3-stage neural network training protocol for generalisable myoelectric control

Myoelectric control methods have undergone rapid evolution since the pre-1960s era. However, a longstanding challenge has been the variability of myoelectric signals across individuals, which impedes the development of universally applicable myoelectric control models. Researchers and companies in the field have been active in exploring various aspects such as different control strategies, pattern recognition methods, signal processing, and decoding. For instance, Meta recently reported a common model for a database of 6700 able-bodied participants. Development of such datasets with people with limb difference, in the higher education sector is unrealistic. But what we believe could be helpful is a scheme to guide researchers in addressing different stages of the process, with the aim of collectively developing a general-purpose, pre-trained, and generalisable myoelectric model. In this paper, we propose a 3-stage neural network training paradigm. Experiments were conducted with able-bodied participants to demonstrate the significance and necessity of each stage in the proposed scheme. Work is in progress to further enhance and verify the method. We aim to share this approach at MEC to receive feedback and invite collaborations for standardising data collection and pulling together our resources.