Bengaluru's rural AI eye check finds a parallel in Australian outback

A team of clinicians and machine-learning researchers in Bengaluru has turned a basic fundus camera into a portable diabetic retinopathy screener that an assistant with no ophthalmology training can operate. Their project, run jointly with Sankara Nethralaya and supported by India's Department of Biotechnology, sends vans into villages hundreds of kilometres from the nearest eye hospital. An onboard model flags patients showing early signs of diabetic retinopathy, India's leading cause of working-age blindness, and triages them for treatment before damage becomes irreversible.

For Australian readers the story lands close to home. Diabetes affects roughly one in twenty Australians, and rates climb steeply in remote and Indigenous communities, where the nearest ophthalmologist may be a full day's drive away. The Bengaluru model, alongside a handful of Australian pilots, points to a future in which preventable blindness is caught early using a phone, a lens attachment and a modest algorithm, rather than a specialist's appointment booked months in advance.

Origins of a village network

The team began with a simple observation: more than seventy per cent of India's population lives in rural areas, but the overwhelming majority of ophthalmologists practise in cities. Patients with uncontrolled diabetes often notice nothing until bleeding inside the eye has begun, and by the time they reach a tertiary hospital treatment is about limiting damage rather than restoring sight.

From 2019 the group collected labelled retinal photographs, building a corpus now in the tens of thousands. Once the algorithm reached acceptable accuracy, they partnered with non-profits to fit out vans with non-mydriatic cameras, a tablet running the model, and trained local-language field workers.

How the model reads the retina

The screening tool is a convolutional neural network trained on images graded by multiple retinal specialists for diabetic retinopathy, macular oedema and glaucoma-like changes. The pipeline first checks image quality, rejecting poor focus or wrong field of view before any diagnosis is attempted. Once accepted, the model returns both a binary referral decision and a confidence score that a clinician can override remotely.

Accuracy in published trials sits in the low to mid-nineties per cent, comparable to consultant ophthalmologists reading the same images. The model runs entirely offline after an initial sync, which matters in districts where 4G coverage is patchy. Workers upload anonymised images at day's end, and a central team audits a random sample to catch drift.

Results from the first trials

The first deployments covered roughly sixty thousand patients across Tamil Nadu, Karnataka and Andhra Pradesh. Audits found more than seventy per cent of flagged patients reached treatment within four weeks, an unusually high compliance rate for rural India where travel costs and lost wages usually scatter follow-up.

That compliance came from the referral chain, not the model. Each van was paired with a coordinating ophthalmologist who reviewed flagged images within twenty-four hours and either confirmed the referral or asked the field worker to repeat the photograph.

Early outcomes from the Tamil Nadu deployment included:

• Roughly one in twelve screened patients referred for clinical review
• About two per cent referred for urgent laser or injection therapy
• Mean turnaround from photograph to specialist review under forty-eight hours
• Screening costs about a tenth of comparable hospital outreach

A familiar geography in Australia

Outback communities across Western Australia, the Northern Territory and western Queensland often have a single resident doctor at best, and an ophthalmologist visit may come only every few months via fly-in services modelled on the Royal Flying Doctor Service. Among Aboriginal and Torres Strait Islander adults the prevalence of diabetes is three to four times the national average, and diabetic retinopathy is a major contributor to avoidable vision loss.

The same logic that drives a van through Mysore district is starting to drive one through the Kimberley and Cape York. Midway through 2024 the Medical Services Advisory Committee agreed to extend Medicare rebates to nurse-led retinal photography in remote settings, opening the door to community screening of the kind the Indian team has refined.

Australian pilots underway

The CSIRO's Australian e-Health Research Centre has trialled an AI retinal screening pathway with Aboriginal community-controlled health organisations in the Kimberley, while the University of Melbourne and the University of Sydney are validating commercial algorithms against Australian cohorts. Findings have been mixed, with lower specificity on older cameras and more pigmented fundi than the original training sets anticipated.

Both sides are also exchanging data, with Bengaluru researchers retraining their model on Australian images shared through formal agreements. Cross-disciplinary science coverage, including broader AI reporting, is clustering such partnerships under one banner.

Limits worth naming

No AI screener is a finished product. The Indian team has discussed cases where early versions of their model missed subtle peripheral lesions or confused diabetic changes with hypertensive retinopathy, and regulators on both sides of the Indian Ocean have moved carefully on clearance. Patients told they are fine may also delay future eye checks, a familiar and unwelcome side effect of any screening tool that promises too much.

There are also questions of data governance. Retinal images are biometric, and a photograph is in many ways a fingerprint that cannot be changed if it leaks. Both countries' programmes have responded by keeping inference local where possible, stripping identifiers before cloud upload, and publishing clear retention policies that patients can read at the point of care.

Lessons that travel

The Bengaluru experience offers a small set of design choices any remote AI screening programme can borrow:

• Train on demographically representative data, then re-test on the population who will use the tool
• Pair every screen with a human explainer who speaks the patient's first language
• Treat the algorithm as a triage assistant rather than a definitive diagnosis in its early years
• Audit referral completion rates, not just screening counts
• Build regulatory and ethical review into the rollout, not as a final hurdle

These are operational rather than technological insights. Successful programmes spend as much effort training community health workers and securing travel subsidies as they do tweaking the underlying model.

The Bengaluru programme now screens more than one hundred thousand patients a year and has redirected around ten per cent to timely treatment. Australia's pilots are smaller, but the architecture is recognisable: portable device, reliable algorithm, trusted local workforce, and a referral pathway that carries a patient from village camera to city laser. Diabetic blindness is mostly a logistics problem.