Speaker
Description
Reliable decoding of cognitive workload from electroencephalography (EEG) signals is essential for adaptive robotics, neuroergonomics, intelligent transportation, and human–machine interaction systems. Despite recent advances in deep learning, EEG-based workload classification remains limited by poor cross-subject and cross-dataset generalization. In this work, we investigate hybrid Riemannian deep learning and self-supervised representation learning frameworks for robust EEG workload estimation across heterogeneous datasets. We developed a comprehensive benchmark pipeline integrating classical signal processing, covariance-based Riemannian geometry, transformer models, self-supervised contrastive learning, and explainable artificial intelligence. Experiments were performed using the publicly available ds007262 arithmetic EEG workload dataset and the STEW simultaneous task EEG workload dataset. Six decoding frameworks were evaluated: FBCSP, EEGNet, Riemannian tangent-space classifiers, transformer networks, self-supervised learning architectures, and a proposed hybrid Riemannian deep learning model. Subject-independent evaluation using leave-one-subject-out cross-validation demonstrated that hybrid covariance-aware models consistently outperformed conventional CNN pipelines. The proposed hybrid framework achieved approximately 61% mean classification accuracy, while transformer and self-supervised learning models demonstrated improved temporal representation learning capabilities. Statistical benchmarking using ROC curves, precision–recall curves, violin plots, and Wilcoxon significance testing confirmed the superiority of hybrid geometric learning approaches. Explainable AI analyses revealed neurophysiologically meaningful workload-related EEG patterns involving frontal theta enhancement and parietal alpha suppression. Channel importance mapping demonstrated dominant contributions from frontal and parietal cortical regions associated with attentional control and working memory processes. To evaluate robustness and transferability, cross-dataset experiments were performed by training on ds007262 and testing on STEW. Results demonstrated substantial performance degradation caused by domain shift, recording variability, and montage mismatch across datasets. These findings highlight the major challenge of EEG transferability in real-world applications and emphasize the need for future domain adaptation and transfer learning strategies. The proposed benchmark framework provides a reproducible pipeline for evaluating EEG workload decoding algorithms under both subject-independent and cross-dataset settings. The integration of Riemannian geometry, explainable AI, and self-supervised learning offers a promising direction for robust cognitive monitoring systems in intelligent robotics and adaptive neurotechnology applications.
