Prediction of proliferative diabetic retinopathy (PDR) from nonproliferative diabetic retinopathy (NPDR) using real-world (RW) electronic health record (EHR) data

Abstract

Purpose : To develop and validate a survival analysis and machine learning algorithm to predict the individual future risk of NPDR worsening to PDR using RW EHR data.

Methods : In this retrospective study of 4408 patients with diabetes, NPDR, and ≥ 365 days of follow-up at Cleveland Clinic Cole Eye Institute from January 2012 to February 2020 (Babiuch A et al. Invest Ophthalmol Vis Sci. 2021;62(8):1120), EHR data from the NPDR index visit (diabetes type, age, race, gender, smoking status, BMI, Charlson Comorbidity Index [CCI] score, diabetic macular edema [DME] status, HbA1c, creatinine, systolic blood pressure, eGFR, BUN, insulin use, anti-VEGF treatment, logMAR letters) was used to build prognostic models of PDR progression. Data was randomly split into 80% development data and 20% test data. Both survival analysis and linear regression models were evaluated. To assess association of baseline characteristics with PDR progression, time-dependent Cox proportional hazards modeled PDR progression as time-to-event data. Median imputation addressed missing data in covariates. Linear regression models were trained for predicting probability of progression to PDR for 6 months (month [M] 6, M12, M18, M24) with respect to index visits. Performance of the Cox model was assessed using time-dependent dynamic area under the curve (AUC) and the linear regression model was evaluated using AUC and 95% CI.

Results : The Cox model performed well, with a dynamic AUC of 0.81 on average over the study period (Figure 1). The linear regression models predicted progression to PDR at M6 with an AUC (95% CI) of 89% (75%, 96%), at M12 with 89% (81%, 95%), at M18 with 86% (78%, 91%), and at M24 with 87% (79%, 92%; Figure 2). The Cox model revealed DME, Black race, and higher baseline CCI and HbA1c as significant positive associations with PDR progression, whereas anti-VEGF treatment and older age had significant negative associations. Linear models showed DME, higher CCI, and lower age as the most important predictors.

Conclusions : Results from 2 prototype models demonstrated feasibility in predicting PDR from individual NPDR patients using RW EHR data, which may inform clinical management of patients with DR; identified risk factors may also help clinical research of DR development in the RW.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Cox Model Dynamic AUC

 

Figure 2. ROC Curves of the Linear Model