Artificial neural networks for HD-sEMG-based hand position estimation: addressing inter- and intra-subject variability

Background: Reliable control of rehabilitation and assistive devices using High-Density surface Electromyography (HD-sEMG) remains limited by poor robustness to electrode shifts, changes in skin condition, and variability across users. Methods: This study evaluates the performance of the Recursive Prosthetic Control Network (RPC-Net)/High-Density Electrode Array (HDE-Array) system, defined in previous studies, under conditions that reflect real-life usage, including electrode repositioning and cross-subject generalization. The first test evaluated whether the RPC-Net/HDE-Array system maintained stable performance when trained without electrode repositioning and evaluated on data from a different session with altered electrode placement. The study further examined whether explicitly incorporating electrode repositioning during training mitigates the performance degradation typically observed when testing is performed in a separate session. Finally, the effects of inter-subject training were assessed. Results: Experimental results demonstrate that the RPC-Net/HDE-Array system is highly sensitive to electrode repositioning and skin condition variability when trained under static conditions. However, robustness improves significantly when such variability is included during training. The results indicate that performance improves with an increasing number of subjects in the training pool, provided the training set includes only data from subjects other than the one tested, suggesting a strong dependency on subject-specific patterns Conclusions: These findings demonstrate that the RPC-Net/HDE-Array system can achieve robust performance across sessions and users when trained under realistic conditions. This work represents a key step toward practical deployment of muscle-computer interfaces.