Assessment of the performance of regression-based machine learning techniques for offline translation of EMG signal to hand kinematics

Introduction: The loss of upper limb functionality is a prevalent issue, impacting millions and showing an increasing trend over time. This condition significantly affects the quality of life of individuals, necessitating urgent solutions. Despite substantial progress in the field, current prosthetic control solutions have limitations, are cumbersome, and are difficult to use, leading to a high rejection rate of prosthetic devices, a well-recognized indicator of patient satisfaction. Additionally, there is a marked disparity between commercial and academic solutions: commercial devices often lack dexterity and use unnatural, non-intuitive control strategies, whereas academic devices typically depend on complex and time-consuming algorithms, challenging real-time applications. Addressing these issues requires a new control system capable of functioning in near real-time across multiple Degrees of Freedom (DoFs), and controlling them proportionally, simultaneously, independently, and continuously. Such an innovation is a critical first step toward meeting user needs, reducing prosthetic abandonment rates, and improving the quality of life of patients. This MSc project represents a collaborative effort between the Laboratory for Engineering of the Neuromuscular System (LISiN), Politecnico di Torino, and the Nuffield Department of Surgical Sciences, University of Oxford, with the main goal of contributing to Giovanni Rolandino's doctoral research by assessing the performance of different neural network architectures in estimating the kinematic state of the hand from the electromyographic signal. Methods: We acquired a new dataset simultaneously recording high-density surface Electromyography (HD-EMG) on the forearm and hand position of subjects performing a variety of hand gestures. This dataset was used to assess the performance of 9 different neural architectures, belonging to one of three groups: Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs). Performance is evaluated using Pearson's correlation coefficient (PCC) and the Mean Distance (MD), assessing prediction accuracy. Results: The findings underscore the feasibility and potential of employing neural networks in prosthetic control, particularly emphasizing simpler network architectures. MLPs stood out for their quick training time, achieving a PCC of 0.8212 ± 0.1415 and a low MD (22.73mm ± 9.96mm) for the top-performing subject. While CNNs and RNNs also demonstrated commendable performance, their complexity may render them less practical for clinical applications. The best outcomes observed for these two latter architectures were as follows: CNNs (PCC: 0.8199 ± 0.1367, MD: 20.21mm ± 10.09mm); RNNs (PCC: 0.7882 ± 0.1351, MD: 32.18mm ± 16.21mm). Conclusions: The results favor the MLP architecture for prosthetic control, contributing to the development of more effective and user-friendly solutions. This research aims to enhance user quality of life and reduce prosthesis abandonment rates, envisioning a future where such control systems are integrated into prosthetic devices, allowing limb control through forearm muscle activity. Potential next steps include the assessment of the performance of these solutions in real-time, experiments on different types of electrodes, and the replication of the results observed with forearm muscles with other, more proximal muscle groups.