An artificial neural network to estimate detailed active-passive spinal loads during static lifting activities for a...

Numerous musculoskeletal models have been developed to estimate spine loads for use in occupational risk assessment programs. These models typically oversimplify the complex intervertebral joints by representing them as pivots, thereby failing to estimate load sharing among the intervertebral discs, facet joints, ligaments and disc annulus/nucleus. Our coupled musculoskeletal-finite-element model integrates both the musculature and detailed geometry of the passive spine to yield comprehensive predictions for a standardized anthropometry. However, this model is time/skill-intensive. We aimed to train an artificial neural network (ANN) to replicate the outputs of this model during sagittally-symmetric static lifting tasks. A training dataset was generated by simulating 500 tasks using the model with different hand-loads (up to 20 kg) and positions (up to 30 cm horizontal distance from the shoulders), trunk flexions (up to 80°), and lumbopelvic ratios (3 levels for each flexion angle). The outputs included T12-S1 intradiscal pressure, maximum annulus principal stress, disc compressive/shear loads, resultant facet joint forces, resultant force of all posterior ligaments, and force in trunk muscles. The ANN successfully mapped the model inputs to its 83 outputs with satisfactory predictive accuracy; normalized-root-mean-squared-errors < 3 % and R² > 0.98. Bland-Altman analyses indicated no consistent bias between the ANN and model predictions, with a narrow 95 % limits of agreement across all outputs, e.g., 0.1 MPa for L4-L5 intradiscal pressure. This ANN is a fast practical tool for risk assessment applications. It can also be used to provide realistic trunk muscle-driven loading scenarios for passive finite element spine models in biomechanical evaluations.