人工智能驱动的眼球震颤分析在前庭评估中的应用：当前证据与未来方向。

文献信息

中文摘要

背景： 前庭疾病常见、异质性强且诊断具有挑战性。眼球震颤是前庭评估中的重要体征，但其解读可能取决于临床医生的专业知识和记录质量。本结构化叙述性综述梳理了人工智能（AI）驱动的眼球震颤和眼动信号分析在前庭评估中的当前证据，重点关注研究特征、输入数据、AI任务、模型架构、验证方法、性能指标、可解释性和临床适用性。方法：通过结构化PubMed检索，识别出开发、验证或评估机器学习、深度学习或相关计算方法用于眼球震颤分析或前庭诊断支持的人体研究。结果：共纳入15项研究，大多数发表于2021年及以后。证据以技术和早期模型开发研究为主，使用VNG/VOG记录、其他眼动记录或衍生特征，如瞳孔轨迹、光流图、眼部和头部轨迹以及慢相速度测量。应用包括眼球震颤检测、瞳孔追踪和信号提取、模式或方向分类、量化或定位，以及与良性阵发性位置性眩晕相关的诊断分类。深度学习方法，尤其是卷积神经网络和CNN-循环混合模型，被普遍使用。然而，验证主要为内部验证，性能报告异质性大，可解释性有限，前瞻性工作流程评估或常规临床实施不常见。结论：AI驱动的眼球震颤分析显示出有前景的临床潜力，特别是在基于VNG/VOG的评估和BPPV相关决策支持方面。要转化为常规实践，需要多中心患者层面的外部验证、透明报告、可解释输出、标准化性能指标以及在真实临床路径中的前瞻性评估。

英文摘要

Vestibular disorders are common, heterogeneous, and diagnostically challenging. Nystagmus is an important sign in vestibular assessment, but its interpretation can depend on clinician expertise and recording quality. This structured narrative review mapped current evidence on artificial intelligence (AI)-enabled nystagmus and eye-movement signal analysis in vestibular assessment, focusing on study characteristics, input data, AI tasks, model architectures, validation approaches, performance metrics, explainability, and clinical applicability. A structured PubMed search identified human studies that developed, validated, or evaluated machine-learning, deep-learning, or related computational methods for nystagmus analysis or vestibular diagnostic support. Fifteen studies were included, most published from 2021 onwards. The evidence was dominated by technical and early-stage model-development studies using VNG/VOG recordings, other eye-movement recordings, or derived features such as pupil trajectories, optical-flow maps, eye and head traces, and slow-phase velocity measures. Applications included nystagmus detection, pupil tracking and signal extraction, pattern or direction classification, quantification or localisation, and benign paroxysmal positional vertigo-related diagnostic classification. Deep-learning approaches, particularly convolutional neural networks and CNN-recurrent hybrid models, were commonly used. However, validation was mainly internal, performance reporting was heterogeneous, explainability was limited, and prospective workflow evaluation or routine clinical implementation was uncommon. AI-enabled nystagmus analysis shows promising clinical potential, particularly for VNG/VOG-based assessment and BPPV-related decision support. Translation into routine practice will require multicentre patient-level external validation, transparent reporting, interpretable outputs, standardised performance metrics, and prospective evaluation in real clinical pathways.