Abstract
Purpose :
To develop a prediction model for the progression of non-proliferative diabetic retinopathy (NPDR) to proliferative diabetic retinopathy (PDR) from patients seen at a safety-net hospital.
Methods :
Using electronic health records (EHR), we identified a cohort of patients with NPDR who were seen at a safety-net hospital in San Francisco (Zuckerberg San Francisco General Hospital [ZSFG]). Patients in the cohort were age 18 years or older, had type 1 or 2 diabetes mellitus, and had no PDR diagnosis prior to the date of first NPDR diagnosis (index date). Patients were excluded if they progressed to PDR or were lost to follow-up ≤ 6 months from the index date. The ZSFG data was split into 75% training and 25% test sets. Cox proportional hazards (Cox), Cox with backward selection (Cox-BW), and Cox with lasso regression (Cox-Lasso) models were used to predict progression. The concordance index (0-1; perfect prediction =1) was used for model evaluation. A similar patient cohort from the University of California, San Francisco (UCSF) was used for external validation of the models.
Results :
The ZSFG cohort of 704 patients had a median age of 64 years (IQR 57-71), of which 332 (47%) were female. The cohort consisted of 39% Hispanic or Latino, 33% Asian, 12% Black, 11% White, and 5% other races. Patient insurance types included 54% Medicaid, 33% Medicare, 10% Commercial, and 3% Self-Pay. The rate of PDR progression from NPDR in this cohort was 0.031 per person year. Table 1 lists the concordance index results from each model for the ZSFG test set and external validation results using UCSF data. The Cox model performed best; significant covariates were younger age, having commercial insurance compared to Medicare, and statin use (p ≤ 0.05).
Conclusions :
A set of prediction models to determine the progression of NPDR to PDR was developed using data from a safety-net hospital. The best-performing model was the Cox proportional hazards model, which achieved reasonable performance on the ZSFG test set and external validation on the UCSF cohort.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.