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

Area of Science:

• Biomechanics
• Gerontology
• Artificial Intelligence

Background:

• Falls are a primary cause of hip fractures in older adults.
• Directly measuring hip impact force in lab settings is difficult due to safety concerns.
• AI-based pose estimation offers a potential method to analyze fall kinematics from video.

Purpose of the Study:

• To evaluate if AI-derived kinematic features from fall videos can predict hip impact forces.
• To determine if combining pose data with demographic and observational data enhances prediction accuracy.
• To compare the predictive performance of different feature sets against IMU-based measurements.

Main Methods:

• Analyzed 69 sideways fall trials from 11 older adults using video recordings.
• Extracted pose-derived kinematic features (hip impact velocity, acceleration, knee flexion) using OpenPose and WHAM.
• Trained and evaluated prediction models using leave-one-participant-out cross-validation, assessing R-squared and Mean Absolute Error (MAE).

Main Results:

• The combined model (demographic + observational + pose data) achieved higher accuracy (R²=0.70, MAE=0.35 BW) than demographic + observational data alone (R²=0.47, MAE=0.48 BW).
• The IMU-only model showed the best performance (R²=0.89, MAE=0.21 BW).
• Pose-derived kinematics significantly improved hip impact force prediction when integrated with other data.

Conclusions:

• Pose-derived kinematics from video are valuable, physics-informed features for estimating hip impact forces during falls.
• Integrating AI kinematics with demographic and observational data enhances the prediction of hip impact forces.
• This approach supports research into fall injury mechanisms in older adults.