Robust and reliable hand gesture recognition for myoelectric control

Surface Electromyography (sEMG) is a physiological signal to record the electrical activity of muscles by electrodes applied to the skin. In the context of Muscle Computer Interaction (MCI), systems are controlled by transforming myoelectric signals into interaction commands that convey the intent of user movement, mostly for rehabilitation purposes. Taking the myoeletric hand prosthetic control as an example, using sEMG recorded from the remaining muscles of the stump can be considered as the most natural way for amputees who lose their limbs to perform activities of daily living with the aid of prostheses. Although the earliest myoelectric control research can date back to the 1950s, there still exist considerable challenges to address the significant gap between academic research and industrial applications. Most recently, pattern recognition-based control is being developed rapidly to improve the dexterity of myoelectric prosthetic devices due to the recent development of machine learning and deep learning techniques. It is clear that the performance of Hand Gesture Recognition (HGR) plays an essential role in pattern recognition-based control systems. However, in reality, the tremendous success in achieving very high sEMG-based HGR accuracy (≥ 90%) reported in scientific articles produced only limited clinical or commercial impact. As many have reported, its real-time performance tends to degrade significantly as a result of many confounding factors, such as electrode shift, sweating, fatigue, and day-to-day variation. The main interest of the present thesis is, therefore, to improve the robustness of sEMG-based HGR by taking advantage of the most recent advanced deep learning techniques to address several practical concerns. Furthermore, the challenge of this research problem has been reinforced by only considering using raw sparse multichannel sEMG signals as input. Firstly, a framework for designing an uncertainty-aware sEMG-based hand gesture classifier is proposed. Applying it allows us to quickly build a model with the ability to make its inference along with explainable quantified multidimensional uncertainties. This addresses the black-box concern of the HGR process directly. Secondly, to fill the gap of lacking consensus on the definition of model reliability in this field, a proper definition of model reliability is proposed. Based on it, reliability analysis can be performed as a new dimension of evaluation to help select the best model without relying only on classification accuracy. Our extensive experimental results have shown the efficiency of the proposed reliability analysis, which encourages researchers to use it as a supplementary tool for model evaluation. Next, an uncertainty-aware model is designed based on the proposed framework to address the low robustness of hand grasp recognition. This offers an opportunity to investigate whether reliable models can achieve robust performance. The results show that the proposed model can improve the long-term robustness of hand grasp recognition by rejecting highly uncertain predictions. Finally, a simple but effective normalisation approach is proposed to improve the robustness of inter-subject HGR, thus addressing the clinical challenge of having only a limited amount of data from any individual. The comparison results show that better performance can be obtained by it compared to a state-of-the-art (SoA) transfer learning method when only one training cycle is available. In summary, this study presents promising methods to pursue an accurate, robust, and reliable classifier, which is the overarching goal for sEMG-based HGR. The direction for future work would be the inclusion of these in real-time myoelectric control applications.