Summary Each year, many workers fall victim to back pain, with the lower back accounting for the great majority of cases. This type of musculoskeletal disorder is usually caused by excessive effort, especially lifting, which causes loading on the spinal tissue. As there is no method for directly measuring these loads in the workplace, they must be estimated through biomechanical models. However, the models used for this purpose so far have been inadequate. This project is a continuation of several studies subsidized by the Institut de recherche Robert-Sauvé en santé et en sécurité du travail (IRSST) and aimed at developing an ambulatory method for continuously estimating low-back loading, i.e., the moment at the lumbar-sacral joint (L5/S1). The proposed approach aims to integrate the electromyographic (EMG) measurement of a minimum number of superficial back muscles with back kinematics obtained from two inertial sensors in order to estimate the low-back load. The idea in itself is not new, but the way of arriving at it – specifically, the use of dynamic calibration to establish the EMG and kinematic relationship to the L5/S1 moment – is innovative. There were essentially three stages to this project. Stage 1 consisted in studying the feasibility of the proposed approach by comparing its estimates with those determined by the validated 3D model normally used in the laboratory. In Stage 2, the method for calibrating the ambulatory model was developed and validated. In Stage 3, the ambulatory model was validated in a load handling situation similar to workplace conditions. The Stage 1 results show that very acceptable estimates of the L5/S1 moment (error of about 10%) for asymmetric handling tasks can be obtained from a limited number of trunk kinematic variables and EMG signals from a limited number of muscles (maximum six), either through a multiple linear regression approach or through an artificial neuron network. The proposed approach therefore shows good potential. The aim in Stage 2 was to develop a methodology for calibrating the EMG-kinematic-moment relationship in the workplace. A simplified model capable of estimating L5/S1 moments had to be developed and validated. This simplified model has only five segments—trunk (including neck and head), two arm segments and two forearm segments—requiring six inertial sensors to measure their orientation. It also requires the use of an instrumented box with handles to measure the forces on the hands and an inertial sensor to measure its orientation. The results show that the model developed can estimate L5/S1 moments with an error of less than 10%, which is very satisfactory for the usage considered. This model can therefore be used to calibrate the EMG-kinematic-moment relationship in the workplace. The purpose of the third and last stage was to validate the ambulatory model, using the calibrating model developed in Stage 2 and the artificial neuron network designed in Stage 1. Conditions similar to an actual use context over a handling period long enough to cause muscle fatigue were simulated. Validation of the ambulatory approach to estimate L5/S1 moments by measuring only trunk kinematics with two inertial sensors and EMG readings from six muscles yielded mixed results, explaining only an average 50% of the variance in L5/S1 moments estimated by a criteria model. The fact that some subjects’ results reached performance levels able to explain 70% of the variance is nevertheless highly promising. Modifications to the neuron network could be explored in order to improve the robustness of the predictions between individuals. It would be worth undertaking further studies to better identify the optimum conditions for using this innovative approach.