Causal Feature Attribution for Artificial Intelligence Triage Models Under Differential Privacy in Community Health Registries: A Healthcare Fairness Estimation

Differential privacy is increasingly proposed as a safeguard for training triage models on community health registry data, yet its interaction with causal feature attribution and downstream fairness remains uncharacterised. The analysis treats the privacy budget as a design parameter that perturbs the joint distribution of features and outcomes, and derives bounds on the disparity in attribution ranks between protected and non-protected groups. A before-after-control-impact design is proposed to evaluate the framework, with operating characteristics characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). Power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design achieves 80% power to detect a 5 percentage point effect with 54 clusters. The framework proposes that fairness estimation under differential privacy must account for attribution instability, and that the intraclass correlation of 0.64 implies that cluster-level randomisation is essential for credible inference. The contribution is a formalisation of the privacy-attribution-fairness trilemma and a design protocol for its empirical evaluation.