Interpretability Audits of Graph Neural Diagnostic Models Under Silent Health Information Exchange Dropouts: A Computer Robustness Evaluation

Abstract

Graph neural networks (GNNs) used for diagnostic prediction in health information exchange (HIE) settings are typically evaluated on complete graphs, yet real exchanges suffer silent dropouts in which entire facilities stop contributing edges without any missingness indicator. the present analysis treats interpretability as a robustness property and asks whether attribution-based audits remain valid when the underlying graph is silently truncated. We formalise silent dropout as a structural perturbation of the message-passing operator and derive a bound on attribution drift in terms of the spectral gap of the graph Laplacian and the fraction of lost edges. The analysis is a design and power study for a before-after-control-impact (BACIP) audit estimator, not an empirical evaluation: no observational data are used or claimed. Under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Fifty clusters were required for 80% power at a 5-percentage-point effect. The results indicate that audit designs with fewer than 24 clusters cannot reliably distinguish genuine attribution instability from sampling noise, and that the minimum detectable effect falls steeply as clusters increase. The contribution is a formal account of why interpretability audits fail silently under dropout and a power-calibrated design rule for detecting that failure.