Comprehensive Summary
This study assessed the effectiveness of artificial intelligence (AI) tools in screening for diabetic retinopathy (DR) in real-world public health settings in India. Researchers first tested five different AI algorithms using low-cost, non-dilated retinal cameras. Three companies agreed to participate, and their systems were validated using data from 250 people with diabetes. The best performing algorithm, called AI 3, achieved a sensitivity of about 68 percent and specificity of 96 percent in detecting DR and was then integrated into a community health center in Punjab. During implementation, AI 3 analyzed 1,372 fundus images from 343 participants and showed excellent ability to identify images suitable for grading and to detect DR with a very high sensitivity of about 99.6 percent, though its specificity dropped to about 65 percent. It also correctly identified most cases needing referral for advanced eye care. Over time, image quality improved as the trained optometrist gained experience and the AI maintained strong performance. However, only 14 percent of patients who were advised to see an eye specialist actually followed through, which revealed gaps in patient follow up.
Outcomes and Implications
For physicians and researchers, this work shows that AI based diabetic eye screening can be successfully introduced into routine public health clinics even in resource limited areas. It demonstrates that well-trained non-physician staff, such as optometrists, can effectively run AI assisted screenings and provide early detection of vision threatening DR. Clinicians should note, however, that patient adherence to referrals remains a major challenge. Strengthening referral systems using reminder calls or messages and ensuring access to follow up care are key steps to fully realize the benefits of AI driven screening. Researchers and health administrators can also see that real world validation is critical before scaling any AI tool. The study highlights that AI performance can shift once deployed, with specificity declining when faced with diverse patient populations and imaging conditions. Regular monitoring, ongoing algorithm updates, and strong health system support will be essential to maintain accuracy and sustainability