Application Value of A Clinical Radiomic Nomogram for Identifying Diabetic Nephropathy and Nondiabetic Renal Disease.

Objective: An ultrasound-based radiomics Machine Learning Model (ML) was utilized to assess non-invasively the conditions of diabetic nephropathy and non-diabetic renal disease in diabetic patients.

Methods: A retrospective examination was conducted on 166 diabetic patients who had undergone renal biopsies guided by ultrasound, with the group comprising 114 individuals diagnosed with diabetic nephropathy and 52 with non-diabetic renal disease. The participants were randomly divided into the training set and the testing set (7:3). Following the extraction of radiomics features from the renal ultrasound images, a univariate analysis was conducted, and the Least Absolute Shrinkage And Selection Operator (LASSO) algorithm was applied to select the most significant features. Three ML algorithms were applied to construct the prediction models. Subsequently, the patients' clinical characteristics were evaluated through both univariate and multivariate logistic regression analyses, which facilitated the development of a clinical model, following a clinical radiomics model was formulated, integrating the radiomics scores (Radscore), along with the independent clinical variables identified through the screening process. The diagnostic performance of the three models constructed was evaluated using the receiver operating characteristic (ROC) curve analysis.

Results: Among the three radiomics ML models, the logistic regression (LR) model achieved the best performance, with the area under the curve (AUC) values of 0.872 (95%CI, 0.800-0.944) and 0.836 (95%CI, 0.716-0.957) for the training set and the testing set, respectively. The decision curve analysis (DCA) verified the clinical practicability of the ML model. Within the same testing set, the AUC of the clinical model was 0.761 (95%CI, 0.606-0.916). The nomogram model based on clinical features plus Radscore showed the best discrimination, with an AUC value of 0.881 (95%CI, 0.779-0.982), which was better than that of the single clinical model and the radiomics model.

Conclusion: The ML model of radiomics based on ultrasound images has potential value in the non-invasive differential diagnosis of patients with diabetic nephropathy. The nomogram constructed based on rad score and clinical features could effectively distinguish DN from NDRD.