Large language models encode clinical knowledge.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) demonstrate advanced capabilities but face high standards for clinical use.
• Current assessments of medical knowledge in LLMs often rely on limited automated benchmarks.

Purpose of the Study:

• To introduce MultiMedQA, a comprehensive benchmark for evaluating LLMs in medical question answering.
• To establish a human evaluation framework assessing factuality, comprehension, reasoning, harm, and bias in LLM responses.
• To assess the performance of Pathways Language Model (PaLM) and Flan-PaLM on the MultiMedQA benchmark.

Main Methods:

• Developed MultiMedQA, integrating six medical QA datasets and the new HealthSearchQA dataset.
• Implemented a human evaluation protocol for LLM-generated medical answers.
• Evaluated PaLM and Flan-PaLM using various prompting strategies on MultiMedQA.
• Introduced instruction prompt tuning for domain adaptation of LLMs.

Main Results:

• Flan-PaLM achieved state-of-the-art accuracy on all MultiMedQA multiple-choice datasets, including 67.6% on MedQA (USMLE-style questions).
• Human evaluation identified significant gaps in LLM performance despite strong automated scores.
• Instruction prompt tuning led to Med-PaLM, which showed improved performance but remained below clinician level.

Conclusions:

• LLM performance in medicine improves with scale and instruction prompt tuning.
• Current LLMs have limitations in clinical applications, underscoring the need for robust evaluation frameworks.
• Further development is crucial for creating safe and effective LLMs for healthcare.