[Paper Review] MEDIC: Comprehensive Evaluation of Leading Indicators for LLM Safety and Utility in Clinical Applications

Praveen K Kanithi, Christophe Clément+9 more|arXiv (Cornell University)|Sep 11, 2024

Scientific Computing and Data Management6 citations

TL;DR

MEDIC introduces a modular framework to evaluate clinical LLMs across five dimensions, plus a Cross-Examination Framework to assess information fidelity, revealing a knowledge-execution gap and safety divergences that discourage one-model dominance.

ABSTRACT

While Large Language Models (LLMs) achieve superhuman performance on standardized medical licensing exams, these static benchmarks have become saturated and increasingly disconnected from the functional requirements of clinical workflows. To bridge the gap between theoretical capability and verified utility, we introduce MEDIC, a comprehensive evaluation framework establishing leading indicators across various clinical dimensions. Beyond standard question-answering, we assess operational capabilities using deterministic execution protocols and a novel Cross-Examination Framework (CEF), which quantifies information fidelity and hallucination rates without reliance on reference texts. Our evaluation across a heterogeneous task suite exposes critical performance trade-offs: we identify a significant knowledge-execution gap, where proficiency in static retrieval does not predict success in operational tasks such as clinical calculation or SQL generation. Furthermore, we observe a divergence between passive safety (refusal) and active safety (error detection), revealing that models fine-tuned for high refusal rates often fail to reliably audit clinical documentation for factual accuracy. These findings demonstrate that no single architecture dominates across all dimensions, highlighting the necessity of a portfolio approach to clinical model deployment. As part of this investigation, we released a public leaderboard on Hugging Face.\footnote{https://huggingface.co/spaces/m42-health/MEDIC-Benchmark}

Motivation & Objective

• Motivate the need for leading indicators beyond static medical exams for real-world clinical utility.
• Define a modular framework (MEDIC) spanning five clinical dimensions to assess LLM performance.
• Incorporate applied task benchmarks (e.g., EHRSQL, MedCalc, MEDEC) and a reference-free Cross-Examination Framework to stress-test fidelity.
• Demonstrate that no single model dominates across dimensions and advocate for a portfolio approach.

Proposed method

• Define five clinical dimensions: Medical reasoning, Ethical and bias concerns, Data and language understanding, In-context learning, and Clinical safety.
• Use deterministic metrics for structured tasks (e.g., execution accuracy, exact match) and a Cross-Examination Framework for open-ended tasks to quantify coverage, conformity, consistency, and conciseness.
• Evaluate across a heterogeneous task suite including MedCalc, EHRSQL, DischargeMe, ACI-Bench, MEDEC, MedQA, MedMCQA, PubMedQA, and others.
• Apply LLM-as-a-judge pairwise comparison with Elo ratings to rank open-ended responses and ensure judge consensus across multiple models.
• Analyze the knowledge-execution gap by contrasting static knowledge benchmarks with operational task performance.

Experimental results

Research questions

• RQ1 Does higher parameter count or domain-specific fine-tuning translate to universal clinical competence across MEDIC dimensions?
• RQ2 How do models perform on operational clinical tasks (e.g., clinical calculations, SQL generation) versus static knowledge tasks (e.g., USMLE-style questions)?
• RQ3 To what extent do safety-oriented alignments for passive refusal translate into effective active safety (error detection) in clinical text auditing?
• RQ4 Can a reference-free cross-examination framework reliably assess factual consistency and coverage in open-ended clinical outputs?

Key findings

• Model capability is heterogeneous and task-dependent; no architecture dominates across all MEDIC tasks.
• Static knowledge proficiency does not reliably predict functional execution in operational tasks such as MedCalc or EHRSQL.
• Larger models may exhibit lower conformity to source documents, indicating more hallucinations despite greater fluency.
• Passive safety (refusal) saturates across models, but active safety (error detection) shows sharp performance degradation and variability.
• Open-ended inquiry rankings are robust to judge selection, with high inter-judge agreement in pairwise model comparisons.