Tracing high transductive cohort AUC to same-site supervision in a site-aware population GNN for multisite fMRI

Population-graph models can use cohort-level context, complicating interpretation of strong multisite neuroimaging performance. We investigated which information pathways accounted for high transductive cohort discrimination in a site-aware heterogeneous population graph neural network for autism classification. Using a frozen ABIDE-I cohort (871 participants, 20 sites) and evaluation protocol, we applied controlled graph, feature, supervision, and architecture interventions to a clean-room re-implementation across C-PAC and NIAK preprocessing pipelines. Cohort out-of-fold AUC was approximately 0.94. A site-only graph retained similarly high discrimination (0.948 versus 0.941 for the full model), with no statistically significant difference detected. Removing same-site supervision under a site-wise intervention reduced AUC to 0.480, whereas a supervision-budget control retaining same-site supervision achieved 0.921. Among the alternative heads evaluated, the high cohort AUC was observed only with the sex-heterogeneous dual-channel head; a canonical topology-only Parisot-GCN did not reproduce it. A demographic classifier using site, sex, and their interaction achieved a best pooled AUC of 0.51. Under leave-one-site-out evaluation, AUC was 0.522 for C-PAC and 0.532 for NIAK, with confidence intervals including 0.5 and no clear evidence of unseen-site discrimination; an imaging-only reference achieved 0.653 and 0.588. Because leave-one-site-out evaluation jointly changes supervision availability, target-site topology, and training-domain composition, the decrease cannot be attributed to a single factor. These findings distinguish cohort-visible transductive performance from unseen-site generalization and motivate explicit site-only, no-graph, and site-held-out controls.