Google's ARDA has Screened One Million Patients for a Leading Cause of Blindness

Diabetic retinopathy is a complication of diabetes in which prolonged high blood sugar damages the small blood vessels that nourish the retina. It develops silently, often without symptoms until vision has already begun to deteriorate, and it is one of the leading causes of preventable blindness in working age adults. At least 537 million adults worldwide are estimated to be living with diabetes, and nearly half of them will develop some form of the eye disease. The tragedy is that blindness from diabetic retinopathy is largely avoidable when it is caught early and treated in time. The problem has never been the treatment. It has been the screening.

An AI system from Google known as ARDA, short for Automated Retinal Disease Assessment, is quietly closing that gap. According to a Comment published in Nature Medicine on 23 September 2026, the system has now been used to screen more than one million patients across health systems in India, Thailand and Australia. A technician takes a photograph of the patient's retina with a specialized camera, the model analyzes the image, and a result indicating whether the patient needs referral for eye care comes back without an ophthalmologist ever needing to look at the picture.

A Screening Gap Measured in Millions

The scale of the unmet need is staggering. The International Diabetes Federation and the International Agency for the Prevention of Blindness estimate that at least 537 million adults live with diabetes, with nearly 227 million of them in the Asia Pacific region alone. India's national survey on diabetic retinopathy, covering 2015 to 2019, found the condition affects between 10% and 16.9% of people with diabetes in recent studies. The same body of research identified an acute shortage of retina specialists in India, with approximately one for every 1.26 million people.

The arithmetic is unforgiving. Even if every retina specialist in India did nothing but screen diabetic patients, the majority of the population at risk would never reach them. People in rural districts may live hours from the nearest hospital with an eye department, and the standard model of care, in which every diabetic patient receives a dilated eye examination by a specialist every year, is simply not achievable with the current workforce. This is the gap that automated screening was designed to fill.

From Research to One Million Patients

ARDA began nearly a decade ago as a research collaboration between Google Research and clinical partners including Aravind Eye Hospital in India and Rajavithi Hospital in Thailand. The early question was whether a deep learning model could interpret retinal photographs as accurately as trained specialists. The published validation work answered that question affirmatively, and the harder work began: moving the model out of the laboratory and into routine clinical care.

Prospective deployments across clinics in India and Thailand showed the system achieving 91.4% sensitivity and 95.4% specificity for detecting vision threatening diabetic retinopathy, the stage at which prompt treatment is needed to prevent vision loss. By late 2024, Google reported that the model had supported more than 600,000 screenings worldwide. The one million patient milestone, reported in Nature Medicine in September 2026, means the system has kept scaling through real health systems rather than stalling after the pilot phase, a fate that has befallen many promising medical AI tools.

A Licensing Model Built for Scale

What makes ARDA's expansion notable is not only the technology but the distribution strategy. Rather than operating screening services itself, Google licensed the model to three local partners: Forus Health and AuroLab in India, and Perceptra in Thailand. These partners are responsible for securing local regulatory approvals and deploying the system inside existing clinical care pathways. Over ten years, the partners aim to deliver a combined six million AI supported screenings to communities with limited access to eye care, at no cost to patients.

In Thailand, Google has also worked with the Department of Medical Services, the agency responsible for the country's national diabetic retinopathy screening program, on implementation research and cost effectiveness analysis. That collaboration paved the way for a partnership between Perceptra and the Department to apply the model in public sector hospitals, integrating AI screening into national health infrastructure rather than running it as a parallel service.

The licensing approach reflects a lesson from a decade of global health technology: the model is often the easy part. Distribution, regulatory clearance, maintenance and trust within the health system are what determine whether a tool reaches one million patients or dies in a pilot. By handing the technology to companies that already understand the local healthcare landscape, Google traded control for reach.

What Happens in the Clinic

The patient experience is deliberately simple. A person with diabetes sits down in front of a retinal camera, which is significantly cheaper and more widely available than a ophthalmologist's time. The photograph is taken, the AI grades it, and the patient either receives reassurance or a referral for treatment such as anti VEGF injections or laser therapy, interventions that can preserve vision if delivered before irreversible damage occurs. The whole encounter can happen in a primary care setting, far from any specialist hospital.

Independent reviews of real world adoption, including a 2025 analysis in Clinical and Experimental Ophthalmology, note that large scale ARDA implementations in India and Thailand have accumulated hundreds of thousands of screening sessions, with prospective studies supporting the safety of the automated pathway when paired with appropriate referral systems. The evidence base is now one of the deepest of any deployed medical AI tool.

The Road to Six Million

One million patients screened is a milestone, but the target is six million across India and Thailand over the next decade, and the global need extends far beyond two countries. If the partnership model holds, the same playbook, license the model to local operators, integrate with national screening programs, keep screenings free at the point of care, could be replicated wherever diabetes is rising faster than the specialist workforce can grow.

The deeper significance of ARDA may be what it proves about medical AI itself. A decade ago, automated diabetic retinopathy screening was a research demo. Today it is a regulated, licensed, nationally integrated service that has reached a million people who might otherwise never have seen an eye specialist. That is the trajectory that matters for every AI tool aiming to matter in the real world.