In early 2018 Nature Biomedical Engineering published “Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning” by Ryan Poplin, Avinash Varadarajan, and colleagues at Google. The result was striking because it found that a photograph of the back of the eye carries signals that doctors had not routinely read from it.
The team trained deep-learning models on retinal images from 284,335 patients and validated on two independent groups. The models predicted a person’s age to within about 3.3 years on average, their sex with an area under the curve of 0.97, smoking status at 0.71, and systolic blood pressure to within roughly 11 mmHg. They also predicted the onset of major adverse cardiac events at an AUC of about 0.70. Using attention techniques, the authors showed the models often focused on anatomical features such as blood vessels and the optic disc.
These are associations, not a deployed clinical test, and predicting a coarse risk score is not the same as diagnosing disease. But the work suggested that a cheap, non-invasive eye photograph might one day contribute to cardiovascular risk assessment, which today relies on blood tests and other measurements.
For a general reader, this paper is a good illustration of how machine learning can surface unexpected, subtle patterns in existing medical data, raising both opportunity and the careful follow-up needed before such findings become trustworthy tools.