A neurophysiology-constrained joint network for multi-finger force estimation from surface electromyogram

Le résumé fourni par la source

Decoding multi-finger forces from surface electromyogram (sEMG) is critical for dexterous prosthetic hand control. While recent deep learning-based end-to-end decoders have shown promising performance, they often lack physiological interpretability. Conversely, motor unit (MU) firing information provides a more accurate characterization of forces but is challenging to integrate effectively. Therefore, this study proposes a neurophysiology-constrained joint network (NCJ-Net), which explicitly integrates MU firing information by coupling a gated recurrent unit (GRU)-based decomposition network that estimates MU firing-rate representations with a temporal force regression network. A stage-wise training strategy, featuring initial pretraining followed by joint fine-tuning and MU quality-aware supervision on intermediate decomposition representations, was adopted to tightly couple the two subnetworks beyond a mere structural cascade. High-density sEMG and multi-finger forces were recorded from ten healthy participants performing multi-finger, multi-level isometric force tasks. The proposed NCJ-Net achieved the best performance (root mean squared error = 3.59 ± 0.82 % MVC (maximum voluntary contraction); coefficient of determination = 0.947 ± 0.031) and significantly outperformed all other baseline methods. An additional preliminary test involving one bilateral upper-limb amputee with visually presented target trajectories as surrogate reference outputs showed that NCJ-Net achieved the best overall performance. These results demonstrate that embedding neurophysiological hierarchy into a jointly optimized architecture can improve multi-finger force estimation while preserving the network's interpretability aligned with the neurophysiology prior of descending motor control pathway. Overall, the proposed framework offers a principled route toward intelligent and physiologically grounded decoding systems for human-machine interface applications such as dexterous prosthetic hand control.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.