Searching for Optimal EMG Latent Subspace With Discriminant DoF-Wise Distributions for Subject-Generic Model

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Surface electromyography (sEMG) is vital for neural interfaces and human-machine interaction.
• Subject-generic models for sEMG gesture recognition face challenges due to inter-subject variability and overlapping muscle activation patterns.
• Existing methods struggle to effectively disentangle features related to degrees of freedom (DoFs) and individual subjects.

Purpose of the Study:

• To develop a subject-generic model for robust sEMG-based hand gesture recognition.
• To disentangle sEMG features into subject-invariant and DoF-specific latent subspaces.
• To improve the accuracy and generalizability of gesture recognition across different users and days.

Main Methods:

• Introduction of a multi-branch autoencoder (AE) architecture.
• Disentanglement of sEMG features into a DoF-specific (subject-invariant) latent space and a subject-specific (DoF-invariant) latent space.
• Systematic comparison against established methods like PCA, KPCA, LDA, KDA, conventional AE, CCA, and SRDA.

Main Results:

• The multi-branch AE significantly improved Degrees of Freedom (DoF) discrimination.
• The proposed method demonstrated superior subject invariance compared to baseline methods.
• Consistently higher inter-subject classification accuracy was achieved across various classifiers.

Conclusions:

• The multi-branch AE architecture offers a promising approach for robust, user-independent sEMG-based gesture recognition.
• Disentangling features into subject-invariant and DoF-specific subspaces is key to overcoming inter-subject variability.
• This method has significant potential for advancing neural interfaces and human-machine interaction systems.