主要成果和措施：基于计算机算法的IRIS的灵敏度和假阴性率与阅读中心解释具有相似的图像。结果:共总纳入15015例连续患者（年龄18~98岁）；平均有54.3年被知悉有Ⅰ型或Ⅱ型糖尿病并接受非散瞳眼底照相用于糖尿病性视网膜病扫描检查。IRIS算法在检测视力威胁糖尿病的敏感性与读取中心算法相比为66.4%（95% CI， 62.8%-69.9%），假阴性率为2%。特异性为72.8％（95％CI，72.0％-73.5％）。在一个人口中有15.%有威胁视力的糖尿病眼病的人群中，IRIS算法阳性预测值为10.8% (95% CI， 9.6%-11.9%)，阴性预测率为97.8%(95% CI， 96.8%-98.6%)。
IMPORTANCE:Diabetic retinopathy is a leading cause of blindness， but its detrimental effects are preventable with early detection and treatment. Screening for diabetic retinopathy has the potential to increase the number of cases treated early， especially in populations with limited access to care.
OBJECTIVE:To determine the efficacy of an automated algorithm in interpreting screening ophthalmoscopic photographs from patients with diabetes compared with a reading center interpretation.
DESIGN， SETTING， AND PARTICIPANTS:Retrospective cohort analysis of 15015 patients with type 1 or 2 diabetes in the Harris Health System in Harris County， Texas， who had undergone a retinal screening examination and nonmydriatic fundus photography via the Intelligent Retinal Imaging System (IRIS) from June 2013 to April 2014 were included. The IRIS-based interpretations were compared with manual interpretation. The IRIS algorithm population statistics were calculated.
MAIN OUTCOMES AND MEASURES:Sensitivity and false-negative rate of the IRIS computer-based algorithm compared with reading center interpretation of the same images.
RESULTS:A total of 15 015 consecutive patients (aged 18-98 years); mean 54.3 years with known type 1 or 2 diabetes underwent nonmydriatic fundus photography for a diabetic retinopathy screening examination. The sensitivity of the IRIS algorithm in detecting sight-threatening diabetic eye disease compared with the reading center interpretation was 66.4% (95% CI， 62.8%-69.9%) with a false-negative rate of 2%. The specificity was 72.8% (95% CI， 72.0%-73.5%). In a population where 15.8% of people with diabetes have sight-threatening diabetic eye disease， the IRIS algorithm positive predictive value was 10.8% (95% CI， 9.6%-11.9%) and the negative predictive value was 97.8% (95% CI， 96.8%-98.6%).
CONCLUSIONS AND RELEVANCE:In this large urban setting， the IRIS computer algorithm-based screening program had a high sensitivity and a low false-negative rate， suggesting that it may be an effective alternative to conventional reading center image interpretation. The IRIS algorithm shows promise as a screening program， but algorithm refinement is needed to achieve better performance. Further studies of patient safety， cost-effectiveness， and widespread applications of this type of algorithm should be pursued to better understand the role of teleretinal imaging and automated analysis in the global health care system.