Abstract
Purpose :
To develop a prediction model for progression of non-proliferative diabetic retinopathy (NPDR) to proliferative diabetic retinopathy (PDR) and to determine if incorporating additional clinical time points improves model performance.
Methods :
Using de-identified electronic health records from an academic medical center, we identified a cohort of patients with NPDR at the index date. This cohort included patients with age ≥18 years, presence of type 1 or 2 diabetes mellitus, and no prior diagnosis of PDR. Patients who progressed to PDR or were lost to follow-up (censored) ≤6 months from the index date were excluded. Figure 1 illustrates our study design. Four types of models were compared: Cox proportional hazards, Cox with backward selection, Cox with lasso regression, and Random Survival Forest (RSF). Covariates were included in the models if they were significant (p-value <0.05) in the univariable Cox models or clinically relevant. For each model, three sets of covariates were compared: covariates at the index date (static model 1), covariates updated at 6-month follow-up (static model 2), and covariates at the index date plus change of time-varying covariates during the 6-month observation period (dynamic model). The data were split into 80% training and 20% testing. The model performance was evaluated by the concordance index (0-1; perfect prediction= 1).
Results :
The cohort of 1107 patients had a median age of 66 years (IQR 56-74) and 547 (49%) were female. The cohort consisted of 32% white, 30% Asian, 14% Latinx, 11% black and 11% other races. The insurance types included 48% Medicare, 23% PPO/HMO, 15% Medi-Cal/Self-Pay and 14% other. 89 (8.0%) patients progressed to PDR with a median event time of 22.8 months (IQR 11.7-40). The concordance index of the trained models assessed on the test dataset are shown in Table 1. The dynamic model using RSF had the best predictive performance at a concordance index of 0.848. In this model, the variables ranked as most important were insurance, age, number of outpatient visits before the index date, mean Systolic Blood Pressure (SBP) and change in mean SBP during the observation period.
Conclusions :
We developed a set of prediction models for progression of NPDR to PDR that achieved high performance (concordance index >0.7). The dynamic model that incorporated the change in clinical variables during the 6-month observation period had the best performance.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.