On the Deployment of Edge AI Models for Surface Electromyography-Based Hand Gesture Recognition

Robotic-based therapy has emerged as a prominent treatment modality for the rehabilitation of hand function impairment resulting from strokes. Aim: In this context, feature engineering becomes particularly important to estimate the intention of upper limb movements by utilizing machine learning models, especially when a hardware embedded-on-board implementation is expected, due to the strong computational, energy, and latency constraints. Methods: The present study details the implementation of four cutting-edge feature engineering techniques (random forest, minimum redundancy maximum relevance (MRMR), Davies–Bouldin index, and t-tests) in the context of machine learning algorithms (neuronal networks and bagged forests) deployed within a resource-constrained autonomous embedded system. Results: The findings of this study demonstrate that by assigning relative importance to features and removing redundant or superfluous information, it is possible to enhance the system’s execution by up to 31% while preserving the model’s performance at a comparable level. Conclusions: This work proves the usefulness of TinyML as an approach to properly integrate AI into constrained edge embedded systems to support complex strategies such as the proposed hand gesture recognition for the smart rehabilitation of post-stroke patients.