Artificial neural networks for predicting ground reaction forces, feet centers of pressure, spine loads, and trunk muscle forces during load-handling activities.

Excessive spinal loading is a key contributor to occupational musculoskeletal injuries. Numerous biomechanical models, ranging from simple to highly sophisticated, are available to evaluate these loads. While simple models often lack accuracy, advanced ones are too complex for practical uses. We aim to develop a set of easy-to-use artificial neural networks (ANNs) that replicate the outputs of a detailed musculoskeletal model, i.e., AnyBody Modeling System (AMS). Kinematics data were collected from 20 participants each performing 204 load-handling tasks, combined with 9 hand-load magnitudes, resulting in 36,720 simulated tasks in AMS. The ANNs were trained to map inputs (hand-load position and magnitude, body weight and height, lifting-technique, and handling-technique) to AMS outputs (T12-S1 spine loads, trunk muscle forces, ground reaction forces (GRFs), and feet centers of pressure (CoPs)). The ANNs demonstrated good-to-excellent predictive performance, with normalized-root-mean-squared-errors (nRMSEs) ranging from 2.6% to 11.2% and (R2) between 0.59 and 0.97. The ANNs for compressive and anteroposterior shear loads, vertical GRFs and CoPs (2.6% < nRMSEs < 6.4% and 0.88 < R2 < 0.97) outperformed those for horizontal GRFs and mediolateral shear loads (7.2% < nRMSEs < 9.9% and 0.59 < R2 < 0.68). Among the predicted muscle forces, the erector spinae exhibited the lowest error (nRMSE = 6.8%, R2 = 0.88), followed by the external oblique (nRMSE = 8.0%, R2 = 0.84) and the multifidus (nRMSE = 8.2%, R2 = 0.81). External validation on two unseen subjects demonstrated good generalization of the proposed ANNs, with strong agreement for CoPs, vertical GRFs, and lumbar anteroposterior shear loads (R2 ≈ 0.86-0.93; nRMSE ≈ 4.8-7.1%), while acceptable performance was observed for major trunk muscle forces (R2 ≈ 0.57-0.77). These findings suggest that ANNs offer promising alternatives to complex biomechanical simulations, potentially facilitating their integration into occupational risk assessment frameworks.