Kim Iris M, Radgoudarzi Niloofar, Chen Evan M, Liu Jocelyn, Moussa Kareem, Shantha Jessica G, Tsui Edmund, Gore Charlotte, Baxter Sally L, Porco Travis C, Arnold Benjamin F, Sun Catherine Q
Department of Ophthalmology, University of California, San Francisco, San Francisco.
Francis I. Proctor Foundation, University of California, San Francisco, San Francisco.
JAMA Netw Open. 2025 Jul 1;8(7):e2521150. doi: 10.1001/jamanetworkopen.2025.21150.
Risk stratification of patients at risk of progression to proliferative diabetic retinopathy (PDR) enables earlier detection and more efficient allocation of health care resources.
To develop survival models to predict progression to PDR in patients with type 2 diabetes.
DESIGN, SETTING, AND PARTICIPANTS: This prognostic study used deidentified electronic health record data from the University of California (UC) Health Data Warehouse (UCHDW) from March 1, 2012, to July 3, 2024. The UCHDW consists of data on 10 million patients receiving care from the UC Health system. Patients aged 18 years or older with nonproliferative diabetic retinopathy (NPDR) or diabetic macular edema (DME) and type 2 diabetes were included. Patients who had a diagnosis of PDR before the index date or failed to meet the look-back criterion (≥1 outpatient visit at least 1 year before index date) were excluded. The cohort was divided into a development set and an external test set by UC Health site. The development set was split into an internal training set (75%) and test set (25%). The 3 survival models used (Cox proportional hazards regression, Cox with least absolute shrinkage and selection operator [LASSO] regression, and random survival forest [RSF]) were trained on the internal training set and validated on the internal and external test sets.
Risk factors included demographics, eye-related diagnoses, systemic comorbidities, laboratory data, vital signs, medications, and procedures.
The primary outcome was time from the index date (first diagnosis of diabetic retinopathy recorded in the UCHDW) to the first PDR diagnosis or the last in-person visit. Harrell concordance index (C index) was used to assess the model's predictive performance.
Among 7739 participants (mean [SD] age, 66.7 [11.4] years; 4116 males [53.2%]; 1548 Asian [20.0%], 744 Black or African American [9.6%], 1665 Hispanic or Latino [21.5%]; 3278 White [42.4%], 1286 other [16.6%] individuals) were included. Of these patients, 723 (9.3%) had PDR progression with a mean (SD) progression time of 1.89 (2.09) years. All survival models had good predictive discrimination (C index, 0.73-0.75) for the internal and external test sets. The Cox proportional hazards regression and RSF models had good calibration for up to 2 years. Key independent risk factors summarized across all models were baseline age, race, ethnicity, DME, NPDR severity, mean hemoglobin A1c level, and diabetic nephropathy.
The survival models developed and validated in this study demonstrated good discrimination across internal and external test sets and good calibration up to 2 years. These models have the potential to enable earlier identification, personalized follow-up, and targeted resource allocation for high-risk patients.
对有进展为增殖性糖尿病视网膜病变（PDR）风险的患者进行风险分层，有助于更早发现并更有效地分配医疗资源。
建立生存模型以预测2型糖尿病患者进展为PDR的情况。
设计、设置和参与者：这项预后研究使用了来自加利福尼亚大学（UC）健康数据仓库（UCHDW）的去识别化电子健康记录数据，时间跨度为2012年3月1日至2024年7月3日。UCHDW包含1000万接受UC健康系统护理的患者的数据。纳入年龄在18岁及以上、患有非增殖性糖尿病视网膜病变（NPDR）或糖尿病性黄斑水肿（DME）以及2型糖尿病的患者。排除在索引日期之前被诊断为PDR或未达到回顾标准（索引日期前至少1年有≥1次门诊就诊）的患者。该队列按UC健康站点分为一个开发集和一个外部测试集。开发集又分为内部训练集（75%）和测试集（25%）。所使用的3种生存模型（Cox比例风险回归、带最小绝对收缩和选择算子[LASSO]回归的Cox模型以及随机生存森林[RSF]）在内部训练集上进行训练，并在内部和外部测试集上进行验证。
风险因素包括人口统计学特征、眼部相关诊断、全身合并症、实验室数据、生命体征、药物和治疗程序。
主要结局是从索引日期（UCHDW中首次记录的糖尿病视网膜病变诊断）到首次PDR诊断或最后一次亲自就诊的时间。使用Harrell一致性指数（C指数）评估模型的预测性能。
共纳入7739名参与者（平均[标准差]年龄为66.7[11.4]岁；男性4116名[53.2%]；亚洲人1548名[20.0%]，黑人或非裔美国人744名[9.6%]，西班牙裔或拉丁裔1665名[21.5%]；白人3278名[42.4%]，其他1286名[16.6%]）。在这些患者中，723名（9.3%）出现了PDR进展，平均（标准差）进展时间为1.89（2.09）年。所有生存模型在内部和外部测试集上均具有良好的预测区分度（C指数为0.73 - 0.75）。Cox比例风险回归模型和RSF模型在长达2年的时间内具有良好的校准度。所有模型总结出的关键独立风险因素为基线年龄、种族、民族、DME、NPDR严重程度、平均糖化血红蛋白水平和糖尿病肾病。
本研究中开发并验证的生存模型在内部和外部测试集上均显示出良好的区分度，且在长达2年的时间内具有良好的校准度。这些模型有可能实现对高危患者的更早识别、个性化随访和针对性的资源分配。
