[Paper Review] Large Language Models Encode Clinical Knowledge

Karan Singhal, Shekoofeh Azizi+28 more|arXiv (Cornell University)|Dec 26, 2022

Topic Modeling256 citations

TL;DR

Flan-PaLM achieves state-of-the-art results on several medical QA benchmarks; instruction prompt tuning (Med-PaLM) improves alignment with clinicians, though gaps remain compared with human clinicians.

ABSTRACT

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but the quality bar for medical and clinical applications is high. Today, attempts to assess models' clinical knowledge typically rely on automated evaluations on limited benchmarks. There is no standard to evaluate model predictions and reasoning across a breadth of tasks. To address this, we present MultiMedQA, a benchmark combining six existing open question answering datasets spanning professional medical exams, research, and consumer queries; and HealthSearchQA, a new free-response dataset of medical questions searched online. We propose a framework for human evaluation of model answers along multiple axes including factuality, precision, possible harm, and bias. In addition, we evaluate PaLM (a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM, on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA, MedMCQA, PubMedQA, MMLU clinical topics), including 67.6% accuracy on MedQA (US Medical License Exam questions), surpassing prior state-of-the-art by over 17%. However, human evaluation reveals key gaps in Flan-PaLM responses. 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, recall of knowledge, and medical reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal important limitations of today's models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLM models for clinical applications.

Motivation & Objective

• Create a diverse, open benchmark (MultiMedQA) for medical question answering spanning exams, research, and consumer queries.
• Develop and apply a human-evaluation framework to assess factuality, harm, bias, and usefulness of LLM responses.
• Evaluate PaLM and Flan-PaLM on the benchmark and propose instruction prompt tuning to align models to medical domain requirements.
• Demonstrate that model scale and prompting strategies improve medical knowledge recall and reasoning, while identifying remaining safety and alignment gaps.

Proposed method

• Curate MultiMedQA by combining six existing medical QA datasets and adding HealthSearchQA (3375 consumer-searched questions).
• Use few-shot, chain-of-thought, and self-consistency prompting to evaluate PaLM and Flan-PaLM on the benchmark.
• Report state-of-the-art accuracies on MedQA, MedMCQA, PubMedQA, and MMLU clinical topics for Flan-PaLM (e.g., 67.6% on MedQA).
• Introduce instruction prompt tuning to align Flan-PaLM to medical domain tasks, producing Med-PaLM with improved clinical alignment.
• Develop a pilot human evaluation framework with clinician and lay-user assessments across axes like agreement with scientific consensus, potential harm, and bias.
• Discuss key limitations and propose directions for future research to enhance safety and usefulness of medical LLMs.

Experimental results

Research questions

• RQ1 To what extent do large language models encode and utilize clinical knowledge across diverse medical QA tasks?
• RQ2 Can instruction prompt tuning align LLMs with medical domain requirements to reduce harm and improve factuality and usefulness?
• RQ3 How do model scale and prompting strategies affect comprehension, recall, and medical reasoning in LLMs?
• RQ4 How do clinician and lay evaluations compare with automated benchmarks in assessing medical AI outputs?

Key findings

• Flan-PaLM achieves state-of-the-art accuracy on MedQA (67.6%), MedMCQA (57.6%), PubMedQA (79.0%), and MMLU clinical topics.
• MedQA performance improves the previous SOTA by over 17%.
• Clinician evaluation: 61.9% of Flan-PaLM long-form answers aligned with scientific consensus, vs 92.6% for Med-PaLM and 92.9% for clinician-generated answers.
• Potential harm: 29.7% of Flan-PaLM answers were rated as potentially leading to harm, compared with 5.8% for Med-PaLM and 6.5% for clinician-generated answers.
• Med-PaLM, the instruction-tuned version, shows encouraging results in aligning with medical consensus and reducing harm, yet remains inferior to clinicians in some axes.
• HealthSearchQA introduces 3,375 consumer-focused medical questions to evaluate real-world question answering.