Wearable fatigue detection system for wheelchair propulsion: Identifying key muscles via an attention-based LSTM Model.

Fatigue induces adverse short- and long-term health risks in wheelchair users. This study aimed to develop a machine learning approach to predict fatigue onset during wheelchair incremental exercise via surface electromyography (sEMG). sEMG from eight upper-limb muscles and oxygen uptake data were collected from nine wheelchair users (3 females, 6 males) performing incremental wheelchair propulsion on an ergometer. Ventilatory threshold (VT) derived from oxygen uptake was defined as the fatigue onset label, with sEMG signals as predictive features. A dynamic weighted attention long short-term memory (DWA-LSTM) model was built to distinguish non-fatigued and fatigue-transition propulsion cycles and identify critical muscle contributors. The model achieved an average classification accuracy of 94.82% using eight-muscle sEMG intensity. Notably, single-channel pectoralis major sEMG still yielded 89% accuracy. These findings indicate wearable fatigue monitoring systems may rely on single-muscle EMG to detect propulsion-related fatigue onset.