Neuromorphic Motoneuron Spike Train Decoding for Finger Force Predictions

Abstract

Continuous neural control of robotic hand at individual finger level remains a challenge in robotics. Decoding of motor unit (MU) spiking activity from surface electromyogram (sEMG) signals has shown the promise to address this challenge, but computational efficiency limits real-time implementations. Spiking neural networks (SNN) offer an event-driven and computationally efficient alternative that is well aligned with the sparse and temporal nature of MU discharge activity. In this study, we developed an SNN model framework to decode MU spike trains from high-density sEMG (HD-sEMG) signals and further predicted dexterous finger forces. Specifically, we acquired HD-sEMG data from the forearm extensor and flexor muscles during single- and multi-finger isometric force tasks. Ground-truth MU spike trains were obtained using an established MU decomposition pipeline, and root mean square EMG features were used as SNN inputs. The decoded spike trains were then transformed to neural command for finger force predictions. Our results showed that the developed SNN model achieved competitive spike train decoding performance compared with other deep learning models, while exhibiting lower latency and substantially lower computational cost. In addition, the SNN model enabled accurate and robust finger force predictions by transforming the decoded MU spike trains into finger-specific neural command representations, achieving an ${{R}^2}$of 0.90 ± 0.01 with recorded true forces and a force prediction error of 4.27 ± 0.42% maximal voluntary contraction. These findings support SNN-based decoding as an efficient framework for future real-time and continuous control of assistive robotic hands.