Community-based screening for diabetic macular edema runs into a fundamental problem: the optometrists conducting OCT scans at the point of care frequently refer patients who do not have clinically significant DME. The downstream cost is appointment backlogs at specialist centers and delayed care for patients who genuinely need treatment.
A randomized controlled trial published in JAMA offers a direct, numerically large solution. Investigators at 12 community optometry clinics in China (ChiCTR2300075087) enrolled 276 patients with diabetes undergoing OCT screening and randomly assigned them to AI-assisted grading (n=138) or standard grading by community optometrists (n=138).
The primary outcome was the false-referral rate—the proportion of patients without true DME who were sent to specialist care anyway. AI-assisted grading produced a false-referral rate of 24.1% (95% CI, 14.6%–37.0%), against 69.1% (95% CI, 61.0%–76.1%) in the standard grading arm. The absolute reduction of 45.0 percentage points (95% CI, 32.1–56.2; p<0.001) is large enough to be practically significant in any system struggling with optometry-to-ophthalmology referral volume.
Sensitivity was preserved: both arms detected true DME at a rate of 97.1%, meaning the AI system did not achieve its specificity gain by missing disease—it became more accurate at ruling out non-disease. Median time to specialist appointment was reduced by 18 days in the AI group.
The AI system evaluated in the trial is not commercially deployed in China or elsewhere; the study was conducted at academic-affiliated community clinics with standardized OCT equipment, and generalizability to uncontrolled community settings or different imaging devices would require further validation. The investigators note that the model was not trained on the study population, which strengthens the external validity argument.
For healthcare systems investing in diabetic retinopathy screening infrastructure, the efficiency signal from this trial will be difficult to ignore.