Artificial Neural Networks for Estimating Postural Risk Levels During Work Task Cycles

Musculoskeletal disorders (MSDs) are a common occupational health concern caused by poor posture that reduces productivity. Traditional postural assessments rely on expert evaluation, which is time consuming and fails to capture the complexity of human movement, limiting its real–time applicability. This study proposes a data–driven approach to 3D–postural risk assessment using inertial measurement unit (IMU) sensors and eight Multilayer Perceptron Neural Network (MLPNN) models. A forecasting model is developed for real–time risk prediction. Sixty–nine participants performed repetitive tasks. Using MATLAB® scripts, the joint angle (JA) in all three planes was calculated for the upper-body segments. MLPNN models were developed for postural risk level classification using SPSS®. The performance of the MLPNN models was evaluated using precision, accuracy, F1–score, and recall, both per class and as macro–averages. Agreement between the classified and observed postural risk levels was evaluated using Weighted Cohen’s kappa (WCK), sign test, and Bland–Altman analysis. The models performed well, particularly in flexion–extension postures. Rotational movements such as pronation–supination showed low performance. The WCK values were consistently high (>0.98), indicating a strong agreement. The sign test revealed significant differences (p<0.05), with a high proportion of ties. The Bland–Altman analysis revealed minimal bias and narrow limits of agreement. These results support the feasibility of using MLPNN for real–time postural risk assessment and forecasting. Future studies should address class imbalances and explore regularization techniques using diverse datasets. Integrating these findings with predictive models, real–time feedback, and multisensor fusion can improve workplace interventions and support occupational health monitoring.