Towards Real-Time Minimum-Input Prediction of Lumbar Moment Based on Flexible Sensors and Machine Learning

Abstract

Knowledge of low-back loading is essential for understanding and mitigating the risk of low-back overexertion injuries. Conventional data acquisition methods for estimating joint loading are limited to laboratory settings, whereas wearable sensors can provide a mobile and cost-effective alternative. This study investigated the feasibility of learning prediction of L5S1 flexion moment based on kinematics and electromyography (EMG) measurements from flexible sensors. Four machine learning methods were compared, and different subsets of sensor inputs were explored. Results indicated that the support vector machine (SVM) method outperformed others, and a subset of four out of seven sensor locations, namely sacrum, thigh, shank, and thoracic erector spinae, yielded the best predictive accuracy. The study demonstrates that machine learning can unlock the potential of mobile miniaturized flexible sensors in field biomechanics or ergonomics studies.