Task-State EEG-Based Mental Stress Recognition Using Multi-Band Dynamic Attention Network

Task-state mental stress detection aims to identify stress levels during cognitive tasks using electroencephalogram signals and is key in brain-computer interface and mental health research. However, current deep learning methods struggle to extract frequency band-specific features and often overlook inter-band interactions, leading to poor neurophysiological representation. To address this issue, the authors of this paper propose a multiband dynamic attention network that combines multifrequency decomposition, frequency-domain attention, and cross-band interaction. First, wavelet packet transform adaptively extracts key time-frequency features across electroencephalogram rhythms. Then, a frequency-domain attention mechanism emphasizes stress-related frequency components. Finally, a cross-band interaction module with multi-head attention and gating explores intrinsic inter-band relationships, enhancing feature representation. Experiments show that the multiband dynamic attention network significantly improves accuracy and robustness, outperforming existing methods in feature extraction and inter-band modeling.