Publication: Deep neural network (DNN) for Predicting spinal loads using video Based motion capture.
No Thumbnail Available
Date
2025-04
Authors
Jeevinthiran A/L Karunagaran
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Spine loads are often used as a measure to evaluate the risk of injury to the lower back. Obtaining spine loads typically involves musculoskeletal modelling techniques with data inputs from full body motion capture and ground reaction forces (GRFs).Despite promising advances in data-driven biomechanics, current methods remain constrained by reliance on intermediate musculoskeletal models or task-specific regression equations. No end-to-end deep neural network (DNN) has yet been developed to directly map vision-based motion capture and GRF inputs to spine loads.This study evaluated the validity of a DNN in predicting the L5/S1 axial compression and anterior-posterior (AP) shear force using kinematics from a marker less video based open-source motion capture tool (OpenCap) and GRFs. A data collection experiment with 12 participants comprising of students performing basic lifting (with dumbbells), isometric pushing, and isometric pulling tasks were carried out. The DNN model was trained on reference spine load values derived from the AnyBody Modelling system.
Model validity was assessed on a previously unseen data subset. It effectively replicates the waveform of the spine loads across time-continuous data segments for many of the tasks.The model shows robust performance, achieving coefficient of determination (R2)values of 0.92 and 0.89 for compression and shear forces, respectively, along with root mean square errors (RMSE) of 197N for compression and 85N for shear. Bland-Altman analysis supports this finding with limits of agreement of –332 N to 420 N for axial
compression and –124 N to 182 N for AP shear; indicating minimal systematic and proportional errors in the system. NIOSH sets upper thresholds of 3400 N for compression and 1000 N for shear; subtracting the Bland–Altman lower limits (332 N and 124 N, respectively) yields effective limits of 3060 N and 870 N. Caution is advised when evaluating activities resulting in spine loads beyond these limits when using the spine load prediction model.