方法：收集2020-07/2021-07就诊于宁夏银川两个社区卫生服务中心的2型糖尿病患者1 358例2 707眼的眼底彩照，采用Eye Wisdom AI眼病辅助筛查和诊断系统自动检测分析出血、微动脉瘤以及视网膜内微血管异常等DR的特征性改变。根据其国际分期的标准对眼底彩照检测结果进行自动分级，由人工分析组进行图像判读后反馈结果，分析AI辅助筛查系统诊断DR的灵敏度、特异度、误诊率及漏诊率，比较AI与人工分析的一致性，对AI筛查系统与人工分析的结果做Kappa一致性检验。

结果：与人工分析相比，AI诊断有无DR的灵敏度为91.84%，特异度为99.06%，漏诊率为8.16%，误诊率为0.94%，对于二者诊断结果的一致性分析Kappa值为0.817(P<0.001)。与人工分析相比，AI组检测无DR的灵敏度为99.06%，特异度为91.84%； 检测轻度NPDR的灵敏度为85.36%，特异度为98.52%； 中度NPDR的灵敏度为81.53%，特异度为98.55%； 重度NPDR的灵敏度为70%，特异度为99.51%； PDR的灵敏度为86.67%，特异度为99.63%。二者对DR分期诊断一致性分析的Kappa值为0.878(P<0.01)。

结论：AI辅助筛查系统与人工分析的结果一致性好，可以满足DR筛查的需求，为基层社区DR患者提供了一种新的有效防治模式。

METHODS:From July 2020 to July 2021, fundus photograph of 2 707 eyes from 1 358 diabetic patients with type 2 diabetes in two communities of Ningxia and Yinchuan were included in this study. The Eye Wisdom AI assisted screening and diagnosis system was used to analyze automatically and detect the characteristic changes of DR, such as hemorrhage, microaneurysms and retinal microvascular abnormalities. The results of fundus photograph were automatically graded according to the standard of DR international stage standard. The manual analysis group gave feedback after image interpretation, analyzed the sensitivity, specificity, misdiagnosis rate and missed diagnosis rate of the AI-assisted screening system for DR diagnosis, and compared the consistency between AI and manual analysis. Kappa consistency test was performed for the results of AI screening system and manual analysis.

RESULTS:Compared with manual analysis, the sensitivity, specificity, missed diagnosis rate and misdiagnosis rate of AI were 91.84%, 99.06%, 8.16% and 0.94% respectively. The Kappa value of consistency analysis of the two diagnosis results was 0.817(P<0.001). Compared with manual analysis, the sensitivity and specificity of AI group to diagnose non-DR were 99.06% and 91.84% respectively. The sensitivity and specificity of mild NPDR were 85.36% and 98.52% respectively. The sensitivity and specificity of moderate NPDR were 81.53% and 98.55% respectively. The sensitivity and specificity of severe NPDR were 70% and 99.51% respectively. The sensitivity and specificity of PDR were 86.67% and 99.63% respectively. The Kappa value of the consistency analysis of DR staging diagnosis was 0.878(P<0.01).

CONCLUSION: The AI remote screening system adopted in this study showed good consistency with the results of manual analysis, which can meet the needs of DR screening and provide a new effective prevention and treatment mode for DR patients in the community.