A hybrid framework for predicting fall impact forces using pose-estimated kinematics and descriptive information

Falls are the leading cause of hip fractures among older adults; however, direct measurement of hip impact force during laboratory falls is challenging due to safety constraints. Artificial intelligence (AI)-based pose estimation (e.g., OpenPose, WHAM) can extract kinematics from video, which may be used to estimate ground reaction forces. This study evaluated whether pose-derived kinematic features from video-captured falls, alone or combined with demographic and observational data (e.g., initial contact region), can accurately predict hip impact forces. We analyzed 69 sideways fall trials from 11 older adults (9 male; 62.6 ± 3.8 years) using video recordings of the falls, along with hip-worn IMUs and force plates to measure hip impact force. Pose-derived kinematic features (pose data) included hip impact velocity and acceleration extracted using OpenPose, and knee flexion angles obtained using WHAM. Four feature sets were evaluated for predicting hip impact force: demographic + observational, pose-only, demographic + observational + pose, and IMU-only. Models were trained using leave-one-participant-out cross-validation and evaluated using R² and mean absolute error (MAE). MAE differences across feature sets were assessed using a linear mixed effects model with Bonferroni-adjusted post-hoc comparisons. The combined model (demographic + observational + pose) demonstrated improved prediction accuracy (R² = 0.70; MAE = 0.35 BW) compared to the demographic + observational model (R² = 0.47; MAE = 0.48 BW), but did not reach the performance of the IMU-based model (R² = 0.89; MAE = 0.21 BW). Pose-derived kinematics provide meaningful, physics-informed features for predicting hip impact forces from video-captured falls. Integrating these features with demographic and observational data improves predictive accuracy, supporting their use in studying injury mechanisms in falls.