BRENNEN

Industries / Health / Healthcare / Optometry

Screening for disease

Screening is checking the eyes of people who have no complaint, to find disease early and send the right ones on to a specialist.

7

Milestones achieved, from 1980 to 2024.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1980

    Iceland screens people with diabetes

    Iceland began yearly eye examinations for people with diabetes in 1980, and the Icelandic ophthalmologists Ólafsdóttir and Stefánsson call it the first country to start systematic screening for diabetic eye disease. They report that blindness among Icelanders with diabetes fell from 2.4% to 0.5%, largely because of the program. From 1994 people with no retinopathy, the damage diabetes does to the retina's small blood vessels, were examined every other year; of 296 followed for 10 years, everyone who developed sight-threatening disease had first shown mild changes and moved to yearly checks, with no undue delay in treatment, and the authors estimate the change cuts screening visits by more than 25%. England now also invites people with no retinopathy at their last two tests every 2 years.

    Lessonpeople at low risk can be checked less often without losing sight.

    pmc.ncbi.nlm.nih.govgov.uk

    Credit

    E. Ólafsdóttir and E. Stefánsson (University Hospital Landspitali, Reykjavik)

  2. 2009

    Glaucoma referrals swamp clinics

    Glaucoma slowly damages the optic nerve, often with raised pressure inside the eye (intraocular pressure, or IOP), and gives no warning until sight is lost. After NICE, England's guideline body, issued glaucoma guidance in April 2009, community optometrists sent more people to hospital on a pressure reading alone. In four English referral schemes (1,086 patients), referrals on IOP alone from optometrists without glaucoma training rose from 19% to 45.1% of their referrals, and the share of patients discharged at their first hospital visit rose from 21.9% to 35.4% by 2011. NICE's 2017 guideline set the referral line at an IOP of 24 mmHg or more on a Goldmann-type tonometer and asked commissioners to consider referral filtering services, in which a trained optometrist repeats the tests first. In the US, the Preventive Services Task Force found in 2022, as in 2013, too little evidence to say whether screening adults for glaucoma does more good than harm.

    Lessona lower referral line without a second check fills hospital clinics with false alarms.

    pmc.ncbi.nlm.nih.govnice.org.ukuspreventiveservicestaskforce.org

    Credit

    NICE; Ratnarajan and colleagues; US Preventive Services Task Force

  3. 2011

    Software clears normal photos

    Scotland's diabetic eye screening program passes every retinal photo through an autograder, software that checks the photo can be graded and looks for microaneurysms, the tiny bulges that are the earliest sign of retinopathy. Photos with none are given a final grade of no disease without a person looking; in 2018 that was 50.4% of just over 182,000 images, and a senior grader rereads 500 of its images a year, as for any human grader. England's program still grades by hand. In 2025 Rudnicka and colleagues ran eight commercial systems on 1.2 million images from 202,886 North East London screening visits, without the makers seeing the human grades: sensitivity for referable retinopathy ranged from 83.7% to 98.7% by maker, and the share of people without retinopathy wrongly flagged from 4.3% to 61.4%.

    Lessonthe false-alarm rate, more than sensitivity, decides how much human grading software saves.

    pmc.ncbi.nlm.nih.goveuropepmc.orgeuropepmc.orgtheophthalmologist.com

    Credit

    Scotland's diabetic retinopathy screening program; Alicja Rudnicka (City St George's, University of London), Kingston University and Homerton Healthcare NHS Trust

  4. 2014

    Home tests catch wet AMD

    Age-related macular degeneration (AMD) can turn "wet", when leaky blood vessels grow under the retina, and injections work best if started before much sight is lost. The HOME trial gave half of 1,520 people at high risk, at 44 clinical centers, a ForeseeHome device, a daily home vision test that sends results to a monitoring center. When wet AMD was found, people with the device had lost a median of 4 letters on the eye chart against 9 with usual care, and 87% against 62% still saw 20/40 or better; the trial's safety committee stopped it early because the device worked.

    pmc.ncbi.nlm.nih.gov

    Credit

    Chew and colleagues, AREDS2-HOME Study Research Group; Notal Vision

  5. 2017

    Screen every child once

    Amblyopia, or lazy eye, is poor vision in one eye that develops because the brain ignores a blurred or turned eye, and treatment with glasses, patches or drops works less well as the child grows. In September 2017 the US Preventive Services Task Force recommended vision screening at least once for all children aged 3 to 5 and found the evidence insufficient under age 3. It lists photoscreeners, cameras that judge focusing errors and eye alignment from light reflected out of a child's eyes, among accurate tools, with high testability in children as young as 1. In Sweden, Kvarnström and colleagues followed all 3,126 children born in 1982 in three municipalities to age 10: screening found most conditions before age 6, with far less amblyopia than in unscreened populations.

    uspreventiveservicestaskforce.orgeuropepmc.org

    Credit

    US Preventive Services Task Force; G. Kvarnström, P. Jakobsson and G. Lennerstrand

  6. 2020

    Clinic photos defeat a model

    In 2016 Gulshan and colleagues at Google reported a deep learning model, software that learns a task from examples, trained on 128,175 graded retinal photos, that found referable retinopathy with 90.3% sensitivity and 98.1% specificity on 9,963 test images. When Google researchers watched nurses use it in 11 clinics in Thailand, more than a fifth of photos were rejected as too poor to grade, often taken fast in poor light, and those patients were told to see a specialist elsewhere on another day. The team changed the protocol so eye specialists reviewed ungradable photos instead. In the study the fieldwork ran alongside, covering 7,651 patients at nine Thai primary care sites, the system found vision-threatening retinopathy with 91.4% sensitivity and 95.4% specificity, against 84.8% sensitivity for retina specialists reading the same images.

    Lessonaccuracy on clean test sets says little about how many patients a model turns away in a real clinic.

    europepmc.orgresearch.googletechnologyreview.comeuropepmc.org

    Credit

    Gulshan, Beede, Ruamviboonsuk and colleagues; Google Health; Thailand's Ministry of Public Health; Rajavithi Hospital

  7. 2024

    A handheld camera, no doctor

    After IDx-DR in 2018, the FDA cleared more software that decides on its own whether a person with diabetes has more than mild retinopathy: Eyenuk's EyeArt in August 2020, and AEYE Health's AEYE-DS in November 2022 and, in April 2024, with the handheld Optomed Aurora camera. In AEYE's two studies of the handheld version, operators who had never photographed an eye got a result for 99% of patients, about nine in ten without dilating drops; sensitivity was 92% and 93%, and 68% and 53% of positive results were correct. A US billing code for such tests, CPT 92229, began in January 2021. Medicare's public data show 430 of these tests in 2021 and 5,113 in 2024, at an average Medicare payment of US$27.08 in 2024, and in records of 4.96 million US patients with diabetes Shah and colleagues found that 0.09% had had one.

    Lessonclearance and a billing code did not make the test common.

    accessdata.fda.govaccessdata.fda.govdata.cms.govpmc.ncbi.nlm.nih.gov

    Credit

    Eyenuk; AEYE Health; Optomed; US Food and Drug Administration; Shah and colleagues

Watch for

  1. England weighs AI grading

    A new evidence summary concludes that some automated grading systems are as good as the first human grader and could be tested in real NHS services. It found most as good as people at finding severe disease but worse at clearing people with no disease, with results that varied by maker and setting, and the committee does not yet recommend them.

    Achieved whenthe committee recommends automated grading and England's program uses it in routine screening.

    nationalscreening.blog.gov.uk

    Claimed by

    UK National Screening Committee, in a consultation open from May to 3 August 2026

  2. Glaucoma from a retinopathy photo

    AEYE Health reports that its software found glaucoma damage to the optic nerve in the same macula-centered photos used for retinopathy screening, with 90.6% sensitivity and 90.5% specificity in about 10,900 EyePACS patients, and 96.4% and 85.3% in 100 patients from Wills Eye Hospital. The reference was ophthalmologists grading photos, four authors are company employees, and the paper is a preprint that has not been peer reviewed.

    Achieved whena regulator clears it, or an independent prospective study reports its accuracy at real screening prevalence, about 2% of US adults by the Task Force's estimate.

    pmc.ncbi.nlm.nih.gov

    Claimed by

    AEYE Health with Wills Eye Hospital, January 2026

  3. Heart risk from retina photos

    Toku says its CLAiR software, reading retinal photos taken at 10 US eye care and primary care sites, picked out adults whose 10-year risk of heart attack or stroke is 7.5% or more with 91.1% sensitivity and 86.2% specificity against the standard risk calculator, in 874 people aged 40 to 75; 94% of images were usable. Toku funded the study, the lead author is its chief health officer, and the FDA has given CLAiR a Breakthrough Device designation, which speeds review but is not approval. Earlier research models, from Google (2018) and RETFound (2023, trained mostly on Moorfields Eye Hospital images), predicted heart events with an area under the curve of 0.70 and 0.74, where 0.5 is chance and 1.0 is perfect.

    Achieved whenthe FDA authorizes CLAiR, or a group independent of Toku repeats the result in a new population.

    news-medical.neteuropepmc.orgpmc.ncbi.nlm.nih.gov

    Claimed by

    Toku, at the American College of Cardiology meeting, March 2026

Open

  1. 01

    Count ungradable eyes as misses

    A public table of every FDA-cleared autonomous retinopathy device, with each pivotal study's sensitivity, specificity, positive predictive value and imageability worked out two ways: as the maker reported it, and with eyes the device could not grade counted as missed cases or false alarms.

    The brief
    Credit · unclaimed
  2. 02

    Handheld cameras through small pupils

    A bench test of the handheld and phone-mounted retinal cameras sold for screening, photographing a model eye that holds a resolution chart through artificial pupils of 2, 3, 4 and 6 mm, and reporting the field of view in degrees and the finest line pattern resolved at each size.

    The brief
    Credit · unclaimed
  3. 03

    Who bills autonomous eye tests

    A yearly public tally of autonomous AI eye tests in US Medicare from the free CMS files: CPT 92229 tests by state, by the kind of clinician who billed them (primary care, optometrist, ophthalmologist), and per 1,000 Medicare patients with diabetes.

    The brief
    Credit · unclaimed

Open 01 · the brief

Count ungradable eyes as misses

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of every FDA-cleared autonomous retinopathy device, with each pivotal study's sensitivity, specificity, positive predictive value and imageability worked out two ways: as the maker reported it, and with eyes the device could not grade counted as missed cases or false alarms.

Why it is still open

Studies report ungradable eyes in different ways, and the gap can be large. EyeArt's FDA summary gives 91.7% sensitivity for vision-threatening retinopathy in one primary care cohort, and 78.6% when ungradable eyes are counted. A 2023 systematic review found ungradable images "analysed and reported inconsistently", and a 2025 meta-analysis of 82 studies of 25 regulator-approved systems called for post-market audits "with standardized gradability metrics"; both pool published studies, not the FDA summaries device by device. Lee and colleagues (Diabetes Care, 2021) ran seven systems on 311,604 Veterans Affairs images regardless of image quality, so ungradable photos were in their counts, but only one of the seven had FDA approval and the makers were not named.

accessdata.fda.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov

Built with what exists

The FDA 510(k) and De Novo databases and their summary PDFs, pdftotext, a vision-capable AI model such as Claude to pull the tables from scanned pages (see the Note "Seven more flagged"), every number checked by hand, and a short Python script for the arithmetic.

  1. First step

    List the clearances under the FDA's diabetic retinopathy product code, PIB, and rework the EyeArt and AEYE-DS tables.

  2. Done looks like

    Every cleared device in one public table, each figure linked to its page in the summary, a script that reproduces both versions, and two people's separate extractions agreeing on every cell.

Open 02 · the brief

Handheld cameras through small pupils

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A bench test of the handheld and phone-mounted retinal cameras sold for screening, photographing a model eye that holds a resolution chart through artificial pupils of 2, 3, 4 and 6 mm, and reporting the field of view in degrees and the finest line pattern resolved at each size.

Why it is still open

Screening increasingly skips dilating drops, so small pupils decide whether a photo can be graded. In a 2017 study of two undilated cameras in people at the University of Tübingen, 82% of images were of good quality when the pupil was 3.0 mm or wider and 25% when it was smaller. A 2023 bench test of five handhelds against a desktop camera used a Zeiss model eye that emulates a fully dilated pupil, so it did not test small pupils, and it came before the Optomed Aurora was cleared with AEYE-DS. A 2025 phantom study with one fixed artificial pupil compared six wide-field systems, not handheld screening cameras.

pmc.ncbi.nlm.nih.govdiscovery.ucl.ac.ukpmc.ncbi.nlm.nih.gov

Built with what exists

A commercial or 3D-printed model eye, a USAF 1951 resolution target (Edmund Optics), laser-cut pupil disks, the cameras themselves, and ImageJ for measurement.

  1. First step

    Fit a model eye with pupil disks and a resolution target, and measure one camera at all four sizes.

  2. Done looks like

    At least six cameras tested at four pupil sizes, three repeats each by two operators, with the raw images and a results table published.

Open 03 · the brief

Who bills autonomous eye tests

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A yearly public tally of autonomous AI eye tests in US Medicare from the free CMS files: CPT 92229 tests by state, by the kind of clinician who billed them (primary care, optometrist, ophthalmologist), and per 1,000 Medicare patients with diabetes.

Why it is still open

CMS's public file shows 92229 rising from 430 Medicare tests in 2021 to 5,113 in 2024, but lists only 27 states for 2024 and gives no breakdown by specialty. Shah and colleagues counted from licensed TriNetX records, and Wu and colleagues from a nationwide database of 11 billion claims, both ending in 2023. A November 2026 opinion piece states that optometrists, not primary care offices, billed most of the AI exams Medicare paid for in 2023, but publishes no table or method.

data.cms.govdata.cms.govpmc.ncbi.nlm.nih.goveuropepmc.orgmedcitynews.com

Built with what exists

The CMS Medicare Physician & Other Practitioners files "by Geography and Service" and "by Provider and Service" (data.cms.gov API), CDC estimates of diabetes prevalence, Python with pandas.

  1. First step

    Download the 2021 to 2024 "by Provider and Service" files and count 92229 lines by provider type.

  2. Done looks like

    A public notebook that rebuilds CMS's national 92229 totals exactly for 2021 to 2024, splits them by state and clinician type, and is rerun when the 2025 file appears.