Abstract
Purpose :
Diabetic kidney disease (DKD) is a chronic condition of advanced diabetes, and there is a pressing need to detect DKD at an early stage. Past studies have typically used invasive methods such as serum creatinine and urine protein along with microvascular changes in the retina to predict the risk of DKD. In this study, a deep learning (DL) based automated workflow was developed to predict DKD stage, and progression from retinal images along with other non-invasive clinical tests.
Methods :
A CNN classifier based on VGGnet was first trained with annotated in fundus images, such as cotton wool spots, exudates, aneurysms, and microvascular changes, as inputs to predict diabetic retinopathy (DR) stages (Fig. 1A). The inferred DR stage from this CNN, patient demographics, duration of diabetes, comorbidities (e.g., cardiac disease, sepsis, urinary infections, peripheral neuropathy, etc), history of hypertension, and non-invasive nephrological data of patients such as 24-hour urine protein and creatinine were used to train a machine learning (ML) random forest classifier to predict the stage of DKD. A 2D CNN was used to predict the referral, and a 3 class CNN classifier for baseline progression based on the staging. Data of 643 patients were used for training, 168 for validation, and 159 for testing.
Results :
The sensitivity of DR classfifer was 88.71%. The sensitivity and specificity of the referral workflow were 79%, and 86%, respectively. The diagnostic efficacy of the referral algorithm was AUC: 86% (Fig. 1B). Further, the sensitivity of baseline progression was found to be 82%. The results indicate that there exists a strong correlation between DR stages found in fundus and DKD stages.
Conclusions :
The presented novel workflow predicts the DKD stage in patients with diabetes based on fundus images. This method shall provide ophthalmologists with an additional decision-making support tool on whether to refer the patient to a nephrologist for detailed renal assessment based on the predicted stage of kidney disorder. Further investigations shall focus on categorizing progression rate of the kidney disease into stable, slow, and rapid progressor shown in Fig. 2.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.