Unmasking data leakage in EEG-ADHD literature: a rigorous, interpretable SOTA framework (DSAEN).

The field of translational Electroencephalogram-Artificial Intelligence (EEG-AI) faces a significant methodological challenge regarding epoch-wise data leakage. This study aims to quantify the implications of this leakage through a comprehensive methodological critique of existing models and proposes a novel, interpretable Dual-Stream Attention-Enhanced Network (DSAEN) architecture optimized for clinically realistic subject-wise validation. DSAEN processes spatially interpolated 3D raw EEG tensors in parallel with handcrafted neurophysiological features mapped to the exact same topographical grid. This dual-stream framework utilizes 3D Convolutional Neural Networks, Long Short-Term Memory sequences, Transformers, and Bahdanau Attention to capture both localized spatiotemporal microstates and global chaotic dynamics. A replication study on a public dataset of 121 children (61 ADHD, 60 controls) confirmed that the performance of leading published models undergoes a pronounced degradation (from 98.03% to 78.96%, and 86.83% to 55.99%) when shifting from epoch-wise to subject-wise 5-fold cross-validation. Evaluated under the same rigorous subject-wise protocol, DSAEN achieved a state-of-the-art accuracy of 88.4% (Area Under the Curve: 92.5%), significantly outperforming baseline models (p < 0.05). Extensive embedding analyses demonstrated that the dual streams extract mathematically orthogonal features (r^2 approx 2.9%), statistically validating their complementarity, while the 3D spatial pooling ensures high computational efficiency (2.04 MB footprint) suitable for clinical edge deployment. Finally, Explainable AI (XAI) techniques, corroborated by statistical tests on signal variance, confirmed the model's neurophysiological plausibility by highlighting clinically relevant frontal brain regions. By addressing pervasive evaluation artifacts and integrating parameter efficiency with interpretable, topology-aware spatiotemporal modeling, DSAEN establishes a reliable and trustworthy benchmark for diagnostic AI.