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: cohort visibility during training and the context available at inference . We study the training-time axis on an ABIDE-I cohort of 871 subjects (20 acquisition sites), comparing two population-graph implementations: 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 (ΔΔAUC). Across the two preselected outer splits (1024 and 2024), whose training and selection rules were fixed before the run, the direction is consistent and positive, with observed inter-action range 0.06285–0.11967; the earlier audited split 666 shows the same direction and is reported as historical support, because its execution identity cannot be established to the same standard. Uncertainty is assessed within each split by paired subject and site-cluster bootstrap resampling; the intervals are conditional on the fixed partitions and the trained checkpoints. Splits are replication units: they are not pooled and no cross-split average is reported. These findings tie reported performance to the evaluation information regime in the two implementations compared, and motivate explicit reporting of cohort visibility.
Funding Information
  • Natural Science Foundation of Chongqing, China (CSTB2025NSCQ-LZX0057)
  • Natural Science Foundation of Chongqing, China (CSTB2025NSCQ-LZX0073)
  • Innovation and Development Joint Fund of Chongqing Natural Science Foundation (CSTB2024NSCQ-LZX0138)
  • National Social Science Fund of China (23CTJ010)
  • Chongqing Normal University (21XLB020)

This publication has 22 references indexed in Scilit: