Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MediEval: A Unified Medical Benchmark for Patient-Contextual and Knowledge-Grounded Reasoning in LLMs

Zhan Qu, Michael Färber

TL;DR

MediEval tackles the critical gap in evaluating medical LLMs by unifying patient-context grounding with biomedical knowledge through linking MIMIC-IV EHRs to UMLS-based vocabularies. It introduces a four-quadrant, four-way NLI framework to assess truth and grounding, along with novel safety metrics HSR and TIR. The paper presents CoRFu, a DPO-based fine-tuning approach with asymmetric penalties that target safety-critical misclassifications, achieving substantial gains in macro-F1 and reductions in risk across multiple backbones. Together, MediEval and CoRFu offer a rigorous, risk-aware paradigm for evaluating and improving safe medical reasoning in LLMs with real patient contexts.

Abstract

Large Language Models (LLMs) are increasingly applied to medicine, yet their adoption is limited by concerns over reliability and safety. Existing evaluations either test factual medical knowledge in isolation or assess patient-level reasoning without verifying correctness, leaving a critical gap. We introduce MediEval, a benchmark that links MIMIC-IV electronic health records (EHRs) to a unified knowledge base built from UMLS and other biomedical vocabularies. MediEval generates diverse factual and counterfactual medical statements within real patient contexts, enabling systematic evaluation across a 4-quadrant framework that jointly considers knowledge grounding and contextual consistency. Using this framework, we identify critical failure modes, including hallucinated support and truth inversion, that current proprietary, open-source, and domain-specific LLMs frequently exhibit. To address these risks, we propose Counterfactual Risk-Aware Fine-tuning (CoRFu), a DPO-based method with an asymmetric penalty targeting unsafe confusions. CoRFu improves by +16.4 macro-F1 points over the base model and eliminates truth inversion errors, demonstrating both higher accuracy and substantially greater safety.

MediEval: A Unified Medical Benchmark for Patient-Contextual and Knowledge-Grounded Reasoning in LLMs

TL;DR

MediEval tackles the critical gap in evaluating medical LLMs by unifying patient-context grounding with biomedical knowledge through linking MIMIC-IV EHRs to UMLS-based vocabularies. It introduces a four-quadrant, four-way NLI framework to assess truth and grounding, along with novel safety metrics HSR and TIR. The paper presents CoRFu, a DPO-based fine-tuning approach with asymmetric penalties that target safety-critical misclassifications, achieving substantial gains in macro-F1 and reductions in risk across multiple backbones. Together, MediEval and CoRFu offer a rigorous, risk-aware paradigm for evaluating and improving safe medical reasoning in LLMs with real patient contexts.

Abstract

Large Language Models (LLMs) are increasingly applied to medicine, yet their adoption is limited by concerns over reliability and safety. Existing evaluations either test factual medical knowledge in isolation or assess patient-level reasoning without verifying correctness, leaving a critical gap. We introduce MediEval, a benchmark that links MIMIC-IV electronic health records (EHRs) to a unified knowledge base built from UMLS and other biomedical vocabularies. MediEval generates diverse factual and counterfactual medical statements within real patient contexts, enabling systematic evaluation across a 4-quadrant framework that jointly considers knowledge grounding and contextual consistency. Using this framework, we identify critical failure modes, including hallucinated support and truth inversion, that current proprietary, open-source, and domain-specific LLMs frequently exhibit. To address these risks, we propose Counterfactual Risk-Aware Fine-tuning (CoRFu), a DPO-based method with an asymmetric penalty targeting unsafe confusions. CoRFu improves by +16.4 macro-F1 points over the base model and eliminates truth inversion errors, demonstrating both higher accuracy and substantially greater safety.
Paper Structure (41 sections, 11 equations, 6 figures, 3 tables)

This paper contains 41 sections, 11 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. MediEval Framework: Data Construction
  5. Electronic Health Records (EHRs) as Foundation.
  6. Step 1: Context Extraction.
  7. Step 2: Semantic Normalization of Medical Codes.
  8. Step 3: Ontology Graph Construction and Relation Extraction.
  9. Step 4: Quadrant-based Statement Construction.
  10. MediEval Framework: LLM Evaluation
  11. Quadrant Formulation as a Four-Way NLI Task.
  12. Baseline Metrics.
  13. Critical Error Rates.
  14. Interpretation of Metrics.
  15. Counterfactual Risk-Aware Fine-Tuning
  16. ...and 26 more sections

Figures (6)

  • Figure 1: Overview of the current work with a real example; texts in blue indicate the extracted sample.
  • Figure 2: Example of statement verification against patient records (Quadrant 2: True-Unsupported). The statement is medically correct, yet in this case, GERD is treated with omeprazole, a different class of medication, so the statement is not supported by the context.
  • Figure 3: Example of statement verification against patient records (Quadrant 1: True-Supported). The statement is medically correct, and GERD is indeed treated with omeprazole, making it true and supported.
  • Figure 4: Example of statement verification against patient records (Quadrant 3: False-Supported). The crafted statement is medically incorrect, as Atenolol is indicated for hypertension. Since both GERD and Atenolol appear in the patient context, the statement may seem supported, even though it is false.
  • Figure 5: Example of statement verification against patient records (Quadrant 4; False-Unsupported). The statement is medically incorrect, since insulin is not indicated for GERD. Moreover, insulin does not appear in the patient context, making the statement false and unsupported.
  • ...and 1 more figures