文献速递眼底病与葡萄膜炎

台湾真实世界实践中人工智能糖尿病视网膜病变诊断系统的应用评估

REAL-WORLD PRACTICE OF ARTIFICIAL INTELLIGENCE DIAGNOSTIC SYSTEM FOR DIABETIC RETINOPATHY IN TAIWAN.
PubMed 文献 · Retina · 2026-01-01

该回顾性研究纳入台湾糖尿病共享照护网络的11,713例糖尿病患者,评估AI系统VeriSee DR在真实临床流程中筛查需转诊糖尿病视网膜病变(RDR)的表现。VeriSee DR检测RDR的敏感度为0.88、特异度为0.86、AUC为0.87,但性能随年龄增长而下降;AI转诊率高于眼科医师(27.45% vs. 18.15%),却遗漏了青光眼等部分需转诊眼病。

要点

  • VeriSee DR检测RDR敏感度0.88、特异度0.86、AUC 0.87,阴性预测值0.97。
  • AI系统性能随年龄增长而下降,性别分布在各年龄组间保持一致。
  • AI转诊RDR比例高于眼科医师(27.45% vs. 18.15%),但遗漏青光眼等病变。

核心结论

  • VeriSee DR在检测RDR方面具有较高准确度,但在老年患者中性能下降。
  • AI系统转诊率高于眼科医师,却遗漏了青光眼等部分需转诊的眼病,需进一步优化以支持更广泛的眼病检测。

研究设计与样本

  • 回顾性研究,纳入政府主导的糖尿病共享照护网络中11,713例糖尿病患者。
  • 将AI系统VeriSee DR整合入临床工作流程以识别需转诊糖尿病视网膜病变(RDR),并与眼科医师分级在患者层面进行比较。
  • 按年龄和性别进行亚组分析,并记录眼科医师识别的其他转诊疾病。

主要发现

  • VeriSee DR检测RDR的敏感度为0.88、特异度为0.86、准确度为0.86、阳性预测值为0.58、阴性预测值为0.97、AUC为0.87。
  • AI系统性能随年龄增长而下降,性别分布在各年龄组间保持一致。
  • AI系统识别的RDR比例高于眼科医师(27.45% vs. 18.15%)。
  • 除1,818例RDR患者外,眼科医师在4.5%的病例中识别出其他需转诊的眼部疾病。
  • AI系统转诊年龄相关性黄斑变性(2-4级),而对黄斑裂孔和黄斑水肿(1-2级)的转诊决策不一致;临床医师识别的青光眼(0-1级)未被一致转诊。

关键数值与推荐

  • VeriSee DR检测RDR的敏感度0.88、特异度0.86、准确度0.86、阳性预测值0.58、阴性预测值0.97、AUC 0.87。
  • AI系统RDR转诊比例为27.45%,眼科医师为18.15%。
  • 眼科医师在4.5%的病例中识别出其他需转诊的眼部疾病。

临床怎么用

  • VeriSee DR可用于糖尿病视网膜病变筛查,但需关注老年患者中性能下降的问题。
  • AI系统转诊率较高,可能增加转诊负担,且需注意其遗漏青光眼等病变的风险。
  • 在真实临床实践中,AI筛查应结合眼科医师评估以全面发现其他需转诊眼病。

局限性

  • 本研究为回顾性研究,可能存在选择偏倚。
  • AI系统对青光眼等部分眼病的识别不足,需进一步优化。
  • 研究仅纳入台湾糖尿病共享照护网络人群,外推性受限。

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

PURPOSE: The authors evaluated the alterations of applying artificial intelligence (AI) diagnostic system for diabetic retinopathy screening in real-world practice. METHODS: This retrospective study included 11,713 diabetic patients from the government-led Diabetes Shared Care Network. The AI system VeriSee DR was integrated into the clinical workflow to identify referable diabetic retinopathy (RDR). Its performance was compared with ophthalmologist grading at the patient level using sensitivity, specificity, accuracy, positive predictive value, negative predictive value, and area under the receiver operating characteristic curve. Subgroup analysis was performed by age and sex, with additional referral diseases identified by ophthalmologists. RESULTS: VeriSee DR achieved a sensitivity of 0.88, specificity of 0.86, accuracy of 0.86, positive predictive value of 0.58, negative predictive value of 0.97, and area under the receiver operating characteristic curve of 0.87 in detecting RDR. Performance declined with increasing age, whereas sex distribution remained consistent across age groups. The AI system identified a higher proportion of RDR than ophthalmologists (27.45% vs. 18.15%). In addition to 1,818 patients with RDR, ophthalmologists identified other referral-warranted ocular conditions in 4.5% of cases. The AI system referred age-related macular degeneration (Grades 2-4), whereas referral decisions for macular hole and macular edema (Grades 1-2) varied; however, glaucoma (Grades 0-1) identified by clinicians was not consistently referred. CONCLUSION: VeriSee DR demonstrated high accuracy in detecting RDR but exhibited reduced performance in older patients. It had a higher referral rate than ophthalmologists yet missed certain conditions such as glaucoma. Despite effectiveness in diabetic retinopathy screening, further refinement is required to support broader ophthalmic disease detection.

来源:Retina,查看原文。版权归原作者所有。

引用

  1. Lin CC, Cheng CK, Peng PH, et al. REAL-WORLD PRACTICE OF ARTIFICIAL INTELLIGENCE DIAGNOSTIC SYSTEM FOR DIABETIC RETINOPATHY IN TAIWAN. Retina. 2026. PMID: 41818422. doi:10.1097/IAE.0000000000004826

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。