Evaluating Large Language Model Diagnostic Performance on JAMA Clinical Challenges via a Multi-Agent Conversational Framework
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background & Objective
Standard clinical LLM benchmarks use multiple-choice vignettes that present all information up front, unlike real encounters where clinicians iteratively elicit histories and objective data. We hypothesized that such formats inflate LLM performance and mask weaknesses in diagnostic reasoning. We developed and evaluated a multi-AI agent conversational framework that converts JAMA Clinical Challenge cases into multi-turn dialogues, and assessed its impact on diagnostic accuracy across frontier LLMs.
Methods
We adapted 815 diagnostic cases from 1,519 JAMA Clinical Challenges into two formats: (1) original vignette and (2) multi-agent conversation with a Patient AI (subjective history) and a System AI (objective data: exam, labs, imaging). A Clinical LLM queried these agents and produced a final diagnosis. Models tested were O1 (OpenAI), GPT-4o (OpenAI), LLaMA-3-70B (Meta), and Deepseek-R1-distill-LLaMA3-70B (Deepseek), each in multiple-choice and free-response modes. Free-response grading used a separate GPT-4o judge for diagnostic equivalence. Accuracy (Wilson 95% CIs) and conversation lengths were compared using two-tailed tests.
Results
Accuracy decreased for all models when moving from vignettes to conversations and from multiple-choice to free-response (p<0.0001 for all pairwise comparisons). In vignette multiple-choice, accuracy was O1 79.8% (95% CI, 76.9%–82.4%), GPT-4o 74.5% (71.4%–77.4%), LLaMA-3 70.9% (69.5%–72.2%), Deepseek-R1 69.0% (67.5%–70.4%). In conversation multiple-choice: O1 69.1% (65.8%–72.2%), GPT-4o 51.3% (49.8%–52.8%), LLaMA-3 49.7% (48.2%–51.3%), Deepseek-R1 34.0% (32.6%–35.5%). In conversation free-response: O1 31.7% (28.6%–34.9%), GPT-4o 20.7% (19.5%–22.0%), LLaMA-3 22.9% (21.6%–24.2%), Deepseek-R1 9.3% (8.4%–10.2%). O1 generally required fewer conversational turns than GPT-4o, suggesting more efficient multi-turn reasoning.
Conclusions
Converting vignettes into multi-agent, multi-turn dialogues reveals substantial performance drops across leading LLMs, indicating that static multiple-choice benchmarks overestimate clinical reasoning competence. Our open-source framework offers a more rigorous and discriminative evaluation and a realistic substrate for educational use, enabling assessment of iterative information-gathering and synthesis that better reflects clinical practice.
