Abstract
Purpose :
Screening for diabetic retinopathy (DR) is crucial for early detection and timely referral for management. However, the projected increase in disease burden may render existing screening frequencies (1-2 years) unsustainable. We proposed a multimodal artificial intelligence (AI) risk-scoring model (DR-PREDICT) for the development of referable DR within 3 years, among non-referable diabetic patients (ie. no or mild non-proliferative DR (NPDR)).
Methods :
Referable DR was defined as moderate NPDR or worse, or the presence of diabetic macular edema. We used a retrospective population-based DR screening cohort from Singapore over a 10-year period. First, we trained a deep learning (DL) model using baseline 2-field color fundus photographs (CFPs) to predict the development of referable DR, giving a DL CFP image score. The dataset was randomly split at patient-level into training (70%), validation (10%), and test (20%) datasets. Next, using the AutoScore framework, we integrated this DL CFP image score with additional clinical variables (Eg. duration of diabetes, Hba1c, age) to build a point-based risk-scoring model.
Results :
21132 eyes from 11200 diabetic patients with non-referable DR were included. The first DL model based on CFPs alone had an area under the curve (AUC) of 0.899. Predictive performance improved when combined with clinical variables using AutoScore. The combination of DL score with Hba1c and baseline DR severity level (DR-PREDICT) achieved the best performance of AUC 0.927. The point-based risk scoring model further stratified patients as low- and high-risk for conversion to referable DR within 3 years at 3.6% and 59% respectively.
Conclusions :
DR-PREDICT is a multimodal risk-scoring model to predict the development of referable DR within 3 years amongst non-referable diabetic patients. It could potentially be used to personalize screening intervals and maximize resource allocation in large-scale DR screening programs. Patients identified as low-risk using DR-PREDICT can potentially be screened at longer intervals, beyond existing 1-2 yearly intervals. On the other hand, high-risk patients may be targeted for more intensive patient education and counselling.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.