What problem does it solve?
People with diabetes need a retinal exam every year or two to catch diabetic retinopathy before it damages vision. In many places the exam means a separate trip to an eye specialist, and many people never make it. Where photographs are taken in primary care and read by a remote reading network or a retina specialist, results arrive days later and follow up is a second step that is easy to lose.
An AI system that reads the photograph at the clinic moves the result to the same visit. Clinic staff take the images with a fundus camera, the software returns its result in minutes, and only patients with a positive or ungradable result need an eye specialist. The evidence on this page covers a national screening program in Thailand where retina specialists over read every image as a safety check, a randomized trial and a later observational study at Johns Hopkins, and a community health provider in Los Angeles that switched from remote reading networks.
How does it work?
- Take the photographs. A trained member of clinic staff takes retinal images of both eyes with a fundus camera during the diabetes visit. The software checks whether the images are good enough to read.
- Read the images. A cleared or validated image analysis model returns a result in the clinic: no or mild disease, referable disease, vision threatening disease or images that cannot be graded.
- Act on the result. A negative result is recorded and the next exam is scheduled. A positive or ungradable result goes to an eye specialist, and the clinic books the referral and tracks whether the visit happened.
- Monitor the program. The clinic tracks exam completion, referral completion and the share of images that cannot be graded. Programs that start with specialist over reading compare the system with a reference standard before they reduce that checking.
- Audience
- Employee facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Screening completion rate | Too few to pool | 47% to 100% | 2 | 1 organization, 1 vendor |
| Accuracy | Too few to pool | 94.7% | 1 | 1 organization |
Value drivers: Inclusion and access, Risk and loss reduction, Speed and cycle time.
Indicative value
A community health center network with 2,000 adult patients with diabetes in primary care
USD 4000 to USD 18,000
Fee income from additional completed eye exams per year
How this is calculated
Formula: patients * extraExamShare * feePerExam. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Patients with diabetes due for an eye exam each year patients, patients | 2,000 | 2,000 | The reference network. |
| Additional share of patients who complete the exam extraExamShare, fraction of patients | 0.05 | 0.15 | Conservative against the Tarzana Treatment Centers record on this page, where the vendor reports completion rising from 37% to 47% (10 points); the Johns Hopkins youth trial reports a much larger difference in a different setting. |
| Fee received per completed exam feePerExam, USD per exam | 40 | 60 | Editorial assumption, replace with your own payer rates for the exam. |
What it leaves out: Exam fees only. It leaves out the device and software cost, staff time to take images, the referral visits that follow, and the value of vision saved by earlier treatment; the Tarzana Treatment Centers white paper reports that about one in four patients examined with the system had referable disease.
Which organizations already do this?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.
Johns Hopkins Medicine
United States · Healthcare · 2024
Johns Hopkins Medicine put autonomous AI diabetic eye exams into a pediatric diabetes center and into adult primary care clinics. In the ACCESS randomized trial, youth aged 8 to 21 with type 1 or type 2 diabetes were assigned either to an autonomous AI exam at the point of care or to a scripted referral to an eye care provider. A later exploratory study of 3,745 adults seen at Johns Hopkins primary care sites compared referrals to the Wilmer Eye Institute from the AI pathway with referrals made by primary care providers.
- Screening completion rate: 100%, within 6 months, ACCESS randomized trial, youth aged 8 to 21 (n = 81)
"Diabetic eye exam completion rate was significantly higher (100%, 95%CI: 95.5%, 100%) in the intervention group (n = 81) than the control group (n = 83) (22%, 95%CI: 14.2%, 32.4%)(p < 0.001)."
Claimed by: organization
Rajavithi Hospital
Thailand · Healthcare · 2019
Between December 2018 and March 2020 a deep learning system read retinal photographs of patients with diabetes at nine primary care sites in Thailand's national diabetic retinopathy screening program. The system gave a referral recommendation in real time, and regional retina specialists over read every image as a safety measure. A prospective interventional cohort study led by Rajavithi Hospital and published in The Lancet Digital Health compared the system with an adjudicated reference standard set by fellowship trained retina specialists. Of 7,651 patients in the analysis, 2,412 were referred.
- Accuracy: 94.7%, Prospective cohort at nine primary care sites, 2018-12-12 to 2020-03-29
"For vision-threatening diabetic retinopathy, the deep-learning system had an accuracy of 94·7% (95% CI 93·0-96·2), sensitivity of 91·4% (87·1-95·0), and specificity of 95·4% (94·1-96·7)."
Claimed by: organization
Tarzana Treatment Centers
United States · Healthcare · 2022
Tarzana Treatment Centers, a community health provider in Los Angeles, set a goal of 50% completion of the annual diabetic eye exam across its six primary care clinics in 2022. It replaced remote reading networks with an autonomous AI exam taken at the clinic. The vendor's white paper reports the completion rate before and after and says one in four patients examined was found to have referable diabetic retinopathy.
- Screening completion rate: 47%, Q4 2022, six primary care clinics (875 out of 1,875 patients)
"Overall rates for EED across the six primary care clinics improved from 37% (511 out of 1,398 patients) in Q2 2022 with remote reading networks to 47% (875 out of 1,875 patients) in Q4 2022 with autonomous AI."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Registry or population list of patients with diabetes and their last exam date
- Referral pathway and capacity at a named eye care provider
- A reference standard or specialist over reading for the first months, to compare the system with
Systems to integrate
- Electronic health record for orders, results and exam due dates
- Fundus camera and the vendor's image analysis software
- Referral and scheduling system of the eye care provider
Complexity: Medium
The image analysis is a regulated medical device bought from a vendor, so the build is the fundus camera and software installation in each clinic, an order and result interface in the electronic health record, a referral and tracking workflow to eye specialists, and training for the staff who take the images.
- 1
Confirm the clearance and the intended use
Check that the device is cleared or approved in your jurisdiction for the camera, population and setting you plan to use, and that your staff work within that intended use.
- 2
Run a supervised start
Keep specialist over reading of every image for the first months, as the Thai program did, and compare the system with the specialists' reference standard before reducing that checking.
- 3
Build the referral loop
Book the eye specialist visit from the result and track whether it happened. The Johns Hopkins trial measured follow through as a result of its own.
- 4
Audit on a schedule
Review ungradable image rates, referral rates and any missed cases found later, by clinic and by patient group.
Guardrails
- Use only within the cleared intended use, with the specified camera and trained image takers
- Patients with an ungradable or positive result always reach an eye specialist
- Clinicians keep the right to refer a patient whose eyes they are worried about whatever the result says
- Separate monitoring of results and follow through by age, ethnicity and clinic
KPIs to instrument
- Share of patients with diabetes who complete the annual exam
- Share of examinations with ungradable images
- Referral completion within a set number of days
- Sensitivity and specificity against specialist reading on a sample
Human in the loop
Clinic staff take the images and book the referral. Eye specialists see every patient with a positive or ungradable result. A clinical lead owns the protocol and the audit, and in supervised starts retina specialists read the images as well.
Common failure modes
- Referrals that are never completed
- A positive result does not help if the visit does not happen. Track every referral to a completed specialist visit and give a named person ownership of the gap.
- Poor image quality
- Small pupils, cataracts and rushed image taking raise the ungradable rate and send avoidable patients to specialists. Train staff and monitor the rate by clinic and operator.
What are the risks and rules?
EU AI Act
High risk
Article 6(1) with Annex I: an AI system that is a medical device under the Medical Device Regulation (EU) 2017/745, or a safety component of one, is high risk when the device needs a third party conformity assessment. Software that analyzes images to diagnose disease is normally class IIa or higher, which needs a notified body. It is not an Annex III use.
Guidance
- Article 6, classification rules for high risk AI systems (European Union, Europe). An AI system that is a medical device, or a safety component of one, is high risk when a third party conformity assessment is required.
- Regulation (EU) 2017/745 on medical devices (European Union, Europe). Software that gives information used for diagnostic decisions is classified under Annex VIII Rule 11 and normally needs a notified body.
- FDA De Novo decision for IDx-DR (DEN180001) (US Food and Drug Administration, North America). The 2018 grant letter for the first autonomous diabetic retinopathy detection device in the United States, with its intended use and special controls.
Controls to put in place
- Documented intended use and clearance status for each device and camera combination
- Audit trail of images, results and referrals per patient
- Periodic comparison with specialist reading on a sample of images
- Equity monitoring of exam completion and follow through
Frequently asked questions
- Does autonomous AI replace the eye specialist?
No. It decides which patients need a specialist. In the Thai national program retina specialists over read every image as a safety mechanism, and in the Johns Hopkins trial patients with an abnormal result were referred to an eye care provider, with 64% of them completing follow through.
- How accurate is the AI in a real clinic?
In the prospective Thai study at nine primary care sites, the deep learning system had an accuracy of 94.7% for vision threatening diabetic retinopathy against specialists' reference standard. The authors say socioenvironmental factors and workflows matter when it is scaled.
- Does point of care screening raise the share of patients who get an eye exam?
The Johns Hopkins ACCESS randomized trial in youth reports exam completion of 100% in the AI group against 22% in the control group within six months. A vendor white paper on Tarzana Treatment Centers reports a rise from 37% to 47% across six clinics. The first is a trial in one pediatric center and the second is a vendor reported before and after figure.
How to cite this page
Blits.ai AI Use Case Library, "Autonomous AI screening for diabetic retinopathy at the point of care", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/diabetic-retinopathy-autonomous-screening. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 18 November 2026: First published