Machine Learning Approaches Reveal Metabolic Signatures of Incident Chronic Kidney Disease in Individuals With Prediabetes and Type 2 Diabetes

Chronic kidney disease (CKD) affects approximately 9.1% of the general population worldwide (1). From 1990 to 2017, the global all-age mortality rate due to CKD increased by 41.5%, resulting in 1.2 million deaths only in 2017 (1).

Among the established risk factors for CKD, diabetes accounts for 30–50% of all CKD cases (2), and its microvascular complication, diabetic nephropathy, is the leading cause of end-stage kidney disease (3). Moreover, undiagnosed diabetes and prediabetes have been related to high prevalence of CKD in U.S., European, and Asian populations (4–7). Early screening of hyperglycemic individuals at risk of developing CKD is therefore crucial for effective prevention and management of incident CKD in the framework of an integrated personalized diabetes management (8).

Increased urinary albumin-to-creatinine ratio (UACR) and reduced estimated glomerular filtration rate (eGFR) are two clinical biomarkers of kidney-related structural damage and functional decline used to diagnose CKD (9). UACR, eGFR, age, and sex were reported to be highly predictive for progression of CKD (10). Albuminuria and eGFR were also found to be the most important variables to predict onset and progression of early CKD in individuals with type 2 diabetes (T2D). However, their predictive ability was modest with an externally validated c-statistic of 0.68 even in combination with age and sex (11). Since the traditional risk factors for CKD are insufficient for reliable prediction of CKD in individuals with T2D, there is an urgent need for more sensitive and specific biomarkers for CKD prognosis in prediabetes and T2D management.

A comprehensive individual profiling by means of metabolomics is a promising approach to discover previously unconsidered associations between metabolic signatures and clinical outcomes such as obesity, prediabetes, and T2D (12–19). Several studies have investigated the metabolite profiles of CKD, both in the general population and populations with T2D (20–22). However, to the best of our knowledge, none of them have explored the metabolites associated with future development of CKD in people with prediabetes or T2D.

In this study, we applied priority-Lasso and multivariate logistic regression (MLR) to identify metabolites associated with incident CKD in the population-based adult cohort KORA (Cooperative Health Research in the Region of Augsburg) (23,24).

Using three machine learning approaches (support vector machine [SVM], random forest [RF], and adaptive boosting [AdaBoost]), we furthermore assessed the predictive power of predictor sets constructed with metabolites and clinical phenotypes and compared their performance with the typically used clinical algorithm for CKD. We finally presented the best set of predictors for incident CKD in individuals with prediabetes or T2D.

We investigated the two follow-ups of the longitudinal cohort KORA survey 4, conducted in the area of Augsburg, Southern Germany. The first follow-up (F4) involved 3,080 individuals (aged 32–81 years) examined between 2006 and 2008. For the second follow-up (FF4), 2,269 participants were examined from 2013 to 2014 (23). Because the metabolomics data and the clinical variables of CKD (eGFR and UACR) were measured in the F4 study, we used F4 as baseline.

Individuals with hyperglycemia and normal glucose tolerance (NGT) were classified according to baseline fasting and 2-h postload glucose (2-h glucose) values with the World Health Organization diagnostic criteria (25). The hyperglycemic group comprised participants with prediabetes and newly diagnosed T2D (i.e., fasting glucose ≥110 mg/dL or 2-h-glucose glucose ≥140 mg/dL), as well as known T2D that was diagnosed by physician-validated self-reporting and/or current use of antidiabetes agents (13,23).

We examined 2,142 individuals who participated in both KORA F4 and FF4. Exclusion criteria were 1) nonfasting samples (n = 5 at F4), 2) missing eGFR and UACR (n = 16 at F4, n = 64 at FF4) or covariate values (n = 19 at F4), and 3) diagnosis for type 1 diabetes (n = 6 at F4), unclear type of diabetes (n = 21 at F4), or CKD (n = 173 at F4). The remaining data set comprised 385 hyperglycemic participants and 1,453 individuals with NGT (Fig. 1 and Table 1). The hyperglycemic participants were used to identify candidate metabolite biomarkers for incident CKD and to develop and evaluate sets of metabolite and clinical predictors. The NGT participants were used for sensitivity analyses of candidate biomarkers.

All study participants gave written informed consent. The KORA study was approved by the ethics committee of the Bavarian Medical Association, Munich, Germany.

The eGFR was calculated from serum creatinine (mg/dL) and cystatin C (mg/dL) (isotope dilution mass spectrometry–standardized and International Federation of Clinical Chemistry and Laboratory Medicine–standardized values) using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation (26). Non-CKD was defined as an eGFR ≥60 mL/min/1.73 m² and an UACR <30 mg/g at both F4 and FF4 (9). Incident cases of CKD were defined as no CKD at baseline (F4) but reduced kidney function (eGFR <60 mL/min/1.73 m²) or kidney damage (UACR ≥30 mg/g) at follow-up (FF4).

The serum samples from participants in the KORA F4 study were measured with the AbsoluteIDQ p150 kit (biocrates life sciences ag, Innsbruck, Austria) (24,27). In total, 3,061 serum samples of the F4 study were quantified for 163 metabolites in 38 randomly distributed kit plates (Supplementary Table 1). Each plate also contained three quality control (QC) samples (sex-mixed human plasma provided by the manufacturer) and one zero sample (PBS).

Identical QC procedures were used (13). Each metabolite met two criteria: 1) average value of the coefficient of variance in the three QCs <25% and 2) 50% of all measured sample concentrations equal to or above three times the median of the 38 zero samples. In total, 125 metabolites passed the criteria and were used in the subsequent analysis (Supplementary Table 1). For minimization of the plate effect, metabolite concentrations were adjusted for the plate normalization factors. For each metabolite, the plate normalization factors were calculated by division of the mean of QC sample values in each plate with the mean of all QC sample values in 38 plates. As shown in Supplementary Fig. 1, plate normalization efficiently corrected the interplate variations in metabolite concentration.

For comparability between different metabolites, their concentrations were natural-log transformed and scaled to a mean value of 0 and SD of 1.

Since feature reduction is an important aspect of predictive modeling, we defined a three-step feature selection procedure.

In order to decrease the false positive rate of the final discovery, we firstly used MLR adjusted for the two sets of covariates based on medical knowledge (11). The basic model was adjusted for age, sex, BMI, systolic blood pressure (BP), smoking status, triglyceride, total cholesterol, HDL cholesterol, and fasting glucose. The full model was additionally adjusted for the use of lipid-lowering, antihypertensive, and antidiabetes medication and for baseline eGFR and UACR (Fig. 1). Metabolites that were significantly associated with incident CKD in the full model (P < 0.05) were retained.

Secondly, we applied the machine learning method priority-Lasso to deal with multicollinearity of included variables and to retain metabolite and clinical variables with nonzero coefficients. Priority-Lasso is a Lasso-based intuitive procedure that uses prior knowledge of the study outcome by defining the blocks of different types of predictor variables (28). We defined 14 clinical variables in the full model as block 1, whereas the metabolites retained after the first-step screen were defined as block 2. The penalization parameters λ in each block were determined as values with maximum area under the receiver operating characteristic curve (AUC) estimated in a 10-fold cross validation.

Thirdly, we used logistic regression with backward stepwise selection according to the Akaike information criterion (AIC) to select for the most strongly associated variables with incident CKD and reduce model complexity (Fig. 1).

After the three-step feature selection, the selected metabolites from the 385 hyperglycemic individuals were regarded as candidate biomarkers.

We conducted four sensitivity analyses to reduce the possibility of chance findings (Fig. 1): 1) A nearest-neighbor propensity score matching in nested case-control study design was used to balance case and control subjects on conventional risk factors of CKD. MLR analysis was used to generate propensity scores using incident CKD as outcome and covariates from the full model. The caliper was defined as 0.1. After one-to-one propensity score matching, we investigated the association of candidate biomarkers with incident CKD by conditional logistic regression. 2) We investigated whether the predictive effect of candidate biomarkers for incident CKD was dependent of the hyperglycemic status. We examined the association of the candidate biomarkers with incident CKD in 1,453 normoglycemic participants by MLR. 3) We explored the interaction effects of candidate biomarkers with glucose levels for incident CKD in 1,838 individuals and performed a stratified analysis by MLR. We next examined the multiplicative interaction effects between candidate biomarkers and glucose groups by adding related multiplicative terms in the MLR models. The significance of interaction terms was tested by ANOVA LRT test. 4) We examined the association of candidate biomarkers with UACR-based (UACR ≥30 mg/g) and eGFR-based (eGFR <60 mL/min/1.73 m²) incident CKD separately in hyperglycemic participants.

We performed the three-step feature selection with 100 random repeats of 10-fold cross validation to develop the sets of metabolite and clinical predictors for incident CKD in hyperglycemia (Fig. 1). Their predictive performances were evaluated using AUC. The AUC values of developed predictors were compared with the established prediction model consisting of age, sex, eGFR, and UACR (10,11). These four clinical variables were used as reference predictors.

In each 10-fold cross validation, the data from 385 hyperglycemic individuals were randomly partitioned into 10 nonoverlapping subsets. Each of these 10 subsets was regarded in turn as testing data, whereas the remaining nine subsets were used as training data (Fig. 1). In each iteration, a set of metabolite and clinical variables for incident CKD was identified with the three-step feature selection procedure using one of the training data sets. The identified predictor set and the reference predictors were used to develop respective prediction models with SVM. In this way, two prediction models were built using one training data set. The AUC values of the respective two models were computed for the testing data only (Fig. 1). The average AUC value over 10 iterations of one 10-fold cross validation was calculated and finally presented. For assessment of the robustness of the predictive results, the predictive models were furthermore built using another two machine learning approaches (i.e., RF and AdaBoost) and the corresponding AUC values were reported.

SVM models were fitted with the R e1071 package (29). The kernel parameter was defined as radial (i.e., Gaussian radial basis function). RF models were fitted with the R randomForest package, which implements Breiman’s classic algorithm (30). The two RF parameters, nTree (i.e., the number of trees to grow for each forest) and mTry (the number of input variables randomly chosen at each split), were set to 600 and the default setting (floor of square root of the number of features), respectively. The R ada package was used to fit the AdaBoost models (31). The three AdaBoost parameters loss (i.e., loss function), type (type of boosting algorithm to perform), and iter (number of boosting iterations to perform) were set to ada (corresponding to the default boosting under exponential loss), discrete (discrete boosting), and 200, respectively.

In total, we performed 100 repeats of 10-fold cross validations including 1,000 times of three-step feature selection. The most frequently selected set of metabolites and clinical variables among these 1,000 selection rounds was subsequently defined as the best set of predictors for incident CKD in hyperglycemia.

All statistical analyses were performed in R (version 3.5.0), and two-sided P value <0.05 was considered as statistically significant.

The KORA F4/FF4 data sets are not publicly available because of data protection agreements but can be provided on request through the KORA-PASST (project application self-service tool [www.helmholtz-muenchen.de/kora-gen]).

Among 1,838 eligible, non-CKD participants of the KORA F4 study, 200 individuals developed CKD during a mean follow-up of 6.5 years (Fig. 1 and Table 1). Incident CKD was diagnosed more frequently in hyperglycemic participants (22.1%) than in individuals with NGT (7.9%) (Table 1). Compared with non-CKD individuals, the incident CKD case subjects in hyperglycemic and NGT groups were significantly older and had significantly higher baseline values of HbA_1c, fasting and 2-h glucose, and UACR, whereas their baseline eGFR values were significantly lower. They also self-reported a significantly higher intake of antihypertensive and lipid-lowering medication (Table 1).

Of 125 analyzed metabolites in 385 hyperglycemic participants, the baseline values of 13 metabolites were nominally associated (P < 0.05) with incident CKD, both in basic and full MLR models (Fig. 2A and Supplementary Table 2). Among the 13 metabolites, nine corresponded to sphingomyelins (SMs) and SM C18:1 remained significant after stringent Bonferroni correction (Fig. 2A and Supplementary Fig. 2). Of the 13 metabolites, 4 metabolites were selected by priority-Lasso and 2 (SM C18:1 and phosphatidylcholine diacyl [PC aa] C38:0) remained significant after stepwise AIC selection (Fig. 1). The relative concentrations of the two metabolites were significantly higher in 85 incident CKD case subjects in comparison with 300 non-CKD individuals (Fig. 2B). For example, a SD increase in the ln-transformed SM C18:1 concentration at baseline was associated with a 122% increased odds of CKD at follow-up (full model P = 3.315E−04) (Supplementary Table 2).

The results of the three-step feature selection thus identified two metabolites, SM C18:1 and PC aa C38:0, as candidate biomarkers of incident CKD in hyperglycemic individuals.

Propensity score matching in 385 hyperglycemic individuals resulted in 62 one-to-one matched incident CKD and non-CKD pairs. All covariates from the full model showed similar characteristics between the case and matched control subjects (Supplementary Table 3), and the two candidate biomarkers showed significant risk associations with incident CKD (Supplementary Table 4).

Both metabolites were not significantly associated with incident CKD in 1,453 normoglycemic individuals, i.e., when 115 incident CKD case subjects were compared with 1,338 non-CKD individuals who were both NGT at baseline (Table 1, Supplementary Table 5, and Fig. 2B). This result indicates that the two candidate biomarkers of incident CKD are specific for hyperglycemia.

Their specificity for hyperglycemia was further confirmed by metabolite-glucose interaction analysis. The risk estimates of SM C18:1 and PC aa C38:0 association with incident CKD were significant only in the hyperglycemic subgroup as well as in the top tertile of fasting and 2-h glucose, respectively (Supplementary Table 5). Moreover, SM C18:1 demonstrated significant multiplicative interaction effects with glycemic status and 2-h glucose (Fig. 3 and Supplementary Table 5).

The fourth sensitivity analysis aimed to address the UACR- and eGFR-based outcomes separately. Among 385 hyperglycemic participants, 32 and 65 developed incident CKD according to UACR and eGFR criteria, respectively. Both metabolites showed consistently significant risk effects for the UACR-based incident CKD in hyperglycemic participants, both in basic and in full MLR (Supplementary Table 6). Moreover, SM C18:1 was a significant predictor for eGFR-based incident CKD in the basic MLR (Supplementary Table 6).

During 100 times of 10-fold cross-validation, the median AUC values of our developed sets of predictors (i.e., metabolites and clinical variables) were stable in all three machine learning algorithms with corresponding values >0.813 (Fig. 4 and Supplementary Table 7). In comparisons with the reference predictors (age, sex, eGFR, UACR), the median AUC value of our developed sets of predictors increased by 2.5% and reached 0.825 (95% CI 0.801–0.849 [SVM algorithm]) (Supplementary Table 7), thereby outperforming the reference predictors in 97 out of 100 times of 10-fold cross validation (Supplementary Table 7). The improvement remained consistent after application of the other two machine learning approaches, RF (2.9% absolute increase in median AUC value) and AdaBoost (1.6%) (Supplementary Table 7). These results suggest that our developed sets of predictors outperform the established clinical predictors for incident CKD.

We further identified the best set of predictors for incident CKD, which consisted of two metabolites (SM C18:1 and PC aa C38:0) and five clinical variables (age, total cholesterol, fasting glucose, eGFR, and UACR). This set was the most frequently selected set: 113 times over 1,000 selection rounds (Supplementary Table 8). Moreover, these seven variables were the most important ones, and metabolites SM C18:1 and PC aa C38:0 were selected 857 and 593 times over these 1,000 rounds (Supplementary Table 9). The mean AUC value of the best set of predictors for incident CKD was 0.857, which was 4.8% higher than the corresponding AUC value of the full model containing 14 clinical variables including two known CKD biomarkers, eGFR and UACR (Supplementary Table 10).

This longitudinal study revealed significant accumulation of sphingo- and glycerophospholipids (SM C18:1 and PC aa C38:0) in individuals with prediabetes and T2D up to 6.5 years before their clinical onset of CKD. These candidate metabolite biomarkers of incident CKD were specific for hyperglycemic state, i.e., individuals with increased fasting and/or 2-h glucose levels. Highly stable performances of the sets of predictors for incident CKD developed from 125 metabolites and 14 clinical variables were furthermore independently confirmed with three machine learning algorithms. The best set of predictors consisted of the two metabolites (SM C18:1 and PC aa C38:0) and five clinical variables (age, total cholesterol, fasting glucose, eGFR, and UACR) and showed the best predictive power for early discrimination of hyperglycemic individuals at high risk of progressing to CKD.

Despite the relatively low coverage of our targeted metabolomics approach, i.e., lack of ceramides and other sphingolipids, our results support evidence on SM accumulation in glomerular diseases of genetic and nongenetic origin (32). Out of 125 analyzed metabolites comprising amino acids, acylcarnitines, hexoses, and glycerophospho- and sphingolipids (Supplementary Table 1), SMs represented the majority of metabolites associated with incident CKD in hyperglycemic participants (P < 0.05) (Fig. 2A). Increased SM levels in relation to CKD were also reported in individuals with type 1 diabetes (33) and T2D (34), except for the nontargeted lipidomic study of type 1 diabetes (35). Isomer annotation of the top significant metabolite, SM C18:1, in our study revealed that it may consist of several sphingoid backbones (d16:1, d18:0, d18:1, d18:2, and d19:1) bound to mainly saturated or monounsaturated fatty acyls with 16–18 carbons (36). A similar preference for saturated fatty acyl chains was found for PC aa C38:0 and PC aa C42:0, two diacyl PCs with positive association trends with incident CKD (Fig. 2A).

Circulatory levels of several other metabolites associated with CKD in our study (SM C16:0, SM C16:1, SM C24:1, and PC aa C38:0) have previously been shown to positively associate with coronary artery disease mortality (37). SM C16:0 and SM C16:1 were also found to be positively associated with myocardial infarction (38). Moreover, higher plasma SMs were found in patients with coronary artery disease and causally related to progression of atherosclerosis lesions in animal models (39,40). The PC aa C32:2 that showed an inverse association with incident CKD in our study was previously found to be protective for coronary artery disease mortality (37). These observations suggest that metabolic alterations associated with incident CKD may also reflect underlying cardiovascular disease, for which CKD is an independent risk factor (41).

Circulatory accumulation in SMs and saturated PCs in individuals with prediabetes and T2D may also reflect early stages of diabetic nephropathy such as mesangial matrix expansion, podocyte injury, and glomerular enlargement (42). The SM (d18:1/16:0) was reported to accumulate in the enlarged glomeruli of diabetic and obese mice and was detected in the glomeruli and vasculature of human kidney (43). The SM (d18:1/16:0) is one of the possible isomers for SM C16:0 that was positively associated with incident CKD in our study (Fig. 2A) and highly correlated with our top hit, SM C18:1 (Pearson correlation coefficient = 0.66, P < 2.2e−16) (Supplementary Fig. 2). Renal accumulation in SM (d18:1/16:0) was related to reduced enzyme activity of AMPK in the diabetic kidney glomeruli, mitochondrial dysfunction, and CKD progression (43).

The altered levels of certain SM and PC species in hyperglycemic individuals at increased risk for CKD could be caused by fluctuations in their fatty acid profile, which influences the first rate-limiting step in de novo SM synthesis, due to nutritional oversupply, dyslipidemia (44), or gut microbiome (45). The severity of CKD correlates with increased levels of saturated and monounsaturated fatty acids (46), and enzymes involved in de novo synthesis and the ceramide-SM homeostasis such as SM synthase 2 (SMS2) show fatty acyl chain specificity and may determine the regional expression of SM species in the kidney (47). Reduction of SM levels in the plasma membranes and lipoproteins improves whole-body insulin sensitivity (48), and SMS2 inhibition was suggested as a potential therapeutic target for controlling inflammatory responses and atherosclerosis (49,50). Whether SMS2 inhibition could prevent the development of CKD in hyperglycemic individuals requires further investigation.

The current predictive models for CKD mainly rely on clinical variables (10,11,51,52). Our study demonstrates that two candidate metabolite biomarkers, in combination with five clinical variables, yield the best performing set of predictors for incident CKD in hyperglycemic individuals. Furthermore, we show the power of appropriate combination of state-of-the-art machine learning and classical statistical approaches to reveal novel biomarkers and improve the performance of classical clinical predictors of CKD. The three-step feature selection, which we define in this study, was able to capture as few predictors as possible but achieve better predictive performance, which fulfills the ideal setting of clinical practice. Many epidemiological studies have used inappropriate ways to evaluate the performance of the identified variables, in which, for example, certain variables were selected from the whole data set and then the predictive performance was only evaluated on those selected variables using resampling approaches on the same data set (53). Consequently, this could have potentially strongly overestimated the predictive performance because the testing data set has been included as part of the whole data set to perform variable selection and it cannot be regarded as the testing data set anymore (53). In our study, we used cross validation in a combination with three-step feature selection and applied stringent internal validation procedures to evaluate the performance of the identified sets of predictors. In each round, the variable selection was only conducted for the training data and the performance evaluation was only performed for the testing data. In this way, we were able to attain accurate and unbiased internal AUC estimates. Given these advantages as described above, the consistent improvement of our developed sets of predictors on top of four established reference predictors in all three machine learning algorithms can be regarded as significant progress.

Our study has several additional advantages. We used a well-characterized, population-based human cohort that allows for adjustment for the influence of demographic parameters, medication, and other clinical variables. Our stringent QC of metabolite profiles and adjustment for plate effects reduced the noise among all 3,061 measured samples. We performed sensitivity analyses to confirm the candidate metabolite biomarkers and investigate their interaction with glycemia.

A limitation of our study is the missing replication (of 10 international human cohorts, none included at least 50 incident CKD cases in hyperglycemia and metabolites we measured). Discriminatory power of the candidate biomarkers and the best set of predictors cannot be generalized due to lack of external validation. Thus, we are aware that larger prospective studies are needed to validate our discoveries.

In summary, we identified two candidate metabolite biomarkers and the best set of predictors for incident CKD that are specific for individuals with prediabetes and T2D. This study demonstrates the value of metabolomics and appropriate combination of predictors in the improvement of accurate detection of hyperglycemic individuals with enhanced risk for CKD. With rising worldwide prevalence and burden of (pre)diabetes-related CKD, combining metabolite and clinical predictors is a promising approach for effective predictions of future CKD in the framework of an integrated personalized diabetes management.

Acknowledgments. The authors express appreciation to all KORA study participants for donating their blood and time. The authors thank the field staff in Augsburg conducting the KORA studies. The authors are grateful to the staff (J. Scarpa, K. Faschinger, N. Lindemann, A. Ludolph, S. Jelic, and B. Langer) from the Institute of Epidemiology and the Genome Analysis Center Metabolomics Platform, and KORA-PASST Platform at Helmholtz Zentrum München – German Research Center for Environmental Health, who helped in the sample and data logistics, and metabolomics measurements. Additionally, the authors thank Dr. Anne-Laure Boulesteix from the Institute of Medical Information Processing, Biometry and Epidemiology, Ludwig-Maximilians-University of Munich, Germany, for tips on statistical methods.

Funding. The KORA study was initiated and financed by Helmholtz Zentrum München – German Research Center for Environmental Health, which is funded by the German Federal Ministry of Education and Research (BMBF) and by the State of Bavaria. Furthermore, KORA research was supported within the Munich Center of Health Sciences (MC-Health), Ludwig-Maximilians-Universität, as part of LMUinnovativ. Part of this project was supported by European Union Seventh Framework Programme (EU FP7) grant HEALTH-2013-2.4.2-1/602936 (Project CarTarDis) and the European Institute of Innovation and Technology (EIT) Health–supported 19076 and 20679 iPDM-GO “Integrated Personalized Diabetes Management Goes Europe” innovation project. EIT Health is supported by the EIT, a body of the European Union. K.S. is supported by Biomedical Research Program funds at Weill Cornell Medicine - Qatar, a program funded by the Qatar Foundation.

Duality of Interest. M.F.S. was employed at Helmholtz Center Munich during his PhD thesis and is currently employed in the CardioRenal Medical Department of Bayer AG. No other potential conflicts of interest relevant to this article were reported.

Bayer AG was not involved in work related to data and manuscript generation.

Author Contributions. J.H. conceived the study, analyzed the data, and wrote the manuscript. C.H. researched cohort data and edited the manuscript. M.C. contributed to pathway analysis and wrote the manuscript. M.T. researched data and edited the manuscript. J. Adam edited the manuscript. S.Z. researched data. C.P. researched metabolomics data. L.W. edited the manuscript. J.N. edited the manuscript. M.F.S. researched data and edited the manuscript. S.N. researched data. G.K. researched metabolomics data. K.S. researched metabolomics data. M.L. reviewed the manuscript. F.S. edited the manuscript. C.G. researched cohort data. J.Adam. researched metabolomics data. M.H.d.A. researched data. A.P. researched cohort data. R.W.-S. designed the study, researched metabolomics data, and wrote the manuscript. R.W.-S. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the 15th Annual Conference of the Metabolomics Society The Hague, the Netherlands, 23–27 June 2019, and at the 7th DZD Diabetes Research School, Barcelona, Spain, 14–16 September 2019.