Specialists guessing the gene behind an eye disease did far better with an AI looking too

Hundreds of different genes can destroy the light-sensing cells at the back of the eye, and the resulting conditions look confusingly alike. A specialist examines the retina, forms a view about which gene is likely, and orders sequencing aimed at that shortlist.

If the shortlist is wrong the testing is slower, more expensive, and sometimes inconclusive. So the guess in the middle of the pathway carries real weight.

How the retinal model was built and trialled

Writing in Nature Medicine, researchers trained a model to predict which of 17 genetic categories a case belonged to, using photographs of the retina and layered depth scans from 1,843 genetically confirmed patients across China, South Korea and Poland.

Then they did the thing medical AI research usually skips. They ran a randomized trial: 300 patients with suspected disease, with the specialist either given the model’s suggestions or not, and the sequencing result as the answer key.

How much the AI improved the specialists’ shortlist

With the model, the correct gene category appeared in the specialist’s top five about 89% of the time. Without it, about 67%.

That is not a marginal gain. It is the difference between a shortlist that usually contains the answer and one that frequently does not.

The first guess improved as well, from about 22% to about 38%, and the top four from about 53% to about 82%.

Whether assisted clinicians made better decisions

Accuracy is the easy thing to measure and not the thing that matters. What matters is whether the clinician then does something different.

The trial scored the downstream management decisions and found the assisted group made better ones. That analysis was done after the fact rather than pre-specified, so it is the weaker of the two findings, and it is the one that connects the model to a patient’s actual care.

Why randomizing the eye specialists is the real finding

Most medical AI arrives as a number: accuracy against a stored dataset, validated on data the model has never seen. That establishes the model works and says nothing about whether a doctor using it does better.

Those are different claims and the gap between them is where a great deal of medical AI quietly fails. A model can be more accurate than the clinician and still change nothing, because the clinician overrides it, or trusts it wrongly, or the workflow never reaches the moment where it would help.

Randomizing the clinicians tests the thing that matters. Very few of these systems have been through it.

What this leaves unsettled about retinal genetics

Median age 33, and the training data came from three countries with particular populations. Inherited retinal disease genetics vary between ancestries, and a model trained largely in East Asia may perform differently elsewhere.

It predicts 17 categories rather than the full landscape of genes, so the difficult rare cases are partly outside its scope. And it reduces uncertainty before sequencing rather than replacing it; everyone in the trial was still sequenced.

What this means for families facing retinal disease

The retina is a layer of tissue in the back of the eye that senses light and sends images to the brain, and when it degenerates the loss is usually permanent.

For families waiting on a genetic answer that determines prognosis, inheritance risk and trial eligibility, shortening that path is worth something concrete. What makes this particular result credible is not the model’s accuracy but that somebody bothered to randomize clinicians to find out whether it helped.