AUTHOR=Wang Yi , Meng Lu , Fan Yuying 

TITLE=CMTS-GNN: a cross-modal temporal-spectral graph neural network with cognitive network explainability

JOURNAL=Frontiers in Neurology

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1700161

DOI=10.3389/fneur.2025.1700161

ISSN=1664-2295

ABSTRACT=Infantile spasms (IS) represent a severe form of epileptic encephalopathy occurring in early infancy. Timely and accurate detection is critical, as delays or misdiagnosis are associated with adverse neurodevelopmental outcomes that can impair perceptual, cognitive, and affective development. Conventional EEG analysis is often challenged by the complexity, heterogeneity, and large volume of IS data, rendering manual review both time-intensive and susceptible to inter-rater variability. To address these challenges, we introduce CMTS-GNN—a Cross-Modal Temporal—Spectral Graph Neural Network. This model integrates complementary information from temporal and spectral EEG representations through bidirectional cross-modal attention and gated fusion mechanisms. It further incorporates explicit modeling of brain-region connectivity to capture functional interactions that underlie perceptual processing, cognitive control, and affective dynamics. By doing so, CMTS-GNN aims to improve both detection accuracy and interpretability. We evaluated the proposed model on an in-house infantile spasms dataset and the publicly available CHB-MIT epilepsy dataset. Evaluation protocols included five-fold cross-validation and subject-independent schemes (leave-one-subject-out/leave-one-patient-out). On our in-house dataset, five-fold cross-validation resulted in an accuracy of 99.02%, precision of 98.96%, recall of 97.47%, F1-score of 98.20%, and AUC of 99.27%. For the CHB-MIT dataset, the same protocol yielded an accuracy of 98.54%, precision of 98.31%, recall of 98.71%, F1-score of 98.47%, and AUC of 98.87, outperforming several recent approaches across most metrics. Subject-independent evaluations further confirmed the model's robustness and generalizability across different patients. Importantly, by modeling connectivity across brain regions, CMTS-GNN provides clinically meaningful explanations for its decisions, enhancing interpretability. In summary, CMTS-GNN offers an accurate, generalizable, and interpretable framework for automated IS detection from EEG. It holds potential to support earlier clinical intervention, thereby helping to mitigate long-term perceptual, cognitive, and affective morbidity in affected infants.