Phase-Guided Self-Supervised Latent Representation Learning for sEMG-Based Gait Decoding.

Surface electromyography (sEMG) provides a non-invasive measure of neuromuscular activity and is widely used for gait decoding in wearable and rehabilitation systems. Existing sEMG-based approaches rely on end-to-end supervised regression with dense kinematic labels, limiting scalability, generalization, and interpretability. This paper proposes a phase-guided latent representation learning framework for multichannel sEMG signals. The framework reduces reliance on continuous joint-angle supervision during encoder training by constructing a continuous phase reference from sparse gait-event annotations. A Transformer-based phase estimator learns a circular phase embedding from sEMG, which is then frozen to condition feature encoding throughout the gait cycle. An anatomically informed graph constrains spatial interactions among sEMG channels, while reconstruction and latent-dynamics consistency objectives promote informative and temporally coherent representations. The learned representations are evaluated through post-hoc knee-angle decoding using a frozen encoder and lightweight probe. Our framework achieved post-hoc knee-angle decoding performance of $R^{2} = 0.93$ under small-sample settings and cross-subject generalization of $R^{2} = 0.88$ on the level-ground walking subset of a public dataset. Ablation experiments supported the contributions of phase conditioning, anatomical graph constraints, and latent-dynamics consistency. Phase-dependent channel analysis revealed activation patterns consistent with biomechanical principles, demonstrating potential for robot-assisted rehabilitation.