2022 LargeLanguageModelsEncodeClinic

From GM-RKB
Jump to navigation Jump to search

Subject Headings: Med-PaLM.

Notes

Cited By

2023

Quotes

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.

1. Introduction

Medicine is a humane endeavor where language enables key interactions for and between clinicians, researchers, and patients. Yet, today’s AI models for applications in medicine and healthcare have largely failed to fully utilize language. These models, while useful, are predominantly single-task systems (e.g., classification, regression, segmentation), lacking expressivity and interactive capabilities [21, 81, 97]. As a result, there is a discordance between what today’s models can do and what may be expected of them in real-world clinical workflows [42, 74]. Recent advances in large language models (LLMs) offer an opportunity to rethink AI systems, with language as a tool for mediating human-AI interaction. LLMs are “foundation models” [10], large pre-trained AI systems that can be repurposed with minimal effort across numerous domains and diverse tasks. These expressive and interactive models offer great promise in their ability to learn generally useful representations from the knowledge encoded in medical corpora, at scale. There are several exciting potential applications of such models in medicine, including knowledge retrieval, clinical decision support, summarisation of key findings, 

...
Automated and Human Evaluation
Medical Question Answering Benchmark
Med-PaLM performs encouragingly on consumer medical question answering triaging patients’ primary care concerns, and more.
...
Figure 1 | Overview of our contributions We curated MultiMedQA, a benchmark for medical question answering spanning medical exam, medical research, and consumer medical questions. We evaluated PaLM and its instructed-tuned variant, Flan-PaLM, on MultiMedQA. With a combination of prompting strategies, Flan-PaLM exceeded SOTA performance on MedQA (USMLE), MedMCQA, PubMedQA, and MMLU clinical topics. In particular, it improved over the previous SOTA on MedQA (USMLE) by over 17%. We next proposed instruction prompt tuning to further align Flan-PaLM to the medical domain, producing Med-PaLM. Med-PaLM’s answers to consumer medical questions compared favorably with clinician-generated answers under our human evaluation framework, demonstrating the effectiveness of instruction prompt tuning.

However, the safety-critical nature of the domain necessitates thoughtful development of evaluation frameworks, enabling researchers to meaningfully measure progress and capture and mitigate potential harms. This is especially important for LLMs, since these models may produce generations misaligned with clinical and societal values. They may, for instance, hallucinate convincing medical misinformation or incorporate biases that could exacerbate health disparities.

To evaluate how well LLMs encode clinical knowledge and assess their potential in medicine, we consider medical question answering. This task is challenging: providing high-quality answers to medical questions requires comprehension of medical context, recall of appropriate medical knowledge, and reasoning with expert information. Existing medical question answering benchmarks [33] are often limited to assessing classification accuracy or automated natural language generation metrics (e.g., BLEU [67]), and do not enable the detailed analysis required for real-world clinical applications. This creates an unmet need for a broad medical question answering benchmark to assess LLMs’ response factuality, use of expert knowledge in medical and scientific reasoning, helpfulness, precision, health equity, and potential harm to humans accepting model outputs as facts.

To address this, we curate MultiMedQA, a benchmark comprising seven medical question answering datasets, including six existing datasets: MedQA [33], MedMCQA [64], PubMedQA [34], LiveQA [1], MedicationQA [2], and MMLU clinical topics [29]. We newly introduce the seventh dataset, HealthSearchQA, which consists of commonly searched health questions. To assess LLMs using MultiMedQA, we build on PaLM, a 540-billion parameter LLM [14], and its instruction- tuned variant Flan-PaLM [15]. Using a combination of few-shot [12], chain-of-thought (CoT) [91], and self-consistency [88] prompting strategies, Flan-PaLM achieves state-of-the-art (SOTA) performance on MedQA, MedMCQA, PubMedQA, and MMLU clinical topics, often outperforming several strong LLM baselines by a significant margin. On the MedQA dataset comprising USMLE questions, FLAN-PaLM exceeds previous SOTA by over 17%.

Despite Flan-PaLM’s strong performance on multiple-choice questions, its answers to consumer medical questions reveal key gaps. To resolve this, we propose instruction prompt tuning, a data- and parameter- efficient alignment technique, to further adapt Flan-PaLM to the medical domain. The resulting model, Med-PaLM, performs encouragingly on the axes of our pilot human evaluation framework. For example, a panel of clinicians judged only 61.9% of Flan-PaLM long-form answers to be aligned with scientific consensus, compared to 92.6% for Med-PaLM answers, on par with clinician-generated answers (92.9%). Similarly, 29.7% of Flan-PaLM answers were rated as potentially leading to harmful outcomes, in contrast with 5.8% for Med-PaLM, comparable with clinician-generated answers (6.5%).

While these results are promising, the medical domain is complex. Further evaluations are necessary, particularly along the dimensions of fairness, equity, and bias. Our work demonstrates that many limitations must be overcome before such models become viable for use in clinical applications. We outline some key limitations and directions of future research in our study.

Our key contributions are summarized below:

  • Approaches for evaluation of LLMs in medical question answering
    • Curation of HealthSearchQA and MultiMedQA We introduce HealthSearchQA, a dataset of 3375 commonly searched consumer medical questions. We present this dataset alongside six other existing open datasets for medical question answering, spanning medical exam, medical research, and consumer medical questions, as a diverse benchmark to assess the clinical knowledge and question answering capabilities of LLMs (see Section 3.1).
    • Pilot framework for human evaluation We pilot a framework for physician and lay user evaluation to assess multiple axes of LLM performance beyond accuracy on multiple-choice datasets. Our evaluation assesses answers for agreement with scientific and clinical consensus, likelihood and possible extent of harm, reading comprehension, recall of relevant clinical knowledge, manipulation of knowledge via valid reasoning, completeness of responses, potential for bias, relevance, and helpfulness (see Section 3.2).
  • State-of-the-art results on medical question answering benchmarks On the MedQA, MedMCQA, PubMedQA and MMLU clinical topics datasets, FLAN-PaLM achieves SOTA performance via a com- bination of prompting strategies, surpassing several strong LLM baselines. Specifically, we reach 67.6% accuracy on MedQA (more than 17% above prior SOTA), 57.6% on MedMCQA, and 79.0% on PubMedQA (see Section 4).
  • Instruction prompt tuning to align LLMs to the medical domain We introduce instruction prompt tuning, a simple, data- and parameter-efficient technique for aligning LLMs to the safety-critical medical domain (see Section 3.3.3). We leverage this to build Med-PaLM, an instruction prompt-tuned version of Flan-PaLM specialized for the medical domain. Our human evaluation framework reveals limitations of Flan-PaLM in scientific grounding, harm, and bias. However, Med-PaLM significantly reduces the gap (or even compares favorably) to clinicians on several of these axes, according to both clinicians and lay users (see Section 4.5).
  • Key limitations of LLMs revealed through our human evaluation While our results demonstrate the potential of LLMs in medicine, they also suggest several critical improvements are necessary in order to make these models viable for real-world clinical applications. We outline future research directions and mitigation strategies to address these challenges (see Section 6).

2 Related work

Large language models (LLMs) Over the past few years, LLMs have shown impressive performance on natural language processing (NLP) tasks [12, 14, 15, 30, 69, 70, 73, 89, 91, 99]. They owe their success to scaling up the training of transformer-based models [84]. It has been shown that model performance and data-efficiency scales with model size and dataset size [37]. LLMs are often trained using self-supervision on large scale, using general-purpose text corpi such as Wikipedia and BooksCorpus. They have demonstrated promising results across a wide range of tasks, including tasks that require specialized scientific knowledge and reasoning [17, 29]. Perhaps the most interesting aspect of these LLMs is their in-context few-shot abilities, which adapt these models to diverse tasks without gradient-based parameter updates [12, 40, 43, 89]. This allows them to rapidly generalize to unseen tasks and even exhibit apparent reasoning abilities with appropriate prompting strategies [14, 47, 79, 91].

Several studies have shown that LLMs have the capacity to act as implicit knowledge bases [29, 35, 79]. However, there is a significant risk of these models producing hallucinations, amplifying social biases present in their training data, and displaying deficiencies in their reasoning abilities. To examine the current limitations of LLMs and to quantify the large gap between human and LLM language capabilities, BIG-bench was introduced as a community-wide initiative to benchmark on tasks that were believed at time of publication to be beyond the capabilities of current language models [78].

LLMs for science and biomedicine]] Recent studies, such as SciBERT [5], BioNLP [46], BioMegatron [76], BioBERT [44], PubMedBERT [25], DARE [66], ScholarBERT [31], and BioGPT [56], have demonstrated the effectiveness of using curated scientific and biomedical corpora for both discriminative and generative language modeling. These models, although promising, are typically small in scale and scope compared to LLMs such as GPT-3 [12] and PaLM [14]. While the medical domain is challenging, specific proposals for LLMs have already included examples as varied as augmenting non-critical clinical assessments to summarisation of complex medical communications [3, 41, 75].

The closest precedents to our work are Taylor et al. [79], who introduced a LLM for science named Galactica, and Liévin et al. [50], who studied the reasoning capability of LLMs in the medical question answering context. In particular, Liévin et al. [50] used Instruct GPT-3, an instruction-tuned LLM [63], and applied chain-of-thought prompting [91] on top to improve the results on the MedQA, MedMCQA, and PubMedQA datasets.

3 Methods

...

...

References

;

 AuthorvolumeDate ValuetitletypejournaltitleUrldoinoteyear
2022 LargeLanguageModelsEncodeClinicYossi Matias
Blaise Agüera y Arcas
Greg S. Corrado
Martin Seneviratne
Aakanksha Chowdhery
Hyung Won Chung
Karan Singhal
Shekoofeh Azizi
Tao Tu
S Sara Mahdavi
Jason Wei
Nathan Scales
Ajay Tanwani
Heather Cole-Lewis
Stephen Pfohl
Perry Payne
Paul Gamble
Chris Kelly
Nathaneal Scharli
Philip Mansfield
Dale Webster
Katherine Chou
Juraj Gottweis
Nenad Tomasev
Yun Liu
Alvin Rajkomar		Large Language Models Encode Clinical Knowledge2022
Retrieved from "http://www.gabormelli.com/RKB/index.php?title=2022_LargeLanguageModelsEncodeClinic&oldid=862116"