AI-based eye health screening holds great potential for community application. However, its real-world implementation performance, coverage capacity, and overall impact on chronic disease management require systematic evaluation.

To evaluate the effectiveness of an AI-based ophthalmic telemedicine diagnostic platform in community eye health screening and to explore its value for broader implementation and adoption.

A quasi-experimental study design was adopted. Eye health screening was conducted in 2025 among residents with chronic diseases or at high risk in Baiyun Street Community, Yuexiu District, Guangzhou. The screening employed an AI-powered ophthalmic telemedicine diagnostic platform with an "AI initial screening + ophthalmologist verification" model and a "general practitioner-specialist co-management" approach. The first follow-up was scheduled at 6 months after the initial screening. Evaluation metrics included the positive detection rate of ocular diseases, referral rate, AI diagnostic performance, and changes in biochemical indicators. The review diagnosis by three ophthalmologists served as the gold standard. A cross-sectional survey on satisfaction was conducted using a self-designed questionnaire among 200 randomly selected participants.

A total of 1 343 individuals were screened, accounting for 45.3% of the target chronic disease or high-risk population in the community (target population: 2 965 individuals). The positive detection rate of ocular diseases was 64.93% (872/1 343) for the right eye and 63.66% (855/1 343) for the left eye. The referral rate during the study period was 0.67% (9/1 343). The screening results showed that the main abnormalities in the fundus were cataracts and leopard print fundus, accounting for about 30%. AI diagnostic performance evaluation: The gradable rate was 92.55% (1 243/1 343) for AI, which was broadly comparable to the 93.52% (1 256/1 343) achieved by ophthalmologist review. The AI abnormality detection rate was 44.53% (598/1 343), versus 59.20% (795/1 343) for the ophthalmologist review, suggesting that the AI adopted a relatively conservative strategy for abnormality detection. The overall diagnostic agreement rate of AI was 89.5% (95%CI=87.6%-91.2%), with a sensitivity of 83.7%, a specificity of 96.1%, and an AUC of 0.899. A total of 100 individuals were classified by AI as "ungradable"; among these, 96.0% (96/100) were confirmed as abnormal by the ophthalmologist review, indicating that such results should be prioritized as positive for referral. The overall consistency (Kappa coefficient) between AI and ophthalmologists for all fundus images was 0.925 (P<0.001). The proportion of ungradable images determined by AI was 7.45% (100/1 343), higher than that determined by ophthalmologists [6.48% (87/1 343)] (χ²=13.00, P<0.001). At the 6-month follow-up, the levels of glycated hemoglobin (HbA_1c), fasting blood glucose (FPG), total cholesterol (TC), and triglycerides (TG) in the screened population decreased compared to baseline (P<0.05). The overall satisfaction rate was 92.50% (185/200).

The AI-based ophthalmic telemedicine diagnostic platform demonstrates high screening efficiency and diagnostic accuracy. This platform enables comprehensive community-based eye health coverage, addresses the primary-care management of certain ophthalmic conditions, facilitates the implementation of tiered healthcare delivery, and may help reduce unnecessary referrals. Furthermore, it enables dynamic monitoring of chronic disease management outcomes, thereby demonstrating substantial potential for broader application and scalability.