Distilling Trial-Level EEG Variability into Participant Representations: Neurophysiologically Constrained Graph Learning in ADHD

Extracting stable individual-level representations from noisy and variable trial-wise EEG remains a major challenge in computational cognitive neuroscience. We developed a spatiotemporal graph attention network to analyze Go/NoGo EEG from children and adolescents with and without ADHD. The framework combines ERP-informed temporal windows, phase-synchrony graph priors, and trial-to-participant transfer learning, making both information routing and representation geometry explicitly testable. The model learned reproducible attention trajectories, yet ADHD and control participants did not differ reliably in spatial attention allocation. Instead, group-related structure emerged in compact participant-level embeddings that distilled trial-level variability and retained discriminative information beyond convolutional representations. Performance exceeded a convolutional neural network comparison, but the central advance is methodological: neurophysiologically constrained graph learning can reveal where diagnostically relevant information emerges within the analytical pipeline. This provides a framework for developing auditable cognitive-neurophysiological markers of neurodevelopmental variation.