Large language models encode clinical knowledge
Большие языковые модели кодируют клинические знания
2023-07-12
SCID: 54.1/7fa75k75
Discuss with AI
HealthSearchQAMed-PaLMMultiMedQA benchmarkinstruction prompt tuningmedical question answering
Figures from the paper
Abstract (AI)
Abstract Large language models (LLMs) have demonstrated impressive capabilities, but the bar for clinical applications is high. Attempts to assess the clinical knowledge of models typically rely on automated evaluations based on limited benchmarks. Here, to address these limitations, we present MultiMedQA, a benchmark combining six existing medical question answering datasets spanning professional medicine, research and consumer queries and a new dataset of medical questions searched online, HealthSearchQA. We propose a human evaluation framework for model answers along multiple axes including factuality, comprehension, reasoning, possible harm and bias. In addition, we evaluate Pathways Language Model 1 (PaLM, a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM 2 on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA 3 , MedMCQA 4 , PubMedQA 5 and Measuring Massive Multitask Language Understanding (MMLU) clinical topics 6 ), including 67.6% accuracy on MedQA (US Medical Licensing Exam-style questions), surpassing the prior state of the art by more than 17%. However, human evaluation reveals key gaps. To resolve this, we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, knowledge recall and reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal limitations of today’s models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLMs for clinical applications.
Key Findings
1
Flan-PaLM achieved state-of-the-art accuracy on every evaluated MultiMedQA multiple-choice dataset, including 67.6% on MedQA, exceeding the prior state of the art by more than 17%.
2
Instruction prompt tuning produced Med-PaLM, which showed encouraging clinical performance but remained inferior to clinicians in human evaluation.
3
Model scale and instruction prompt tuning improved comprehension, knowledge recall, and reasoning, while evaluations still identified safety and reliability gaps requiring further development.
4
The MultiMedQA benchmark combines six existing medical question-answering datasets with the newly introduced HealthSearchQA dataset spanning professional, research, and consumer queries.
5
The proposed human evaluation framework assesses model answers for factuality, comprehension, reasoning, potential harm, and bias, addressing limitations of automated benchmark-only evaluations.
Research Object
Large language models (LLMs) evaluated on medical question answering datasets (MultiMedQA and HealthSearchQA), including PaLM, Flan-PaLM 2 and Med-PaLM
Research Subject
clinical knowledge, including factuality, comprehension, reasoning, knowledge recall, harm and bias, and the effects of model scale and instruction prompt tuning
Publication Details
Publication Date
2023-07-12
Journal
Publisher
ISSN
Cited by
5021
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai12
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer2019
Best Practices for Developing and Validating Scales for Health, Social, and Behavioral Research: A Primer2018
Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing2022
Training language models to follow instructions with human feedback2022
On the Opportunities and Risks of Foundation Models2021
PaLM: Scaling Language Modeling with Pathways2022
Scaling Laws for Neural Language Models2020
Deep learning-enabled medical computer vision2021
BioGPT: generative pre-trained transformer for biomedical text generation and mining2022
Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models2022
Ethical Machine Learning in Healthcare2021
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model2022
Cited by20
A Survey of Large Language Models2026
A Survey on Evaluation of Large Language Models2024
Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum2023
BloombergGPT: A Large Language Model for Finance2023
Benefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine2023
The future landscape of large language models in medicine2023
Benchmarking Retrieval-Augmented Generation for Medicine2024
A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation2025
ChatGPT and the rise of large language models: the new AI-driven infodemic threat in public health2023
Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs2024
A framework for human evaluation of large language models in healthcare derived from literature review2024
Almanac — Retrieval-Augmented Language Models for Clinical Medicine2024
A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges2024
The ethics of ChatGPT in medicine and healthcare: a systematic review on Large Language Models (LLMs)2024
AgentClinic: a multimodal benchmark for tool-using clinical AI agents2026
Empowering biomedical discovery with AI agents2024
Exploring Large Language Model‐Based Intelligent Agents: Definitions, Methods, and Prospects2026
Diagnostic reasoning prompts reveal the potential for large language model interpretability in medicine2024
Large language models in medical and healthcare fields: applications, advances, and challenges2024
A Comprehensive Review on Synergy of Multi-Modal Data and AI Technologies in Medical Diagnosis2024