Predicting the risk of developing diabetic retinopathy using deep learning

AI Generated Summary*

Two deep-learning versions, one using three-field and one using single-field fundus photographs, were built to estimate whether eyes free of retinopathy at screening would develop mild or worse disease within two years. Development used 575,431 US teleretinal eyes, though only 28,899 had known outcomes. In 3678 internal eyes, the three-field version reached an AUC of 0.79. The single-field version reached 0.70 on 2345 Thai eyes. Adding the model (three-field internally, single-field externally) raised AUC from 0.72 to 0.81 over a four-factor clinical set internally, and from 0.62 to 0.71 over glycated haemoglobin alone externally. High-risk groups developed disease significantly faster than low-risk groups in survival analysis. Saliency maps sometimes, not always, highlighted areas that later developed lesions.

*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.