Information regimes in population-graph neuroimaging: cohort context and implementation-dependent evaluation sensitivity

Abstract

Population-graph models aggregate a query subject with an unlabeled cohort, so their reported performance can depend on the evaluation regime rather than on predictive information alone. Evaluation context has two separable axes: the visibility of the cohort during training, and the context available at inference. We study the training-time axis under a protocol fixed before analysis, on an ABIDE-I cohort of 871 subjects (20 acquisition sites). Two population-graph implementations are compared: a learned/adaptive graph (EV-GCN) and a static weighted graph (Parisot-GCN). Within each implementation the cohort, connectivity features, outer splits and training and selection rules are held fixed, and only the information available to the graph varies. The primary estimand is the implementation-by-training-context interaction: the difference between the P0-to-P1 AUC changes of the two implementations ({Delta}{Delta}AUC). Across three preselected outer splits (666, 1024, 2024) the direction is consistent and positive, with observed interaction range 0.06285-0.13090; uncertainty is assessed separately within each split by paired subject and site-cluster bootstrap resampling and is reported per split. Splits are replication units: they are not pooled and no cross-split average is reported, so the result is not a claim of invariance to arbitrary splits. These findings tie reported population-graph performance to the evaluation information regime and motivate explicit reporting of cohort visibility. The contribution is methodological: a two-axis taxonomy of information regimes, preselected cross-split evidence that the sensitivity of reported performance to cohort context depends on the implementation, and a reusable reporting checklist.