WL-GraphTrax: A Graph-Transformer Framework for EEG-Based Cognitive Workload Classification

Vishal Pandey, Nikhil Panwar, Partha Pratim Roy, Sushil Chandra

Abstract: Accurately measuring cognitive workload is critical in the development of adaptive human-machine systems, particularly in high-stakes environments such as driving. This study introduces a novel EEG-based framework for cognitive workload classification using graph neural networks. EEG data were collected during a controlled yet ecologically valid simulated driving task designed to elicit low, medium, and high workload states. Graphs were constructed using four distinct functional connectivity measures—Coherence (COH), Mutual Information (MI), Granger Causality (GC), and Directed Transfer Function (DTF)—across five canonical frequency bands. A dual-stage architecture combining Graph Attention Networks (GAT) and Transformer encoders was employed to model intra-graph and inter-graph dependencies. Three configurations were evaluated: per-graph GAT models, intra-matrix band fusion models for each connectivity type, and all-matrix fusion integrating all graphs across spectral and topological dimensions. Subject-independent evaluation reveals that directed measures, particularly DTF, offer superior discriminative power. The DTF Band-Fusion model achieves an accuracy of 94.14%, while the All-Matrix Fusion model reaches a peak accuracy of 95.68%. Attention-based analyses further highlight the dominant role of theta and alpha frequency bands in workload discrimination. These findings advance EEG-based workload classification and demonstrate the potential for interpretable, scalable cognitive monitoring in real-world applications.

External IDs:dblp:journals/access/PandeyPRC25