Accurately estimating physical fatigue is critical for preventing injuries, enhancing performance and supporting rehabilitation, yet current approaches are often limited to specific tasks with no interpretability. In this work, we present a systematic framework for predicting perceived fatigue in upper limb movements using wearable inertial measurement units (IMU) and surface electromyography (EMG). We processed two publicly available datasets into structured repetition level data with extracting over 3,700 biomechanical features across time, frequency and time–frequency domains. We model perceived fatigue using the Borg Rating of Perceived Exertion (RPE) scale. A six-level hierarchical evaluation was designed to progressively test models’ performance and generalisation, from single-task conditions to cross-family transfer. The results show that the models achieve the highest accuracy in task-specific contexts (R2 to 0.69), but the performance declines with increasing task diversity, with cross-family transfer resulting in a negative R2. Fewshot domain adaptation significantly improved performance by turning negative R2 into positive values across all family combinations. Feature importance analyses found that task specific models relied more heavily on IMU based kinematic features, while EMG features grew in importance as models generalised. Sensor reduction experiments showed that close to optimal accuracy can be achieved with as few as four to six sensors compared to eleven sensors for the full dataset. This study provides a unified methodology for fatigue estimation, exposes the challenges of cross-task generalisation and provides guidelines for sensor configurations that balance performance with minimal burden.
Qirtas et al. (Thu,) studied this question.