Deep-Learning Based Lumbar Moment Estimation During Exosuit Augmented Lifting with Variable Loading Conditions

Wearable robotic devices have shown promise in aiding the mitigation of lower back injuries by reducing strain on muscles within the posterior chain, predominantly the erector spinae. Being a determinant of muscle strain, lumbar moments represent valuable measurements in assessing the efficacy of such devices and could further provide a more granular control input than kinematically derived heuristics. To date, computing lumbar moments is a cumbersome process, largely due to the time-intensive setup and processing requirements associated with optical motion capture (OMC) based inverse dynamics. Despite recent advances in wearable sensor-based alternatives, these limitations complicate studies that investigate real-time assistance adaptations to variations in task or loading conditions, which could ultimately provide valuable insight into how differences in control strategies affect spine kinetics and injury risk. Here, we explore the potential of using body-worn, inertial measurement units in combination with lab-integrated force plates, instead of a fully OMC based approach to estimate lumbar moments within participants. To this end, we examine two deep learning architectures, a baseline fully connected neural network (FCNN) and a long-short-term memory (LSTM) network, particularly suited for capturing temporal dependencies within the input data. We validated our approach on experiment conditions and external loads that were not present within the training set. Both models achieved high accuracy (1.58 ± 1.02 Nm RMSE) and excellent correlation (r = 0.95 ± 0.06) with OMC-based lumbar moment estimates.