Detection of signs of disease in external photographs of the eyes via deep learning

AI Generated Summary*

A deep learning system trained on external eye photographs from 145,832 patients with diabetes at 301 California screening sites was evaluated in 48,644 patients at 198 further sites across 18 US states. For HbA1c of 9% or higher, AUCs fell between 67.6% and 73.4%, against 59.8% to 66.5% for logistic regression on self-reported characteristics such as age, sex and, where recorded, diabetes duration. AUCs for moderate or worse retinopathy, macular oedema and vision-threatening disease were 75.0%, 77.9% and 79.2% in undilated eyes and 84.0% to 86.7% in one dilated set. Exploratory lipid predictions were weaker (57.9% to 67.1%) and did not reliably beat baseline. Saliency maps favoured conjunctival and pupil regions; performance with consumer cameras was not tested.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.