Dataset for 'Reducing Sensor Configuration for Data-Driven Shoulder Load Estimation for Exoskeleton Control'

Abstract

This paper investigates the role of various sensors in estimating shoulder loads during manual manipulation tasks in the context of exoskeleton control. The sensors examined include textile-integrated electromyography (EMG) sensors forthe trapezius, deltoids, biceps, and forearm muscles; inertial measurement units (IMUs) on key body segments such as the pelvis, shoulder, upper arm, and forearm; and pressure-sensing insoles. The objective is to reduce the sensor configuration for predicting the internal torque exerted on the shoulder in the sagittal plane. To achieve this, a study involving nine subjects manipulating dumbbells in the sagittal plane was conducted. The protocol encompasses four static positions as well as isolated elbow and shoulder flexions. Further, four distinct machine learning model architectures were trained, systematically omitting one sensor at a time. The significance of each sensor was evaluated by assessing the impact of its omission on the predictive correlation using cross-validated R2 scores. Consequently, a top-five sensor configuration was identified and compared against configurations based solely on domain knowledge and the full sensor array. The configuration proposed in this study achieved a correlation of R2 = 0.83 in predicting shoulder loads, slightly surpassing the performance of the full sensor setup (R2 = 0.82) and outperforming the domain knowledge-based (DKB) setup (R2 = 0.63). All metrics are determined in a leave-one-subject-out cross-validation (loso-cv) training strategy.