Predicting Diabetic Retinopathy from Risk Factor Data and Digital Retinal Images

摘要 糖尿病视网膜病变是美国20岁至74岁成年人失明的主要原因。 激光光凝手术已被确定为治疗视网膜病变的一种有效方法 很早就发现了。每年一次的视网膜筛查是降低糖尿病发病率的有力工具。 糖尿病视网膜病变致盲，因为它们为糖尿病患者提供及时的诊断和 因此，及时治疗的可能性。初级保健安全网诊所提供监测和其他 为糖尿病患者提供的服务，但他们往往没有装备提供特殊护理服务，如视网膜 放映。通过使用可提供视网膜筛查的专家，可以增加与专家的联系 远程医疗作为一种糖尿病视网膜病变筛查手段在美国和 在国际上。查尔斯·德鲁大学的研究人员进行了一项试点研究，总共进行了2876次远程筛查 为糖尿病视网膜病变进行手术，来自6个南洛杉矶安全网的2,732名独特的糖尿病患者 对诊所进行了筛选。本研究的目的是在这项先前工作的基础上：(A)开发新的软件， 利用临床记录中的信息来检测未患糖尿病患者的潜伏性糖尿病视网膜病变 仍接受年度眼科检查，以及(B)设计加速糖尿病视网膜病变的方法 通过部分自动化获取数字视网膜图像的糖尿病患者的检测过程 过程中使用了图像处理和机器学习技术。具体来说，我们建议： 1.利用从患者临床记录中收集的危险因素，开发糖尿病视网膜病变的预测模型。 2.开发使用患者组合的自动糖尿病视网膜病变评估的预测模型 风险因素数据和以前由专家评估的数字视网膜图像数据。 3.评估以下方面的预测准确性：a)为具体目标开发的模型2，以及b)评估 验光师根据护理标准对验光师进行的扩张性视网膜检查 眼科医生使用一个新的洛杉矶县阅读中心为300名糖尿病患者服务。 4.根据在具体目标1中开发的预测模型创建基于Web的软件工具，该模型可以 用于在资源不足的情况下启动对高危患者的外联，提高这些患者的发现率 糖尿病视网膜病变风险最高的患者。 5.建立有针对性的外展方法，以促进对预测模型来自的患者的筛查 特定目标1确定为潜在的未被发现的糖尿病视网膜病变。

Abstract Diabetic retinopathy is the leading cause of blindness among US adults between the ages of 20 and 74 years. Laser photocoagulation surgery has been established as an effective way of treating retinopathy if it is detected early. Yearly retinal screening examinations are a potent tool in the battle to reduce the incidence of blindness from diabetic retinopathy because they provide diabetic patients with timely diagnoses and consequently, the potential for timely treatment. Primary care safety net clinics provide monitoring and other services for diabetic patients but they are often not equipped to provide specialty care services such as retinal screenings. Access to specialists who can provide retinal screenings can be increased through the use of telemedicine, which has shown great promise as a means of screening for diabetic retinopathy in the US and internationally. A pilot study by Charles Drew University investigators had a total of 2,876 teleretinal screenings performed for diabetic retinopathy, with 2,732 unique diabetic patients from six South Los Angeles safety net clinics screened. The present study aims to build on this prior work by: (a) developing novel software that utilizes information from clinical records to detect latent diabetic retinopathy in diabetic patients who have not yet received an annual eye examination, and (b) devising methods to speed up the diabetic retinopathy detection process for diabetic patients who have had digital retinal images taken by partially automating the process using image processing and machine learning techniques. Specifically, we propose to: 1. Develop predictive models for diabetic retinopathy using risk factors collected from patient clinical records. 2. Develop predictive models for automated diabetic retinopathy assessment using a combination of patient risk factor data and data from digital retinal images previously evaluated by experts. 3. Evaluate the predictive accuracy of: a) the models developed for specific aim 2, and, b) the assessments of optometrist readers against standard of care dilated retinal examinations by board certified ophthalmologists for 300 diabetic patients utilizing a new Los Angeles County reading center. 4. Create web-based software tools based on the predictive models developed in specific aim 1 that can be used to initiate outreach to high-risk patients in under-resourced settings, boosting detection rates for those patients who are most at risk for diabetic retinopathy. 5. Establish targeted outreach methods to promote screening for patients that the predictive models from specific aim 1 identify as potentially having undetected diabetic retinopathy.