Enhancing Myocontrol of the Hannes Prosthetic Hand with Continual Learning to Tackle Data Distribution Shift

In the field of prosthetics, researchers have recognized the potential of utilizing the remaining muscles in the residual limb, which can contract and provide valuable information for predicting human intentions. Electromyography (EMG) is currently the most widely adopted technique for extracting such information, which can then be employed to control prosthetic devices. While Machine Learning (ML) algorithms are frequently utilized for EMG signal analysis (e.g., diagnostics), their application in the realm of prosthetics is largely limited to laboratory settings. This limitation arises from the dynamic nature of EMG signals, which continuously evolve over time due to many factors such as muscle fatigue, sensor position shifts, and sweating [1, 2]. These phenomena introduce a significant data distribution shift, which may degrade the ML model accuracy and thus negatively affect prosthesis usage. Traditional batch ML algorithms excel when the testing data distribution closely aligns with the training dataset, a scenario rarely encountered in myoelectric control. Consequently, these algorithms require periodic retraining, which implies storing large amounts of data and long training times. These are incompatible with the time and memory constraints of the real-world application [3]. To address this issue, this thesis explores Continual Learning (CL) algorithms capable of adapting learning models online on a stream of incoming data to enhance prosthesis performance during multi-day usage. This class of methods represents an active research frontier, focusing on sequentially updating models while mitigating the detrimental effects of catastrophic forgetting [4]. Catastrophic forgetting indicates the decrease in performance of a neural network when new tasks or classes are introduced and the data distribution shifts in time. The contributions of this thesis are outlined as follows: 1. Highlighting the Need for CL: Experimental investigation of the significance of CL as an effective solution to address the challenges posed by data distribution shifts and catastrophic forgetting in prosthetic control. 2. Exploring Normalization Strategies: Highlighting the impact of various normalization strategies on CL for myoelectric control, empirically analyzing their effectiveness in enhancing the performance of the CL model. 3. Comparative Analysis of CL Algorithms: Comprehensive evaluation of a representative selection of CL algorithms. The performances are assessed using established CL metrics such as average accuracy (ACC), backward transfer (BWT), and forward transfer (FWT) [5]. This thesis is part of an ongoing collaborative project between the Istituto Italiano di Tecnologia (IIT) and Politecnico di Torino aimed at improving the performance of the Hannes prosthesis. Hannes is a cutting-edge multi-joint prosthetic system, encompassing a highly articulated hand with three active joints that facilitate wrist pronation/supination, wrist flexion/extension, and hand opening and closing [6]. In summary, this thesis aims to bridge the gap between ML for prosthetic research and practical applications by harnessing CL to improve the accuracy, flexibility, and adaptability of real-world myoelectric prosthetic systems.