated electrode channel loss up to 50% channels caused relatively minor, non-significant differences in prediction performance.Significance. These findings suggest robust multi-day performance and resilience to electrode loss, thereby supporting the use of neural-drive-based approaches for continuous user control of assistive devices.