Subject-Specific Parameter Optimization for Event-Driven Encoding of sEMG in Regression-Based Motor Control

Myoelectric control of upper-limb prostheses using surface electromyography (sEMG) has been widely adopted for decades. While control strategies have evolved from simple on/off and proportional control schemes to pattern recognition-based methods enabling multiple gestures, achieving natural and continuous control remains a challenge. Regression-based approaches have recently made it possible to infer hand kinematics from sEMG for more fluid control. In this paper, we propose a brain-inspired, neuromorphic computing approach for regression-based motor control using spiking neural networks (SNNs). We investigate how the choice of spike encoding parameters and network architecture of the decoding model influences regression performance. Using a Leaky Integrate-and-Fire (LIF) encoding scheme and a multilayer spiking perceptron, we evaluate our method on the publicly available MoveR dataset, achieving a mean absolute error of 6.01 ± 3.27° across five degrees of actuation. Furthermore, we demonstrate that subject-specific tuning of encoding parameters leads to improved performance, and we show that decoding accuracy correlates with the structure of the underlying spiking activity. These findings highlight the potential of SNN-based models for energy-efficient, real-time motor decoding in wearable neuroprosthetic systems.