Selecting the appropriate input variables in a regression approach to estimate actively generated muscle moments around L5/S1 for exoskeleton control

Abstract

Back support exoskeletons are designed to prevent work-related low-back pain by reducing mechanical loading. For actuated exoskeletons, support based on moments actively produced by the trunk muscles appears a viable approach. The moment can be estimated by a biomechanical model. However, one of the main challenges here is the feasibility of recording the required input variables (kinematics, EMG data, ground reaction forces) to run the model. The aim of this study was to evaluate how accurate different selections of input variables can estimate actively generated moments around L5/S1. Different multivariate regression analyses were performed using a dataset consisting of spinal load, body kinematics and trunk muscle activation levels during different lifting conditions with and without an exoskeleton. The accuracy of the resulting models depended on the number and type of input variables and the regression model order. The current study suggests that third-order polynomial regression of EMG signals of one or two bilateral back muscle pairs together with exoskeleton trunk and hip angle suffices to accurately estimate the actively generated muscle moment around L5/S1, and thereby design a proper control system for back support exoskeletons.