Pre-training, personalization, and self-calibration: all a neural network-based myoelectric decoder needs.

Abstract

Myoelectric control systems translate electromyographic signals (EMG) from muscles into movement intentions, allowing control over various interfaces, such as prosthetics, wearable devices, and robotics. However, a major challenge lies in enhancing the system's ability to generalize, personalize, and adapt to the high variability of EMG signals. Artificial intelligence, particularly neural networks, has shown promising decoding performance when applied to large datasets. However, highly parameterized deep neural networks usually require extensive user-specific data with ground truth labels to learn individual unique EMG patterns. However, the characteristics of the EMG signal can change significantly over time, even for the same user, leading to performance degradation during extended use. In this work, we propose an innovative three-stage neural network training scheme designed to progressively develop an adaptive workflow, improving and maintaining the network performance on 28 subjects over 2 days. Experiments demonstrate the importance and necessity of each stage in the proposed framework.