#3527 NOVEL MODELS FOR THE PREDICTION OF NON-DIABETIC RENAL DISEASE IN TYPE 2 DIABETES MELLITUS

Abstract

Abstract Background and Aims Accurately distinguishing non-diabetic renal disease (NDRD) from patients with diabetic kidney disease (DKD) in absence of kidney biopsy remains big challenge. We aimed to develop and validate a prediction model to guide the kidney biopsy among DKD patients in clinical practice. Method A total of 3,205 patients with type 2 diabetes mellitus who underwent kidney biopsy between 2000 and 2021 were included. Random sampling approach stratified by predictive outcome was used to acquire the training set (N = 2,244) and testing set (N = 961). The primary outcome was the occurrence of biopsy-proven NDRD. TheLASSO model and XGBoost model using 80 candidate predictors were developed and compared with a Clinical model using currently recommended indicators. The performance of predictive models was evaluated by the areas under receiver operating characteristic (ROC) curves and precision and recall (PR) curves, calibration plot, the categorical net reclassification improvement (NRI) and the integrated discrimination improvement (IDI). Results The training set and testing set included 1,456 and 624 patients with NDRD, respectively. In the testing set, the predictive performance of XGBoost model (ROC-AUC, 0.905; 95% CI, 0.886-0.923) and LASSO model (ROC-AUC, 0.845; 95% CI, 0.820-0.871) were significantly better than Clinical model (ROC-AUC, 0.737; 95% CI, 0.705-0.771) (P <2.2e-16 for both). The LASSO model and XGBoost model also had higher PR-AUC values and significant improvement in reclassification as assessed by NRI and IDI. Utilizing the XGBoost mode in the full dataset based on the low- and high- risk thresholds, we could screen out 559 (49.7%) patients with isolated diabetic nephropathy (DN) to avoid kidney biopsy, and identify 1221 (58.7%) patients with NDRD before kidney biopsy. Conclusion Our models outperformed recommended indicators for guiding the kidney biopsy among patients with DKD. Physicians could use these models to screen out patients with isolated DN to avoid kidney biopsy, and identify patients with high suspicion of NDRD for recommended kidney biopsy.