Clinical prediction score for diagnosing non-diabetic renal disease in patients with type 2 diabetes mellitus: a cohort study

Background: When non-diabetic renal disease (NDRD) is suspected, kidney biopsy is used for denite diagnosis; however, this is not always easily available and may lead to complications. A clinical prediction score may help selecting appropriate patients for kidney biopsy. Methods: A retrospective cohort study was conducted in type 2 diabetes mellitus (T2DM) patients with atypical features of diabetic nephropathy (DN), who had kidney biopsy at Thammasat University Hospital from 2011-2019. We divided patients into NDRD alone, coexisting NDRD and DN, and DN alone, conrmed by pathological diagnosis. We developed a clinical prediction score by weighing coecients of predictors in a multivariable logistic model. Internal validation was performed with bootstrapping. Results: We included 81 patients: 28 (34%) had NDRD alone, 15 (18%) had coexisting NDRD and DN, and 38 (41%) had DN alone. Primary membranous nephropathy, primary focal and segmental glomerulosclerosis (FSGS), and secondary FSGS were prevalent in any NDRD. Absence of diabetic retinopathy (DR) showed a signicant association with any NDRD (OR 3.72; 95% CI, 1.28-10.8; p=0.02). The prediction score, AUC of 0.75 (95% CI, 0.63-0.86), had four predictors: duration of DM <10 years, eGFR >30 ml/min/1.73m 2 , HbA1c <8%, and absence of DR. Higher scores were associated with higher probability of NDRD. Conclusions: This clinical prediction score appears to be a useful tool to determine NDRD probability. T2DM patients atypical DN scores defer kidney biopsy.


Background
Diagnosing non-diabetic renal disease (NDRD) in patients with type 2 diabetes mellitus (T2DM) is often problematic and typically requires a kidney biopsy for de nite diagnosis. The common pathological diagnoses of NDRD are membranous nephropathy (MN), focal segmental glomerulosclerosis (FSGS) and IgA nephropathy [1][2][3]. Unfortunately, it remains di cult to diagnose as NDRD may present alone or in combination with diabetic nephropathy (DN); the latter obscures the classical presentation of each disease. Thus, kidney biopsy is the standard investigation for de nite diagnosis [4]. However, this invasive procedure may have bleeding complications and its availability is limited in some resourceconstrained hospitals.
The decision to perform a kidney biopsy depends on the likelihood of NDRD. According to the Kidney Disease Outcomes Quality Initiative (KDOQI) guideline [4], patients with atypical presentations of DN with the absence of diabetic retinopathy (DR), low or rapidly declining glomerular ltration rate (GFR), rapidly increasing proteinuria or nephrotic syndrome, the presence of active urinary sediment, or signs and symptoms of systemic disease, should be evaluated for NDRD. Even using these criteria, NDRD prevalence was found in only half of some biopsy reports [1,[5][6][7]. It appears there is a knowledge gap toward improving NDRD diagnostic performance.
Previous studies have only reported predictors, such as absence of DR, duration of diabetic mellitus (DM), degrees of proteinuria, Hemoglobin A1c (HbA1c), or levels of creatinine or GFR associated with NDRD, but none have mentioned the utility of a combined predictive probability [3,[7][8][9]. A clinical prediction score combines predictors in the model, informs clinicians and patients about disease probability, and can aid in decision-making [10]. In this study, we aimed to develop a simpli ed clinical prediction score for NDRD to help determine the appropriate clinical setting for kidney biopsy in T2DM with atypical presentation.

Methods
Data were collected by retrospective chart review; we included all patients with T2DM who had undergone kidney biopsy ³18 years old at Thammasat University Hospital between January 2011 and December 2019. The diagnosis of T2DM was obtained from history and criteria established by the American Diabetes Association [11]. We excluded patients with history of kidney transplantation and inadequate specimen for interpreting pathological diagnosis from kidney biopsy. The study was approved by the Human Research Ethics Committee of Thammasat University No 1 (Faculty of Medicine), certi cate of approval 123/2020.
We collected all clinical parameters including age, gender, body mass index (BMI), duration of DM, presence of DR, hypertension, established cardiovascular diseases (CVD), also indications for kidney biopsy. The presence of DR was examined and recorded by an ophthalmologist. The duration of DM refers to the time from rst diagnosis to kidney biopsy. Hypertension was de ned as either having a recorded past history of hypertension or systolic blood pressure ≥140 mmHg and/or diastolic blood pressure ≥90 mmHg, determined at clinical evaluation for kidney biopsy in the absence of any antihypertensive drug. CVD was a past history of myocardial infarction or congestive heart failure in previous medical records.
Laboratory results were collected at the time of biopsy including complete blood count (CBC), blood urea nitrogen (BUN), serum creatinine using enzymatic method with estimated glomerular ltration rate (eGFR) using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) [12], urinary analysis, urine protein creatinine ratio (UPCR), fasting plasma glucose (FPG), and HbA1c. Hematuria was de ned as a red blood cell (RBC) count of ≥3 cells in urine examination. UPCR was calculated by dividing urine protein (mg/dL) by urine creatinine (mg/dL).
All kidney specimens were examined using light microscope and immuno uorescence assay. Because our hospital does not routinely perform electron microscope, very few cases had these results. DN diagnosis and classi cation were made using criteria established by the Renal Pathology Society in 2010 [13]. The pathological diagnosis was divided into three groups consisting of DN, NDRD, and coexisting DN and NDRD (DN+NDRD).

Statistical analysis
All statistical analyses were performed using Stata software, v16.0 (StataCorp). All p values were twosided; p value <0.05 was considered statistically signi cant. Continuous variables data were presented as mean ± standard deviation (SD) or median and interquartile range (IQR). Categorical variables were expressed as frequency and percentages.
In the processes of model speci cation, we reviewed literature, which reported predictors of NDRD, and explored associations with any type of NDRD (NDRD alone or DN+NDRD) in univariable logistic model. Any predictors, in which p value <0.2 in univariable analyses, were selected for multivariable logistic regression. Then, we tested the performance of the nal model with Hosmer-Lemeshow test and area under receiver operating characteristic curve (AUC).
To generate the clinical prediction score, coe cients of all predictors were weighed by dividing the lowest coe cient and rounded into integers. The scores of individual patients were matched with the estimated probability of it being NDRD in the nal model, also referred to as the developed model. Next, these individual scores were tested for predictive performance using AUC.
We performed internal validation using a bootstrapping procedure with a 1,000-bootstrap sample. This procedure quanti ed the optimism of the developed model. We revised the new model, now called the optimism-adjusted model using the uniform shrinkage factor. Then, the scores of individual patients were matched again with estimated probability in this optimism-adjusted model. AUC was again tested.
For clinical implications, we now proposed an algorithm by categorizing the clinical prediction scores into three groups: low probability of NDRD, intermediate probability, and high probability with our suggestions for kidney biopsy decision.
The baseline clinical characteristics of all patients, as well as DN, NDRD, and DN+NDRD groups, are shown in Table 1. Mean age of all patients was 56.0 ± 13.1 years. The amount of males and females were proportional. Median duration of DM was highest in the DN group being ~10 years, with an average of ve years for both NDRD and DN+NDRD group. Presence of DR was predominantly high in the DN group, accounting for 63% of those cases, with 53% having DN+NDRD, and 25% NDRD. Mean eGFR was lowest in the DN group as 31.6 ml/min/1.73m 2 , with 40.6 ml/min/1.73m 2 in DN+NDRD, and 42 ml/min/1.73m 2 in NDRD. Hematuria was found in high proportion in all three groups, but urine RBC >30 cells/HPF predominated in the NDRD group. Mean HbA1c was 7.8% for the DN group, with 7.4% in the DN+NDRD group, and 6.9% in NDRD group.
Pathological ndings of NDRD Forty-three patients were diagnosed with some type of NDRD: 28 NDRD alone and 15 DN+NDRD ( Table  2). Primary membranous nephropathy, primary focal segmental glomerulosclerosis and secondary focal segmental glomerulosclerosis were the three most common lesions, each accounting for 14%. Pauciimmune glomerulonephritis, IgA nephropathy, and postinfectious glomerulonephritis were also prevalent in these groups.
In Table 4, we compared the estimated probability of being NDRD from the developed and optimismadjusted models. The score ranged from 0 to 8, with a higher score associated with greater probability of NDRD. The score was divided into three categories: low probability of NDRD (score 0-2), intermediate probability (score 3-5) and high probability (score 6-8) (Figure 1). The clinical prediction score had an AUC of 0.75 in the developed model and 0.70 in the optimism-adjusted model.

Discussion
This study aimed to develop a tool for diagnosing NDRD in T2DM patients with atypical presentations of DN. We found some predictors that indicated when it was more likely for NDRD to occur, such as duration of DM, eGFR, HbA1c, and absence of DR. Clinical predictive scores from these can be used as a risk strati cation tool whether kidney biopsy should be performed or not.
The majority of our cases had kidney biopsies due to sudden onset of proteinuria or nephrotic syndrome or rapid eGFR decline; this conformed with standard indications in the KDOQI guideline [4]. Unfortunately, we found only around half (53%) of the study population had any type of NDRD (NDRD alone or DN+NDRD), and the remainder had DN alone. This may mean the kidney biopsy was unnecessarily performed in those cases and atypical features can be found in DN. Previous kidney biopsy studies reported varying prevalence of NDRD. In a literature review, Kanodia et al [1], found the percentage of NDRD ranged from 45-75%. Sharma et al [3] reviewed 620 biopsies in patients with diabetes and noted 63% had NDRD. Because of this moderate yield in NDRD detection, there is a room for a diagnostic tool to differentiate T2DM patients with low or high probability of NDRD.
Previous research has reported some predictors are associated with NDRD. Similar to our ndings, the absence of DR was the strongest predictor of NDRD in many studies [2,5,6,8,9,14,15]. In other words, when DR is present, it is a suggestive of DN in T2DM, because both are microvascular complications. Longer duration of DM is inversely associated with NDRD. Dong et al [8] found DM history ≤5 years had OR of 4.6 (95% CI, 1.7-12.5), Kritmetapak et al [7] reported a duration of DM >8 years with an OR of 0.15 (95% CI, 0.04-0.49), and nally, Yang et al [9] showed a duration of DM <10 and >5 years had OR of 0.06 (95% CI, 0.97-0.75). In our opinion, the exact duration of T2DM seems di cult to obtain. Therefore, we used a cutoff of <10 years for simplicity and found signi cant associations in the univariable model but not within the multivariable model: this latter phenomenon may be due to low power. Higher renal function was found to be a signi cant predictor in a recent study [9], showing eGFR ≥90 ml/min/1.73m 2 had OR of 6.38 (95% CI, 1.58-25.7). High HbA1c or fasting blood sugar were reported in a few studies [5,7]. Unlike previous reports [6,9,15], we could not nd any association of proteinuria levels with NDRD as the majority of our population already had macroalbuminuria.
This study may be the rst to combine predictors into a comprehensive risk strati cation score. For clinical purpose and ease of use, we divided these scores into low, intermediate and high probabilities of NDRD. We have also proposed an algorithm of care for each category, especially in the low probability of NDRD group, where patients may defer kidney biopsy (Figure 1). This could be useful in settings with limited resources for kidney biopsy and prevent patient discomfort and surgical complications. However, with intermediate and high probability, a kidney biopsy is still suggested for de nite diagnosis of NDRD.
We must still point out that there are some essential limitations to consider in our work here. First, the small sample size is likely to have affected the statistical power, potentially creating a risk of over tting and optimism bias. We attempted to manage this by choosing predictors from previous studies and using a p value of 0.20 to select predictors into the multivariable analysis, without backward elimination if the predictors were not statistically signi cant. Internal validation was done with bootstrapping to adjust the developed model for optimism. Second, the inherent nature of retrospective data collection can affect data quality in the veri cation of outcomes or missingness. As around 5% of our data was missing for the essential predictors, complete-case analysis was used.

Conclusions
A clinical prediction score for NDRD is a useful risk strati cation tool for kidney biopsy in T2DM patients with atypical presentations. Using multiple predictors as opposed to a sole one appeared to improve the predictive ability; hopefully, this kind of score can lead to the deferment of unnecessary kidney biopsy. It would be interesting to apply this prediction score in other populations at other centers to observe if we can ameliorate its external validity.

Declarations
Ethics approval and consent to participate The study was approved by the Human Research Ethics Committee of Thammasat University No 1 (Faculty of Medicine), certi cate of approval 123/2020.

Consent for publication
Not applicable.

Availability of data and materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests Funding Not applicable.
Authors's contributions E.S. and P.T. designed the study. E.S. collected data. P.T. and Th.E. analyzed the data and made the gure. All the authors interpreted and gave suggestions for additional analyses. E.S. and P.T. drafted the manuscript. All the authors revised the manuscript and approved the nal version of the manuscript. with useful data regarding kidney biopsies and other information. We also thank Debra Kim Liwiski for the English editing.     years", score of 2 for "eGFR >30 ml/min/1.73m 2 " and "HbA1c <8%", and score of 3  Figure 1 Proposed algorithm using the clinical prediction score A proposed algorithm for clinical use of the clinical prediction score, composed of four predictors and categorized into three groups. The low probability (score 0-2) of NDRD group means less likelihood of NDRD (or likely DN), which may defer kidney biopsy.