MedGUIDE: Benchmarking Clinical Decision-Making in Large Language Models

TL;DR

MedGUIDE provides a first-of-its-kind benchmark to test whether large language models can follow structured clinical guidelines, using 7,747 high-quality MCQs derived from 55 NCCN decision trees across 17 cancers. The authors implement a two-stage quality filtering (expert reward models and LLM-judge ensembles) to ensure clinically plausible, well-formed items and evaluate 25 diverse LLMs. Results reveal that even domain-specific models struggle with guideline-based sequential reasoning, with closed-source models generally outperforming open models and medical LLMs not reliably outpacing general counterparts. They further explore guideline grounding via in-context prompting and continued pretraining, finding substantial but uneven improvements, underscoring MedGUIDE’s value for evaluating safe, guideline-consistent clinical decision support in real-world settings.

Abstract

Clinical guidelines, typically structured as decision trees, are central to evidence-based medical practice and critical for ensuring safe and accurate diagnostic decision-making. However, it remains unclear whether Large Language Models (LLMs) can reliably follow such structured protocols. In this work, we introduce MedGUIDE, a new benchmark for evaluating LLMs on their ability to make guideline-consistent clinical decisions. MedGUIDE is constructed from 55 curated NCCN decision trees across 17 cancer types and uses clinical scenarios generated by LLMs to create a large pool of multiple-choice diagnostic questions. We apply a two-stage quality selection process, combining expert-labeled reward models and LLM-as-a-judge ensembles across ten clinical and linguistic criteria, to select 7,747 high-quality samples. We evaluate 25 LLMs spanning general-purpose, open-source, and medically specialized models, and find that even domain-specific LLMs often underperform on tasks requiring structured guideline adherence. We also test whether performance can be improved via in-context guideline inclusion or continued pretraining. Our findings underscore the importance of MedGUIDE in assessing whether LLMs can operate safely within the procedural frameworks expected in real-world clinical settings.

Figures (11)

• Figure 1: Overview of the MedGUIDE Benchmark Pipeline. Stage (1): selecting and processing 55 NCCN clinical decision tree guidelines for 17 common cancers to generate 16K synthetic MCQs. Stage (2): annotating and filtering these MCQs based on clinical and general quality criteria using expert-labeled data and a 5-head reward model, resulting in a curated set of 7,747 high-quality MCQs. Stage (3): evaluating 25 LLMs (open-source, closed-source, and medical) on the benchmark and applying two improvement methods—guideline-aware prompting (Method I) and guideline-based continued pretraining (Method II).
• Figure 2: NCCN Acute Myeloid Leukemia (AML) Guideline. The orange boxes illustrate the workflow through which the sample QA dataset is generated. Red and green annotations represent the correct and incorrect options.
• Figure 3: Distributions of cancer types (left), MCQ option counts (top right), and LLMs used for question generation (bottom right).
• Figure 4: Benchmark evaluation results
• Figure 5: Weighted accuracy before and after using Method I