Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms

AI Generated Summary*

A VGG16 deep-learning model, trained from scratch, was built for each of 47 systemic biomarkers using 236,257 retinal photographs from 72,890 participants in Korean, Beijing, Singapore and UK Biobank cohorts. Training used one Korean screening centre, with four external test sets. Internally, sex (AUC 0.96) and age (R² 0.83) were predicted well, while muscle mass (R² 0.52), height (0.42), bodyweight (0.36) and creatinine (0.38) were moderate. External results were weaker. In the UK Biobank, R² was 0.08 or lower for height, bodyweight and creatinine, and muscle mass was only testable in a second Korean set (0.33). Thirty-seven biomarkers, including thyroid and inflammation measures, were poorly predicted. Only age, sex and blood pressure generalised consistently.

*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.