What question did this study set out to answer?

The aim is to develop a wearable system for real-time monitoring and classification of lifting postures to reduce workplace injuries.

July 2, 2026Open Access

An intelligent wearable insole system for machine learning-based detection of high-risk postures in static symmetrical load-lifting tasks

Key Points

The aim is to develop a wearable system for real-time monitoring and classification of lifting postures to reduce workplace injuries.
Implemented a two-phase protocol with 23 participants performing 36 static lifting tasks under three load conditions.
Collected a 13-dimensional feature vector including plantar-pressure readings and trunk flexion angles for classification.
Trained five machine learning models with a participant-wise train-test split, and validated the best model in real time with ten additional participants.
Logistic regression achieved 91.17% offline classification accuracy, with 93.75% sensitivity and 87.83% specificity.
Real-time validation on the insole reached 88% accuracy, correlating with estimated UTAH back compressive forces (r = 0.69).
The system enables continuous monitoring and immediate feedback for lifting posture, enhancing ergonomic risk management.

Abstract

Work-related musculoskeletal disorders (WMSDs) remain a major occupational health concern worldwide, with manual material handling, particularly lifting, being a primary contributor to low-back pain. Conventional ergonomic interventions often rely on observational assessments and lack continuous, objective, and automated monitoring in real-time conditions. This study presents an intelligent wearable insole system that integrates plantar-pressure and an inertial measurement unit with embedded machine-learning algorithms for the automated classification of low-risk/high-risk lifting postures in real-time, providing immediate visual/auditory feedback. A two-phase protocol was implemented. In Phase 1, twenty-three participants performed thirty-six static symmetric lifting tasks under three load conditions, with a 13-dimensional feature vector recorded per trial, comprising 12 force-sensitive resistor plantar-pressure readings and one inertial measurement unit-derived trunk flexion angle, and ground-truth labels assigned using the UTAH back compressive force criteria. In Phase 2, five machine learning models (logistic regression, support vector machine, k-nearest neighbors, decision tree, and random forest) were trained on the 13-dimensional feature vectors using a participant-wise 78%/22% train-test split with 5-fold cross-validation. The best-performing model was then deployed on the assembled intelligent insole and validated in real time with ten independent participants. Among the evaluated models, logistic regression demonstrated the best overall performance, achieving 91.17% offline classification accuracy, 93.75% sensitivity, 87.83% specificity, and an AUC of 0.94, with trunk flexion angle as key predictor. Upon deployment on the smart insole, real-time validation achieved 88% accuracy, with a moderate to strong correlation (r = 0.69) between classified postures and estimated UTAH back compressive forces. The proposed system enables automated, continuous monitoring of static lifting posture and timely feedback without the need for expert observation. This approach offers a practical and scalable solution for ergonomic risk assessment and prevention of lifting-related WMSDs in industrial and manufacturing environments.

KI fragen

Bookmark

View Full Paper