Intelligent Diagnosis of Diabetic Kidney Disease Based on Heterogeneous Data

Diabetic kidney disease (DKD) is a complication of diabetes, with high incidence rate, low prevention rate, and low cure rate. The diagnostic process for DKD is difficult and most of the screening techniques for DKD are invasive. In this paper, we try to use the non-invasive heterogeneous data consisting of fundus images and urine albumin-to-creatinine ratio (UACR) to develop a deep learning method to diagnose DKD. More specifically, we design an image feature module to extract the image information from fundus images while creating a biochemistry feature module to extract the information from UACR. After concatenating the two kinds of information, we propose a decision module to classify DKD or non-DKD. We apply the method to the heterogeneous dataset, where the results show that our method has 82.4% accuracy, 85.7% area under the receiver operating characteristic curve (AUC), 87.2% sensitivity, and 12.8% false negative rate (FNR). The results prove that the method we proposed helps diagnose DKD by using non-invasive heterogeneous data, such as fundus images and UACR. Besides, the ablation experiments indicate that fundus images contain much information about DKD and can be an important reference for diagnosing DKD.