Between August 2021 and June 2022, researchers from the Centre for Eye Research Australia placed a self-imaging fundus camera in two general practice clinics in Victoria. An artificial intelligence (AI) model read each photograph for glaucoma. General practices do not normally own a fundus camera, and the authors chose the setting for that reason. They write that general practitioners "typically lack both the equipment and training necessary to diagnose glaucoma" [1]. Patients were approached while they waited to see their doctor.
The model was first trained in 2016 on colour fundus photographs from a Chinese population. In its original validation it reached a sensitivity of 95.6 percent and a specificity of 92.0 percent. However, it did less well in the two clinics.
- One in three people could not be photographed. Of 414 adults aged 50 or over who consented, 277 had an image captured from at least one eye. The people who could not be imaged were older, 71.4 years against 64.8, and more often reported a cataract diagnosis, 40.9 percent against 19.1 percent.
- Sensitivity was 65.0 percent and specificity was 94.6 percent, counted per person. The reference standard was two therapeutically endorsed optometrists grading the same photographs, with an ophthalmologist adjudicating. It included no optical coherence tomography (OCT) and no visual field test.
- Most flagged patients did not attend. Among 161 people who had seen an eye care provider in the previous twelve months with no glaucoma diagnosis, the model flagged 18. Sixteen could be contacted. Six attended. One reported a glaucoma diagnosis at telephone follow-up.
The authors wrote that "too many patients in our study did not adhere to referral" [1].
Data as a table
| Group | Stage | People |
|---|---|---|
| Everyone who consented | Consented | 414 |
| Everyone who consented | Image captured, at least one eye | 277 |
| Subgroup: eye care in 12 months, no glaucoma diagnosis | People in this subgroup | 161 |
| Subgroup: eye care in 12 months, no glaucoma diagnosis | Flagged as referable by the AI | 18 |
| Subgroup: eye care in 12 months, no glaucoma diagnosis | Contactable | 16 |
| Subgroup: eye care in 12 months, no glaucoma diagnosis | Attended the referral | 6 |
| Subgroup: eye care in 12 months, no glaucoma diagnosis | Reported a glaucoma diagnosis at follow-up | 1 |
| Measure | Original validation | Clinic | Exact 95% interval, clinic | Interval as printed |
|---|---|---|---|---|
| Sensitivity | 95.6 | 65.0 | 40.8 to 84.6 | 59.4 to 70.6 |
| Specificity | 92.0 | 94.6 | 91.0 to 97.0 | 91.9 to 97.2 |
The study appeared in npj Digital Medicine in 2025. The title calls it a pragmatic trial. The authors describe the design as prospective observational, with no comparator arm and no trial registration. They declare that one senior author is chief medical officer of Eyetelligence, an Australian retinal AI company. It is still the paper I would hand a colleague first, because it reports what product literature leaves out. It counts who could not be imaged, what the sensitivity was in a real clinic, and how many flagged patients turned up.
In my own practice, the referrals I remember are the ones that came back marked "did not attend".
AI is sold to optometrists on its accuracy, and this series looks at what happens after that number. A tool can be accurate in a validation study and still not improve care in a primary care practice. A tool can also be in wide use without any evidence that it helps patients. Both apply to the tool I sell, and I come back to that at the end.
One number in the paper does not reproduce
The sensitivity of 65.0 percent is printed with a 95 percent confidence interval (CI) of 59.4 to 70.6, and the specificity with 91.9 to 97.2. The paper's own table shows 15 people with referable glaucoma on manual grading and 5 more whose photographs the graders could not grade. The published figures reproduce from that table only when those 5 are counted as referable, which makes the sensitivity 13 of 20. The methods say that such participants were excluded, so the table and the methods do not agree.
An exact interval on 13 of 20 runs from 40.8 to 84.6 percent by my calculation. The published interval has the width that a denominator of 277 would give. Only the authors can settle it. Until they do, the true sensitivity in those clinics could plausibly be anywhere from about 40 to 85 percent.
Where AI now sits in an eye examination
The glaucoma camera is one tool, used at one point in an eye examination. Most of the AI now sold to optometrists and ophthalmologists comes in at one of five points.
- Before the clinician: software on a fundus camera reads the photograph and returns a referral decision during the visit, with no clinician looking at the image.
- Beside the clinician: software reads the photograph or OCT scan the clinician has already taken, and adds a diagnosis, a referral suggestion or a measurement.
- Past the eye: software reads the retina for cardiovascular risk, or for the gene behind an inherited retinal disease.
- At the phoropter and in surgical planning: a deep learning refraction, and an intraocular lens (IOL) calculator.
- With the patient: a chatbot that answers questions about an operation.
Before the patient reaches a clinician
The tool is software attached to a fundus camera, usually in a general practice, diabetes or endocrinology clinic. It reads the photograph for diabetic retinopathy and returns a result while the patient is still there. The person taking the photograph does not have to be an eye care professional. In the Pilbara, Lions Outback Vision ran Google's Automated Retinal Disease Assessment (ARDA) from a van. A research officer with no formal healthcare qualification took the photographs, and patients were told the result on the spot [2].
Four systems are cleared by the United States (US) Food and Drug Administration (FDA) to make the decision without a clinician. They are LumineticsCore from Digital Diagnostics, formerly IDx-DR, EyeArt from Eyenuk, AEYE-DS from AEYE Health, and iPredict-DR from iHealthscreen, which was cleared on 2 July 2026 [3, 4].
The examination moves to where the patient already is, and the result comes back during that visit instead of at an eye appointment.
- In a US trial, 100 percent of young people with diabetes offered the AI examination in their own diabetes clinic completed an eye examination within six months. Of those given a scripted referral to an eye care provider and education, 22 percent did [5].
- In specialist eye clinics in Bangladesh, clinic days randomised to an FDA-authorised autonomous system completed 1.59 patient encounters per specialist hour, against 1.14 without it [6].
- In England, eight CE-marked systems were tested on the stored images from 202,886 encounters in the North East London screening programme, against the programme's own human grading. The question was whether software could take the first read and leave the human graders only the images it flags. Sensitivity for moderate to severe retinopathy ran from 96.7 to 99.8 percent across the eight. The false-positive rate ran from 4.3 to 61.4 percent, so the choice of vendor would decide how much human grading the filter saves [7].
For an optometrist, the referral in this pathway arrives from a diabetes or general practice clinic, and the finding behind it is the software's. The patient still has to make the trip: in the US trial, 16 of 25 young people with an abnormal result followed through to an eye care provider [5]. The next issue follows these systems out of their pivotal trials.
Beside the clinician
The tool reads the photograph or scan the clinician has already taken and shows its own reading beside the clinician's. The clinician still signs the decision.
- EyeFM is a vision-language model pretrained on 14.5 million ocular images from five imaging modalities, paired with clinical text. In a single-centre, double-masked randomised trial in a Chinese high-risk screening population, 668 participants were allocated to 16 ophthalmologists, half working with the copilot and half without. Ophthalmologists with it reached the correct diagnosis in 92.2 percent of cases against 75.4 percent, and the correct referral in 92.2 against 80.5 percent. Their reports also scored higher on a standardisation measure. At follow-up, 33.7 percent of patients in the copilot arm had acted on a referral suggestion, against 20.2 percent [8].
- The Hong Kong AI-OCT system adds a macular oedema reading to the fundus photograph report a screening programme already produces. It checks image quality, detects diabetic macular oedema and flags the scans it is unsure of. The trial was multicentre, randomised and designed for non-inferiority, in 276 patients referred from the territory-wide screening programme. Referring on both reports cut false-positive referrals to 24.1 percent, against 69.1 percent on the photograph report alone. Sensitivity was 100 percent in both arms [9]. For ethical reasons every participant still saw the specialist, so the trial compared referral decisions and nothing downstream of them.
- Two RetInSight products on the Australian register take measurements instead of making a diagnosis. The Fluid Monitor quantifies the volume of intraretinal fluid, subretinal fluid and pigment epithelial detachment on an OCT scan, to support treatment decisions in neovascular age-related macular degeneration. The GA Monitor, GA standing for geographic atrophy, quantifies the area of retinal pigment epithelium loss and photoreceptor degeneration in geographic atrophy. The register entry says the results must not be the sole basis for a diagnosis or a treatment decision [10]. I found no trial of either in the search window.
- The one copilot tested in community optometry did not do well. In the United Kingdom HERMES study, an OCT referral model called Octane was run on scans from community optometrists referring retinal disease. It could process images from 204 of 396 participants, 51.5 percent. On referral decisions it was less accurate than hospital experts and of similar accuracy to the community optometrists [11].
In the two trials the second reading raised the share of right diagnoses and cut the share of needless referrals. Two other studies looked at what it does to the clinician's own judgement. I come back to them after the gaps.
Beyond the eye
- Mediwhale's software on the Australian register takes a fundus image and indicates suspected cataract, glaucoma and retinal disease, and predicts cardiovascular risk. Both Eyetelligence entries on the register also carry a cardiovascular risk output [10]. I found no randomised trial of cardiovascular risk read from the retina in the search window, so what it changes for a patient is not yet known.
- Retina4IRD reads colour fundus photographs and OCT scans and predicts which of 17 genotype categories lies behind a suspected inherited retinal disease, before genetic testing. In a multicentre trial, 300 people with suspected disease were randomised to a specialist with or without it, and 295 had sequencing reports to compare against. The correct genetic category was in the top five predictions in 88.5 percent of cases with the tool, against 67.3 percent without. The top prediction was right in 37.8 against 22.4 percent [12]. In practice the specialist would order the genetic test with a shortlist already in hand.
Refraction and surgical planning
- Tongren Digital Sight is a portable autorefractor: an infrared photorefraction camera and a deep learning model trained on 362,000 images. In a randomised crossover trial of 282 adults at Beijing Tongren Hospital, its spherical equivalent differed from subjective refraction by a mean of 0.07 dioptres. Corrected acuity with its prescription was non-inferior on the day, with a mean difference of 0.03 on the logMAR (logarithm of the minimum angle of resolution) scale. Agreement was weaker in high myopia [13]. The authors built it for places with no phoropter and no one trained to use one. Every participant in the trial also had a subjective refraction to compare it with.
- The Zeiss calculator on the Australian register predicts the post-operative residual refraction for a given eye biometry and IOL power. It also helps the surgeon choose the lens model and its axis for corneal astigmatism [10]. It is listed as AI-enabled. No trial of it appears in the search window.
With the patient
- A cataract surgery chatbot at the university hospital in Brussels answered only with clinician-validated responses. In a pilot randomised trial of 64 patients, brochures plus the chatbot made no detectable difference to knowledge, anxiety or satisfaction compared with brochures alone. Only 35 of the 64 returned the questionnaire. Seventeen of the 33 patients given the chatbot never used it, and those who did were younger on average, 64 years against 74 [14]. A pilot this small cannot show that the two approaches are equivalent either.
What none of this has shown
No trial followed vision over time. I found seven randomised trials of an AI tool in ophthalmology or optometry published between January 2023 and September 2026 [5, 6, 9, 12, 8, 13, 14]. None had a vision outcome as its primary endpoint. Three measured a process of care, three measured accuracy or agreement, and one measured what patients knew and how anxious they were. The autorefraction trial measured corrected acuity on the day as a secondary outcome, and it was non-inferior.
Data as a table
| Trial | Participants | Setting | Primary endpoint | Result, AI arm vs comparator | Endpoint group |
|---|---|---|---|---|---|
| ACCESS | 164 young people | Two paediatric diabetes sites, Baltimore | Eye examination completed within 6 months | 100 vs 22 percent | Process of care |
| B-PRODUCTIVE | 993 patients, 105 clinic days | Retina clinics, Bangladesh, randomised by clinic day | Completed encounters per specialist hour | 1.59 vs 1.14 | Process of care |
| AI-OCT for macular oedema | 276 patients | Multicentre, Hong Kong screening programme | False-positive referral rate | 24.1 vs 69.1 percent | Process of care |
| Retina4IRD | 300 randomised, 295 analysed | Multicentre, inherited retinal disease | Top-5 genetic category accuracy | 88.5 vs 67.3 percent | Accuracy or agreement |
| EyeFM copilot | 668 participants | Single centre, 16 ophthalmologists, China | Correct diagnosis, and correct referral | 92.2 vs 75.4 percent, and 92.2 vs 80.5 percent | Accuracy or agreement |
| Deep-learning autorefraction | 282 adults | Crossover, Beijing Tongren Hospital | Agreement with subjective refraction, spherical equivalent | Mean difference 0.07 dioptres, r = 0.96 | Accuracy or agreement |
| Cataract education chatbot | 64 patients | Single-centre pilot, Brussels | Knowledge, anxiety and satisfaction | No significant difference detected | Patient-reported |
The wider AI trial literature to early 2024 looks much the same. A scoping review of 86 randomised trials of AI in clinical practice found that 70, or 81 percent, had positive primary endpoints. In 46 of the 86, or 54 percent, the endpoint was diagnostic yield or performance. Its full text records that all three ophthalmology trials reported a reduction in operational time [15].
The count comes from the evidence map behind this series. The map was built from PubMed, ClinicalTrials.gov, the Australian New Zealand Clinical Trials Registry and the regulator lists between 27 and 30 September 2026. It is not a registered systematic search, so a trial indexed under terms I did not use would be missing.
Two more gaps run through the sections above. Attendance after a positive result is the step the software cannot take for the patient: 16 of 25 in the US trial, 6 of 18 in the glaucoma clinics, and one in three in the EyeFM copilot arm. And outside diabetic retinopathy the evidence is thin. The glaucoma model that opened this issue was far less sensitive in general practice than in its original validation. The one OCT copilot tested in community optometry was no more accurate than the optometrists.
When the clinician and the AI disagree
Two studies measured what an AI reading does to the clinician's own decision.
Explaining a glaucoma AI's recommendation did not measurably improve optometrists' accuracy, and it increased reliance on the AI's mistakes. One user study from the search period put 87 optometrists in front of a glaucoma referral model built from routine eye care data. Accuracy was 51 percent unaided (95 percent CI 46 to 56). It was 58 percent with the AI recommendation alone, 60 percent with a feature-importance explanation and 62 percent with a scorecard explanation. The three AI-supported conditions did not differ significantly from one another, so the study found no gain from the explanations. Post-hoc explanations "increased over-reliance on incorrect AI recommendations" [16].
The AI was right on 80 percent of the study cases. Those were 20 cases selected for the study, so the 80 percent describes the case set and not the model in a clinic.
A second study asked 45 trainee and expert ophthalmologists to grade 120 fundus photographs. A week later they read the same photographs again with an AI whose judgements were deliberately wrong 30 percent of the time. Accuracy rose from 48.4 to 59.6 percent overall. It was 63.9 percent when the AI was right and 47.9 percent when it was wrong [17].
However, the paper tested the readings made with a wrong AI against the first-session readings and found no significant difference. The 47.9 percent covers a subset of the photographs and the 48.4 percent covers all 120, so those two figures are not a matched pair. The gain was larger for trainees, 14.2 percentage points against 8.6, and the effect of a wrong AI was less pronounced in experts. The standard deviations run from 20 to 27 percentage points, and a repeat reading a week apart carries its own learning effect.
Data as a table
| 87 optometrists | Accuracy, percent | Published 95% interval |
|---|---|---|
| Optometrists alone | 51 | 46 to 56 |
| With AI, no explanation | 58 | 56 to 60 |
| With AI, feature importance | 60 | 55 to 64 |
| With AI, scorecard | 62 | 57 to 66 |
| AI alone, on the 20 study cases | 80 | not reported |
| 45 ophthalmologists | Photographs | Accuracy, percent |
|---|---|---|
| First reading, unaided | All 120 | 48.4 |
| Second reading, with AI | All 120 | 59.6 |
| Second reading, with AI | Photographs where the AI was right | 63.9 |
| Second reading, with AI | Photographs where the AI was wrong | 47.9 |
What is available in Australia
The Therapeutic Goods Administration (TGA) publishes a list of AI-enabled devices in the Australian Register of Therapeutic Goods (ARTG). Updated 10 September 2026, it carries eight ophthalmic entries, and the register says what each does [10].
- TeleMedC's DR Grader screens and grades diabetic retinopathy from colour fundus photographs, in people with diabetes and in people not yet diagnosed.
- Mediwhale's software reads a fundus image for suspected cataract, glaucoma and retinal disease, and predicts cardiovascular risk.
- RetInSight's Fluid Monitor and GA Monitor measure fluid volumes and atrophy areas on OCT, to support treatment decisions in neovascular and atrophic age-related macular degeneration.
- The iCare RETCAD software reads colour fundus photographs for age-related macular degeneration, diabetic retinopathy and glaucoma, as decision support.
- A Zeiss calculator predicts the post-operative residual refraction for a given biometry and IOL power.
- Two Eyetelligence products flag referable age-related macular degeneration, diabetic retinopathy, glaucoma and cardiovascular risk from colour fundus photographs. The register notes that the Australian configuration of Assure Plus excludes proliferative diabetic retinopathy and neovascular age-related macular degeneration.
None of the eight is described on the register as deciding without a clinician. Each entry names the health professional it supports. The four FDA-cleared autonomous diabetic retinopathy systems are LumineticsCore, EyeArt, AEYE-DS and iPredict-DR [3, 4]. None is on the TGA list.
The TGA states that its list captures devices where AI or machine learning terms appear "in the product name or intended purpose". It also captures devices whose sponsor reported AI after March 2025. Absence from the list is therefore not absence from the register [10]. SELENA+, a Singapore retinopathy grader, has held ARTG entry 418401 since 17 August 2023 and is not on the list [18]. I therefore searched the public ARTG for each of the four products and its maker on 30 September 2026, and found no matching entry. A name search cannot rule out an entry under a different product name or sponsor [19].
One Australian programme
One documented Australian programme began in Western Australia and now reaches beyond the state. Lions Outback Vision ran Google's ARDA model from a van across nine Pilbara communities between February and August 2024. Seventy-eight patients consented, 10.3 percent of photographs showed referable retinopathy, 8.4 percent were ungradable, and screening rose eleven-fold on 2023 [2].
In October 2025 the partners wrote that "the regulatory steps for implementation of new software as a medical device need to be undertaken" [20]. By April 2026 the spin-out, Ninox Vision, had supplied the tool to a Perth general practice, two clinics in Albany, three Kimberley Aboriginal medical services and a primary care service in Darwin. That list comes from the Australian Medical Association (WA) [21]. Neither source says how the software is regulated today.
The rule that changed in January
The guidance that matters most for a product like mine was rewritten on 29 January 2026. The TGA's guidance on clinical decision support software (CDSS) now states that "an AI-enabled CDSS will not meet the exemption criteria". It reads the third exemption criterion as requiring software to be transparent in how it generates recommendations [22]. Fixed rules do not make a product exempt. All three criteria have to be met. A later issue covers the rules in full.
Conflict of interest, funding and how this was written
I own OptoGuide Pty Ltd, which sells OptoGuide, a clinical decision support product for optometrists. Its clinical recommendations come from fixed rules, and no AI model runs when a clinician uses it. I used AI tools to write its code, to draft its learning cases and to generate its clinical illustrations, which are labelled in the product as AI-generated. It has no ARTG entry. On 30 September 2026 I notified the TGA that it is supplied as exempt clinical decision support software, which is my own determination and not a ruling. It is live in Australia only. Versions for other countries have been researched and are not released. It has no published evaluation, no trial data and no outcome data, so every point in this issue about the gap between validation and outcome applies to it.
No one funded this issue. I have no financial relationship with Eyetelligence, the Centre for Eye Research Australia, Lions Outback Vision, Ninox Vision, Google, or any maker of a diabetic retinopathy system named here. I checked every figure against the paper or the register it is cited to, and the errors are mine. This series will cover the field without promoting OptoGuide, and where a finding bears on rule-based decision support I will say so.
Before I believed OptoGuide changed a patient's outcome, I would want to see referral decisions made with and without it, in clinics I do not run, and followed through to what happened to each patient.
The next issue follows the diabetic retinopathy systems out of their pivotal trials: what held, what moved, who could be photographed and who came back after a positive screen. The issue after that covers the rules for Australian optometrists using AI: the guidance from the Australian Health Practitioner Regulation Agency (Ahpra) and the Code of conduct, the TGA's treatment of software as a medical device, the privacy obligations, including the automated decision-making transparency requirement that starts on 10 December 2026, and where liability sits when a tool is wrong.