A multimodal neural network that distinguishes between type 1 and type 2 diabetes in young persons using MRI and clinical data

Early diabetic kidney disease (DKD) is common in young persons with type 1 (T1D) and type 2 diabetes (T2D) and accentuates their lifetime risk of kidney failure, requiring dialysis or a kidney transplant. Although clinical manifestations of DKD are similar in T1D and T2D, the structural lesions may differ, and it remains unclear whether DKD in T1D and T2D represent distinct diseases. Accordingly, the objective of this study is to build machine learning (ML) models using clinical and kidney MRI data to classify diabetes status, as well as structural and functional kidney differences of individuals with T1D versus T2D. We hypothesize that a highly accurate multimodal neural network can be constructed that integrates clinical and functional kidney MRI images.Data were obtained from several studies at University of Colorado in youth with T1D (n=102), T2D (n=91) as well as non-diabetic controls (n=60). A total of 253 participants were included in the analyses. First, we applied to clinical data (CASPER, IMPROVE-T2D, CROCODILE, and RENAL-HEIR trials) logistic regression and 7 ML models: extreme gradient boosting machine (XGBoost), XGBoost with grid search, k-nearest neighbors (KNN), support vector machine (SVM), decision tree, random forest, and a 3-layer neural network (NN-EHR). When a small subset of the clinical data was used as features, the NN-EHR yielded the highest accuracy of 84%. Next, we considered the MRI images (27,000 images from the 253 individuals). We applied three convoluted NN to perform the classification: AlexNet, VGG16, and a 4-layer Neural Network for Diabetes Detection (NN4DD). Considering the images alone, VGG16 and NN4DD both achieved an accuracy of > 80%. Additionally, by integrating the clinical data and the MRI images, the fusion neural network achieved an accuracy of almost 100%. Finally, an interpretability analysis of NN4DD indicated notable differences in kidney structure and function among the three groups.Although the cost of MRI is prohibitive and thus impractical for diabetes diagnosis, these results provide a proof-of-concept that fused neural networks that integrate multimodal data can be a valuable diagnostic tool, and provide structural and functional insight on kidney differences between T1D and T2D. This research is sponsored in part by the Natural Sciences and Engineering Council (Canada) and the National Institutes of Health (USA). This is the full abstract presented at the American Physiology Summit 2023 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.