[Paper Review] Towards Expert-Level Medical Question Answering with Large Language Models

Karan Singhal, Tao Tu+29 more|arXiv (Cornell University)|May 16, 2023

Artificial Intelligence in Healthcare and EducationMedicine332 citations

TL;DR

Med-PaLM 2 surpasses prior medical QA models by leveraging PaLM 2, domain-specific finetuning, and ensemble refinement, achieving state-of-the-art results on multiple benchmarks and favorable human evaluations of long-form answers.

ABSTRACT

Recent artificial intelligence (AI) systems have reached milestones in "grand challenges" ranging from Go to protein-folding. The capability to retrieve medical knowledge, reason over it, and answer medical questions comparably to physicians has long been viewed as one such grand challenge. Large language models (LLMs) have catalyzed significant progress in medical question answering; Med-PaLM was the first model to exceed a "passing" score in US Medical Licensing Examination (USMLE) style questions with a score of 67.2% on the MedQA dataset. However, this and other prior work suggested significant room for improvement, especially when models' answers were compared to clinicians' answers. Here we present Med-PaLM 2, which bridges these gaps by leveraging a combination of base LLM improvements (PaLM 2), medical domain finetuning, and prompting strategies including a novel ensemble refinement approach. Med-PaLM 2 scored up to 86.5% on the MedQA dataset, improving upon Med-PaLM by over 19% and setting a new state-of-the-art. We also observed performance approaching or exceeding state-of-the-art across MedMCQA, PubMedQA, and MMLU clinical topics datasets. We performed detailed human evaluations on long-form questions along multiple axes relevant to clinical applications. In pairwise comparative ranking of 1066 consumer medical questions, physicians preferred Med-PaLM 2 answers to those produced by physicians on eight of nine axes pertaining to clinical utility (p < 0.001). We also observed significant improvements compared to Med-PaLM on every evaluation axis (p < 0.001) on newly introduced datasets of 240 long-form "adversarial" questions to probe LLM limitations. While further studies are necessary to validate the efficacy of these models in real-world settings, these results highlight rapid progress towards physician-level performance in medical question answering.

Motivation & Objective

• Aim to advance physician-level performance in medical question answering using large language models.
• Evaluate model across multi-domain medical QA benchmarks and real-world-style long-form questions.
• Develop and validate prompting strategies to improve medical reasoning and safety.

Proposed method

• Use PaLM 2 as the base LLM.
• Fine-tune with medical-domain data via instruction tuning on MultiMedQA datasets (MedQA, MedMCQA, HealthSearchQA, LiveQA, MedicationQA).
• Introduce ensemble refinement prompting to aggregate multiple reasoning paths and refine answers.
• Evaluate with multiple prompting strategies: few-shot, chain-of-thought, self-consistency, and ensemble refinement.
• Conduct extensive human evaluations (physician and lay-person raters) on long-form answers and adversarial datasets.
• Analyze test-set overlap to assess potential training data contamination.

Figure 1: Med-PaLM 2 performance on MultiMedQA Left: Med-PaLM 2 achieved an accuracy of 86.5% on USMLE-style questions in the MedQA dataset. Right: In a pairwise ranking study on 1066 consumer medical questions, Med-PaLM 2 answers were preferred over physician answers by a panel of physicians across

Experimental results

Research questions

• RQ1 Can Med-PaLM 2 reach or exceed physician-level performance on standard medical QA benchmarks?
• RQ2 Do domain-specific finetuning and advanced prompting strategies improve medical reasoning and safety in long-form answers?
• RQ3 How robust are model outputs when faced with adversarial or equity-focused questions?
• RQ4 What is the impact of training/test overlap on reported benchmark performance?

Key findings

• Med-PaLM 2 achieves up to 86.5% accuracy on MedQA USMLE-style questions, advancing beyond Med-PaLM by over 19%.
• Med-PaLM 2 approaches or exceeds state-of-the-art on MedMCQA, PubMedQA, and MMLU clinical topics.
• In long-form evaluations, physicians preferred Med-PaLM 2 over Med-PaLM on eight of nine clinical-utility axes; lay evaluators found Med-PaLM 2 more helpful and relevant.
• Adversarial datasets show Med-PaLM 2 substantially outperforms Med-PaLM across all axes of safety and limitation probing (e.g., lower harm risk, better alignment with medical consensus).
• A simple ensemble refinement prompting strategy significantly improves performance over baseline few-shot and self-consistency on multiple-choice benchmarks (e.g., MedQA and MMLU variants).
• Overlap analysis indicates limited but non-negligible test-training data contamination, with modest impact on reported performance.

Figure 2: Illustration of Ensemble Refinement (ER) with Med-PaLM 2. In this approach, an LLM is conditioned on multiple possible reasoning paths that it generates to enable it to refine and improves its answer.