HeartSight

Retinal cardiovascular screening

Estimating heart age from a retinal photograph

The vessels at the back of the eye carry markers of cardiovascular ageing. HeartSight reads a standard fundus photo, the kind any optometrist already takes, and returns a predicted heart age, a risk band, and a saliency map showing where the model looked.

4.66 yr

Mean absolute error

held-out test set

0.227

R² (test)

n = 955 eyes

4,450

Training images

ODIR-5K

ViT-B/16

Backbone

768-d embeddings

How the model performs

Evaluated on the ODIR-5K held-out split. The model is a weak-but-honest predictor: it tracks age on average while regressing toward the population mean.

20355065802035506580Chronological age (years)Predicted age (years)
Predicted vs. chronological age

Dashed line is perfect agreement

073146-30-150+15+30Prediction error (years)Patients
Error distribution

Residuals roughly centred on zero

03691220s30s40s50s60s70s80+Raw modelRecalibratedAge groupMean age gap (years)
Bias by age group

Recalibration shrinks the gap in older patients

Try it on an image

Upload a fundus photo, or start from one of the samples.

Drop a retina photo here

or click to browse

Sample images: