NHMT: A NeuroHierarchical Mamba-GCN-Transformer Model for Multilevel EEG Stress and Workload Analysis

This work presents a comprehensive framework for classifying cognitive and stress-related brain states using Electroencephalogram (EEG) signals recorded during relaxation, arithmetic, Stroop, and mirror image tasks. Each condition exhibits unique neural signatures, such as alpha rhythms in resting states, elevated frontal beta during arithmetic stress, frontal theta in Stroop conflict, and parietal activation during visuospatial tasks. To effectively model these diverse spatio-temporal EEG dynamics, we propose NeuroHierarch-Mamba-GCN-Transformer (NHMT), a unified deep learning architecture that integrates adaptive filtering, fractal convolutions, Mamba-based state-space modeling, graph neural networks, and hierarchical Transformers. NHMT captures both short and long range dependencies and performs simultaneous multi-task classification for stress, workload, and task domains using shared feature representations with uncertainty weighted loss. Experimental results show significant performance gains across tasks, with NHMT achieving up to 95.83% accuracy across three benchmark stress datasets, namely SAM-40, STEW, and EEG-MAT. High reliability is confirmed by Cronbach’s and intraclass correlation coefficient (ICC), indicating strong cross-subject consistency. The proposed model offers a robust and interpretable solution for real-time cognitive state monitoring in neurophysiological applications, demonstrating superior performance compared to established benchmark models in terms of accuracy, reliability, and generalization.