OpenSim–Umberger-Based Metabolic Power Stratification During the Sit-to-Walk Transition Using Interpretable Ensemble Learning

Quantifying metabolic cost during short transitional movements is challenging because conventional metabolic measurements have limited temporal resolution. This proof-of-concept study examined whether model-derived metabolic cost during the sit-to-walk (STW) transition could be exploratorily stratified using interpretable ensemble learning. Forty-nine healthy adults completed the STW phase of the Timed Up and Go task with synchronized three-dimensional kinematics, ground reaction forces, and eight-channel surface electromyography. Individually scaled OpenSim gait2392 models and the Umberger metabolic model were used to estimate metabolic power from seat-off to the end of the first complete gait cycle. Window-averaged metabolic power was stratified into low-, medium-, and high-cost levels. Window-level biomechanical features were extracted from kinematic, kinetic, and muscle-state time series. Seven classifiers were trained using a subject-level 7:3 train-test split and stratified five-fold cross-validation within the training set, and their probability outputs were integrated through TOPSIS-weighted classifier fusion. SHapley Additive exPlanations were used for class-specific feature attribution. The fused ensemble achieved an AUC of 0.870, F1 score of 0.703, accuracy of 0.705, and specificity of 0.853 on the independent test set. Discrimination was stronger for the low- and high-cost levels than for the medium-cost level. SHAP-based attribution highlighted force-related changes and knee-angle variability and amplitude measures as prediction-relevant biomechanical features. These findings support a model-derived, interpretable workflow for extending STW assessment from task performance to task cost, while indicating the need for further validation in larger and clinical datasets.