AI Diabetic Retinopathy Screening in India: Evidence and Rollout

AI diabetic retinopathy screening is the most mature clinical AI use case in India, and one of the few anywhere with large-scale real-world evidence. An AI model reads a photograph of the retina and decides whether the patient needs to see an eye specialist. It has been used for hundreds of thousands of patients at Aravind Eye Hospital, and Kerala's government programme now uses AI-enabled cameras across public hospitals.

Diabetic retinopathy causes preventable blindness, and early stages have no symptoms. Annual screening is recommended for people with diabetes, but there are not enough ophthalmologists to examine everyone. AI moves the first read to the point of care: an optometrist or trained technician takes the photo, the AI grades it, and only referable cases go to the specialist.

This guide covers the evidence, how the technology works, what it costs and how a hospital, diabetes clinic or diagnostic centre can start.

How AI retinal screening works

1. Capture. A fundus camera photographs the back of each eye. Some are desktop units; others are handheld or smartphone-based.
2. Quality check. The software checks whether the image is gradable. Poor images are retaken.
3. AI grading. The model classifies the image, typically as referable or non-referable diabetic retinopathy, sometimes with a severity grade and a flag for macular oedema.
4. Decision. Non-referable patients are booked for rescreening; referable or ungradable patients are referred to an ophthalmologist.
5. Record. Images, grades and referrals are stored with the patient record.

Some systems run in the cloud; others run offline on the device, which matters in clinics with patchy internet.

The evidence from India

Aravind and Google's ARDA. Aravind Eye Hospital partnered with Google and Verily in 2018 to use the Automated Retinal Disease Assessment (ARDA). A post-deployment study in JAMA Network Open, published in March 2025, evaluated performance after more than 600,000 patients had been screened in India. On a random sample of 4,537 patients:

The study reported no clinically important misses for severe disease; the misses were generally graded by ARDA as moderate retinopathy, which would still trigger referral. Google licenses ARDA to commercial partners including Forus Health, AuroLab and Perceptra rather than running screening itself.

Kerala's Nayanamritham. Kerala ran the first phase at family health centres in Thiruvananthapuram, with images graded remotely. A published evaluation of 5,307 people screened found 85.3% had gradable images in at least one eye, and an age- and gender-standardised prevalence of 17.4% for any diabetic retinopathy and 3.3% for sight-threatening disease. In February 2025 the state launched Nayanamritham 2.0 with Remidio, using AI-enabled fundus cameras operated by optometrists at community health centres, taluk hospitals and district hospitals, and adding glaucoma and age-related macular degeneration to the screen.

Smartphone-based offline AI. Studies of Remidio's offline AI on its smartphone-based fundus camera have reported high sensitivity for referable retinopathy, with specificity varying between studies; a community screening study in JAMA Ophthalmology reported 100% sensitivity and 88.4% specificity for referable disease.

Autonomous AI in the US

In the US, LumineticsCore (formerly IDx-DR) was the first autonomous AI diagnostic system authorised by the FDA, meaning it gives a result without a clinician interpreting the image. It is cleared with a specific camera. Banner Health rolled it out to 24 primary care locations in Arizona in September 2025. Other systems, including EyeArt, have since received authorisations for autonomous screening with their own specified cameras.

Autonomous use changes the model of care: the screening happens in the diabetes or primary care clinic, with no eye specialist involved unless the result is positive.

Reading the numbers correctly

Two points trip up many buyers.

• Positive predictive value is often modest. In the Aravind study, about half of severe-disease flags were confirmed as severe. That is acceptable for screening, because the cost of a false referral is an eye examination, while a miss can mean vision loss.
• Ungradable images are a real category. Small pupils, cataract and inexperienced operators produce images the AI cannot read. In the Kerala pilot, nearly 15% of people lacked a gradable image in either eye. Your programme needs a plan for them, such as dilation or direct referral.

Where AI screening fits in a hospital or clinic

For a multi-specialty hospital, the simplest high-value setup is to screen every diabetic patient attending medicine or endocrinology OPD once a year, with same-visit referral to the eye OPD when positive. Our guide to multi-specialty hospital management software covers cross-department referrals.

Costs and what drives them

Indicative cost drivers:

• Camera: desktop, handheld or smartphone-based; non-mydriatic models avoid dilation in most patients.
• AI licence: per-screen fees or bundled with the camera.
• Staff time: a technician can learn image capture in days, but quality improves with practice.
• Connectivity: offline AI avoids dependence on internet at the point of capture.
• Integration: linking images and grades to the patient record and generating referrals.
• Follow-up capacity: laser and anti-VEGF treatment capacity for referred patients.

Building the workflow into your HIS

Screening fails most often at the hand-offs, so build them into the system:

1. Identify eligible patients. Flag patients with a diabetes diagnosis or on diabetes medication who have no retinal screen in the last 12 months.
2. Order the screen at registration or consultation, with a token for the screening room.
3. Capture and grade. Images and AI results return to the record automatically.
4. Auto-refer positives and ungradables to the eye OPD, with a booked slot where possible.
5. Recall negatives for rescreening in a year, using SMS or WhatsApp reminders.
6. Track the loop. Report how many referred patients were seen and treated.

Reminders and recalls can run on the same messaging stack you use for appointments; see our guide to WhatsApp automation for hospitals and clinics. Algon builds this kind of screening workflow and dashboard inside hospital management systems.

Beyond diabetic retinopathy: glaucoma and macular degeneration

The same fundus photograph can be screened for more than diabetic retinopathy. Kerala's Nayanamritham 2.0 added glaucoma and age-related macular degeneration, and Google's ARDA programme and Remidio's offline AI have both been studied on multiple conditions. Multi-disease screening raises the value of every photograph, especially for older diabetic patients who carry risk for all three.

It also raises the stakes. Each condition has its own sensitivity and specificity, and glaucoma in particular usually needs pressure measurement and visual fields to confirm. Treat multi-disease AI as a way to find patients who need a full eye examination, not as a replacement for one.

Questions to ask a retinal AI vendor

1. Which conditions are graded, and what are the published sensitivity and specificity for each?
2. Has the system been validated in Indian patients, on the camera we will use?
3. What share of images is ungradable in typical use, and how does the system handle them?
4. Does the AI run offline on the device, in the cloud, or both?
5. What is the regulatory status in India for the AI and the camera?
6. How do results reach our HIS: API, HL7, FHIR or PDF?
7. Where are retinal images stored, for how long, and are they used to train models?

Ask to see a live screening session with your own staff operating the camera. Image quality in a busy OPD is different from a vendor demo.

Regulation and data

In India, AI that grades retinal images to support diagnosis is medical device software, and CDSCO's guidance asks developers to show performance across Indian sub-populations. Check the regulatory status of any product you buy. Retinal images are health data under the DPDP Act, so consent, purpose and storage location matter, especially if images go to a vendor cloud. See our DPDP Act checklist for hospitals and our guide to AI medical device regulation.

How to start

1. Count your diabetic patients and how many had a retinal exam last year. This is your gap.
2. Pick a camera and AI combination with published validation, ideally including Indian data, and an offline option if connectivity is unreliable.
3. Train two or three technicians and audit image quality weekly at first.
4. Run a parallel phase: an ophthalmologist grades a sample alongside the AI for a few weeks.
5. Wire up referrals and recalls in your HIS before going live.
6. Report monthly: screened, gradable, referred, seen, treated.

Mistakes to avoid

• Buying cameras without a referral pathway or treatment capacity.
• Ignoring ungradable images, which then fall through the cracks.
• Paper-based recall lists that are never worked.
• Screening the same patients repeatedly while unscreened patients are missed.
• No audit of image quality by operator.

Conclusion

AI diabetic retinopathy screening in India has moved past pilots. The Aravind data show high sensitivity at very large scale, and Kerala's programme puts AI cameras into public hospitals. For a hospital or clinic, the technology is the easy part. The work is in finding eligible patients, capturing good images, and making sure every referred patient is seen and treated.

Start with your own diabetic population, wire screening into the OPD flow, and measure the full loop. For another screening use of AI that fits the same OPD model, see our guide to AI ECG screening for heart failure.