The AI that can see your child's myopia coming

The vision screening letter comes home in a plastic folder, somewhere between the spelling list and the fun fair form. Right eye, left eye, referral or no referral. Most Singapore parents have read one and filed it, because myopia here is treated less like a diagnosis and more like weather. The Singapore National Eye Centre puts the scale of it plainly: more than 60% of children in Singapore become myopic by Primary 6, and it estimates that more than 80% of young adults are myopic.

Which makes “will my child need glasses” a fairly uninteresting question. The one worth asking is who ends up at the severe end, where eyesight stops being an inconvenience and starts being a medical risk. SNEC’s own summary of that risk: “High myopia is associated with potentially sight-threatening eye diseases later in life. These include glaucoma, cataract, and myopic maculopathy.”

Nobody has been able to tell you that in advance. A model built here now claims it can.

A photo of the back of the eye, taken at seven

Researchers at SNEC and the Singapore Eye Research Institute published a deep learning system in npj Digital Medicine in January 2023 that does something unglamorous and rather useful. Feed it a retinal photograph of a child aged six to twelve, plus routine clinical details (age, sex, race, current prescription, the length of the eyeball), and it estimates the risk that the child will have high myopia five years later, at eleven to seventeen.

It was trained on Singapore children: 998 of them from the Singapore Cohort of Risk factors for Myopia, and 7,456 baseline retinal images, with a separate set of eyes held back for testing. The combined model, images plus clinical data, reached an area under the curve of 0.97 to 0.99 on internal validation and 0.95 to 0.98 on the external test. High myopia was defined strictly: a spherical equivalent of minus six dioptres or worse, or an eyeball measuring 26.5mm or longer. For a five year forecast on a seven year old, those are strong numbers.

The authors are candid about the soft spots. Few children in any cohort actually reach high myopia, so the classes are badly imbalanced. All the images came from one type of camera. And there is a shortage of untreated, long follow up cohorts in other countries to test it against. What is published is a research tool, not a service you can ask for at a polyclinic tomorrow.

The prediction is the easy half

A larger Chinese system called DeepMyopia, published in the same journal in August 2024, went one step further and asked the question that actually matters: does knowing change anything?

It was validated across 22,060 participants at seven sites, and then used in an emulated trial on 3,303 children who were not yet myopic, average age 7.8. The intervention was not a device or a drug. It was daylight: at least 120 minutes outdoors a day, aimed specifically at the children the model flagged as high risk. Compared with the standard approach of sorting children by basic background data, targeting by model produced a 17.8% relative reduction in myopia onset (95% CI 6.4% to 29.4%).

The line I keep coming back to is the authors’ own: “The prevention of myopia and its consequences come from the benefits of the intervention and not the DeepMyopia directly.”

That sentence belongs above every AI product pitched at children. The model repairs nothing. It aims a limited supply of adult attention at the children most likely to need it, and the repairing is done by daylight.

What the aiming actually buys

Here honesty costs the story its drama. The interventions we have are real but small.

A systematic review published in Frontiers in Medicine in January 2026 pooled nine randomised placebo-controlled trials of 0.01% atropine eye drops, covering 1,091 children. The drops slowed myopic shift by 0.14 dioptres a year (95% CI 0.04 to 0.24) and slowed the eyeball’s lengthening by 0.05mm a year (95% CI 0.01 to 0.08). The authors describe the effects as modest and heterogeneous, and note the prediction intervals crossed zero. That is a real effect, compounding over years of childhood, and it is not a cure.

The free half is daylight. The Health Promotion Board’s advice to parents is to have children “spend at least 2 hours a day outdoors where there is sunlight exposure”, and it is upfront that there is no cure, only slowing.

Unexciting as all this is, the national picture is quietly encouraging. The Ministry of Health told Parliament in March 2024 that myopia prevalence among Primary 1 children stood at 26% in 2023, within the target of 30% or lower, and that in the schools sampled, high myopia fell from 3% to 2% among primary students and from 11% to 7% among secondary students between 2013 and 2023. That is population data across a decade of screening, education and outdoor play, so no single cause can claim the credit. The trend runs the right way, which is more than most childhood health trends manage.

What a parent can do before any of this reaches the clinic

Three things, none of which require the AI to exist.

Treat a school screening referral as an appointment rather than a suggestion. The models are trained on children aged six to twelve because that is when the trajectory is set, and a referral acted on in Primary 2 is worth more than the same one acted on in Primary 5.

At the eye appointment, ask for the axial length, not just the prescription. That millimetre figure is what the research models use and what tracks progression most honestly, and it is measured routinely.

And protect the two hours outside as if it were a subject. It is the only intervention in this entire piece that is free, has no side effects, and works on a child who has not yet needed a single dioptre of correction.

The pattern is worth noticing, because it keeps recurring. The AI that helps a child most is rarely the one talking to the child. It was true of the reading tools that help some children and slow others, and it is true here: a model that never meets your kid, quietly telling the adults where to spend the effort.

Disclosure: we build an AI mentoring product for children, so we have a commercial interest in AI being useful to families. We have none in eye care, and nothing above is medical advice. That conversation belongs with your child’s doctor or optometrist.