Zhang AI-OCT (DME)

An AI-Based OCT System to Detect Diabetic Macular Edema: A Prospective Validation and Noninferiority Randomized Clinical Trial

Patient / Population Intervention / Exposure Comparison Outcome

In 276 patients with suspected diabetic macular oedema referred from a territory-wide diabetic retinopathy screening programme in Hong Kong, adding an AI-based OCT report to the referral decision reduced false-positive referrals for diabetic macular oedema compared with automatic referral on the fundus-photograph screening report alone.

276

Design

Prospective silent-mode validation + multicentre non-inferiority RCT, Hong Kong

Endpoint

False-positive DME referral rate (non-inferiority margin 20%)

Relevance

ResultFalse-positive DME referral 24.1% vs 69.1% (absolute difference −45%, 95% CI −58.2% to −31.9%; P<.001 for non-inferiority). Sensitivity for DME referral 100.0% in both groups; specificity 86.5% vs 0.0%. Silent-mode validation (603 patients): sensitivity 98.8%, specificity 90.7%.

Zhang S, et al. JAMA. 2026;336(3):215-223. 10.1001/jama.2026.7025

Discussion & critique

A model of stepwise evaluation — silent-mode validation first, then a randomised trial — showing that an AI-OCT second-stage screen can cut unnecessary specialist referrals for DME substantially without missing a case. The trial is small, drawn from a single screening programme, and because every participant was ultimately seen by a specialist it measures referral accuracy rather than downstream visual outcomes.