Answers are grounded in this paper and cite their sources.
Dyslexia is characterized by atypical neural synchronization across functional brain networks. While Electroencephalography (EEG) captures these interactions, modeling multi-channel connectivity across heterogeneous populations remains a significant challenge. This study proposes a Gated Multi-View Graph Co-Attention Network for automated dyslexia diagnosis, advancing both multi-view connectivity fusion methodology and population-level biomarker analysis. Trained on an EEG dataset of 53 subjects, our framework integrates three complementary connectivity representations, namely Pearson Correlation (PCC), Phase-Locking Value (PLV), and Coherence (COH), using a hierarchical attention mechanism. At the spatial level, Graph Attention Network version 2 (GATv2) backbones extract view-specific features, which are then refined via a cross-modal co-attention layer to capture inter-view dependencies. A Global Gating Network dynamically weights these representations, allowing for adaptive modality selection. The framework is evaluated under a nested subject-grouped cross-validation protocol: band, architecture and epoch are selected on inner subject-grouped folds, and final performance is estimated with AUC = 0.807 ± 0.190 exclusively from locked outer-test predictions. For interpretability, we focus on the Theta-band DualView_GatedCoAttn configuration, the joint configuration most frequently selected under the nested procedure. At the primary XAI confidence threshold ( τ = 0.70 ), F8 ( p = 0.00058 ) and F4 ( p = 0.04480 ) showed significant model-derived saliency. The analysis also identified a concentration of directed GATv2 attention coefficients around Cz and a PLV gate contribution of approximately 51.4%. The theta/PLV pattern is compatible with the Temporal Sampling Framework but does not establish a specific phonological mechanism. Overall, the proposed multi-view framework provides a subject-independent and interpretable approach to EEG-based dyslexia classification and demonstrates competitive, but not statistically superior, performance relative to the comparator architectures on the present dataset.
No takes yet. Share an insight, caveat, or question.
Zakraoui et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: