基于小波尺度图的眼震分类用于良性阵发性位置性眩晕亚型的自动诊断

文献信息

中文摘要

背景： 良性阵发性位置性眩晕（BPPV）是前庭性眩晕最常见的原因，但若无专科经验，准确诊断仍具挑战性。本研究旨在开发并评估一种基于视频眼震图（VNG）记录的BPPV人工智能（AI）诊断算法。方法：在一家三级转诊医院共收集了1352份VNG记录，来自后半规管（PSCC）BPPV患者（n=234）、水平半规管（LSCC）BPPV患者（管石症，n=188；嵴帽结石症，n=58）以及健康对照（n=224），共产生1738个信号片段（1312个由两名耳鼻咽喉科医师标注的眼震片段和426个正常片段）。使用基于深度学习的VNG眼动追踪算法提取眼轨迹信号，通过连续小波变换（CWT）转换为二维（2D）尺度图表示，并在患者水平数据划分下通过系统比较选择任务特异性骨干架构进行分类。消融分析显示，基于CWT的2D表示显著优于原始一维（1D）时间序列基线。最终模型在水平眼震分类和扭转眼震分类中分别达到受试者工作特征曲线下面积（AUROC）0.974和0.961。对于BPPV亚型诊断，记录水平准确率在PSCC评估中为0.935，在LSCC评估中为0.938；所有PSCC类别的F1分数均超过0.84，LSCC类别的F1分数范围为0.59至0.98，其中管石症变异型优于嵴帽结石症变异型。结论：这些发现表明基于VNG记录进行AI辅助BPPV诊断具有可行性。

英文摘要

Benign paroxysmal positional vertigo (BPPV) is the most common cause of vestibular vertigo, yet accurate diagnosis remains challenging without specialist expertise. This study aimed to develop and evaluate an artificial intelligence (AI)-based diagnostic algorithm for BPPV using video-nystagmography (VNG) recordings. A total of 1352 VNG recordings were collected from patients with posterior semicircular canal (PSCC) BPPV (n = 234), lateral semicircular canal (LSCC) BPPV (canalolithiasis, n = 188; cupulolithiasis, n = 58), and healthy controls (n = 224) at a single tertiary referral hospital, yielding 1738 signal segments (1312 nystagmus segments annotated by two otolaryngologists and 426 normal segments). Eye trajectory signals were extracted using a deep learning-based VNG eye-tracking algorithm, transformed into two-dimensional (2D) scalogram representations via continuous wavelet transform (CWT), and classified using task-specific backbone architectures selected through systematic comparison under a patient-level data split. Ablation analysis showed that CWT-based 2D representations significantly outperformed raw one-dimensional (1D) time-series baselines. The final models achieved area under the receiver operating characteristic curve (AUROC) values of 0.974 for horizontal and 0.961 for torsional nystagmus classification. For BPPV subtype diagnosis, recording-level accuracy was 0.935 for the PSCC and 0.938 for the LSCC evaluation; F1-scores exceeded 0.84 for all PSCC classes and ranged from 0.59 to 0.98 for LSCC classes, with canalolithiasis variants outperforming cupulolithiasis variants. These findings suggest the feasibility of AI-assisted BPPV diagnosis from VNG recordings.