Multimodal Evaluation of Exoskeleton Assistance Using Optimized Machine Learning

Abstract

Repetitive manual material handling (MMH) remains a leading occupational risk factor for work-related musculoskeletal disorders (WMSD), and passive shoulder-support exoskeletons have emerged as promising assistive technologies. Most evaluations rely on group-level statistics from single sensor modalities, leaving open whether their effects can be detected at the individual level. Thirty-five adults performed low-risk (4.54 kg, Lift Index (LI) < 1.0) lifting tasks, derived from the Revised NIOSH Lifting Equation, with and without the passive shoulder exoskeleton support. Surface electromyography (sEMG), inertial measurement unit (IMU) kinematics, and NASA Task Load Index (NASA-TLX) ratings were derived into a 34 feature vector and classified using four models under Leave-One-Subject-Out Cross Validation to ensure our models were not trained on already seen data. SVM with radial basis kernel reached 89.4% accuracy (AUC = 0.932) on low-risk lifting, Logistic Regression reached 84.8% (AUC = 0.904) on high-risk lifting. NASATLX physical demand proved to be the most important feature in permutation importance, outranking all biomechanical features. Perceived workload may be a more consistent cross-subject signal than objective biomechanics, with practical implications for low-burden field evaluation.