Development of a metabolite-based deep learning algorithm for clinical precise diagnosis of the progression of diabetic kidney disease

Diabetic kidney disease (DKD) is one of the most important microvascular complications of diabetes mellitus (DM). Early recognition and intervention in the treatment of DKD may delay the progression to end-stage kidney disease, but it is still a challenging task of early diagnosis of DKD and monitoring of its progression. Blood metabolites reflect the complex situation of human beings and provide a possible direction for disease diagnosis in clinics. In this paper, inspired by the recent success of deep neural networks in medical data understanding, we designed a learning-based method to capture the metabolic complexity and diagnose DKD accurately. Particularly, we trained a convolutional neural network (CNN) and fully connected network (FC) based on the metabolite dataset from 1521 clinical participants acquired by GC–MS. We proposed a novel data pre-processing method that translates the collected metabolite data into corresponding images followed by a normalization function, which is efficient for the deep neural network to extract robust features from metabolite data and mining the potential biomarkers of diseases. A metabolite-based deep neural network was firstly constructed, which is mainly used for the early stage of DKD (accuracy: 83.3%, sensitivity: 83.2%, specificity 80.2%) and the advanced stage of DKD (accuracy: 83.3%, sensitivity: 82.9%, specificity 81.0%) diagnosis. Meanwhile, it can also be applied to diagnosis of DM (accuracy: 83.3%, sensitivity: 83.4%, specificity: 80.4%) and CKD (accuracy: 87.0%, sensitivity: 83.7%, specificity: 81.2%). Our work presents the potential of using the metabolite data to construct an AI-enabled disease diagnosis system and finally being applied in real-world clinics, as well as provides new data types for AI application in medicine.