Beyond Semantic Similarity: A Component-Wise Evaluation Framework for Medical Question Answering Systems with Health Equity Implications

Current evaluation methods are inadequate: Most medical AI systems are evaluated only on semantic similarity (how closely answers match expected text), which fails to capture actual medical accuracy or identify health equity risks—a critical gap for systems advising patients on health matters.

A new evaluation framework reveals hidden problems: The VB-Score framework evaluates four separate components (entity recognition, semantic similarity, factual consistency, and information completeness), and when applied to three major LLMs on 48 public health topics, it uncovered major discrepancies between how similar answers seemed and whether they were actually medically accurate.

All tested models showed severe performance failures: Across the board, the three well-known LLMs examined demonstrated "almost uniformly severe performance failures" when measured against rigorous medical accuracy standards, suggesting current medical AI systems may be less reliable than their semantic similarity scores suggest.

Health equity disparities are significant: The models showed 13.8% lower performance on public health topics related to chronic conditions affecting older adults and minority populations compared to their overall average, indicating that vulnerable populations may receive systematically worse medical information from these systems.

Semantic similarity scores are misleading: The research demonstrates that an AI system can appear to perform well on traditional metrics while actually providing inaccurate or incomplete medical information, highlighting the danger of relying on surface-level evaluation methods for high-stakes applications like healthcare.