多模态人工智能在耳鼻咽喉科临床应用的最新综述
本综述检索六大数据库至2025年3月4日,纳入44项多模态AI在耳鼻咽喉科应用的研究,其中55%发表于2024年、18%发表于2025年。头颈癌是最常见亚专科(75%),所有研究用于临床教育(9%)或决策支持(91%),但多数限于小样本单中心且缺乏前瞻性验证。
要点
- 头颈癌是主要应用领域(75%),图像与文本组合最常见(80%)。
- 模型用于决策支持(91%)和教育(9%),性能差异大且缺乏前瞻验证。
核心结论
- 多模态AI在耳鼻咽喉科应用迅速扩展,尤以头颈癌为焦点,但需严格验证与伦理监管以确保安全、公平、有效。
研究设计与样本
- 遵循PRISMA-ScR指南的范围综述,检索MEDLINE、EMBASE、SCOPUS、Cochrane Library、Web of Science和CINAHL数据库,从建库至2025年3月4日,纳入44项研究。
主要发现
- 44项研究中55%(24/44)发表于2024年,18%(8/44)发表于2025年。
- 图像与文本是最常见组合模态(80%,35/44),视频与向量数据组合占2%(1/44),组学与文本和/或图像组合占14%(6/44)。
- 头颈癌是最常见亚专科(75%,33/44),其次为普通耳鼻咽喉科(11%,5/44)。
- 所有研究用于临床教育(9%,4/44)或决策支持(91%,40/44),评估领域包括 board-style 考试表现(准确率37%—86%)和疾病分类与预后(AUC 0.65—0.96)。
- 多数研究限于小样本单中心且缺乏前瞻性验证,模型错误、数据集偏倚和语言局限需进一步改进。
临床怎么用
- 临床医生应理解多模态大语言模型的能力与局限,在耳鼻咽喉科诊疗中审慎应用。
- 需进行严格验证和伦理监督,以确保多模态AI的安全、公平和有效采纳。
局限性
- 多数研究为小样本、单中心设计,缺乏前瞻性验证。
- 存在模型错误、数据集偏倚和语言局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: Artificial intelligence (AI) has advanced to simultaneously process visual, auditory, and textual inputs, providing users with "multimodal" AI. Given the clinical integration potential of these tools, otolaryngologists must stay informed. This study reviews current literature on applications of multimodal AI in otolaryngology. DATA SOURCES: The MEDLINE, EMBASE, SCOPUS, Cochrane Library, Web of Science, and CINAHL databases. REVIEW METHODS: Databases were searched from the date of inception to March 4, 2025, following Preferred Reporting Items for Systematic Reviews and Meta-analyses extension for scoping reviews (PRISMA-ScR) guidelines. Studies on any application of multimodal AI in otolaryngology were included. CONCLUSIONS: Forty-four studies were included, with 55% (24/44) published in 2024 and 18% (8/44) in 2025. Image and text were the most commonly combined modalities (80%, 35/44), with emerging combinations including video with vector data (2%,1/44) and omics with text and/or image (14%, 6/44). Head and neck cancer was the most common subspecialty of focus (75%, 33/44), followed by general ear, nose, and throat (ENT) (11%, 5/44). All studies applied the models for clinical education (9%, 4/44) or decision support (91%, 40/44), assessing performance in areas such as board-style examination performance (accuracy: 37%-86%) or disease classification and prognostication (area under the receiver operating characteristic curve [AUC] 0.65-0.96). However, most studies were limited to small, single-institution samples and lacked prospective validation. Model error, data set bias, and language limitations underscore the need for further refinement. IMPLICATIONS FOR PRACTICE: The application of multimodal large language models (LLMs) in otolaryngology is rapidly expanding. Clinicians must understand both the capabilities and limitations of these systems. Rigorous validation and ethical oversight will be essential to ensure the safe, equitable, and effective adoption in otolaryngologic care.
引用
- Li YJ, Su F, Banyi N, et al. Clinical Applications of Multimodal Artificial Intelligence in Otolaryngology: A State-of-the-Art Review. Otolaryngol Head Neck Surg. 2026. PMID: 42117403. doi:10.1002/ohn.70285
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。