Deep Learning Approaches for Fatigue Detection: A Traditional Review of Models, Datasets, and Applications

Fatigue is an important factor affecting human safety, effectiveness and health in various industries, but it still is a challenge to accurately detect it. Recent advances in deep learning fuel the emergence of vision-based, physiology-based, and hybrid fatigue detection systems with better accuracy and real-time monitoring. We cover work up until 2025 and perform a meta-analysis of recent models, datasets, and methodologies. Face- and appearance-based—building system upon facial expression, eye closure, and head pose information—is highly effective but sensitive to image variation caused by illumination change (lighting condition change) or occlusions. Physiological systems based on EEG, ECG and HRV approaches are more sensitive to subtle fatigue markers but lack practicality in natural environment due to their intrusiveness. Hybrid models fusing multimodal data sources can achieve better accuracy (often more than 92%) and robustness, while at a sufficient computational cost and use of resources. Challenges identified include small and non-diverse datasets, dependence on simulated environments, as well as unaddressed ethical implications of continuous monitoring. Huge practical dataset, non-invasion sensing techniques and privacy-preserving methods under the data governance norms are necessary to address these issues. The findings of this work provide a guided path toward scalable, valid and ethically responsible deep learning fatigue prediction systems.