AI-driven correction of video-derived gait kinematics in healthy older adults.

Abstract

Video-based gait analysis offers a cost-effective and scalable method for remotely assessing human walking patterns. However, issues such as noise and joint misidentification in the video-based gait analysis severely reduce its precision, particularly during perturbed walking (PW) in older adults. To address this challenge, this study developed and validated an automated correction method to enhance the accuracy of the video-based approach in gait analysis. Perturbed (slip) and unperturbed walking (UPW) trials were collected from 307 healthy older adults for model development, and another 126 healthy older adults were used for evaluation. Joint kinematics (segment angles) were calculated and compared across three methods: gold standard methods using data from motion capture system (Mocap), conventional video-based gait analysis (Video), and video-based gait analysis with proposed gait correction model (Model). Intra-class correlations (ICC) and RMSE were used to assess the agreement of Video-angle and Model-angle compared to Mocap-angle. Results indicated that Video-angle alone showed moderate reliability (ICC 0.66-0.77) during PW and good reliability (ICC 0.77-0.85) during UPW compared to Mocap-angle. The developed automated correction model could significantly improve the ICC by an average of 0.082 for both trial types, leading to good to excellent reliability (ICC 0.73-0.85 for PW and 0.87-0.92 for UPW). This study provides insights into optimizing video-based gait analysis protocols and proposes potential solutions to enhance accuracy and applicability for assessing mobility and fall risk in older populations.