A retrospective multicenter cohort study that has been performed in eighteen nephrology departments from the Spanish Group for the Study of Glomerular diseases (GLOSEN), the Catalonian Group for the Study of Glomerular diseases (GLOMCAT) and the Spanish Group of Diabetic Nephropathy (GEENDIAB). This study was conducted according to STROBE statement for cohort studies.24 Data from consecutive kidney biopsies performed in patients with diabetes from 2002 to 2014 was collected. The indications for kidney biopsy were: nephrotic syndrome, abrupt reduction of eGFR in patients with previous stable renal function, acute kidney injury (AKI), nephrotic proteinuria without DR, signs or symptoms of systemic disease, proteinuria >1g with DM <5 years of evolution, micro/macrohaematuria and nephrotic proteinuria with DM <5 years of evolution. Patient identification was performed by reviewing risk factors for NDKD in diabetes histopathological charts and clinical histories. Patients included in the study correspond to those published in the before mentioned work.8

The Healthcare Ethics Committee of Parc de Salut Mar, Barcelona, Spain approved the study protocol; the approval number is CEIC2013/5468/I.

A total of 112 variables were studied: 58 (51.8%) were clinical and 54 (48.2%) were laboratory data. Patient demographic characteristics were recorded (age, gender and race), along with history of hypertension, dyslipidemia, duration of DM, presence or absence of diabetic retinopathy (DR) (diagnosed by retinography), diabetic neuropathy, ischaemic heart disease, stroke, peripheral vasculopathy (diagnosed according to the criteria of each center), malignancy and systemic diseased. Furthermore, treatment with renin–angiotensin–aldosterone system blockers (RAASB), oral antidiabetics, insulin, statin and aldosterone antagonists were collected. At the time of renal biopsy, weight, height, systolic blood pressure (SBP) and diastolic blood pressure (DBP) were recorded. Regarding laboratory data, renal function [serum creatinine in milligram per deciliter and estimated glomerular filtration rate (eGFR) according to Modification of Diet in Renal Disease-4 in mL/min/1.73m2], urea in milligram per deciliter, basal level of blood glucose levels in milligram per deciliter, proteinuria (g/24h), urine albumin/creatinine ratio in milligram per gram, urine protein-to-creatinine ratio in milligram per gram, microhaematuria, autoimmune markers [antinuclear antibodies (ANAs), anti-double stranded DNA (Anti-DsDNA), anti-neutrophil cytoplasmic antibodies (ANCAs), anti-glomerular basement membrane (anti-GBM) and cryoglobulins] and viral serology (anti-hepatitis C virus (HCV), surface antigen of the hepatitis B virus (HBsAg) and anti-human immunodeficiency virus (anti-HIV)) were also collected in 683 patients (82%). The indications of renal biopsy were gathered and classified in these categories: nephrotic syndrome, acute kidney injury (AKI), nephrotic proteinuria in patients with diabetes and less than five years of evolution, nephrotic proteinuria without DR, abrupt decrease in eGFR, presence of micro/macrohaematuria, signs or symptoms of systemic disease and proteinuria more than 1g/24h (excluded nephrotic) in patients with diabetes and less than five years of evolution were also recorded. Renal biopsies were reviewed for this study at every participating center. The morphological characteristics found in the biopsy (number of glomeruli, diffuse or nodular mesangial expansion, global or segmental sclerosis, percentage of glomerulosclerosis and increase of basement glomerular membrane) and the final diagnoses were collected. The diagnosis “unclassified” was for kidney biopsies with an insufficient number of glomeruli to interpret the histological findings.

Based on the diagnoses, the renal biopsies were classified into these categories: isolated DN, NDKD or DN-superimposed NDKD (DN plus NDKD). Finally, the follow-up was assessed at first, third, fifth- and tenth-year post-kidney biopsy. The variables collected were renal function (serum creatinine level and eGFR), urea concentration, baseline level of blood glucose, 24-h proteinuria, urine albumin/creatinine ratio, urine protein/creatinine ratio, need for renal replacement therapy (RRT) and death.

Statistical analysis was performed using IBM's SPSS Statistics version 20.0 and STATA. The quantitative variables are expressed in mean and standard deviation and the qualitative variables in percentages. The distribution of variables was assessed using the Kolmogorov–Smirnov test. Univariate comparisons between groups were performed using a Chi-squared test for categorical variables and one-way analysis of variance test for comparing means. We performed a multivariate analysis using binary logistic regression to identify potential predictors of DN vs. NDKD (dependent variable). DN is given the value 0 and the NDKD is given the value 1. The variables for the multivariate analysis have been selected with an automatic “stepwise” method, the LASSO method (“Least Absolute Shrinkage and Selection Operator”, Tibshirani, 1996). This method, unlike stepwise methods, penalized some coefficients to 0 from all to none using a lambda parameter that was maximized. As lambda increased, the variables that were 0 increased. The method automatically selected the best lambda value and the number of variables that best fit the result. The best cut point based on Youden index (Sensitivity+Specificity−1). To use the model in a predictive way, a nomogram was represented that allowed calculating the probability of NDKD using OR obtained in multivariate analysis for each variable included in model. After fitting the model, we evaluated the performance of the model. Three issues were evaluated: (a) Discrimination: The AUC, equivalent to the C-statistic for logistic regression model and their 95% confidence interval will be calculated. It can be interpreted as the probability of correct classification of each pair of subjects with and without NDKD. A value near to 1 means a good classifier. (b) Calibration: It refers to the agreement between observed and predicted endpoints. A graph will be presented and the Hosmer_Lemeshow goodness of fit test will be calculated. (c) Clinical usefulness: The decision-curve analysis provides information of the net true-positive classification rate by using a model over a range of thresholds.

A total of 832 patients with diabetes and consecutive kidney biopsy were included in this study (Fig. 1). The most relevant clinical and analytical data at the time of renal biopsy has been previously published8 (see supplementary Table 1). The histological diagnosis of renal biopsy has also been previously reported.8 A total of 26 cases (3.1%) were “unclassified”. The indications for renal biopsy were: nephrotic syndrome (n=261, 31.4%), abrupt reduction of eGFR in patient with stable CKD (n=173, 20.8%), acute kidney injury (AKI) in patients with previous normal renal function (n=118, 14.2%), nephrotic proteinuria without DR (n=89, 10.7%), signs or symptoms of systemic disease (n=53, 6.4%), proteinuria >1g with DM <5 years of evolution (n=46, 5.5%), micro/macrohaematuria (n=42, 5%) and nephrotic proteinuria with DM <5 years of evolution (n=18, 2.2%).8

Fig. 1.

Flowchart of patients included in the present study.

(0.1MB).

A univariate analysis was performed in order to find differences between histological groups including the next variables: sex, age, presence of DR, ischemic heart disease, peripheral vascular disease, stroke, serum creatinine level, eGFR, proteinuria, presence of microhematuria and time of evolution of DM (Table 1). Patients with DN and NDKD were included in NDKD group to perform this analysis, because this group differs from DN isolated with NDKD lesions in kidney biopsy. With these results and with the purpose of identifying the predictive factors for NDKD, a multivariate binary logistic regression analysis was performed. The variable “time of evolution of DM” was excluded from analysis for missing data. The adjustment of the coefficients for the variables selected according to the Lasso model and readjusted alone were performed (Table 2).

The independent risk factors for NDKD were the presence of microhaematuria, older age, the absence of DR and the absence of peripheral vascular disease. The model's discriminatory capacity obtained a ROC curve with an area under the curve of 0.7242 (Fig. 2) and the C-statistics is 0.742.

Fig. 2.

ROC curve: the model's discriminative ability showed an ROC curve with an area under the curve of 0.742.

(0.05MB).

The best cut point based on Youden index (Sensitivity+Specificity−1) is 58%. The sensitivity of the model to detect non-diabetic nephropathy is 76.9 and the specificity 58.9 (supplementary Table 3).

A nomogram was created that allows for calculating the probability of presenting NDKD with five variables (Fig. 3).

Fig. 3.

Nomogram for calculating the probability of NDKD in diabetic patients.

(0.08MB).

Depending on the value of the variables used for the model, a score number was obtained, which corresponded to a percentage of probability that patients with diabetes present NDKD in the renal biopsy. In our new score, the number obtained increases as does the probability of NDKD. The observed results suggest that the hypothesis of good calibration of the model cannot be rejected as the p value of the test is over 5%. As a clinical example of the use of our nomogram in a daily practice: a 50-year-old patient (5 points), without DR (5.5 points) or microhematuria (0 point) or vasculopathy (2 points) and serum creatinine of 2.5mg/dl (1 point), the total SCORE is 13.5 (60–65% probability of NDKD).

The internal calibration of nomogram is shown in Fig. 4A. Another graph for test calibration is plot the predicted probability vs. the observed frequency. Again, no deviations from the diagonal were observed (Fig. 4B). Finally, with the purpose to evaluate the clinical usefulness, the probability threshold vs. the net benefit (true positive) can be evaluated in the decision curve. NDKD means all positive, and DKD means all negative. The red line that separates from the NDKD and DKD all indicates the probability threshold on model has a net benefit (supplementary Fig. 1).

Fig. 4.

(A) Calibration belt for the logistic regression model. (B) Expected predictions vs. observed frequencies the logistic regression model.

(0.27MB).

In order to obtain an internal validation of score, Bootstrap was used and calculated a shrinkage coefficient. If this coefficient is near 1 one can say that the internal validation is good. In our case, after 1000 bootstrap sample simulations, no differences have been evidenced between the model and bootstrap coefficients. The shrinkage coefficient is almost 1 (0.959), near 1 (supplementary Table 3).