Proceedings of the Annual Meeting of the Cognitive Science Society

Mental fatigue and high cognitive loads reduce cognitive efficiency and task performance. Electroencephalogram (EEG) can track these states with high temporal resolution, but informative patterns span time, frequency, and distributed scalp topology. Most decoders handle these domains separately or fuse them only at the output, which can miss cross-domain dependencies. We propose TD-CAT, a tri-domain cross-attention Transformer that models interactions among temporal, spectral, and spatial representations. TD-CAT builds domain-specific features via temporal and spectral convolutions and a topology-aware graph module, then integrates the three streams with bidirectional cross-attention and low-rank fusion. Under leave-one-subject-out evaluation on SADT and SEED-VIG (driver fatigue) and MAT (cognitive load), TD-CAT reaches 80.7% and 92.6% accuracy on SADT and SEED-VIG, and 82.4% on MAT, demonstrating strong cross-subject performance across multiple datasets and task settings.