Physical fatigue significantly impacts safety and performance across industrial, athletic, and medical domains, yet its detection remains challenging due to individual variability and limited generalizability of existing methods. This study introduces a multimodal fatigue detection system integrating surface electromyography (sEMG) and inertial measurement unit (IMU) signals, processed through a hybrid convolutional neural network–long short-term memory–attention (CNN-LSTM-Attention) model. Fatigue was induced in 35 healthy participants via step-up-and-down exercises, with gait data collected during natural walking before and after fatigue. The model leverages sEMG from the gastrocnemius lateralis and IMU-derived jerk signals from the tibialis anterior and rectus femoris to classify fatigue states. Evaluated using leave-one-subject-out cross-validation (LOSOCV), the system achieved an accuracy of 87.94% with bilateral EMG signals and a balanced recall of 87.94% for fatigued states using a combined IMU-EMG approach. These results highlight the system’s robustness for personalized fatigue monitoring, surpassing traditional subject-dependent methods by addressing inter-individual differences.

Reliable stress assessment is fundamental to effective management, but conventional physiological approaches require contact-based sensors and relatively long recording periods. Eye tracking offers a practical non-contact alternative, yet conventional fixation, blink, and pupil features may not fully capture subtle gaze instability during fixation. In this study, we propose a gaze sway feature framework inspired by postural sway analysis and evaluate its utility for short-duration stress assessment during a brief fixation task in virtual reality (VR). Data from 34 participants were collected before and after VR-based stress induction, and conventional eye, gaze sway, and pulse rate variability (PRV) features were extracted. Classification was evaluated under a leakage-resistant subject-level cross-validation protocol with bootstrap 95% confidence intervals (CIs). Gaze sway showed broad post-stress changes across spatial, geometric, frequency, and stability domains. Using only gaze sway features, XGBoost achieved an accuracy of 0.926 (95% CI: 0.868–0.985) and an area under the ROC curve of 0.939 (95% CI: 0.849–0.997), yielding higher point-estimate performance than conventional eye features alone. Although PRV remained the strongest single-modality comparator, gaze sway achieved competitive performance using only non-contact eye tracking with stable, interpretable feature subsets. These findings support gaze sway as a feasible and interpretable approach for short-duration, non-contact stress assessment.