FairGuard: Blockchain-Enforced Continuous Fairness Governance for Demographically Equitable LLM-Based Emergency...

Demographic bias in AI-driven clinical decision support systems (CDSS) represents one of the most consequential and least-addressed risks in healthcare AI. Large Language Models (LLMs) applied to emergency triage can silently perpetuate or amplify existing health disparities across thousands of patient decisions. This paper introduces FairGuard, a continuous bias detection and governance framework that embeds demographic fairness auditing directly into the LLM inference pipeline using four integrated mechanisms: (1) an equity-enforcing consent gate that applies demographic-blind access control; (2) a RAG corpus bias analyzer that traces class-directional misclassification to retrieval layer composition; (3) per-subgroup confusion matrix stratification computing ΔF1 across demographic groups; and (4) a blockchain-anchored continuous monitoring layer that enforces a ΔF1 ≤ 0.05 governance threshold across every inference batch. Evaluated on the MIMIC-IV Full Emergency dataset (N=300-1,000), FairGuard achieves a gender ΔF1 of 0.020, well within threshold, while identifying RAG corpus composition as the root cause of conservative triage bias (50.9% Urgent → Non-Urgent misclassification), establishing that the observed bias is class-directional rather than demographically concentrated. FairGuard provides the first continuous, blockchain-enforced fairness governance mechanism for LLM-based emergency triage CDSS.