Estimation of dynamic spinal loads during manual lifting using smartphone-based markerless motion capture

This study evaluated the feasibility of dynamic spinal load estimation using kinematic data collected with a smartphone-based markerless motion capture system. Twenty-one participants performed various lifting tasks while marker-based and markerless kinematics were recorded simultaneously. Marker-based kinematics were used to drive a conventional optimization-based musculoskeletal (MSK) modeling workflow to obtain reference estimates of lumbosacral (L5/S1) compression and shear forces in OpenSim. Two markerless-driven approaches were then developed and evaluated against the conventional approach: (1) a markerless-driven MSK modeling workflow and (2) a machine learning model that uses intermediate outputs from MSK modeling (e.g., joint angles and moments) to directly predict L5/S1 compression and shear forces, removing the need for the time-consuming optimization step. Both approaches estimated compression forces with reasonable accuracy relative to reference values, with normalized root mean square error (nRMSE) of 12% for markerless-driven MSK modeling and 9% for the ML-based approach. However, larger errors were observed for shear force estimation, especially with markerless-driven MSK modeling. Overall, the findings support the feasibility of spinal load estimation during manual lifting using smartphone-based markerless kinematics and highlight the advantage of integrating data-driven predictive algorithms with conventional physics-based MSK modeling. Such a hybrid modeling framework, combined with the markerless system, can potentially enable fast, low-cost, and reasonably accurate spinal load estimation in field settings. Future studies should focus on improving the kinematic accuracy of the markerless system and enhancing the generalizability and accuracy of the ML model, especially for shear force prediction. • Demonstrated the feasibility of spinal load estimation using smartphone video-based kinematics. • Markerless-driven musculoskeletal modeling and the machine learning model both estimated L5/S1 compression force with reasonable accuracy. • Errors were higher for shear force estimation. • Markerless motion capture combined with data-driven predictive algorithms has the potential to enable rapid, low-cost, and accurate estimation of low back biomechanical loads in the field.