シミュレーション由来の生体力学的特徴量を用いた解釈可能な機械学習によるハムストリングス-大腿四頭筋不均衡検出

Hamstrings-quadriceps (H-Q) imbalance represents a biomechanical marker of knee instability and injury risk in running. This in silico (simulation-based) study introduces a digital machine learning framework designed to estimate H-Q imbalance using biomechanical features derived entirely from synthetic running trials and conceptually mappable to inertial-sensor domains. Key biomechanical predictors included the dynamic hamstrings-to-quadriceps ratio (H:Qdyn), the knee moment limb symmetry index (LSI), and the early-stance co-contraction index (CCI), all standard indicators of muscular balance and coordination in sports medicine. A reduced musculoskeletal framework emulating flexor-extensor balance, limb symmetry, and co-contraction patterns generated 573 synthetic running trials for 160 virtual subjects across three speeds. These interpretable features trained a calibrated gradient-boosting classifier evaluated via ROC-AUC, PR-AUC, balanced accuracy, F1, and Brier score. Across all conditions, the model achieved ROC-AUC 0.933 (95% CI 0.908-0.958), balanced accuracy 0.943 (95% CI 0.924-0.962), PR-AUC 0.918 (95% CI 0.892-0.943), F1 0.940 (95% CI 0.919-0.958), and Brier 0.056 (95% CI 0.041-0.072), outperforming the logistic baseline. Dynamic H:Q ratio and knee moment symmetry were the dominant predictors, while co-contraction provided complementary biomechanical nuance. These results demonstrate that simulation-derived frameworks can reproduce IMU-relevant biomechanical variability, enabling interpretable machine learning for transparent assessment of muscular balance in sports medicine.