Auditing Evidence Use in Medical LLM Diagnosis

2026-07-22

AI Summary

Current evaluations of medical large language models predominantly focus on diagnostic accuracy, offering limited insight into whether these models appropriately leverage clinical evidence. This work proposes a role-aware evidence utilization auditing framework that decomposes patient information into discrete evidence units, evaluates diagnoses under controlled subsets of evidence, and uncovers low-order interactions to distinguish clinically sound differential reasoning from suspicious evidence usage patterns. The approach effectively disentangles genuine interaction effects from spurious attributions, revealing evidence misuse often masked by high overall accuracy. Experiments on three datasets, including DDXPlus, demonstrate that most identified interactions align with clinical reasoning; blinded review further indicates that misuse primarily stems from improper handling of negative or localized clinical findings.

Abstract

Medical LLMs are often evaluated by whether they select the correct diagnosis, but diagnostic accuracy alone does not show whether the model used the case evidence appropriately. We present a behavioral audit of evidence use in medical diagnosis. For each case, we decompose patient information into evidence units, score candidate diagnoses under controlled evidence subsets, and mine low-order interactions in diagnostic margins. Because medical evidence is diagnosis-relative, the audit separates interaction discovery from failure assignment: large or negative interactions can reflect plausible differential diagnosis, while suspicious interactions require robustness checks and clinical review. We evaluate five open-weight LLMs on DDXPlus, CupCase, and MedCase. Across datasets, faithful support and differential conflict or cancellation account for most interaction strength, showing that many evidence interactions are clinically plausible rather than failures. In a DDXPlus-focused blinded five-reviewer 130-item enriched review sample, invalid or shortcut-like cases concentrate in negated or absent findings and clinically local evidence. These results show that accuracy can hide candidate evidence-use failures and motivate role-aware audits for medical LLM evaluation.

Problem

• Medical LLM
• evidence use
• diagnostic accuracy
• behavioral audit
• clinical plausibility

Innovation

• evidence auditing
• diagnostic reasoning
• interaction mining
• medical LLM evaluation
• differential diagnosis