Fundus photograph–based diabetic retinopathy screening is widely used but provides 2-dimensional retinal information, which can lead to over-referral for diabetic macular edema evaluation. In prior studies, false-positive diabetic macular edema referral rates with fundus photograph–based screening have ranged from 71% to 86%, increasing the burden on specialist eye clinics.
In a study published in JAMA, researchers evaluated whether an artificial intelligence–based optical coherence tomography (AI-OCT) system could improve diabetic macular edema (DME) triage when used as an add-on secondary screening tool. The system was not tested as a fully autonomous diagnostic system.
Researchers conducted a stepwise evaluation in Hong Kong that included a prospective silent-mode validation study followed by a multicenter noninferiority randomized clinical trial. The randomized trial included 276 patients with suspected DME identified through a territory-wide diabetic retinopathy screening program. Patients were randomly assigned to referral decisions based on standard fundus photograph–based screening reports alone or on standard screening reports plus AI-OCT reports.
The prespecified primary outcome was the false-positive DME referral rate, defined as the proportion of referred patients who were ultimately determined not to have DME. The trial used a prespecified noninferiority margin of 20%. Referral sensitivity, referral specificity, and referral rate were secondary outcomes that were derived from prospectively collected data but were not specified in the original protocol. The between-group superiority analysis was exploratory.
The false-positive DME referral rate was 24% in the AI-OCT group compared with 69% in the standard-care group, meeting the prespecified criterion for noninferiority. In the standard-care group, all 139 patients were referred for specialist evaluation because enrollment required a fundus photograph–based report indicating suspected DME, which triggers referral under the existing pathway. As a result, the comparator was an AI-augmented review pathway vs a refer-everyone rule rather than a head-to-head comparison with individualized clinician judgment.
Under the AI-OCT pathway, 54 of 137 patients would have been referred for DME evaluation compared with 139 of 139 patients in the standard-care group, representing a referral-rate reduction from 100% to 39%.
Referral sensitivity was 100% in both groups. In the AI-OCT group, 41 referred patients were confirmed to have DME, and no DME cases occurred among patients whose referral was deferred. Referral specificity was 87% in the AI-OCT group compared with 0% in the control group, reflecting the automatic referral of all control patients under standard care.
The randomized trial was preceded by a prospective silent-mode validation study involving 603 patients with diabetes who underwent OCT imaging at a tertiary eye hospital. Among 1,200 scans, 7% were classified as ungradable. Among 1,114 gradable scans that proceeded to the DME detection model, 4% were classified as uncertain. With uncertain cases counted as positive predictions, the AI-OCT system achieved 99% sensitivity and 91% specificity for DME detection.
The system incorporated automated image-quality assessment, DME detection, and uncertainty flagging. In the randomized trial, ophthalmologists did not rely solely on the AI output. They reviewed AI-generated probability scores along with clinical information available from screening reports, including visual acuity. Uncertain and ungradable scans also underwent review by optometrists before ophthalmologists determined whether referral was warranted.
Several limitations should be considered. The validation study enrolled patients from a tertiary eye hospital rather than directly from a diabetic retinopathy screening population. The study evaluated OCT scans from a single device manufacturer, which may limit generalizability to other imaging systems. The randomized trial included only patients with suspected DME, so the findings may not apply to broader diabetic retinopathy screening populations or to other retinal diseases.
The trial also did not test real-world implementation. For ethical reasons, all patients ultimately underwent specialist assessment, meaning referral decisions in the AI-OCT group were effectively hypothetical and no patient had care withheld. The study evaluated referral metrics, not clinical or visual outcomes, and further studies are needed to determine whether use of the system in routine care can safely reduce referrals, preserve vision, or reduce vision loss attributable to DME.
Disclosures: Several researchers reported industry relationships, patents, or interests in eye disease–related digital health companies outside the submitted work. The study was funded by the Innovation and Technology Fund and General Research Fund in Hong Kong.
Source: JAMA