A deep-learning based tool for prediction of chronic kidney disease from retinal images in people with type 2 diabetes

Background

The burden of chronic kidney disease (CKD) is increasing due to exponential increase in diabetes and hypertension globally. The association between diabetic retinopathy (DR) and diabetic kidney disease (DKD) suggests common pathways of microangiopathy.  Application of Artificial intelligence (AI) and deep learning (DL) in medicine aids early detection, disease diagnoses and predictions and enables appropriate timely intervention/ management. Retinal colour photography can be a non-invasive approach for identification of early microvascular alterations of chronic systemic complications especially CKD prior to the onset of overt clinical presentation.

Objectives:

To develop and validate a DL based algorithm to detect and prognosticate the risk for development of CKD using retinal images and clinical data in individuals with type 2 diabetes.

Methods:

Anonymized clinical metadata and corresponding retinal images of 2000 individuals with type 2 diabetes with and without DR will be utilized for development of DL model to detect CKD. Systemic parameters such as age, gender, duration of diabetes, glycated haemoglobin (HbA1c) and hypertension (HT) will be utilized along with retinal colour photographs for this model. Longitudinal data will be used for developing the prognostic tool to detect people at risk of stage 3 CKD (defined as estimated glomerular filtration rate [eGFR] < 60 ml/ min/ 1.73m2).  Convolution neural network-support vector machine (CNN-SVM) DL model will be used.  After the assessment of accuracy (sensitivity and specificity) of the DL algorithm in individuals with type 2 diabetes, the external validation will be carried out prospectively among 400 patients.

Expected outcome:

The tool would aid early detection and prediction of CKD (before it reaches stages 4 or 5), easily manageable with glycemic control and blood pressure control; hence reduce the morbidity and healthcare costs due to CKD.  Holistic screening for DR and CKD in individuals with diabetes would be possible through non-invasive retinal imaging with use of AI.