Accurate estimation of hip range of motion using MediaPipe and inertial sensors with machine learning models.

Accurate, objective assessment of hip joint range of motion (ROM) is essential for orthopedic diagnosis and rehabilitation. Conventional tools, such as goniometers, are limited by subjectivity, inter-observer variability, and poor compatibility with telemedicine applications. This study aimed to develop and validate a markerless, video-based system for estimating hip joint ROM by integrating human pose estimation (MediaPipe) with machine-learning models using inertial sensor data as the reference standard. Twenty healthy adult males performed hip flexion/extension, abduction/adduction, and internal/external rotation movements. Skeletal coordinates extracted from the videos were converted into geometric features and used to train five regression models (linear regression, ElasticNet, support vector regression, random forest, and LightGBM). Model performance was evaluated using the coefficient of determination (R²) and mean absolute error (MAE), and Shapley additive explanations (SHAP) were used to interpret feature contributions. Among the evaluated models, tree-based methods showed high predictive accuracy, and LightGBM was used for subsequent interpretability analyses across all directions of hip ROM (R² up to 0.94; MAE 4-6°), while inertial sensor validation confirmed high measurement accuracy of the reference data. SHAP analysis revealed that distinct geometric descriptors dominated flexion/extension, abduction/adduction, and internal/external rotation, indicating direction-specific biomechanical determinants of hip motion. These findings demonstrate that markerless, video-based estimation of hip ROM is feasible with promising accuracy and highlight the potential of this approach for telemedicine, remote rehabilitation monitoring, and future integration into gait and musculoskeletal modeling workflows.