Thomas Savage, John Wang, Robert Gallo, Abdessalem Boukil, Vishwesh Patel, Seyed Amir Ahmad Safavi-Naini, Ali Soroush, Jonathan H Chen
{"title":"大语言模型不确定性代理:医学诊断和治疗的鉴别与校准。","authors":"Thomas Savage, John Wang, Robert Gallo, Abdessalem Boukil, Vishwesh Patel, Seyed Amir Ahmad Safavi-Naini, Ali Soroush, Jonathan H Chen","doi":"10.1093/jamia/ocae254","DOIUrl":null,"url":null,"abstract":"<p><strong>Introduction: </strong>The inability of large language models (LLMs) to communicate uncertainty is a significant barrier to their use in medicine. Before LLMs can be integrated into patient care, the field must assess methods to estimate uncertainty in ways that are useful to physician-users.</p><p><strong>Objective: </strong>Evaluate the ability for uncertainty proxies to quantify LLM confidence when performing diagnosis and treatment selection tasks by assessing the properties of discrimination and calibration.</p><p><strong>Methods: </strong>We examined confidence elicitation (CE), token-level probability (TLP), and sample consistency (SC) proxies across GPT3.5, GPT4, Llama2, and Llama3. Uncertainty proxies were evaluated against 3 datasets of open-ended patient scenarios.</p><p><strong>Results: </strong>SC discrimination outperformed TLP and CE methods. SC by sentence embedding achieved the highest discriminative performance (ROC AUC 0.68-0.79), yet with poor calibration. SC by GPT annotation achieved the second-best discrimination (ROC AUC 0.66-0.74) with accurate calibration. Verbalized confidence (CE) was found to consistently overestimate model confidence.</p><p><strong>Discussion and conclusions: </strong>SC is the most effective method for estimating LLM uncertainty of the proxies evaluated. SC by sentence embedding can effectively estimate uncertainty if the user has a set of reference cases with which to re-calibrate their results, while SC by GPT annotation is the more effective method if the user does not have reference cases and requires accurate raw calibration. Our results confirm LLMs are consistently over-confident when verbalizing their confidence (CE).</p>","PeriodicalId":50016,"journal":{"name":"Journal of the American Medical Informatics Association","volume":null,"pages":null},"PeriodicalIF":4.7000,"publicationDate":"2024-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Large language model uncertainty proxies: discrimination and calibration for medical diagnosis and treatment.\",\"authors\":\"Thomas Savage, John Wang, Robert Gallo, Abdessalem Boukil, Vishwesh Patel, Seyed Amir Ahmad Safavi-Naini, Ali Soroush, Jonathan H Chen\",\"doi\":\"10.1093/jamia/ocae254\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p><strong>Introduction: </strong>The inability of large language models (LLMs) to communicate uncertainty is a significant barrier to their use in medicine. Before LLMs can be integrated into patient care, the field must assess methods to estimate uncertainty in ways that are useful to physician-users.</p><p><strong>Objective: </strong>Evaluate the ability for uncertainty proxies to quantify LLM confidence when performing diagnosis and treatment selection tasks by assessing the properties of discrimination and calibration.</p><p><strong>Methods: </strong>We examined confidence elicitation (CE), token-level probability (TLP), and sample consistency (SC) proxies across GPT3.5, GPT4, Llama2, and Llama3. Uncertainty proxies were evaluated against 3 datasets of open-ended patient scenarios.</p><p><strong>Results: </strong>SC discrimination outperformed TLP and CE methods. SC by sentence embedding achieved the highest discriminative performance (ROC AUC 0.68-0.79), yet with poor calibration. SC by GPT annotation achieved the second-best discrimination (ROC AUC 0.66-0.74) with accurate calibration. Verbalized confidence (CE) was found to consistently overestimate model confidence.</p><p><strong>Discussion and conclusions: </strong>SC is the most effective method for estimating LLM uncertainty of the proxies evaluated. SC by sentence embedding can effectively estimate uncertainty if the user has a set of reference cases with which to re-calibrate their results, while SC by GPT annotation is the more effective method if the user does not have reference cases and requires accurate raw calibration. Our results confirm LLMs are consistently over-confident when verbalizing their confidence (CE).</p>\",\"PeriodicalId\":50016,\"journal\":{\"name\":\"Journal of the American Medical Informatics Association\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":4.7000,\"publicationDate\":\"2024-10-12\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Journal of the American Medical Informatics Association\",\"FirstCategoryId\":\"91\",\"ListUrlMain\":\"https://doi.org/10.1093/jamia/ocae254\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, INFORMATION SYSTEMS\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of the American Medical Informatics Association","FirstCategoryId":"91","ListUrlMain":"https://doi.org/10.1093/jamia/ocae254","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, INFORMATION SYSTEMS","Score":null,"Total":0}
Large language model uncertainty proxies: discrimination and calibration for medical diagnosis and treatment.
Introduction: The inability of large language models (LLMs) to communicate uncertainty is a significant barrier to their use in medicine. Before LLMs can be integrated into patient care, the field must assess methods to estimate uncertainty in ways that are useful to physician-users.
Objective: Evaluate the ability for uncertainty proxies to quantify LLM confidence when performing diagnosis and treatment selection tasks by assessing the properties of discrimination and calibration.
Methods: We examined confidence elicitation (CE), token-level probability (TLP), and sample consistency (SC) proxies across GPT3.5, GPT4, Llama2, and Llama3. Uncertainty proxies were evaluated against 3 datasets of open-ended patient scenarios.
Results: SC discrimination outperformed TLP and CE methods. SC by sentence embedding achieved the highest discriminative performance (ROC AUC 0.68-0.79), yet with poor calibration. SC by GPT annotation achieved the second-best discrimination (ROC AUC 0.66-0.74) with accurate calibration. Verbalized confidence (CE) was found to consistently overestimate model confidence.
Discussion and conclusions: SC is the most effective method for estimating LLM uncertainty of the proxies evaluated. SC by sentence embedding can effectively estimate uncertainty if the user has a set of reference cases with which to re-calibrate their results, while SC by GPT annotation is the more effective method if the user does not have reference cases and requires accurate raw calibration. Our results confirm LLMs are consistently over-confident when verbalizing their confidence (CE).
期刊介绍:
JAMIA is AMIA''s premier peer-reviewed journal for biomedical and health informatics. Covering the full spectrum of activities in the field, JAMIA includes informatics articles in the areas of clinical care, clinical research, translational science, implementation science, imaging, education, consumer health, public health, and policy. JAMIA''s articles describe innovative informatics research and systems that help to advance biomedical science and to promote health. Case reports, perspectives and reviews also help readers stay connected with the most important informatics developments in implementation, policy and education.