Development of a tool for predicting HNF1B mutations in children and young adults with congenital anomalies of the kidneys and urinary tract

Background We aimed to develop a tool for predicting HNF1B mutations in children with congenital abnormalities of the kidneys and urinary tract (CAKUT). Methods The clinical and laboratory data from 234 children and young adults with known HNF1B mutation status were collected and analyzed retrospectively. All subjects were randomly divided into a training (70%) and a validation set (30%). A random forest model was constructed to predict HNF1B mutations. The recursive feature elimination algorithm was used for feature selection for the model, and receiver operating characteristic curve statistics was used to verify its predictive effect. Results A total of 213 patients were analyzed, including HNF1B-positive (mut + , n = 109) and HNF1B-negative (mut − , n = 104) subjects. The majority of patients had mild chronic kidney disease. Kidney phenotype was similar between groups, but bilateral kidney anomalies were more frequent in the mut + group. Hypomagnesemia and hypermagnesuria were the most common abnormalities in mut + patients and were highly selective of HNF1B. Hypomagnesemia based on age-appropriate norms had a better discriminatory value than the age-independent cutoff of 0.7 mmol/l. Pancreatic anomalies were almost exclusively found in mut + patients. No subjects had hypokalemia; the mean serum potassium level was lower in the HNF1B cohort. The abovementioned, discriminative parameters were selected for the model, which showed a good performance (area under the curve: 0.85; sensitivity of 93.67%, specificity of 73.57%). A corresponding calculator was developed for use and validation. Conclusions This study developed a simple tool for predicting HNF1B mutations in children and young adults with CAKUT. Graphical abstract A higher resolution version of the Graphical abstract is available as Supplementary information Supplementary Information The online version contains supplementary material available at 10.1007/s00467-023-06262-9.


Introduction
Mutations in the HNF1B gene are the leading monogenic cause of congenital anomalies of the kidneys and urinary tract (CAKUT) [1][2][3][4], which, in turn, is a common cause of kidney failure [5].The product of this gene, i.e., hepatocyte nuclear factor 1β (HNF1β), is expressed in many organs, predominantly the kidney, liver, intestine, pancreas, and urogenital tract.Patients with HNF1B-related disease may exhibit organ involvement and a wide range of biochemical disturbances.Thus, the spectrum of phenotypes in patients harboring pathogenic HNF1B variants is pleotropic [6].The kidney phenotype is highly variable, yet it is the earliest and most frequent manifestation [6][7][8].The involvement of other organs makes patients very vulnerable.Given this, patients with HNF1B-associated disease may have special medical needs and require multispecialized care.In this respect, the molecular diagnosis of HNF1B-related disease is crucial, as early diagnosis allows for personalized care.Accurate patient selection is also essential in terms of a cost-effective approach, which is of special importance in areas with Extended author information available on the last page of the article limited resources.However, selecting appropriate patients for genetic testing remains a challenge for clinicians [9,10].
The benefits of timely detection and the challenges involved in making the diagnosis of HNF1B disease have driven clinicians to develop tools to predict the likelihood of an HNF1B mutation in patients with CAKUT.Faguer et al. [9] proposed a scoring system based on 17 parameters, with kidney hyperechogenicity and/or cysts and genital and pancreatic abnormalities scoring highest, followed by other features, such as abnormal fetal kidney ultrasound, positive family history, kidney hypoplasia or dysplasia, and hypomagnesemia.However, perhaps due to its complexity, this scoring system is not widely used in clinical practice [6] and requires revision in the pediatric population [11].One of the most important limitations of the score is that it was established in patients of all ages, which calls into question its use in children.To improve the score, Raaijmakers et al. [12] proposed to modify Faguer's criteria, which were validated in a prospective manner.However, this approach yielded a low detection rate (10%).Therefore, the authors, based on a post hoc analysis of their results, suggested consideration of HNF1B mutations in patients with bilateral kidney anomalies, with kidney dysplasia and cysts, as well as concomitant hypomagnesemia.Importantly, the above criteria proposed by Faguer and Raaijmakers were developed in a mixed population of adults and children by using adult standards to define disease-specific disorders, such as hypomagnesemia and hyperuricemia.In fact, the reference values for serum magnesium (sMg) and serum uric acid (sUA) are highly dependent on age [13][14][15].Considering the age-dependent variability in biochemical parameters and the dynamic nature of the disease, pediatric patients, for whom a uniform definition of biochemical disorders should not be applied, warrant a special approach.In fact, Kołbuc et al. [16] showed that hypomagnesemia is underdiagnosed in children with HNF1B due to applying an inaccurate lower limit of normal for sMg.
Thus, this study aimed to establish an accurate and simple predictive model based on clinical and laboratory data and to design a corresponding calculator of HNF1B probability for clinical use.Our goal was to develop a tool dedicated to children with CAKUT.Hence, we utilized the most recent age-appropriate norms to define biochemical abnormalities.

Study population
This retrospective multicenter study was designed to collect data from children and young adults with known HNF1B mutation status.The main source of the study cohort was the Polish Registry of Inherited Tubulopathy (POLtube), which oversees the recruitment and data collection of tubular disorders and other rare kidney diseases including HNF1B nephropathy in Poland.Additionally, to increase the number of individuals to make statistical analysis possible, and to eliminate a possible selection bias, we identified study subjects by approaching physicians and research groups from other countries, who had collaborated with us in previous studies.A spreadsheet with requested information was sent to clinicians with a request to provide clinical, biochemical, and genetic results of patients with known HNF1B status.Patient selection for molecular testing was at clinicians' discretion and was not subject to analysis in our study.Informed consent for the study was obtained individually according to local regulations in each participating center/ country.Ethical approval was waived by the local Ethics Committee of the University of Zielona Gora in light of the retrospective nature of this study.The exclusion criteria included age > 25 years, chronic kidney disease stage 5 or post-kidney transplant status, absence of anomaly of kidney and urinary tract, insufficient clinical data, or lack of genotype.Between January 2021 and October 2022, data from 234 patients were collected.Of these, the records of 21 subjects were discarded (Fig. 1).The study cohort ultimately comprised 213 patients, of whom 109 subjects were positive for HNF1B mutations (mut +) and 104 were negative (mut −).The patients were from Poland (66.7%), Germany (12.7%),South Korea (12.7%),Russia (4.7%), North Macedonia (1.7%), and Lithuania (1.5%).

Clinical assessment and definitions of analyzed parameters
Anonymized data provided by referring physicians were retrospectively reviewed.Data from the time of genetic testing were used, and if unavailable, from the latest visit.The CAKUT phenotype was defined according to the sonographic criteria suggested by Faguer et al. [9].Patients were also classified as having a major or minor kidney anomaly, according to Raaijmakers et al. [12].
Hypomagnesemia and hyperuricemia were defined according to age and sex (for sUA) reference values [15].To demonstrate the effect of an appropriate cutoff value for the definition of hypomagnesemia, we used a conservative lower limit for sMg of 0.7 mmol/l.Hypokalemia was defined as a serum potassium level < 3.5 mmol/l, and hyperparathyroidism was defined as a parathyroid hormone (PTH) level > 65 pg/ml.Patients receiving supplements of magnesium or potassium or those being treated with allopurinol were also considered to have hypomagnesemia, hypokalemia, or hyperuricemia, respectively.The raw sMg, serum potassium, and sUA values of these patients and those who were receiving diuretics or renin-angiotensin-aldosterone system inhibitors were not included in the analysis.
Hypermagnesuria was defined as a fractional excretion of magnesium (FEMg) > 5% [16][17][18], and hypocalciuria as a fractional excretion of calcium < 1%.The estimated glomerular filtration rate (eGFR) was calculated using the Schwartz equation for patients < 20 years of age [19,20] or the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation for those who were > 20 years of age [21].The subject's severity of chronic kidney disease (CKD) was classified according to the National Kidney Foundation Kidney Disease Outcomes Quality Initiative CKD staging system, considering physiologically lower GFR in children < 2 years [22].Standard deviation scores (SDS) for height and body mass index (BMI) were calculated based on World Health Organization growth charts (http:// www.who.int/ growt href/ en).Growth impairment was defined as height-SDS < − 2, overweight as BMI-SDS > 1, or obesity when BMI-SDS > 2.
Genetic testing was conducted locally, and molecular results were reported by the referring physician.In the vast majority of patients (79.8%), mutational analysis was performed by targeted Sanger sequencing of exons 1-9 of the HNF1B gene, and by MLPA analysis to detect copy number variations.In a small proportion of patients, CAKUT/PKD multigene panels were performed (16.9%) or whole-exome sequencing (3.3%).All variants were evaluated using online databases, including HGMD Professional, OMIM, PubMed, ClinVar, and VarSome databases.The pathogenicity of the variants was classified using the American College of Medical Genetics and Genomics criteria.

Statistical analysis
Data are presented as means (95% confidence interval) for continuous variables and as counts (percentages) for categorical variables.Hypotheses of differences between mut + and mut − groups were tested using Welch's independent-samples t test for continuous variables and Fisher's exact test for categorical variables.Cohen's d and odds ratios (OR) were used as effect size coefficients.Pearson's correlation coefficient was used to investigate the strength of correlations.
The subjects were randomly divided into a training set (70%) and a test (validation) set (30%).To choose the best algorithm for the construction of a predictive model for the HNF1B mutation, various machine learning algorithms were tested.The recursive feature elimination (RFE) algorithm was used for feature selection into the model, and receiver operating characteristic (ROC) curve statistics were used to verify its predictive effect.As a global statistical significance level, α = 0.050 was used.All analyses were performed using R, SQL, and Python programming languages supported by multiple packages.

Patient characteristics
The clinical characteristics of the two cohorts (mut + and mut −) are highlighted in Table 1.Both groups were comparable with respect to age, gender distribution, and anthropometric parameters.The majority of patients had CKD stage 1 or stage 2 (88.24% and 83.67% of mut − and mut + , respectively).The mean eGFR was slightly lower in the mut + group compared to the mut − group (p = 0.025).However, impaired kidney function was not suggestive of the HNF1B mutation.
Not surprisingly, we found hypermagnesuria (64.90%), hypomagnesemia (59.00%), hyperuricemia (40.57%), liver involvement (12.84%), and diabetes mellitus (DM; 11.11%) to be more common in mut + than in mut − subjects.Pancreatic anomalies were proven in 11.93% of patients and were almost exclusively found in those with HNF1B mutations (OR 13.95, p = 0.001).Organ involvement is presented in more detail in Supplementary Table S1.None of the subjects presented with hypokalemia; however, the mean serum potassium level was lower, and the fractional Table 1 Characteristics of the study groups BMI-SDS body mass index standard deviation score, CKD chronic kidney disease, eGFR estimated glomerular filtration rate, FECa fractional excretion of calcium, FEK fractional excretion of potassium, FEMg fractional excretion of magnesium, FEUA fractional excretion of uric acid, Height-SDS height standard deviation score, mut + HNF1B positive patients, mut − HNF1B negative patients, PTH parathyroid hormone, sMg serum magnesium, sUA serum uric acid a Mean (95% confidence interval) for continuous; n (%) for categorical b Cohen's d for continuous, odds ratio for categorical c Welch two-sample t test for continuous, Fisher's exact test for categorical d According to reference values suggested by Ridefelt et al. [15] e Described in detail in Suppl.The kidney phenotype is presented separately for the two cohorts (Table 2).Cystic kidney disease and multicystic dysplastic kidney were the most common phenotypes in both cohorts.Their frequencies, similar to the other kidney phenotypes, were similar between groups.Bilateral anomalies were observed more frequently in the mut + group and had a high discriminative value (OR 4.85, p < 0.001).Notably, the percentage of major and minor kidney anomalies did not differ between groups.

Genetic results
Detailed information on the status of the detected variants is presented in Supplementary Table S2.Whole gene deletion was the most common mutation (55.04%) (Supplementary Table S3).One hundred and four patients (95.4%) had a pathogenic or likely pathogenic variant including five alterations classified as variants of uncertain significance (VUS).The clinical phenotype of patients carrying VUS is shown in Supplementary Table S4.Nine variants were novel.In patients who had multigene panels or whole-exome sequencing, no other CAKUT-related genes were found.

Assessment of magnesium homeostasis
Hypomagnesemia and hypermagnesuria were the most common abnormalities in mut + patients.Low sMg levels were tenfold more frequent in the mut + group when age-appropriate reference values were applied (59.00% vs. 5.75%).Applying the conservative cutoff for sMg (0.7 mmol/l) reduced the frequency of hypomagnesemia in the mut + cohort by more than 20%.Likewise, hypomagnesemia, according to age-dependent norms, had a better discriminatory value (OR 23.60 vs. 10.05,respectively).There was no change in the mut − group in this regard.
As expected, the correlation analysis in the entire cohort showed the effect of age on sMg (r = − 0.38, p = 0.001 for females; r = − 0.28, p = 0.004 for males) (Fig. 2).Thus, hypomagnesemia was also analyzed in two age subgroups (i.e., 0.5-3-and 4-17-year intervals) [14].The frequency of hypomagnesemia remained higher in mut + in the analyzed age intervals (Table 3).At the age of 0.5-3 years, hypomagnesemia was exclusively found in the mut + group.Importantly, when using the age-adjusted reference values, it was observed in 42.31% of mut + patients, while when using the cutoff of 0.7 mmol/l, hypomagnesemia was 11-fold lower (3.85%).
A similar effect of applying an appropriate limit of sMg in the older age interval (4-17 years) was observed, but the difference was not that significant (67.65% vs. 51.47%).The abovementioned differences were not noticeable in the mut − group, regardless of the cutoff applied.

Construction of a random forest-based predictive model
All subjects' data were randomly divided into a training (70%) and a test set (30%).Columns with more than 50% missing data (i.e., fractional excretion of potassium) were excluded from further analysis to ensure sufficient quality of imputation.Missing values in the data were imputed using the Multiple Imputation by Chained Equation (MICE) algorithm for each dataset separately.The imputation of missing data resulted in 24 different training datasets.To avoid overfitting, they were not combined into one large dataset but were analyzed independently, and the results were then pooled.
To choose the best algorithm for the construction of a predictive model, various machine learning algorithms were tested (Supplementary Figure S1).The results for all 24 datasets were pooled according to Rubin's rules to obtain an overall summary of algorithm performance.The algorithm with the highest accuracy was random forest (RF), which was used for further analysis and building of the predictive model.
Feature selection for the model was performed using the RFE algorithm.For every variable used for modeling, the number of occurrences in the 24 datasets and the relative  variable importance in each dataset were assessed to perform subsequent feature selection.We assumed that variables with a higher number of occurrences and higher relative importance were more likely to accurately predict the occurrence of HNF1B mutation.When plotting these two variables on a scatterplot, a clear pattern emerged, which allowed us to choose the possible predictors (Fig. 3).As the variables sMg and hypomagnesemia tended to explain the same portion of the variance in the model, two separate models were tested, and the one with better accuracy was chosen (data not shown).The final variables selected for the modeling were FEMg, hypomagnesemia (better accuracy in the model than sMg level), bilateral kidney anomalies, pancreatic malformations, and serum potassium.Supplementary Figure S2 shows the importance of each variable.We conducted hyperparameter tuning and fitting of the final model with a predefined formula.The model was fitted to each imputed dataset separately, and then the votes of all subsequent models were counted and summarized.Then, predictions on the training and test datasets were made.The final model performance is presented in Table 4.The ROC curve on the validation dataset is shown in Fig. 4. The area under the curve (AUC) of our model was 0.85, with a sensitivity of 93.67% and a specificity of 73.57%.For comparison, the performance results of the HNF1B (Faguer's) score are also shown.
Finally, based on our model, we designed a corresponding user-friendly online calculator for clinicians to use in clinical practice and to facilitate independent external validation of our model (http:// www.hnf1b proje ct.com).The result of calculation is a probability of having HNF1B mutation.

Discussion
The diagnosis of HNF1B-associated disease is a challenging task due to high phenotypic variability, incomplete penetrance, frequent absence of family history, and possible overlap with other conditions [9].Given that, an appropriate selection of those patients who likely have a high diagnostic Fig. 3 Scatterplot of relative importance and frequency of occurrences for each variable derived from the recursive feature elimination algorithm.In the top right-hand corner, a well-differentiated cluster of relevant variables is depicted.B, bilateral kidney affection; C, kidney cysts; DD, developmental delay; DM, diabetes mellitus; FC, fractional excretion of calcium value; FC1, hypocalciuria (FECa < 1%); FM, fractional excretion of Mg value; FM5, hypermagnesuria (FEMg > 5%); FU, fractional excretion of uric acid; G, geni-tal tract anomaly; hP, hyperparathyroidism; hU, hyperuricemia; hM, hypomagnesemia; H, kidney hypodysplasia; I, impaired fasting glucose; IM, hyperglycemia (IFG and/or DM); L, liver involvement; Ma, major kidney anomaly; MC, multicystic kidney disease; Mi, minor kidney anomaly; P, pancreas anomaly; PH, positive history; PTH, parathyroid hormone; sK, serum potassium; sMg, serum magnesium; SK, solitary kidney; sUA, serum uric acid; U, urinary tract malformation yield for targeted genetic testing is crucial.At this moment, there is no accurate tool that can assist clinicians in the pregenetic screening of patients.In this study, we developed a predictive model for HNF1B mutations, and a corresponding calculator, which could be a reliable and efficient tool in the selection of pediatric patients with CAKUT for genetic testing.
The discriminative power of the model is good.Importantly, our model outperformed the existing HNF1B score (AUC 0.85 vs. 0.71).However, it must be remembered that Faguer's score favors negative selection.The negative predictive value (NPV) in the development study for the French cohort was 99% [9], whereas in the validation study performed in a large UK dataset, it was 85% [10].Notably, in our study, NPV of the Faguer's score was 61% compared to 75% achieved by our model.As shown by Clissold et al. [10] and Lim et al. [11], the use of the proposed cutoff of ≥ 8 points in Faguer's score may erroneously indicate the absence of mutations in children.Additionally, in the study by Raaijmakers et al. [12], the authors retrospectively applied this score in their cohort and found that it would have missed some patients carrying causative mutations.This might be due to the dynamic  Our predictive model, compared with the screening criteria of Faguer and Raaijmakers, is more applicable to routine clinical practice.The HNF1B score is very complex and requires much data, some of which are not available on a routine basis [12].Raaijmakers' criteria, however, have not been developed in the form of scoring, which restricts their use and validation.In this respect, our model and a corresponding online calculator are user friendly.It includes only five items that are easily available.The development of a calculator should make the model readily accessible and commonly used in practice at the bedside.In contrast to Faguer's score, which is based on an arbitrarily developed scoring system and testing the optimal cutoff point, we intended to build a predictive model in which a decision is made by an algorithm.For this purpose, we applied efficient and advanced statistics for feature selection and model building.
Our results are not surprising since, like Faguer's study [9], we showed that hypomagnesemia and pancreatic anomalies are relevant parameters.The selection of bilateral kidney anomalies in our model is also in concordance with the findings of Raaijmakers et al. [12], who considered this feature to be the most predictive in terms of kidney phenotype.In the same study, hypomagnesemia had a high relative risk for HNF1B-associated disease.Similar to our observation, Raaijmakers et al. suggest that HNF1B analysis should be considered in all patients with bilateral kidney anomalies and hypomagnesemia, irrespective of age at presentation, extrarenal symptoms, or the presence of an affected family member.In fact, bilateral defects are typical for HNF1B nephropathy [7,8,11,12,[23][24][25].In line with Raaijmakers et al. [12], a family history of kidney disease and/or DM was not discriminatory in our study.
Importantly, hypomagnesemia was found to be a highly selective parameter for HNF1B mutations, which is in concordance with other studies [9,12].In this study, we provide a proof of concept developed in an earlier study [16].Here, we show that hypomagnesemia, according to ageappropriate norms, had better discriminatory value than the age-independent cutoff of 0.7 mmol/l.We also demonstrated that hypomagnesemia may be an early and common manifestation in individuals with HNF1B and, as such, could serve as a predictor at a young age.Adalat et al. [23] were the first to find an association between hypomagnesemia and HNF1B mutations.In a later study, the authors claimed that this abnormality develops with increasing age, and it is usually diagnosed at a later age [26].Here, we confirm that the prevalence of hypomagnesemia increases with age, but noteworthy, we found hypomagnesemia in a substantial proportion of HNF1B patients in younger ages.The explanation for these conflicting results is the utilization of appropriate lower limits of sMg in our study instead of one cutoff value for sMg.Furthermore, we analyzed FEMg and proposed this parameter as a predictor of HNF1B.Considering the high selective value of hypomagnesemia, one could expect that hypermagnesuria may add to the prediction.It seems clear that impairment of tubular transport of magnesium should be assessed in pairs (i.e., sMg and FEMg), as it provides more information about magnesium handling [16].For example, some patients may have hypermagnesuria with an intact sMg level, as the relationship between sMg concentration and FEMg is dynamic [27].Therefore, magnesium wasting in the presence of normomagnesemia may indicate HNF1B nephropathy.
In this study, pancreatic anomalies were reported in 11.93% of pediatric patients and were selected as a valuable parameter.In this respect, our finding is quite unexpected.In a study of a large group of children, pancreatic anomalies were detected in only 0.9% of patients, while in adults, this feature was much more common (9.8%) [10].Similarly, Raaijmakers et al. [12] reported this manifestation in only one patient, and this item did not receive any recognition.The inconsistency in reports could be partially explained by the fact that anatomical anomalies may be detected more commonly in adults, as abdominal computed tomography (CT) or magnetic resonance imaging is conducted more easily in this population.In a case series, pancreatic malformations were detected in three out of four patients, as CT was performed either to visualize kidney cysts or during an episode of liver derangement as part of differential diagnostics [28].Currently, there are no data to advocate for CT scans in patients suspected of HNF1B to detect pancreatic anomalies.Therefore, the pancreas should be formally assessed by ultrasound in every patient with CAKUT.Additionally, in children with loose stools and/or failure to thrive, it is suggested that fecal elastase-1 be included in the work-up, which could serve as a screening marker of pancreas involvement [29].
There are conflicting results for DM, which in our study and in the study by Clissold et al. [10] were discriminative based on OR.In contrast, other authors [12,25] did not find this parameter useful for selection.This discrepancy may possibly be due to the age of the studied population.Due to the higher prevalence of DM in adults, its selective power may be diminished.Importantly, maturity-onset diabetes in the young type 5 (MODY5), which is associated with HNF1B mutations, typically appears in adolescence and early adulthood; thus, this manifestation is usually absent when considering molecular testing in young children.Despite being discriminative, this parameter was not selected as an important feature and was not included in our model.On the other hand, this is not an unexpected finding, as the highest likelihood of HNF1B mutation identification in children with both MODY5 and kidney malformations is only 38% [30].It is also no surprise that hyperuricemia was found to be irrelevant for prediction.In a previous study, we demonstrated its limited value due to a strong association of sUA with kidney function and PTH [17].
Not surprisingly, hypokalemia was absent in our cohort, yet serum potassium levels were associated with HNF1B mutations.Although this biochemical abnormality may not be frequently present in HNF1B nephropathy (3.5%), hypokalemia was also recognized by Faguer et al. [9].As demonstrated by Adalat et al. [26], low potassium levels develop with age; thus, the chances of identifying it are higher at older ages.Given this, we shall not expect hypokalemia in young children.Despite this, potassium level was selected in our model, as it seems that discrete changes in the level of serum potassium may be informative.
We acknowledge some limitations of this study.For some patients, the data were not complete, and parameters were obtained/analyzed at unequal time points.As patients were recruited from different centers, the method of assessment of some laboratory parameters was nonuniform, and extrarenal abnormalities were not studied formally.Since the majority of patients had targeted HNF1B testing, we may not eliminate the possibility of the presence of other gene mutations responsible for polycystic kidney disease or CAKUT.The studied cohort had well-preserved kidney function, so our model may not be applicable to patients with more advanced CKD.We are also aware that this model has not been validated externally.The aim for future studies is the development of a model to predict HNF1B disease at a younger age, preferably for those < 2 years of age, when kidney anomalies may be the first and only manifestation of the disease, and the clinical utility would be the most appreciated.Unfortunately, the cohort size in this study was too small to generate a separate model for the young population of patients, which could be regarded as a limitation.
In conclusion, our clinical prediction tool for HNF1B mutations was developed in a large pediatric cohort using supervised machine learning methods.Based on the RF algorithm, this study established an accurate and simple predictive model that calculates an individual's probability of having HNF1B mutation.Our tool would facilitate the selection of children and young adults with CAKUT with normal or mildly compromised kidney function for genetic testing, thereby improving the detection rate and reducing the costs of testing.We also showed that abnormalities in magnesium handling are important features of HNF1B-associated disease.Hence, the utilization of ageappropriate lower limits of sMg is highly recommended to achieve good model performance.Further work is needed to validate the model in a prospective study.

Fig. 2
Fig. 2 Correlations between serum magnesium concentrations and age shown separately for males (M) and females (F)

Table S1 f
Family member with either diabetes mellitus and/or structural kidney disease

Table 2
Kidney phenotype [12]an (95% confidence interval) for continuous; n (%) for categorical b odds ratio c Fisher's exact test d According to classification suggested by Raaijmakers et al.[12]

Table 3
[15]uency of hypomagnesemia according to age intervalssMg serum Mg level a Comparison between the groups when hypomagnesemia defined by age-dependent norms[15]