#5147 PREDICTING NON-DIABETIC KIDNEY DISEASE IN PATIENTS WITH TYPE 2 DIABETES MELLITUS: EXTERNAL VALIDATION OF FOUR PREDICTIVE MODELS

Abstract

Abstract Background and Aims Several prediction models have been published for estimating the probability of non-diabetic kidney disease in kidney biopsies performed in patients with diabetes mellitus. External validation and comparison of these models is essential prior to their clinical application. The aim of this study was to validate and compare the predictive capacity of four previously described clinical models as a tool to predict non-diabetic kidney disease (NDKD) in diabetic patients. Method We retrospectively reviewed kidney biopsies performed in patients with type 2 diabetes mellitus between January 1999 and December 2022 in our centre. The probability of presence of NDKD was calculated by using the clinical and laboratory parameters at the time of the decision to perform a biopsy, according to 4 models; Surinrat's model, Liu's model, Li's model and Garcia-Martin's model (see Table 1). Overall model fit was assessed, calibration curves were plotted and discrimination for each model was assessed by using the receiver operating characteristic (ROC) curve in line with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines. Results The study included 280 patients, 68.9% were males and the mean age was 65.4 ± 11.9 years. 172 patients (61.4%) had a diagnosis of non-diabetic kidney disease. On calculating the risk of NDKD, Surinrat's, Li's and Garcia-Martin's models predicted 33.6%, 64.8% and 53.4% of NDKD when using the “probable NDKD” threshold, and 60.7%, 81% and 90.7% respectively when including the “grey zones”. Fleiss’ Kappa test showed a fair agreement between the four models when using the “probable NDKD” threshold (k = 0.23, 95% CI 0.13-0.33, p<0.001) and when including the “grey zones” (k = 0.39, 95% CI 0.29-0.49, p<0.001). On using the “probable NDKD” threshold, the overall accuracy for Surinrat's, Liu's, Li's and Garcia-Martin's models was 50.7%, 62.7%, 63.6% and 65.3% respectively, while on including the “grey zones” the accuracy slightly increased in the first to models to 54.3% and 66.4% respectively while decreasing in Garcia-Martin's model to 64.2%. Area under the ROC curve (AUC) was higher in Garcia-Martin's model (AUC 0.66, 95% CI 0.58-0.74, p<0.001) compared to Lin's model (AUC 0.59, 95% CI 0.50-0.67, p = 0.044), while Surinrat's and Liu's models did not discriminate (AUC 0.55, 95% CI 0.47-0.64, p = 0.214 and AUC 0.58, 95% CI 0.49-0.63, p = 0.084). Conclusion There is a fair agreement between the four models for predicting non-diabetic kidney disease in type 2 diabetic patients. Of the models studied, Garcia-Martin's model most accurately predicted non-diabetic kidney disease in type 2 diabetes patients, although discrimination remained poor. These models may not be suitable for guiding clinicians on indicating a kidney biopsy in diabetic patients in our population.