Multi-Axial Analysis of Clinical Reasoning in Large Language Models: Inter-Verifier Disagreement and Its Implications...

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Clinical Decision Support

Background:

• Evaluating clinical reasoning in Large Language Models (LLMs) faces challenges with existing metrics and the LLM-as-judge approach.
• Assessing LLM-generated clinical justifications requires methods that verify evidence-conclusion coherence.

Purpose of the Study:

• To evaluate the clinical reasoning capabilities of three generator LLMs on MIMIC-IV cases.
• To assess the reliability of independent LLM verifiers in judging evidence-conclusion coherence.
• To determine the necessity of human oversight in LLM-based clinical reasoning evaluation.

Main Methods:

• Three generator LLMs were assessed on 1,000 MIMIC-IV hospital-stay cases.
• Evaluation used four axes: medical concept grounding, semantic similarity, semantic uncertainty, and evidence-conclusion coherence.
• Evidence-conclusion coherence was independently judged by three frontier verifier LLMs (Claude Sonnet 4.6, Gemini 2.5 Pro, GPT-5.4 mini).

Main Results:

• Coherence analysis revealed LLMs can score well on other metrics yet provide unsupported conclusions.
• Inter-verifier agreement on coherence was consistently low (Fleiss' κ 0.087-0.223), with high disagreement rates (62.2%-74.3%).
• Physician adjudication of 50 cases showed variable agreement with different verifiers, indicating no single LLM reliably replaces clinical assessment.

Conclusions:

• A single LLM verifier lacks sufficient reliability for stand-alone judgment of clinical reasoning at scale.
• Structured human oversight remains essential for evaluating LLM clinical reasoning.
• Selective automation via unanimous-agreement tiers is possible but requires further validation.