A Wearable Platform for Real-Time Control of a Prosthetic Hand by High-Density EMG

Abstract

This study presents a heterogeneous embedded architecture that addresses a fundamental gap in wearable myoelectric systems: the inability of existing platforms to simultaneously provide high-density signal acquisition, computational flexibility, and autonomy. The platform integrates two 64-channel RHD2164 front-ends (128 channels total) with a Zynq UltraScale+ multiprocessor system-on-chip for heterogeneous processing. A PYNQ-based Python/Linux framework enables scalable algorithm development. Experiments with 21 healthy subjects performing eight motor tasks (finger flexion/extension, thumb opposition, and grasp patterns) at two frequencies (0.50 and 0.75 Hz) demonstrated the platform’s capability in high-density surface electromyography (HD sEMG) recording and real-time control of a single degree of freedom (1-DoF). Signal quality exceeded recommended thresholds (Signal-to-Noise Ratio: <inline-formula> <tex-math notation="LaTeX">$13.93 \pm 7.51$ </tex-math></inline-formula> dB; Signal-to-Motion-artifact Ratio: <inline-formula> <tex-math notation="LaTeX">$25.18 \pm 5.18$ </tex-math></inline-formula> dB), confirming the effectiveness of the dual-front-end architecture. The processing pipeline combined reinforced electrode signal adaptation (RESA), non-negative matrix factorization (NMF), and Kalman filtering, resulting in strong agreement between estimated and reference signals, with maximum normalized cross-correlation (<inline-formula> <tex-math notation="LaTeX">$\mathrm {XC}_{\max }$ </tex-math></inline-formula>) values from <inline-formula> <tex-math notation="LaTeX">$0.54 \pm 0.22$ </tex-math></inline-formula> to <inline-formula> <tex-math notation="LaTeX">$0.80 \pm 0.16$ </tex-math></inline-formula>. The coefficient of determination (R2) for HD sEMG reconstruction ranged from <inline-formula> <tex-math notation="LaTeX">$0.87 \pm 0.09$ </tex-math></inline-formula> to <inline-formula> <tex-math notation="LaTeX">$0.95 \pm 0.03$ </tex-math></inline-formula>, with higher values for prehension tasks. End-to-end latency from acquisition to command output ranged from <inline-formula> <tex-math notation="LaTeX">$63.3 \pm 1.0$ </tex-math></inline-formula> ms (30 ms buffer) to <inline-formula> <tex-math notation="LaTeX">$219.1 \pm 4.5$ </tex-math></inline-formula> ms (150 ms buffer), maintaining temporal alignment (<inline-formula> <tex-math notation="LaTeX">$\mathrm {XC}_{\max }$ </tex-math></inline-formula> lag: <inline-formula> <tex-math notation="LaTeX">$0.02 \pm 2.09$ </tex-math></inline-formula> s). The heterogeneous architecture supports full local processing, with the FPGA handling acquisition and the Arm Cortex-A53 cores performing motor intention decoding, providing a scalable foundation for adaptive multi-DoF prosthetic control.