Fatigue assessment in multi-activity manual handling tasks through joint angle monitoring with wearable sensors

• Using joint motion and coordination data in an FFNN showed reasonable accuracy in detecting performance fatigue. • Performance fatigue resulted in reduced coordination of adjacent joints, assessed by mutual information. • Comparing end-to-end and feature-engineering models showed motion and coordination data’s potential in fatigue detection. • Segmenting manual handling tasks revealed activity-related joint behavior and its importance for fatigue analysis. Performance fatigue is a primary contributor to work-related musculoskeletal disorders and understanding its impact during manual handling tasks (MHT) is crucial to preventing such issues. This study evaluated fatigue during prolonged MHT by analyzing body joint angles kinematics and their coordinative variability using readouts from inertial measurement units (IMUs). Eight individuals participated in the MHT experiment, continuously reporting their fatigue levels. The MHT was further segmented into repetitive activities of lifting, carrying, and lowering, and kinematic metrics (average, maximum joint excursions, variability) were extracted from the body joint angles in the sagittal plane. During lifting and lowering repetitions, mean and peak joint angles increased with fatigue levels across all joints except the knee, where both decreased, with average Spearman’s ρ values of −0.24 and −0.16 during lifting, respectively. Furthermore, as fatigue progressed, coordination among adjacent joints decreased, indicated by reduced information transmission measured by mutual information theory. Particularly, the knee-hip mutual information during carrying activity decreased with fatigue (average correlation coefficient: −0.47). Finally, using the proposed features, a feed-forward neural network model achieved a subject-independent accuracy of 66 % in detecting five stages of perceived fatigue. Comparatively, a multi-head convolutional neural networks and long-short-term memory networks using the normalized raw joint angle data achieved 74 % accuracy while requiring significantly greater data and computational resources. These findings provide insights into how fatigue affects joint kinematics and coordination, enhancing our understanding of fatigue-related risk of work-related musculoskeletal disorders. Further investigations are needed to characterize such risks.