Equity Considerations in Artificial Intelligence-Driven Triage Systems

Abstract

Artificial intelligence and algorithmic tools increasingly determine the order in which patients are seen, escalated, or allocated scarce resources such as ventilators, raising the question of whether these tools reproduce the very inequities they are often marketed as correcting. This policy analysis reviews three documented mechanisms of bias in AI-driven and algorithmic triage: cost-based proxy labels that understate need in historically underserved populations, pulse-oximetry measurement error that masks hypoxemia in patients with darker skin, and crisis-standards-of-care scoring rubrics that have been challenged as discriminatory toward people with disabilities. Each mechanism enters the triage pipeline at a different point -- training data, point-of-care instrumentation,
and protocol design -- which is why no single technical fix addresses all three simultaneously. We examine the governance responses that have followed, including regulatory inquiries into algorithmic bias, disclosure requirements for AI-enabled medical devices, and calls for continuous post-deployment auditing rather than one-time certification. We conclude that defensible AI-driven triage requires equity to be treated as a design specification from the outset, sustained subgroup-specific monitoring after deployment, and explicit acknowledgment that some triage decisions encode value judgments that no algorithm can neutrally resolve.