A medium-density EMG system for real-time proportional and simultaneous neuroprosthetic hand control using lightweight neural networks.

OBJECTIVE: Restoring intuitive and natural hand control remains a key challenge in neuroprosthetics. A promising approach is to decode continuous finger-joint angles from EMG signals, enabling dexterous interaction. Deep neural networks (DNNs) show strong potential for this task, but are often too computationally demanding for deployment on embedded systems and are rarely validated under real-time conditions. APPROACH: This study presents a two-phase framework for realtime decoding of 11 finger-joint angles from medium-density EMG (MD-EMG) signals using lightweight neural networks. We compare a CNN baseline and an adapted version of the Dual Predictive Attractor-Refinement Strategy (DPARS) model across both offline and real-time scenarios. Models were genetically optimized across different preprocessing and downsampling configurations to identify deployment-efficient setups. To assess realtime applicability, participants controlled a virtual hand using EMG signals while imitating target gestures. This setup enabled us to evaluate model accuracy, latency, and the ability to sustain correct joint trajectories over time. MAIN RESULTS: Our adapted DPARS achieved real-time performance comparable to CNN (R2 = 0.750 vs. 0.777), while reducing model size by 7× (227.8 KB) and forward pass latency by 120× (0.25 ms). SIGNIFICANCE: This work highlights the feasibility of combining MD-EMG acquisition with genetically optimized lightweight neural networks for efficient and reliable real-time decoding of finger joint angles under embedded constraints, ultimately advancing the development of intuitive and dexterous neuroprosthetics.