Physiology-aware latent shift adaptation for sEMG-based hand kinematics decoding across able-bodied and amputee...

Objective. Decoding continuous hand kinematics from surface electromyography (sEMG) across able-bodied and amputee populations remains a significant challenge, as amputation induces systematic alterations in neural drive and muscle activation patterns that reshape the generative relationship between motor intentions and observed sEMG signals. Such physiology-driven changes are not explicitly modeled by purely statistical domain alignment approaches, limiting their ability to achieve robust and interpretable cross-population decoding.Approach. We develop an integrated approach, termed physiology-aware latent shift adaptation (PALSA), which combines biophysical modeling with latent-space domain adaptation to characterize cross-population discrepancies at the level of latent neuromuscular representation. PALSA first learns a multidimensional, physiologically interpretable synergy latent space from able-bodied subjects (Ninapro DB2) by integrating a differentiable, motor-unit-based physiology layer. For amputee adaptation (Ninapro DB3), PALSA freezes both the neural encoder and kinematic regressor while introducing a compact, learnable latent shift vector to compensate for population-specific deviations within the learned neuromuscular manifold.Main results. Since Ninapro DB3 does not provide paired kinematics, we evaluate population-level plausibility rather than subject-specific accuracy, using DB9-derived calibrated joint angles as a standardized kinematic reference. PALSA outperforms the classic UDA algorithms CORAL and DANN in transfer learning for motor intention recognition of amputees while effectively characterizing the physiological differences between the two population groups. Across 40 gestures, PALSA achieves a mean cosine similarity of 0.6238 and a mean joint-angle root mean square error of 22.60∘with respect to this reference, consistently demonstrating higher performance for extrinsic-dominant (macro) movements than for intrinsic-dependent (micro) gestures.Significance. By integrating neuromuscular modeling priors with deep transfer learning algorithms, this paper proposes a biologically meaningful and statistically robust solution for hand motion intention decoding targeted at amputees in cross-population scenarios.