One eye-scan model flagged heart attack and Parkinson's risk

Researchers at Moorfields Eye Hospital and UCL trained a model called RETFound on 1.6 million retinal images without disease labels, letting it first learn the structure of a healthy retina before being fine-tuned on much smaller labeled sets for specific conditions.

With a fraction of the labeled data other models needed, RETFound matched or beat them at diagnosing eye diseases, and once fine-tuned on outcomes data it also predicted future heart attacks and Parkinson's disease from retinal photos alone, with moderate but above-chance accuracy for both.

Its Parkinson's prediction was the weaker of the two signals and needs confirmation in larger, more varied groups before it means much for an individual patient. The researchers released RETFound publicly so other hospitals and research teams could fine-tune it for their own conditions rather than building a foundation model from scratch.

An eye scan also read the body.