Diabetic retinopathy screening commonly relies on fundus photographs, but this approach can lead to a high number of false-positive referrals for diabetic macular edema, or DME.
According to a recent study published in JAMA, those referrals can place added demand on specialist eye clinics. The researchers evaluated whether adding an artificial intelligence-based optical coherence tomography system, or AI-OCT, to the screening pathway could improve how patients with suspected DME are referred for further evaluation.
Methodology
This study was conducted in the Hong Kong Special Administrative Region and used a stepwise design. First, researchers performed a prospective silent-mode validation from February 2020 to July 2023. That phase included 603 patients with diabetes recruited at a tertiary hospital triage unit.
The second phase was a multicenter noninferiority randomized clinical trial conducted from September 2023 to April 2025, with follow-up completed in May 2025. This phase included 276 patients with suspected DME who had been referred from a territory-wide diabetic retinopathy screening program.
Participants in the randomized trial were assigned to one of two groups. In the intervention group, referral for DME evaluation was based on both fundus photograph-based screening reports and AI-OCT reports. In the control group, referral was automatic and based only on fundus photograph-based screening reports. The intervention group included 137 patients, and the control group included 139 patients.
The AI-OCT system included image-quality assessment, DME detection and uncertainty flagging. The primary outcome was the false-positive DME referral rate, defined as the proportion of referred participants who were found not to have DME. The prespecified noninferiority margin was 20%. Secondary outcomes included sensitivity and specificity for DME detection and DME referral. For ethical reasons, all participants ultimately received specialist evaluation.
Results
In the silent-mode validation phase, 86 of 1,200 scans, or 7.2 percent, were identified as ungradable. Among the 1,114 gradable scans, 49, or 4.4 percent, were classified as uncertain. The AI-OCT system achieved a sensitivity of 98.8 percent and a specificity of 90.7 percent for DME detection.
In the randomized trial, DME prevalence was similar between groups: 30.9 percent in the intervention group and 29.9 percent in the control group. The false-positive DME referral rate was 24.1 percent in the intervention group and 69.1 percent in the control group. The absolute difference was minus 45 percentage points, and the upper bound of the confidence interval was below the prespecified noninferiority margin of 20 percent.
Sensitivity for DME referral was 100 percent in both groups. Specificity for DME referral was 86.5 percent in the intervention group and 0 percent in the control group. The study also reported that no cases of DME occurred among nonreferred participants in the intervention group.
What this means for eyecare providers
For eyecare providers involved in diabetic retinopathy screening and referral pathways, this study suggests that adding an AI-OCT system as a secondary screening tool may reduce potentially unnecessary referrals for DME evaluation while maintaining referral sensitivity.
In this trial, the intervention pathway was associated with fewer false-positive referrals than referral based on fundus photographs alone. The findings may be relevant for practices and systems looking to manage specialist clinic demand while preserving detection of DME in patients flagged through screening.
Read more of the latest AI in Eye Care news here
Author
