Interpretable Biomechanical Feature Selection for VR Exercise Assessment Using SHAP and LDA.

Virtual reality (VR) technologies are increasingly applied in rehabilitation, offering interactive physical and spatial exercises. A major challenge remains the objective assessment of human movement quality (HMQA). This study aimed to identify biomechanical features differentiating correct and incorrect execution of a lateral lunge and to determine the minimal number of sensors required for reliable VR-based motion analysis, prioritising interpretability. Thirty-two healthy adults (mean age: 26.4 ± 8.5 years) performed 211 repetitions recorded with the HTC Vive Tracker system (7 sensors + headset). Repetitions were classified by a physiotherapist using video observation and predefined criteria. The analysis included joint angles, angular velocities and accelerations, and Euclidean distances between 28 sensor pairs, evaluated with Linear Discriminant Analysis (LDA) and SHapley Additive exPlanations (SHAP). Angular features achieved higher LDA performance (F1 = 0.89) than distance-based features (F1 = 0.78), which proved more stable and less sensitive to calibration errors. Comparison of SHAP and LDA showed high agreement in identifying key features, including hip flexion, knee rotation acceleration, and spatial relations between headset and foot or shank sensors. The findings indicate that simplified sensor configurations may provide reliable diagnostic information, highlighting opportunities for interpretable VR-based rehabilitation systems in home and clinical settings.