The following is a summary of “A Multimodal AI Risk Scoring Model to Predict the Development of Referable DR Within 3 and 5 Years: DR-PREDICT,” published in the issue of Ophthalmology 2023 issue by Tan et al. 

Referable diabetic retinopathy (DR) is defined as moderate Non-Proliferative DR (NPDR), a more severe condition, and Diabetic Macular Edema. Researchers performed a prospective study to develop an AI model to predict which people with mild or no DR will likely need treatment within 3 or 5 years. 

The study trained a deep learning model using initial color fundus photographs (CFPs) to predict the likelihood of developing referable DR (referred to as the DL score). Subsequently, they integrated this DL score with additional clinical factors such as diabetes duration, HbA1c levels, and age using AutoScore to create an interpretable risk-scoring model.

The study included 21,132 non-referable Diabetic Retinopathy (DR) eyes from 11,200 diabetic patients. Initially, a deep learning (DL) model using only color fundus photographs achieved an area under the curve (AUC) of 0.794 for the 3-year task and 0.731 for the 5-year task. Performance improved when this DL model was combined with clinical variables using AutoScore. The best AUC of 0.817 for the 3-year task and 0.816 for the 5-year task was achieved by integrating the DL score, HbA1c, and baseline DR severity, referred to as DR-PREDICT. With DR-PREDICT, the risk of conversion to referable DR was below 2.6% for low-risk and above 49.9% for high-risk individuals within 3 years, and within 5 years, the risk ranged from below 4% for low-risk to above 59.1% for high-risk patients.

The study found that DR-PREDICT can identify people at high risk of developing diabetic retinopathy so they can be screened more often. 

Source: aao.apprisor.org/apsSession.cfm?id=PO487