弥合耳科学与神经耳科学人工智能的转化差距：临床就绪度、局限性与研究优先事项

文献信息

中文摘要

背景与目标： 人工智能（AI）在耳科学与神经耳科学领域已产生大量概念验证应用，但相对较少进入常规临床使用。本叙述性综述不仅评估AI在听力学与前庭医学中能做什么，还评估各项应用距离临床转化有多近，以及哪些障碍仍在限制其落地实施。
材料与方法： 检索PubMed/MEDLINE、Scopus和Web of Science，时间自建库至2026年7月31日，使用与AI、机器学习、深度学习、听力损失、人工耳蜗植入、耳鸣、前庭疾病、眼震、颞骨影像、数字表型、康复和机器人手术相关的术语组合。同时筛查相关出版物的参考文献列表。证据按主题综合，并依据任务成熟度、外部验证、临床效用、工作流程整合和患者安全要求进行解读。
结果： 最成熟的应用是狭窄、定义明确的任务，涉及图像或信号分类和解剖分割，包括听力图模式识别、眼震自动提取、颞骨分割和前庭神经鞘瘤体积测量。相比之下，多模态预后模型、AI引导的治疗选择、数字表型、自主决策支持以及机器人或术中系统在临床上仍较不成熟。跨领域而言，转化受限于小型回顾性数据集、中心和设备特异的采集模式、不一致的参考标准、有限的外部验证与前瞻性验证、校准和失效模式报告不足，以及改善患者结局或工作流程效率的证据稀缺。
结论： 耳科学与神经耳科学中AI的主要挑战已不再是技术可行性，而是证明可推广的临床价值。未来研究应优先考虑具有代表性的多中心数据、独立验证、前瞻性影响研究、具有临床意义的对照、互操作性、可解释性、公平性以及部署后监测。AI应作为可审计、有人类监督的技术实施，以增强专科判断和以患者为中心的照护。

英文摘要

Background and Objectives: Artificial intelligence (AI) has generated numerous proof-of-concept applications in otology and neurotology, but relatively few have progressed to routine clinical use. This narrative review evaluates not only what AI can do across audiovestibular medicine, but also how close individual applications are to clinical translation and which barriers continue to limit implementation. Materials and Methods: PubMed/MEDLINE, Scopus, and Web of Science were searched from database inception to 31 July 2026 using combinations of terms related to AI, machine learning, deep learning, hearing loss, cochlear implantation, tinnitus, vestibular disorders, nystagmus, temporal bone imaging, digital phenotyping, rehabilitation, and robotic surgery. Reference lists of relevant publications were also screened. Evidence was synthesized thematically and interpreted according to task maturity, external validation, clinical utility, workflow integration, and patient-safety requirements. Results: The most mature applications are narrow, well-defined tasks involving image or signal classification and anatomical segmentation, including audiogram pattern recognition, automated nystagmus extraction, temporal-bone segmentation, and vestibular schwannoma volumetry. By contrast, multimodal prognostic models, AI-guided treatment selection, digital phenotyping, autonomous decision support, and robotic or intraoperative systems remain less clinically established. Across domains, translation is constrained by small retrospective datasets, center- and device-specific acquisition patterns, inconsistent reference standards, limited external and prospective validation, inadequate reporting of calibration and failure modes, and scarce evidence of improved patient outcomes or workflow efficiency. Conclusions: The principal challenge for AI in otology and neurotology is no longer technical feasibility, but demonstration of generalizable clinical value. Future research should prioritize representative multicenter data, independent validation, prospective impact studies, clinically meaningful comparators, interoperability, explainability, fairness, and post-deployment surveillance. AI should be implemented as an auditable, human-supervised technology that strengthens specialist judgment and patient-centered care.