Let's go back to the mid 1800s. There's a cholera outbreak in London and its source is unknown.
At the time it was accepted that cholera was transmitted through the air. A doctor named John Snow believed that it was instead spread through contaminated water. He just couldn't prove it. So he went about plotting the deaths on a map street by street, house by house.
A pattern emerged. Every death was clustered around a single water pump. Proximity to the pump also seemed to correlate to mortality risk. Snow convinced the local council to remove the pump's handle, so nobody could draw water from it anymore. The outbreak stopped almost immediately.
Now jump forward to today. Think about the last time you heard a song but didn't know the name. You opened Shazam, and a few seconds later you have the artist, title, and album.
The founders of Shazam were told it was impossible to reliably pick a song interwoven with the noise of real life. Their trick was to focus on the song's loudest, sharpest peaks, the frequencies that spike above everything else. Background noise is comparatively flat and messy. So the peaks survive the noise, and the pattern they form becomes the song's fingerprint. That fingerprint is unique enough to identify it in a database of millions.
For both of these examples there were patterns hiding in plain sight. What else is hiding in plain sight, waiting for the right tool to find it?
Hidden Signals
Snow and the Shazam team both understood their pattern well enough to build a specific method for finding it. The next example is different. Researchers had a hunch, but no method, so they handed data to an AI and let it find the specific signal itself.
Mayo Clinic researchers suspected something in a person's voice might predict cardiovascular disease, they just didn't know what. In a small study, they recorded voice samples and fed them to an AI, which identified voice features linked to coronary artery disease. It produced a score and patients with a high score turned out to have roughly three times the risk of a heart attack or a confirmed blockage over the next two years. That's what we call a hidden signal, a pattern that was invisible until the AI found it.
Interestingly, with the Mayo study the AI found the signal but we don't know exactly what it is, and that's typical. The answer is usually locked somewhere inside the model, and even its creators can't read it. We touched on this black box problem in a recent Blink, and it raises important ethical questions in health care when an AI supported clinical decision can't be explained. We'll circle back to this in future posts.
What AI Found in the Retina
Consider a clinician who has spent years looking at fundus images. Of course they can spot diabetic retinopathy, AMD, choroidal naevi etc.
Now instead ask them if the image is from a male or female. I would expect most ophthalmologists and optometrists would say "not possible".
In 2018, a team at Google had a hunch that a retina might reveal more than eye disease, so they decided to test it. They fed a model thousands of fundus images, each one labelled with patient information such as age, blood pressure, smoking status, and sex. The model's job was to find patterns matching each label. Once trained, it was tested on new images it had never seen. It predicted age to within about three years on average. It picked up signs of smoking and high blood pressure. And it also identified the patient's sex correctly with striking reliability.
No computer scientist had told the model what to look for. The patterns were so deeply buried that no clinician had ever noticed them. This is the strange thing about this kind of AI. It's not that it's faster or more consistent than a clinician, it's that it sees things nobody knew were there in the first place.
There's a reason the eye keeps turning up in this research. The retina is the only place in the body where you can look directly at blood vessels and nerve tissue in a living person, without cutting anything open. Every other organ requires a scan or a scalpel.
We've always known the retina reveals more than eye disease. Hypertensive changes, papilloedema suggesting raised intracranial pressure or a Hollenhorst plaque hinting at carotid disease. We're trained to spot all of these. What's new is a name for it. This emerging field is known as oculomics, a term coined in 2020, and it's something we'll be hearing more about in coming years.
Doctors currently use blood pressure, cholesterol and smoking history to calculate someone's risk of a heart attack or stroke. The 2018 Google study predicted the same thing just as well, from the photograph alone. It's plausible that in the future a routine eye exam could identify a cardiovascular risk that nobody had picked up.
Neurodegenerative disease is another active area of research in oculomics. In 2022, researchers trained an AI model to recognise patients with Alzheimer's disease from retinal photographs. The retina develops from the same embryonic tissue as the brain, so it makes sense that the same disease process leaves a signature in both.
Parkinson's disease has a similar story. A 2023 study found retinal changes in patients on average seven years before their clinical Parkinson's diagnosis. From an OCT scan most optometrists already have in their practice.
Kidney disease, anaemia, thyroid disorders all have emerging retinal signals too. There is also early research looking at retinal signals for conditions like ADHD and autism. New papers turn up on this every few months.
It's worth staying sceptical though. These findings come from specific datasets, and those datasets don't always represent the people we see. If the volunteers are mostly healthy, or mostly white, the results may not apply to the patient sitting in your chair.
Optometry's scope of practice has never really stood still. It began with refraction, then expanded into diagnosing and managing ocular disease, with therapeutic prescribing following. Screening for systemic health problems through oculomics could be the next shift.
So what does this mean for us at the moment in practice? Not a whole lot currently. Most of this is still emerging but the direction is clear, and it's moving fast.
Where this leaves us
The routine eye exam is becoming more useful than it has ever been. AI models are finding connections between the eye and the rest of the body that we never knew were there.
Next time, we look at what AI actually sees when it looks at a scan, and at heatmaps, which go some way to answering the black box problem.