Diabetic retinopathy remains a leading cause of blindness worldwide, yet routine screening remains inconsistent across healthcare systems. The integration of artificial intelligence into clinical workflows could transform how millions of diabetic patients receive preventive eye care, potentially catching vision-threatening changes before irreversible damage occurs.
The Verisee AI system demonstrated 88% sensitivity and 86% specificity when screening 11,713 diabetic patients within Taiwan's government-led Diabetes Shared Care Network. The system identified referable diabetic retinopathy in 27.45% of patients compared to 18.15% flagged by ophthalmologists, suggesting the AI may catch cases that human reviewers initially miss. Performance metrics included 97% negative predictive value and 0.87 area under the curve, indicating strong reliability for ruling out disease.
This real-world validation addresses a critical gap in AI medical device research, which often relies on controlled datasets rather than messy clinical environments. The system's performance declined with patient age, reflecting the increased complexity of aging eyes with multiple pathologies. Notably, ophthalmologists identified additional referral-worthy conditions in 4.5% of cases that the AI missed, including glaucoma and various macular disorders.
The findings suggest AI screening could serve as an effective first-line filter in healthcare systems with limited ophthalmology access, though human oversight remains essential. For health-conscious adults with diabetes, this technology could enable more frequent, accessible screening while reducing the burden on specialist services. However, the 42% false positive rate indicates room for algorithmic refinement to minimize unnecessary referrals.