Abstract
Fatigue remains a leading contributor to cognitive failure in safety-critical environments, yet current detection systems struggle with rare event sensitivity, individual variability, and delayed response times. We introduce NeuroFatigueNet, a revolutionary deep learning architecture that integrates temporal convolutional networks with fatigue-prioritized transformers, augmented by cross-session memory tokens. This framework pioneers spectral-temporal fusion for joint processing of raw EEG and instantaneous frequency features, multi-dilation convolutions capturing micro-to-macro fatigue patterns, and clinical- guided sparse attention that focuses computation on high-risk segments. Validated on the DROZY dataset (15 subjects, synchronized EEG/EOG/video) and UCI EEG Eye State dataset (14,980 samples), NeuroFatigueNet achieves unprecedented performance.