Case Study 01 — Medical artificial intelligence
AI Diabetic Retinopathy Screening in the USA
How autonomous artificial intelligence brought eye screening into primary care
The screening gap
~60%
of people with diabetes in the United States do not receive an annual eye examination.
Centers for Disease Control and Prevention, 2024a
For millions of Americans living with diabetes, the greatest threat to their sight may not be the absence of treatment. It may be the delay before the disease is discovered.
Diabetic retinopathy damages blood vessels in the retina and can eventually cause vision loss or blindness. During its early stages, however, it may produce no noticeable symptoms. A person can believe their sight is healthy while changes are already developing inside the eye (National Eye Institute, 2025).
Regular screening is therefore essential. Yet around 60 per cent of people with diabetes in the United States do not receive an annual eye examination (Centers for Disease Control and Prevention, 2024a).
This gap is not always caused by a lack of awareness. Patients may need to find a specialist, arrange transport, take time away from work and attend a separate appointment. These barriers can be particularly significant in rural and underserved communities.
In 2018, the United States Food and Drug Administration authorised an artificial intelligence system that offered a different approach. Instead of asking every patient to visit an eye specialist for initial screening, the technology made it possible to complete a defined diabetic eye assessment inside a primary care clinic.
A landmark in medical artificial intelligence
The first autonomous AI diagnostic system authorised by the FDA in any field of medicine.
The system was originally called IDx DR and is now marketed as LumineticsCore. Its role is specific. A trained member of staff captures images of the patient’s retinas using a compatible camera. The system checks whether the images are clear enough and then analyses them for signs of diabetic eye disease.
Within seconds, it produces one of two clinical recommendations. The patient should either be referred to an eye care professional or return for screening at the appropriate time.
The word autonomous is important. An ophthalmologist does not need to review every image before the system provides its result.
This does not mean that the technology replaces ophthalmologists or performs a complete eye examination. It completes one clearly defined screening task. Patients with a positive result still require professional evaluation and, where necessary, treatment.
The evidence behind FDA authorisation
Performance from a prospective, multi-site primary care trial.
The pivotal clinical trial enrolled 900 adults with diabetes across ten primary care sites in the United States. None of the participants had previously been diagnosed with diabetic retinopathy.
Primary care staff received standardised training to operate the retinal camera and artificial intelligence system. The results were then compared with assessments produced by a specialist reading centre using detailed retinal photography and optical coherence tomography.
These results were strong enough to exceed the performance requirements established for the trial. They contributed to the FDA granting authorisation through its De Novo classification pathway in April 2018 (Food and Drug Administration, 2018a).
The FDA concluded that the probable benefits of the system outweighed its probable risks when used for its intended purpose and patient population (Food and Drug Administration, 2018b).
87.2%
Sensitivity
Abràmoff et al., 2018
90.7%
Specificity
Abràmoff et al., 2018
96.1%
Assessable images
Abràmoff et al., 2018
Sensitivity measures how successfully the system identifies people who have the target level of disease. Specificity measures how successfully it identifies people who do not.
Why primary care matters
Screening can happen where the patient already is.
The importance of the system lies not only in its ability to analyse retinal images. Its greater value may be where that analysis takes place.
A patient attending a routine diabetes appointment can potentially complete eye screening during the same visit. The result can be discussed immediately, and a referral can begin before the patient leaves the clinic.
This creates a more connected patient journey. Screening, communication and referral can become part of the same care experience.
The Centers for Disease Control and Prevention states that more than 90 per cent of vision loss related to diabetes can be avoided through early detection and treatment (Centers for Disease Control and Prevention, 2024a). Moving screening closer to the patient could therefore have meaningful consequences.
However, installing an artificial intelligence system does not guarantee that more patients will receive treatment. The complete pathway still matters. Patients with positive results must be able to reach an eye specialist. Clinics need systems for following referrals, communicating results and supporting people who face financial or practical barriers to care.
Artificial intelligence can strengthen the pathway, but it cannot repair every weakness surrounding it.
More than 90% of diabetes-related vision loss can be avoided through early detection and treatment.
Understanding the limitations
The system is autonomous in interpretation, but the service around it remains deeply human.
IDx DR is designed for a specific population and purpose. The original FDA authorisation covered adults aged 22 or older who had diabetes but had not previously been diagnosed with diabetic retinopathy.
It does not provide a complete assessment of eye health. A negative result does not rule out glaucoma, cataracts, age related macular degeneration or other conditions.
Patients experiencing blurred vision, floating spots, flashes, blind areas or other visual changes still need professional attention, regardless of the artificial intelligence result.
The quality of the retinal images is another important factor. Cataracts, small pupils, movement and camera positioning can make images difficult to assess. In the pivotal trial, sufficient images were obtained for 96.1 per cent of participants, although some required pupil dilation (Abràmoff et al., 2018).
Clinics therefore need trained staff and clear procedures for patients whose images cannot be assessed.
Accuracy and fairness
Performance should continue to be monitored as the system is introduced into new clinics and populations.
Bias is a serious concern in medical artificial intelligence. A system developed or tested on an unrepresentative population may not perform equally well for every community.
The pivotal trial included participants from different racial and ethnic backgrounds. Approximately 29 per cent were African American and 16 per cent were Hispanic. The researchers reported that diagnostic accuracy remained robust across race, ethnicity, sex, lens status and metabolic control, although specificity differed by age (Abràmoff et al., 2018).
This is encouraging, but no single trial can answer every question about fairness.
Performance should continue to be monitored as the system is introduced into new clinics and populations. Healthcare organisations need to examine unsuccessful assessments, positive results, completed referrals and differences between demographic groups.
Trust should not end with regulatory authorisation. It must be maintained through continued evaluation.
“Trust should not end with regulatory authorisation. It must be maintained through continued evaluation.”
Reading the evidence responsibly
The number of images analysed is not the final measure of success. Preserved sight is.
The pivotal trial had several important strengths. It was prospective, conducted across multiple primary care sites and used an expert reference standard.
It also had limitations. The study was funded by IDx LLC, the company that developed the system, and several authors were affiliated with the business. The trial assessed diagnostic performance in a defined population. It did not prove that national deployment would automatically reduce blindness or eliminate inequalities in access (Abràmoff et al., 2018).
Those broader outcomes depend on what happens after screening.
Does the patient understand the result? Is specialist care available? Is the referral completed? Does treatment begin when disease is confirmed?
The number of images analysed is not the final measure of success. Preserved sight is.
What healthcare leaders can learn
Successful adoption begins with a clearly defined problem.
This case offers a practical lesson about medical artificial intelligence.
Successful adoption begins with a clearly defined problem. The technology should have a specific purpose, a suitable patient population and a result that leads to a clear clinical action.
Healthcare organisations must also evaluate the complete service. Staff training, patient communication, image quality, referral capacity and continued monitoring are as important as the algorithm itself.
Patients should understand what the system is examining, what its recommendation means and what it cannot detect. Clear routes to human care must remain available whenever a result is positive, uncertain or inconsistent with the patient’s symptoms.
The World Health Organization argues that artificial intelligence in healthcare should protect autonomy, promote safety, support transparency and advance equity (World Health Organization, 2021).
These principles become even more important when an algorithm moves from assisting a clinician to producing a diagnostic result.
AI in healthcare should protect autonomy, promote safety, support transparency and advance equity.
Bringing care closer to the patient
IDx DR did not replace the eye specialist. It changed where the first meaningful assessment could happen.
That distinction captures one of the most promising possibilities for medical artificial intelligence. Carefully designed systems may bring selected forms of clinical expertise into primary care and other community settings where specialist services are not immediately available.
The technology is not perfect, and it cannot stand alone. It requires trained people, honest communication, reliable referrals and continued evaluation.
But for a patient who might otherwise miss screening, bringing the assessment into primary care could mean discovering disease before sight begins to disappear.
The most valuable future for medical artificial intelligence may not be one in which machines practise medicine alone. It may be one in which technology brings the opportunity for timely care closer to everyone who needs it.
“The most valuable future for medical artificial intelligence may not be one in which machines practise medicine alone. It may be one in which technology brings the opportunity for timely care closer to everyone who needs it.”
References
Sources and further reading
Abràmoff, M.D., Lavin, P.T., Birch, M., Shah, N. and Folk, J.C. (2018) ‘Pivotal trial of an autonomous AI based diagnostic system for detection of diabetic retinopathy in primary care offices’, NPJ Digital Medicine, 1, article 39. Available at: National Library of Medicine.
Centers for Disease Control and Prevention (2024a) Promoting eye health. Available at: Centers for Disease Control and Prevention.
Centers for Disease Control and Prevention (2024b) Vision loss and diabetes. Available at: Centers for Disease Control and Prevention.
Food and Drug Administration (2018a) Device classification under Section 513(f)(2): IDx DR. Available at: FDA medical device database.
Food and Drug Administration (2018b) De Novo classification request for IDx DR. Available at: FDA decision summary.
National Eye Institute (2025) Diabetic retinopathy. Available at: National Eye Institute.
World Health Organization (2021) Ethics and governance of artificial intelligence for health. Geneva: World Health Organization. Available at: World Health Organization.
More case studies