Selective Correction of Condition-Sensitive Parameters for Cross-Condition Torque Prediction in an EMG-Driven Elbow Musculoskeletal Model

Electromyography-driven musculoskeletal models estimate muscle forces and joint torque. A shared parameter set can leave errors that vary with movement speed and load. This study examined selective condition correction in an EMG-driven elbow model. The model retained a common parameter set and added speed- and load-dependent terms only to selected parameters. Ten healthy participants performed constrained isokinetic elbow flexion-extension. Subject-specific MuJoCo models were constructed from OpenSim arm26. Speed-Load had the lowest mean validation RMSE among three calibration combinations (1.37 N·m) and was used for common calibration. Length-range constraints were included during fitting. Full reduced mean RMSE by 10.0-18.6% across held-out repetitions, interpolation and extrapolation conditions. The largest reduction occurred in extrapolation: RMSE decreased from 2.08 to 1.70 N·m, while mean R² increased from 0.71 to 0.81. The ablation models had intermediate mean errors. The full two-stage procedure improved within-subject torque prediction across the tested speed and target-torque conditions with a limited number of additional coefficients.