机器学习算法识别与抗癫痫药物氨己烯酸相关的视野缺损——一项概念验证研究
本研究开发了两种机器学习模式识别算法,依照欧洲药品管理局批准的方案(FF135筛查与C30-2T阈值检测)客观识别氨己烯酸相关视野缺损(VAVFL)。算法与金标准临床判读高度一致,FF135敏感度22/23、特异度30/30,C30-2T敏感度17/18、特异度48/51;对称化处理有助于识别单眼学习或疲劳影响及合并同向性缺损时的VAVFL。
要点
- 两种机器学习算法依照欧洲药品管理局批准方案客观识别氨己烯酸相关视野缺损。
- FF135算法与金标准临床判读一致:敏感度22/23,特异度30/30。
- C30-2T算法与金标准临床判读一致:敏感度17/18,特异度48/51。
核心结论
- 可直接解释的机器学习结果正确识别了氨己烯酸相关视野缺损,可能有助于社区(神经)眼科的患者管理。
研究设计与样本
- 研究开发两种机器学习模式识别算法,依照欧洲药品管理局批准的氨己烯酸相关视野缺损检测方案(FF135筛查与C30-2T阈值检测,Humphrey视野分析仪)进行。
- 每个算法将每只眼实测视野与按严重程度匹配的参考模式比较相似性,参考模式客观来源于既往描述的123名成人病例系列。
- 算法可选加入对称化处理,该技术基于氨己烯酸相关视野缺损的双眼镜像对称性增强信噪比。
- 算法效用在一组89名连续识别个体中评估,该组按六种诊断类别分层,包括同向性和青光眼性视野缺损。
主要发现
- 算法与金标准临床判读高度一致:FF135敏感度22/23、特异度30/30;C30-2T敏感度17/18、特异度48/51。
- 对称化处理在单眼受视野学习或疲劳影响时,以及在合并同向性缺损时,对识别氨己烯酸相关视野缺损特别有用。
关键数值与推荐
- FF135算法与金标准临床判读一致的敏感度为22/23,特异度为30/30。
- C30-2T算法与金标准临床判读一致的敏感度为17/18,特异度为48/51。
临床怎么用
- 原文结论指出,可直接解释的机器学习结果可能有助于社区(神经)眼科的患者管理。
局限性
- 原文未报告局限性。
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
BACKGROUND/AIMS: The antiseizure medication, vigabatrin, is associated with visual field loss (VAVFL). However, the fields can be challenging to interpret due to unfamiliarity with the characteristics of the defect and/or to difficulty in obtaining a reliable examination, particularly in patients with cognitive limitations associated with the epilepsy. Two machine-learning pattern recognition algorithms were developed to identify VAVFL, objectively. METHODS: The algorithms adhered to the European Medicines Agency-approved protocol for the detection of VAVFL (Three Zone Age Corrected Full Field 135 Screening Test (FF135) and the Central C30-2 Threshold Test (C30-2T) with the Humphrey Field Analyzer). Each algorithm compared the similarity of the measured field from each eye to that of modelled reference patterns of VAVFL, matched for equivalent severity, and objectively derived from a previously described case series of 123 adults. The algorithms were augmented by the optional inclusion of symmetrisation, a signal-to-noise enhancement technique based on the between-eye mirror image symmetry of VAVFL. Utility of the algorithms for identifying VAVFL was evaluated against a case series of 89 consecutively identified individuals stratified across six diagnostic categories including homonymous and glaucomatous losses. RESULTS: The algorithms exhibited excellent agreement with a 'gold standard' clinical interpretation (sensitivity and specificity: FF135, 22/23; 30/30; C30-2T, 17/18; 48/51). Symmetrisation was particularly useful in identifying VAVFL when perimetric learning or fatigue influenced the outcome for one eye and for visualisation in the presence of concomitant homonymous loss. CONCLUSION: The directly interpretable machine-learning outcome correctly identified VAVFL and could assist patient management in community (neuro-)ophthalmology.
引用
- Wild JM, Smith PEM, Knupp C, et al. Machine-learning algorithms for the identification of visual field loss associated with the antiseizure medication vigabatrin-a proof of concept. Br J Ophthalmol. 2025. PMID: 40101935. doi:10.1136/bjo-2024-325804
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。