EEG-based ADHD diagnosis models suffer from two persistent issues: data leakage and the lack of physiologically grounded interpretability, limiting clinical adoption. To address these, this paper presents 3D ADHD-Net and DeepTrace. Unlike established 1D and 2D studies, 3D ADHD-Net is a topology-aware spatiotemporal model that preserves scalp geometry by projecting raw EEG onto a grid, preventing spatial information loss. DeepTrace is a novel explainability framework that traces diagnostic information from latent representations back to input electrode space using balanced sign-aligned ensemble Spearman correlation. All experiments utilized strict subject-independent 5-fold cross-validation, explicitly mitigating epoch-level, preprocessing, and optimization leakage. Using a public pediatric EEG dataset (61 ADHD, 60 controls), the framework achieved 84.23% mean accuracy and 91.20% fold-averaged ROC-AUC, significantly outperforming replicated baselines (p0.05), which suffered substantial performance collapses under the same rigorous evaluation. Importantly, DeepTrace identified a consistent fronto-central hypoactivation signature, which was causally validated via phenotype induction and rescue perturbation experiments, confirming reliance on established neurophysiological biomarkers rather than gradient-based artifacts. This study demonstrates that diagnostic performance and interpretability need not be competing objectives, and that physiologically faithful attribution can be integrated into deep learning pipelines for developing clinically trustworthy EEG-based ADHD assessment support systems.
Das et al. (Thu,) studied this question.