Generalization of Knee Joint Moment Prediction During Drop Vertical Jumps Under Graded Visuo-Proprioceptive Conflict: The Role of Multijoint Kinematics Across Validation Frameworks

Accurate estimation of knee joint moment is important for biomechanical monitoring and injury-risk assessment, yet model generalizability under altered sensory environments remains unclear. This study evaluated a support vector regression model for predicting sagittal knee moment during the landing-takeoff cycle of the drop vertical jump (DVJ) under visuo-proprioceptive conflict and examined whether adding hip and ankle kinematics improved performance. Fourteen healthy men performed DVJs under one real and four virtual perturbation conditions with a fixed physical drop height and virtual heights of 0, 10, 30, and 50 cm. Bilateral surface electromyography and three-dimensional lower-limb kinematics were used as inputs, and the inverse-dynamics-derived sagittal knee moment served as the target. Basic and extended feature sets were compared under leave-one-subject-out (LOSO) and leave-one-condition-out (LOCO) frameworks. Within the present experimental design, prediction performance was generally higher under LOCO than under LOSO. Adding hip and ankle kinematics improved prediction mainly under LOCO, whereas gains under LOSO were limited. Waveform similarity showed a non-monotonic decrease-then-recovery pattern across perturbation levels. Residual analysis showed no directional bias, and errors were greater during landing absorption and push-off than during flight. These findings suggest that under the present study design and in this sample, lower performance was observed under LOSO than under LOCO, and that multijoint kinematics may improve prediction robustness under cross-condition settings.