Attention-deficit/hyperactivity disorder (ADHD) is typically diagnosed in childhood based on clinical interviews and behavioral assessments, motivating research into physiologic evidence-based approaches, utilizing neuroimaging techniques like electro...
Attention-deficit/hyperactivity disorder (ADHD) is typically diagnosed in childhood based on clinical interviews and behavioral assessments, motivating research into physiologic evidence-based approaches, utilizing neuroimaging techniques like electroencephalogram (EEG). However, scalp EEG is confounded by volume conduction and provides limited anatomical interpretability. In this work, we present a baseline pipeline that combines EEG source localization with multi-modal graph construction for pediatric ADHD classification. Using an age-matched pediatric MRI template, we estimated source-level activity and built region-of-interest (ROI) connectivity graphs across five frequency bands. Edge attributes were computed from weighted phase lag index (wPLI) and lagged coherence and were further assessed via ablation, and node attributes included time-domain representations extracted from raw EEG via an LSTM module and power spectral density.
Under strict subject-wise cross-validation, the proposed pipeline achieved, the best performance of AUROC 0.805 and PRAUC 0.850, although variability across folds remained, indicating limited robustness in the current setting. To understand these outcomes, we conducted various post-hoc analyses of the constructed graphs. Post-hoc analyses suggested that topology-level summaries tended to attenuate group differences, whereas focusing on task-relevant ROI subsets and interaction-oriented features revealed more meaningful group-level structure. Based on these findings, we summarize key limitations and provide practical future-work directions for improving robustness, anatomical validity of source estimates, and feature design in source-localized EEG graph learning for pediatric ADHD.