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

Area of Science:

• Healthcare AI
• Machine Learning Fairness
• Clinical Informatics

Background:

• Demographic bias in AI clinical decision support systems (CDSS) is a significant risk in healthcare.
• Large Language Models (LLMs) used in emergency triage can inadvertently worsen health disparities.

Purpose of the Study:

• To introduce FairGuard, a novel framework for continuous bias detection and governance in LLM-based emergency triage.
• To embed demographic fairness auditing within the LLM inference pipeline.

Main Methods:

• FairGuard utilizes four mechanisms: an equity-enforcing consent gate, a RAG corpus bias analyzer, per-subgroup confusion matrix stratification (ΔF1), and a blockchain-anchored monitoring layer with a ΔF1 ≤ 0.05 threshold.
• Evaluated on the MIMIC-IV Full Emergency dataset.

Main Results:

• FairGuard achieved a gender ΔF1 of 0.020, meeting the fairness threshold.
• Identified RAG corpus composition as the cause of conservative triage bias (50.9% Urgent → Non-Urgent misclassification).
• The bias was determined to be class-directional, not demographically concentrated.

Conclusions:

• FairGuard offers the first continuous, blockchain-enforced fairness governance for LLM-based emergency triage CDSS.
• The framework effectively detects and mitigates demographic bias, ensuring equitable AI performance in critical healthcare applications.