Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images

Abstract

Regular health screening plays a crucial role in the early detection of common chronic diseases and prevention of their progression. An AI system capable of recapitulating early disease detection, staging and incidence prediction would help to improve healthcare access and delivery, particularly in resource-poor or remote settings. Using a total of 115,344 retinal fundus photographs from 57,672 patients (with data split into mutually exclusive training, internal testing, and external validation sets), we first developed AI models capable of identifying chronic kidney disease (CKD) and type 2 diabetes mellitus (T2DM) based on fundus images. The AI system was shown to be capable of predicting the clinical indicators of CKD and T2DM (including eGFR and blood glucose levels), which indicates its potential for extracting quantitative clinical metrics embedded subtly within retinal fundus images. We further developed an AI system to predict the risk of disease progression using baseline images of 10,269 patients for whom longitudinal clinical data were available for up to 6 years, which demonstrated potential utility in optimizing health screening intervals and clinical management. The generalizability of the AI system in identifying and predicting the progression of CKD and T2DM was evaluated using population-based external validation cohorts. Moreover, a prospective pilot study with 3,081 patients was also conducted to demonstrate the broader applicability of the AI system at the 'point-of-care' using fundus images captured with smartphones. The results provide proof-of-concept for a reliable and non-invasive AI-based clinical screening tool based on fundus photographs for the early detection and incidence prediction of two common systemic diseases.

Artificial Intelligence to Detect Meibomian Gland Dysfunction From in-vivo Laser Confocal Microscopy

“…In addition, due to the low medical and economic resources in underdeveloped and remote areas, the ophthalmologist to patient ratio is low, leading to delayed diagnosis and deterioration of MGD in some patients. For example, in Nigeria, reported physician to patient ratio is as low as 1:2,660 ( 27 ). An efficient and socially effective method for evaluating meibomian gland function is therefore needed urgently.…”

Optical Coherence Tomography Angiography in Type 1 Diabetes Mellitus—Report 2: Diabetic Kidney Disease

“…The implications of this finding are relevant as a blood test is currently required to assess this risk. In this area of research, recent studies have highlighted the potential of artificial intelligence (AI) algorithms to identify CKD categories from retinal photographs using existing datasets from DR screening programs [ 41 , 42 ]. However, the performance of these models is not directly comparable to our results for a variety of reasons.…”

Multi-Task Learning-Based Immunofluorescence Classification of Kidney Disease

“…A study showed that deep learning could diagnose DN based on only six types of IF staining [18]. There have also been studies that predicted systemic conditions and clinical metrics based on fundus photographs, suggesting that deep-learning algorithms can detect subtle associations that are undetectable to human observers [19,20].…”