Optimal Transport and Graph Neural Networks for Cross-Session Mental Workload Classification

Electroencephalography (EEG) offers a noninvasive, high-temporal-resolution modality for estimating mental workload. However, session-to-session variability limits the generalizability of workload classifiers, and few systematic cross-session evaluations are reported in the literature. This study systematically evaluates domain adaptation methods for cross-session mental workload classification using the publicly available COG-BCI dataset within an evaluation framework that may guide future studies on EEG-based classification models. We make four contributions: (i) integration of Optimal Transport (OT) with Graph Neural Networks (GNNs) to model spatial relationships and align feature distributions under strict session-wise separation; (ii) a data-centric evaluation pipeline incorporating Self-Organizing Map (SOM) visualizations for data exploration and a heuristic loss function for model selection; (iii) a strict cross-session protocol examining the effects of graph construction, feature selection, and data splits; and (iv) comparison of OT with CORrelation ALignment (CORAL) and GNN with EEGNet. Incorporating OT improved test accuracies across all experimental configurations. SOM visualizations confirmed enhanced feature alignment after OT. Our results highlight the potential of OT for mitigating session-to-session variability and underscore the importance of a data-centric approach and rigorous cross-session evaluation when developing classifiers for complex cognitive state estimation. Future work should explore semi-supervised OT strategies and scalable implementations for real-time applications.