Feature Significance and Generalizability of Myoelectric Hand Gesture Recognition Under Varying Limb Positions

Myoelectric hand gesture recognition is an effective way to control prosthetic hands. Myoelectric hand gesture recognition often uses Linear Discriminant Analysis (LDA), and its effectiveness heavily depends on the selection of a feature set. For this reason, we investigate the relationship between the choice of a feature set and LDA generalizability, especially under disturbances caused by changes in arm position. Two approaches for assessing feature significance were considered: one model-agnostic approach based on the statistical theory of ANOVA f-statistic and Pearson correlation, called Maximum Relevance Minimum Redundancy, and one LDA-specific approach using the concept of principal angles between subspaces. Our analysis of a large feature set of 14 features indicates that LDA is highly susceptible to overfitting to features with low discriminative power. With an optimized set of 8 features, recognition error on unseen arm positions was reduced by 2 % compared to the set of 14 features. Our results underscore the importance of proper feature selection for optimal myoelectric hand gesture recognition performance.