Developing a Screening Tool to Predict Diabetic Retinopathy

Abstract

Purpose : Current AAO Preferred Practice Patterns recommend that all patients with diabetes (DM) undergo an annual screening for diabetic retinopathy (DR). However previous screenings at Temple University have revealed only a 15% prevalence of DR. Performing a comprehensive eye exam on a normal patient serves no purpose except to increase healthcare costs, reduce appointment availability, and waste time for both patients and physicians. Since 85% of exams showed no sign of DR, performing comprehensive eye exams on everyone appears to be a waste of resources. We present a DR screening tool based on biometric parameters available to a primary care physician to identify patients at a low risk of DR who can safely have less frequent screening exams. Use of this tool could eliminate unnecessary exams and reduce healthcare costs.

Methods : 1,927 patients with DM underwent DR screening from 2016-2020. A retrospective chart review recorded 29 biometric variables based on clinical suspicion of predicting DR. 896 patients with unreadable fundus photos or missing variables were eliminated. Univariate and multivariate analysis identified predictive variables in the remaining 1,031 patients. Logistic regression with maximum likelihood estimates was used to create a series of models that resulted in a score representing the likelihood an individual patient has DR on exam. Models were plotted on receiver operating characteristic (ROC) curves to determine model accuracy and to identify a threshold value that optimizes the accuracy of a diabetic screening exam.

Results : Of the 1,031 patients, 217 (21%) were found to have DR in at least one eye. Eight models were created using the biometric variables. The leading model was formulated using the variables seen in Table 1. The area under the curve (AUC) was 0.74 (Figure 1). Closest to (0,1) Criteria determined the optimal cutoff value of 0.375 with sensitivity and specificity of 48.8% and 82.1% respectively and an accuracy of 0.76.

Conclusions : Our best model demonstrated moderate sensitivity and high specificity to identify DR using a cutoff value of 0.375. The next step will be to use this model, created with readily available biometric data, to predict DR in a new cohort of patients. If successful, we hope this model can be used by primary care physicians to better identify patients more likely to have DR and reduce the number of unnecessary eye exams.

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

 

Table 1

 

Figure 1: ROC curves for models predictive of DR