Department of Ophthalmology and Visual Sciences, Aga Khan University Hospital, National Stadium Road, Karachi, Pakistan.
Department of Ophthalmology and Visual Sciences, Aga Khan University Hospital, National Stadium Road, Karachi, Pakistan.
Diabetes Res Clin Pract. 2023 Nov;205:110943. doi: 10.1016/j.diabres.2023.110943. Epub 2023 Oct 5.
Diabetic retinopathy (DR) is a major cause of blindness globally, early detection is critical to prevent vision loss. Traditional screening that, rely on human experts are, however, costly, and time-consuming. The purpose of this systematic review is to assess the diagnostic accuracy of smartphone-based artificial intelligence(AI) systems for DR detection.
Literature review was conducted on MEDLINE, Embase, Scopus, CINAHL Plus, and Cochrane from inception to December 2022. We included diagnostic test accuracy studies evaluating the use of smartphone-based AI algorithms for DR screening in patients with diabetes, with expert human grader as the reference standard. Random-effects model was used to pool sensitivity and specificity. Any DR(ADR) and referable DR(RDR) were analyzed separately.
Out of 968 identified articles, six diagnostic test accuracy studies met our inclusion criteria, comprising 3,931 patients. Four of these studies used the Medios AI algorithm. The pooled sensitivity and specificity for diagnosis of ADR were 88 % and 91.5 % respectively and for diagnosis of RDR were 98.2 % and 81.2 % respectively. The overall risk of bias across the studies was low.
Smartphone-based AI algorithms show high diagnostic accuracy for detecting DR. However, more high-quality comparative studies are needed to evaluate the effectiveness in real-world clinical settings.
糖尿病视网膜病变（DR）是全球范围内导致失明的主要原因，早期发现对于防止视力丧失至关重要。然而，传统的依赖人类专家的筛查既昂贵又耗时。本系统评价旨在评估基于智能手机的人工智能（AI）系统在 DR 检测中的诊断准确性。
从建库至 2022 年 12 月，我们在 MEDLINE、Embase、Scopus、CINAHL Plus 和 Cochrane 上进行了文献回顾。我们纳入了评估智能手机 AI 算法用于糖尿病患者 DR 筛查的诊断准确性的研究，以专家人工分级作为参考标准。我们使用随机效应模型来汇总敏感性和特异性。分别分析任何 DR（ADR）和可治疗 DR（RDR）。
在 968 篇已识别的文章中，有 6 项符合纳入标准的诊断准确性研究，共纳入 3931 名患者。其中 4 项研究使用了 Medios AI 算法。用于诊断 ADR 的汇总敏感性和特异性分别为 88%和 91.5%，用于诊断 RDR 的汇总敏感性和特异性分别为 98.2%和 81.2%。研究的总体偏倚风险较低。
基于智能手机的 AI 算法在检测 DR 方面具有较高的诊断准确性。然而，需要更多高质量的比较研究来评估其在实际临床环境中的有效性。
