Enhanced robustness of myoelectric pattern recognition to across-day variation through invariant feature extraction

Robust pattern recognition is critical for myoelectric prosthesis (MP) developed in the laboratory to be used in real life. This study focuses on the robustness of MP control during the usage across many days. Due to the variability inhered in extended electromyography (EMG) signals, the distribution of EMG features extracted from several days' data may have large intra-class scatter. However, as the subjects perform the same motion type in different days, we hypothesize there exist some invariant characteristics in the EMG features. Therefore, give a set of training data from several days, it is possible to find an invariant component in them. To this end, an invariant feature extraction (IFE) framework based on kernel fisher discriminant analysis is proposed. A desired transformation, which minimizes the intra-class (within a motion type) scatter meanwhile maximizes the inter-class (between different motion types) scatter, is found. Five intact-limbed subjects and three transradial-amputee subjects participated in an experiment lasting ten days. The results show that the generalization ability of the classifier trained on previous days to the unseen testing days can be improved by IFE. IFE significantly outperforms Baseline (original input feature) in classification accuracy, both for intact-limbed subjects and amputee subjects (average 88.97% vs. 91.20% and 85.09% vs. 88.22%, p <; 0.05).