Source-linked AI summary
Towards Expert-Level Medical Question Answering with Large Language Models
Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Le Hou, Kevin Clark, Stephen Pfohl, Heather Cole-Lewis, Darlene Neal, Mike Schaekermann, Amy Wang, Mohamed Amin, Sami Lachgar, Philip Mansfield, Sushant Prakash, Bradley Green, Ewa Dominowska, Blaise Aguera y Arcas, Nenad Tomasev, Yun Liu, Renee Wong, Christopher Semturs, S. Sara Mahdavi, Joelle Barral, Dale Webster, Greg S. Corrado, Yossi Matias, Shekoofeh Azizi, Alan Karthikesalingam, Vivek Natarajan
TL;DR
Medical question answering still showed gaps in benchmark performance and in comparison with physician answers, especially for nuanced long-form responses. Med-PaLM 2 combines PaLM 2, medical finetuning, and ensemble refinement, achieving strong multi-benchmark and human-evaluation results. The authors nevertheless note that real-world validation, safety, and evaluation across settings remain necessary.
Problem
Prior medical LLMs showed room for improvement in benchmark performance and answer quality compared with physicians, while benchmarks did not capture all aspects of medical question answering.
Method
Med-PaLM 2 combines an improved PaLM 2 base model, medical domain-specific finetuning, and prompting strategies including chain-of-thought self-consistency and ensemble refinement.
Results
Med-PaLM 2 approached or exceeded state-of-the-art across MultiMedQA benchmarks and was preferred to physician answers across eight of nine clinical-utility axes.
Takeaways & Limitations
The results indicate rapid progress toward physician-level medical question answering across multiple-choice and long-form evaluations.
Takeaways & Limitations
The results cannot be considered generalizable to every medical question-answering setting and audience, and further real-world validation is needed.
Abstract
from arXiv · showhide
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.
1 Introduction
Med-PaLM 2 addresses remaining gaps in medical question answering by combining improved language-model foundations, medical finetuning, and ensemble refinement. It achieved strong benchmark and human-evaluation results, including preferences over physician answers on most clinical-utility axes.
- Motivation: Prior Med-PaLM evaluations showed room for improvement in answer quality compared with physicians, despite state-of-the-art multiple-choice benchmark performance.The gaps included long-form answer safety and alignment with human values and expectations.
- Approach: Med-PaLM 2 combines the PaLM 2 base model, medical domain-specific finetuning, and ensemble refinement prompting.The model was developed to improve medical reasoning and bridge limitations identified in prior work.
- Benchmark results: Med-PaLM 2 achieved state-of-the-art results on several MultiMedQA benchmarks, including MedQA USMLE-style questions.The introduction identifies MedQA and other MultiMedQA benchmarks as central evaluation targets.
- Human evaluation: Med-PaLM 2 answers were preferred to physician and Med-PaLM answers across eight of nine axes relevant to clinical utility.The evaluated axes included factuality, medical reasoning capability, and low likelihood of harm.
- Human evaluation: 72.9% of Med-PaLM 2 answers were judged to better reflect medical consensus than physician answers.This was reported as an example from the pairwise evaluation of long-form consumer medical answers.
- Adversarial evaluation: Med-PaLM 2 performed significantly better than Med-PaLM across every axis on two adversarial question datasets.These datasets were introduced to probe model safety and limitations; low risk of harm was reported for 90.6% versus 79.4% of answers.
2 Related Work
Related work established medical question answering as a major LLM challenge and showed rapid benchmark gains from larger general-purpose models. Med-PaLM 2 extends this trajectory by combining general-purpose LLM capabilities with medical alignment and broader human evaluation.
- Prior medical LLMs: Medical question answering has been pursued with both domain-specific language models and larger general-purpose LLMs.These approaches produced steady and, later, leapfrog improvements on medical benchmark datasets.
- Benchmark progress: GPT-3.5, Flan-PaLM, and GPT-4-base reached 60.2%, 67.6%, and 86.1% accuracy, respectively, on MedQA USMLE-style questions.The reported figures illustrate rapid benchmark progress over a short period.
- Clinical evaluation: Prior studies also evaluated general-purpose models on diagnostic and triage tasks using clinical case vignettes.One cited study compared GPT-3 with lay-people and physicians on validated cases.
- Medical alignment: Med-PaLM and Med-PaLM 2 combine general-purpose LLM capabilities with medical question-answering data and physician-written responses for domain alignment.The approach targets safety-critical requirements in medicine.
- Prompting strategy: Ensemble refinement conditions the model on multiple reasoning paths generated by the same model before producing a final answer.The strategy builds on chain-of-thought, self-consistency, and related generation-conditioning methods.
- Evaluation scope: The work evaluates both multiple-choice benchmarks and nuanced long-form answers using physician and lay-person assessment rubrics.Independent and pairwise evaluation supports a broader assessment of model quality and safety.
3 Methods
Med-PaLM 2 was evaluated across multiple-choice and long-form medical question-answering datasets, using instruction finetuning, prompting strategies, overlap analysis, and human assessment. Its evaluation included standard benchmarks, adversarial questions, and blinded physician and lay-person ratings.
- Datasets: Evaluation covered MedQA, MedMCQA, PubMedQA, MMLU clinical topics, MultiMedQA long-form questions, and two adversarial question datasets.The multiple-choice datasets were benchmarked separately from long-form and adversarial evaluations.
- Modeling: Instruction finetuning trained a unified Med-PaLM 2 model on MultiMedQA training splits using empirically determined dataset mixture ratios.The mixture included MedQA, MedMCQA, HealthSearchQA, LiveQA, and MedicationQA.
- Multiple-choice evaluation: Prompting compared few-shot, chain-of-thought, self-consistency, and ensemble refinement strategies for multiple-choice evaluation.Self-consistency used 11 chain-of-thought samplings, while ensemble refinement used staged sampling and refinement.
- Multiple-choice evaluation: Ensemble refinement first samples multiple explanations and answers, then conditions the model on those generations to produce a refined response.The method used 11 first-stage samplings and 33 second-stage samplings in this work, and was applied only to multiple-choice evaluation because of resource costs.
- Overlap analysis: Overlap analysis searched for matching question text between MultiMedQA test questions and the base LLM’s training corpus to assess possible test-set contamination.A question was classified as overlapping when the full question or at least 512 contiguous characters matched a training document, excluding answer options.
- Long-form evaluation: Long-form evaluation used physician and lay-person raters who independently assessed blinded answers from physicians, Med-PaLM, and Med-PaLM 2.Physician answers were generated without time limits and with access to reference materials; model answers were sampled at temperature 0.0.
4 Results
Med-PaLM 2 achieved strong results across multiple-choice benchmarks and long-form evaluations, with improved performance over Med-PaLM on adversarial questions and favorable human ratings across several axes.
- Multiple-choice evaluation: 85.4% accuracy was achieved on MedQA by the unified Med-PaLM 2 model using ensemble refinement.
- Multiple-choice evaluation: 86.5% accuracy was obtained on MedQA by a version instruction-finetuned only on MedQA, establishing a new state-of-the-art result.
- Multiple-choice evaluation: 81.8% accuracy on PubMedQA using self-consistency was reported as state-of-the-art, although the test set contains only 500 examples.
- Multiple-choice evaluation: Med-PaLM 2 achieved state-of-the-art performance on three of six MMLU clinical topics, while GPT-4-base performed better on the other three.
- Long-form evaluation: Physicians rated Med-PaLM 2 answers significantly higher than Med-PaLM answers across all adversarial evaluation axes, including general and health-equity subsets.
- Long-form evaluation: Lay-person raters found Med-PaLM 2 answers more helpful and relevant than Med-PaLM answers, with p ≤0.002 for both dimensions.
- Long-form evaluation: Med-PaLM 2 answers were longer than both Med-PaLM and physician answers, with median lengths of 794, 565.5, and 337.5 characters, respectively.
- Long-form evaluation: Med-PaLM 2 answers were preferred to physician answers on eight of nine MultiMedQA axes and to Med-PaLM answers across every adversarial axis.
5 Discussion
The paper combines broad benchmark evaluation with clinically relevant human assessment to examine progress toward physician-level medical question answering. Its results support substantial gains over Med-PaLM, while highlighting evaluation and deployment boundaries.
- 5 Discussion: Med-PaLM 2 showed strong performance in both multiple-choice and long-form medical question answering, including challenging adversarial datasets.
- 5 Discussion: Med-PaLM 2 approached or exceeded state-of-the-art across MultiMedQA benchmarks and produced substantial long-form gains over Med-PaLM in physician and lay-person evaluations.
- 5 Discussion: Clinically relevant, multidimensional evaluation is needed because structured knowledge benchmarks alone do not fully assess real-world clinical utility.
- 5 Discussion: Longer answers improved performance on missing important content but may trade off against unnecessary detail, and optimal answer length depends on context such as medical history.
- 5 Discussion: The evaluation did not consider multi-turn dialogue or frameworks for actively acquiring additional information.
6 Limitations
The evaluation has important scope and measurement limitations. These include limited generalizability, restricted physician-answer comparisons, and incomplete coverage of safety, bias, and equity.
- Evaluation methodology: Alignment measurement requires continued development because medical information needs are broad and complex.The rubric was not formally validated, although observed inter-rater reliability was high.
- Generalizability: The results cannot be considered generalizable to every medical question-answering setting and audience.Physician answers lacked specific clinical scenarios and nuanced communication requirements, and model answers were often longer.
- Physician comparison: Physicians produced only one answer per question, limiting assessment of the range of possible physician-produced answers.Future evaluations could include multiple answers, explicit scenarios, and inter-physician variation.
- Adversarial evaluation: The current adversarial evaluation is relatively limited and should not be interpreted as a comprehensive assessment of safety, bias, and equity.Future adversarial datasets could expand health-equity coverage and enable disaggregated evaluation over sensitive characteristics.
7 Conclusion
The results indicate rapid progress toward physician-level medical question answering. Broader real-world uptake still requires validation, safety, ethics, and rigorous evaluation across contexts and workflows.
- Conclusion: Med-PaLM 2 demonstrates rapid progress toward physician-level medical question answering.The authors call for further work on validation, safety, ethics, and evaluation in real-world workflows.
A.1 Additional Results
The appendix provides sensitivity analyses, statistical analyses, response examples, pairwise rankings, and answer-length summaries for the evaluation datasets.
- Sensitivity analysis: Appendix Table A.1 reports an overlap sensitivity analysis.It defines overlap using exact question or up to 120-character overlap with the underlying model’s training corpus.
- Independent evaluation: Tables A.2–A.4 provide statistical analyses for physician and lay-person independent evaluations of long-form answers.The analyses cover MultiMedQA 140 and adversarial questions.
- Pairwise rankings: Tables A.5–A.6 report statistical analyses of physician pairwise rankings comparing Med-PaLM 2 with physician and Med-PaLM answers.These comparisons use MultiMedQA 1066.
- Response examples: Tables A.7–A.8 provide long-form response examples and pairwise rankings between Med-PaLM and Med-PaLM 2.The examples connect qualitative answers with comparative evaluations.
- Answer lengths: Table A.9 summarizes answer lengths in characters for Med-PaLM 2, Med-PaLM, and physicians across the evaluation sets.The physician groups include answers from MultiMedQA 140 and adversarial questions.
A.2 Inter-rater Reliability
Inter-rater reliability was assessed on 140 multi-rated question-answer pairs using Randolph’s κ. Agreement was very good for 10 of 12 axes and good for the remaining two.
- Reliability analysis: Inter-rater reliability analysis used 140 question-answer pairs independently rated by three physicians.Agreement was measured with Randolph’s κ because several axes had low baseline positive rates.
- Agreement results: κ exceeded 0.8 for 10 of 12 alignment questions and exceeded 0.6 for the remaining two.The two lower-agreement questions concerned missing important content and unnecessary additional information.
- Figure guide: Figure A.1 shows agreement metrics for all 12 evaluation axes with 95% confidence intervals.The figure marks κ=0.6 as good agreement and κ=0.8 as very good agreement.
A.3 Details of Prompting Strategies
The appendix documents prompting strategies and statistical analyses for long-form answer evaluations. It also identifies tables containing prompts, comparisons, answer-length summaries, and ratings across physician, lay-person, and adversarial-question evaluations.
- A.3.1 Chain-of-Thought prompts: Tables A.10 to A.13 provide Med-PaLM 2 chain-of-thought prompts.
- A.3.2 PubMedQA prompting: PubMedQA evaluation used randomly generated 3-shot prompts without chain-of-thought, with updated instructions selected on the development set.The prompts were generated from the training split for each evaluation question.
- A.3.3 Ensemble refinement prompts: Tables A.15 and A.16 provide Med-PaLM 2 ensemble refinement prompts.
- A.3.4 Long-form question prompts: Table A.17 provides dataset-specific long-form prompts for Med-PaLM and Med-PaLM 2 that were not tuned to improve performance.The HealthSearchQA, LiveQA, and MedicationQA templates match those used in prior work.
- A.3.4 Long-form question prompts: Table A.2 reports statistical analyses for physician ratings on MultiMedQA 140, including 95% bootstrap confidence intervals and pairwise permutation tests.The comparisons were between Med-PaLM 2 and Med-PaLM, and between Med-PaLM 2 and physician answers.
- A.3.4 Long-form question prompts: Table A.3 summarizes physician ratings on adversarial questions overall and for health-equity-focused and general subsets.The subsets contain 182 × 4 and 58 × 4 ratings, respectively.
- A.3.4 Long-form question prompts: Tables A.4 through A.6 report statistical analyses for lay-person and physician pairwise evaluations on MultiMedQA 140 and 1066 questions.The physician comparisons cover Med-PaLM 2 against physician answers and Med-PaLM answers.
- A.3.4 Long-form question prompts: Tables A.7 through A.9 provide example comparisons, pairwise rankings, and answer-length summary statistics for model and physician responses.The answer-length statistics cover MultiMedQA 140 and adversarial question sets.