Multisite Evaluation of BrainNetCNN for ADHD Classification from Static and Dynamic rs-fMRI Functional Connectivity
Catalina Quintero López1, Víctor Daniel Gil Vera1, Juan Pablo Ospina López2
1Luis Amigó Catholic University, Transversal 51A #67B 90 Medellín - Colombia, Postal Code 050001, Colombia.
Abstract:
Attention-Deficit/Hyperactivity Disorder (ADHD) has been associated with altered integration across large-scale brain networks. Temporal fluctuations in functional organization and differences between acquisition centers complicate classification based on resting-state functional magnetic resonance imaging (rs-fMRI). This study used BrainNetCNN as the reference architecture to examine how region-of-interest (ROI) dimensionality, connectivity representation, network design, window configuration and acquisition site influence ADHD discrimination. The sample included 465 participants from four ADHD-200 centers: New York University (NYU; n = 177), Peking University (Peking; n=183), NeuroIMAGE (n=39) and Oregon Health & Science University (OHSU; n=66). Functional connectivity was derived from regional blood-oxygen-level-dependent (BOLD) time series using prespecified 12, 18, 39 and 116-ROI panels from the Automated Anatomical Labeling atlas (AAL116). The reference window was 120 s with an approximately 12-s step. Within each center, classification was assessed through participant-level stratified 10-fold cross-validation repeated five times, with area under the receiver operating characteristic curve (AUC) as the primary metric. BrainNetCNN was configured using NYU data and then kept fixed for evaluation at the remaining centers. Across site-by-ROI combinations, AUC ranged from 50.1% to 64.5%. The 12-ROI configuration reached 63.6% in both NYU and Peking, while the strongest panel differed across centers. Comparisons involving static connectivity, long short-term memory (LSTM), gated recurrent unit (GRU), DeepSets, logistic regression and alternative window settings revealed no consistent advantage. Leave-one-site-out analyses yielded AUCs from 44.4% to 63.9%, with 14 of 16 pointwise 95% confidence intervals including 0.50. Overall, the models showed modest site-dependent discrimination and limited cross-site transportability.
