The retina can be a source of subtle signs of disease. Yet visual inspection of microvasculature, nerves and connective-tissue structures in the retina has only led to a few hallmarks of disease — most notably, of lesions of diabetic retinopathy — that can be incorporated into clinical guidelines as criteria for screening and diagnosis1. In the past few years, the application of deep learning to the analysis of retinal fundus images has shown that retinal tissue can also reveal information about cardiovascular risk (through clinically relevant risk factors2), and that such trained neural networks can be used to predict retinal-vessel calibre3, coronary artery calcium scores4,5, low blood haemoglobin6, risk of chronic kidney disease7 and a host of systemic parameters, such as body mass index (BMI) and creatinine8. This suggests that deep learning could eventually be implemented clinically to examine a patient’s health and for the health screening of populations, conceivably improving affordability and accessibility. However, at present, the development of deep learning for health-screening purposes is at an early stage, and the vast majority of proof-of-concept work has not yet been clinically validated. Writing in Nature Biomedical Engineering, Kang Zhang, Ting Chen, Tao Xu, Guangyu Wang and colleagues now show that deep-learning models can be used to detect chronic kidney disease (CKD) and type 2 diabetes mellitus (T2DM) solely from retinal fundus photographs (collected using standard table-top fundus cameras) or in conjunction with patient metadata9. Crucially, the researchers validated their findings across multiple geographically distinct patient datasets from China, including a dataset prospectively collected under point-of-care (POC) settings using a custom smartphone-based system.