Physiological input design for load-limited graph surrogate modeling of knee joint stress fields: sEMG linear...

Rapid surrogate modeling of knee joint stress fields requires informative input representations, but explicit load information is often difficult to obtain reliably in practical settings. Whether physiological signals can provide complementary information for stress-field prediction when explicit load information is unavailable, and whether low-dimensional muscle synergy is more suitable than surface electromyography (sEMG), remain unclear. Nine healthy adult male participants performed running trials with synchronized kinematics, ground reaction forces, and 12-channel sEMG. Finite element simulations driven by joint posture and reaction forces were converted into fixed-topology graph samples with tissue-level von Mises stress labels. Using a shared MeshGraphNet backbone and grouped threefold cross-subject validation, six input conditions were compared: full pose-load input, pose only, pose + real sEMG, pose + real muscle synergy, and their shuffled physiological-control counterparts. Performance was evaluated using whole-field error, high-stress tail error, hotspot-region overlap, hotspot localization, and real-specific gain metrics. Full pose-load input showed the most favorable overall prediction performance, although its advantage over pose only varied across metrics. Adding sEMG linear envelopes did not improve the pose-only baseline and generally increased prediction errors. In contrast, muscle synergy showed performance generally comparable to pose only and outperformed sEMG linear envelopes across global metrics and several high-stress and hotspot-related measures. Shuffled-control analysis showed no consistent advantage of the real muscle synergy input over its shuffled counterpart; positive median gains in the selected global and hotspot-overlap metrics did not reach statistical significance after correction. The NMF-derived muscle synergy representation was more suitable than sEMG linear envelopes as a physiological auxiliary input for load-limited surrogate modeling of knee joint stress fields. These findings may inform physiological input design for rapid stress-field prediction when explicit load information is unavailable or unreliable.