Personalized electromyography based control algorithm development and training system for hand prosthesis

Abstract

Myoelectric hand prostheses serve as an essential tool for upper limb amputees, facilitating the restoration of hand functionality. Despite notable advancements in the field of pattern recognition algorithms for interpreting surface electromyography (EMG) patterns, there persists a marked paucity of focus on the amputees' motor learning process and the consistency of prosthesis application. This lacuna has engendered control mastery deficits and performance degradation over time among numerous myoelectric prosthesis users. Additionally, the complexity and cumbersome nature of existing prosthesis poses substantial learning difficulties and commercial challenges. Considering these issues, this study primarily addresses the variance in prosthesis performance, develops personalized surface EMG-based prostheses, and standardizes training platforms and prosthesis to enhance the user experience for upper limb amputees. Specifically, five interconnected sub-studies were conducted, including: (1) A modularized multi-channel surface EMG sensing system was developed. The standardized design synchronized the sensing system between the training platform and the prosthesis design. Experimental results validated the system's consistency, from training platform to prosthesis control, facilitated by the modularized design. (2) A study was conducted surveying amputees' preferences between Virtual Reality (VR) and robotic training platforms. A 3D-printed soft-robotic training platform for home application was subsequently developed, aligning with amputees' preferences. Furthermore, a comparative study between the 3D-printed soft-robotic prosthetic hand and traditional prosthetic hands assessed mechanical parameters and wearing comfort. Amputees trailed both prostheses in handling daily objects, with a notable preference towards the soft-robotic prostheses. (3) A novel hand gesture classification algorithm was engineered for prosthesis control, specifically addressing electrode shifting issues occurring inside amputees’ sockets. The shortest-surface-distance-matrix (SSDM) was proposed to characterize channel distribution. Models trained using the SSDM generate a sparser channel distribution compared to traditional models, demonstrating enhanced practicality and robustness for daily applications, without compromising classification accuracy. Additionally, the model incorporating SSDM demonstrated a faster convergence rate, highlighting the algorithm's efficiency. (4) A longitudinal study was conducted examining the variance in amputees' muscle patterns during functional training. Both amputee participants exhibited an increase in accuracy over time, indicating enhanced prosthesis control ability. Concurrently, the muscle activation pattern coalesced as training sessions progressed, illustrating a motor learning process wherein the residual limb's motor map reorganized, and muscle coordination improved. Notably, the motor learning process demonstrated consolidation; during a six-month follow-up session, classification accuracy exceeded the accuracy attained during the initial training sessions, substantiating the enduring effect of the training intervention. (5) An extended study was conducted to investigate the channel feature attribution variance for model development. Both participants presented individualized feature attribution pattern in response to the training, echo the variance in previous study. A decreasing weighted channel dispersion implied a similar motor learning phenomenon and provided an optimization solution for prosthesis control algorithm. In summary, this study offers an in-depth examination of muscle synergy variance and constructs an ecosystem bridging the gap between the training and practical prosthesis application for upper limb amputees. This work represents a pioneering effort in the domain of personalized prosthetics, offering the potential to enhance the living experience of amputees significantly.