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

A previously reported high transductive cohort accuracy in multisite autism functional MRI studies has been traced back to site-specific supervision within a site-aware population graph neural network. Researchers examined various information pathways that contributed to a previously observed high transductive cohort accuracy in a multisite autism fMRI study. The cohort out-of-fold accuracy was approximately 0.94, which depended on site-linked edges.

A site-only graph achieved the same high accuracy (0.948 vs. 0.941) as the full model, indicating that the high performance was directly linked to supervision based on site-specific information. The gain in accuracy required same-site supervision, which was crucial for the model's performance. Masking this supervision reduced the accuracy to a significantly lower 0.480. This effect was not explained by the supervision budget itself.

Among architecture-matched heads, the high accuracy was only observed with the evaluated sex-heterogeneous dual-channel head. In contrast, a canonical topology-only Parisot-GCN did not exhibit the same pattern. Under leave-one-site-out evaluation, the performance fell to near-chance accuracy (0.522 for C-PAC and 0.532 for NIAK). This starkly contrasted with an imaging-only reference, which maintained higher accuracy levels (0.653 for C-PAC and 0.588 for NIAK).

In summary, the high cohort accuracy in this study was primarily dependent on same-site supervision through the transductive graph under the applied protocol. This relationship did not extend to discrimination across different sites, highlighting the site-specific nature of the observed performance.