An evaluation framework for clinical use of large language models in patient interaction tasks

Area of Science:

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

Background:

• Large language models (LLMs) show promise for transforming clinical diagnostics and doctor-patient interactions.
• The real-world readiness and clinical application of LLMs require rigorous, standardized testing.
• Existing evaluation methods often lack the nuance of natural medical dialogues.

Purpose of the Study:

• To introduce the Conversational Reasoning Assessment Framework for Testing in Medicine (CRAFT-MD) for evaluating clinical LLMs.
• To assess the diagnostic capabilities and limitations of prominent LLMs (GPT-4, GPT-3.5, Mistral, LLaMA-2-7b) in a conversational context.
• To propose recommendations for future clinical LLM evaluations.

Main Methods:

• Development and application of the CRAFT-MD framework, utilizing simulated AI agents for natural dialogue-based LLM interaction.
• Evaluation of LLM performance across 12 medical specialties.
• Assessment of both text-based and multimodal (GPT-4V) conversational and visual capabilities.

Main Results:

• Identified significant limitations in current LLMs regarding clinical conversational reasoning, history-taking, and diagnostic accuracy.
• Limitations were observed across various LLMs and persisted even with multimodal capabilities.
• Performance gaps were evident across different medical specialties.

Conclusions:

• Current LLMs require substantial improvement before widespread clinical deployment.
• The CRAFT-MD framework offers a more realistic approach to evaluating clinical LLMs than traditional methods.
• Recommendations emphasize realistic dialogue simulation, comprehensive assessment, and mixed evaluation methodologies for future LLM testing.