Serum ceramides in early pregnancy as predictors of gestational diabetes

Ceramides contribute to the development of type 2 diabetes but it is uncertain whether they predict gestational diabetes (GDM). In this multicentre case–control study including 1040 women with GDM and 958 non-diabetic controls, early pregnancy (mean 10.7 gestational weeks) concentrations of four ceramides—Cer(d18:1/16:0), Cer(d18:1/18:0), Cer(d18:1/24:0) and Cer(d18:1/24:1)—were determined by a validated mass-spectrometric method from biobanked serum samples. Traditional lipids including total cholesterol, LDL, HDL and triglycerides were measured. Logistic and linear regression and the LASSO logistic regression were used to analyse lipids and clinical risk factors in the prediction of GDM. The concentrations of four targeted ceramides and total cholesterol, LDL and triglycerides were higher and HDL was lower among women with subsequent GDM than among controls. After adjustments, Cer(d18:1/24:0), triglycerides and LDL were independent predictors of GDM, women in their highest quartile had 1.44-fold (95% CI 1.07–1.95), 2.17-fold (95% CI 1.57–3.00) and 1.63-fold (95% CI 1.19–2.24) odds for GDM when compared to their lowest quartiles, respectively. In the LASSO regression modelling ceramides did not appear to markedly improve the predictive performance for GDM alongside with clinical risk factors and triglycerides. However, their adverse alterations highlight the extent of metabolic disturbances involved in GDM.

Clinical data.Participants completed background questionnaires about their lifestyles and medical and family histories.Detailed data on pregnancy and delivery were collected from the hospital and maternal welfare clinic records and combined with individually linked register data obtained from the Finnish Medical Birth Register (FMBR).
Data on maternal age at delivery, parity and smoking during pregnancy were obtained from the FMBR.Selfreported maternal height and pre-pregnancy weight were obtained from the maternal welfare clinic records, and BMI (kg/m 2 ) was calculated.Gestational weight gain was calculated as the difference between the pre-pregnancy weight and the weight at the last antenatal visit (≥ 35 weeks of gestation).Based on the questionnaire data, educational attainment was categorised as basic or less, upper secondary, lower-level tertiary or upper-level tertiary.Chronic hypertension was defined as systolic blood pressure ≥ 140 mmHg or diastolic blood pressure ≥ 90 mmHg measured repeatedly or the use of antihypertensive medication before 20 weeks of gestation, while gestational hypertension was defined when hypertension appeared after 20 weeks of gestation.Pre-eclampsia was considered when hypertension appeared after 20 weeks of gestation and was accompanied with proteinuria (≥ 300 mg/day or two ≥ 1 + readings on a dipstick test).Data on previous pregnancies, including prior GDM, were obtained from the questionnaire and the FMBR.The participants' family history of type 2 diabetes was taken from the questionnaire.
Serum samples and laboratory analysis.The maternal early pregnancy serum samples were obtained via the Finnish Maternity Cohort (FMC), a nationwide biobank containing leftover serum samples from routine the early pregnancy routine infectious disease screening.Therefore, fasting before sampling was not required.The samples were stored at − 25 °C in the Biobank Borealis of Northern Finland.
The number of analysed samples was 2020 (91.3%).The background characteristics of those participants with missing samples (total n = 192: n = 124, no sample in the biobank; n = 68, sample was depleted) did not significantly differ from those of the samples included.Samples drawn after 20 weeks of pregnancy (n = 22) were excluded.The samples were drawn, on average, at 10.7 (SD 2.1) weeks of gestation.Finally, 1998 participants (1040 with GDM and 958 controls) were included in the analyses (Fig. 1).All laboratory analyses were performed blinded to the GDM status of the participants and all other phenotypic data.

Statistics.
Statistical analyses were performed using SPSS 28.0 and R software (version 4.2.1).The baseline characteristics of the study participants were described using the unpaired Student's t-test for continuous variables (expressed as means and standard deviations, SDs) and the χ 2 test for categorical variables (expressed as frequencies).The main outcome was GDM.We compared the means of traditional lipids, ceramides and the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio using both linear and logistic regression.To estimate the association of each variable with GDM, mean differences and odds ratios (ORs) with 95% confidence intervals (CI) per SD and per quartile (Q2-Q4) were calculated using the lowest quartile (Q1) as a reference.Model 1 was unadjusted.In Model 2, the results were adjusted for pre-pregnancy BMI, age, parity (dichotomous variable: primipara/ multipara) and gestational weeks at sampling.Model 3 was adjusted for Model 2 and educational attainment, a history of GDM, parental type 2 diabetes and delivery unit.Categorical variables were added as dummy-coded, with a separate dummy variable indicating missing values.The directed acyclic graph summarising the hypothetical causality between ceramides and traditional lipids and GDM, and potential confounding variables used in the regression analyses is shown in Supplemental Fig. 1.
Last, to assess the relative contributions of covariates on the risk of GDM, a least absolute shrinkage and selection operator (LASSO) logistic regression analysis was constructed (using the R package 'glmnet'); this regularization method is useful in selecting parsimonious predictive models, particularly when there is multicollinearity among covariates (as is the case with lipids in this study) 32 .Models with different sets of covariates were considered.First, clinical predictors for GDM (pre-pregnancy BMI, age, parity, a history of GDM, parental type 2 diabetes), educational attainment and delivery unit, altogether 16 covariates, were considered.Then, all four ceramides, the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio and traditional lipids (altogether nine covariates) were included.Furthermore, models including logarithmic, square, cubic and square root transformations of continuous covariates were considered.
The LASSO algorithm shrinks the regression coefficients using a regularization parameter lambda.As lambda increases, the coefficients of covariates deemed less important tend towards zero.The model corresponding to the level of regularization with optimal predictive performance was selected.To achieve this, for a range of values of lambda, tenfold cross validation was repeated 100 times the average area under the curve (AUC) value, and the average root mean squared error (RMSE) was recorded.The model corresponding to the optimal value www.nature.com/scientificreports/(largest AUC or smallest RMSE) of lambda was selected.Furthermore, tenfold cross-validation was also applied to assess the out-of-sample prediction accuracy of various models.Here, the model was repeatedly fitted using 90% of the data, and the accuracy of the model predictions to the actual observed values (according to the AUC or RMSE criteria) for the remaining 10% of the data not used in the model fitting (holdout data) was evaluated.

Power analysis.
The power of the study was sufficient to identify small differences between the study groups.With a power of 0.80, a significance level of 0.05 and an effect size of d 0.13, we were able to detect a difference of 0.13 SD in lipids between women with GDM and the controls.
Ethical aspects.The study was carried out according to the Declaration of Helsinki and approved by the Ethics Committee of Northern Ostrobothnia Hospital District (Reference Number 33/2008), the Finnish Institute for Health and Welfare and the scientific committee of the Northern Finland Biobank Borealis.All participants gave written informed consent after full explanation of the purpose and nature of all procedures used.

Results
Clinical characteristics.The women with GDM were older, less often primiparous and had higher BMI and blood pressure than the controls (Table 1).In the GDM group, 41.2% of multiparous women had a history of GDM, compared with 5.6% in the control group.Of the women with GDM, 195 (19.1%) received antidiabetic medication, including 182 (17.9%) receiving insulin and 22 (2.2%) receiving metformin.A family history of type 2 diabetes was more common in the GDM group than in the control group, 30.5% and 17.5%, respectively.
In the out-of-sample prediction of GDM, the AUC value was 0.796 for the clinical risk factors (Supplemental Fig. 2).The combination of clinical risk factors and triglycerides and/or other traditional lipids increased the AUC to 0.801.Finally, adding four ceramides and the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio with clinical risk factors, traditional lipids resulted in a similar AUC of 0.801.The corresponding out-of-sample RMSEs were 0.430, 0.427, and 0.427.
Ceramides play a lipotoxic role in the development of insulin resistance, type 2 diabetes and cardiovascular disease [17][18][19][20][21][22] .Hypercaloric diet and obesity lead to excess delivery of fatty acids, which causes dysregulation of multiple lipid metabolic pathways and accumulation of numerous lipid subtypes such as ceramides 17,20 .Further, these changes in lipid metabolism promote insulin resistance, mitochondrial dysfunction, oxidative stress and inflammation 17 .Concentrations of several circulating ceramides elevate years before the onset of type 2 diabetes 23,33 .
Although the underlying mechanisms are not fully known, the length of the acyl chain of ceramide seems to play a role in the development of insulin resistance 17,20 .In nonpregnant populations, elevated levels of ceramides containing long acyl chains, such as C16:0 and C18:0, have shown the strongest association with insulin resistance and incident type 2 diabetes 17,18,20,23,25 .Instead, very long chains containing ceramides, such as C24:0, have been suggested as neutral or protective 17,20 ; however, some studies have reported them to be associated with insulin resistance and type 2 diabetes 23,[33][34][35] .When normoglycaemic women were studied 12 weeks after GDM pregnancy, the levels of C22:0 and C24:0 ceramide species were higher among those who develop type 2 diabetes in the long term 33 .www.nature.com/scientificreports/During pregnancy, the serum concentrations of several ceramides, as well as traditional lipids, are known to increase 13,27,[36][37][38] .Maternal hyperlipidaemia is primarily aimed at securing fetal growth and development, especially in the third trimester 38 .Only a few studies have examined the associations of early pregnancy serum ceramide concentrations in subsequent GDM, with conflicting results [26][27][28]36,39,40 . Three prvious studies, a prospective lipidomic study including 492 women with GDM 26 , a prospective cohort study including 53 women with GDM 27 and a nested case-control study including 243 28 women with GDM, reported that higher levels of circulating C14:0 26 , C18:0 27,28 , and C18:1 28 ceramide species in early pregnancy were associated with subsequent GDM.
In line with previous findings 27,28 , we also found that the early pregnancy concentrations of Cer(d18:1/18:0) were higher in women who developed GDM compared with those who did not, but the difference was mostly explained by their higher BMI and age.Further, the difference was attenuated by other clinical risk factors for GDM, such as a history of GDM and a family history of type 2 diabetes.Although Cer(d18:1/18:0) was selected for the LASSO model, it did not improve predictive performance alongside clinical risk factors and/or triglycerides.In line with our findings, in a recent lipidomic study of 336 women with GDM, C18:0 ceramide was not independently associated with GDM 39 .Instead, they detected by the LASSO regression 10 lipid biomarkers in three categories of lipid classes, including one upregulated glycerolipid, five glycerophospholipids and four www.nature.com/scientificreports/downregulated sphingolipids.Furthermore, in two lipidomics studies of 107 40 and 100 36 women with GDM, some di-and triacylglycerides were independent biomarkers for GDM, but ceramides were not 36,40 .
Although the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio has been shown to be an independent predictor of type 2 diabetes 18 , in this study among pregnant women, it did not improve the predictive performance of GDM when clinical risk factors, especially pre-pregnancy BMI or triglycerides, were considered.Further, we found Cer(d18:1/24:0) to be an independent predictor for GDM; in contrast, previous studies with smaller sample sizes found either an inverse association 28,41 or no association between C24:0 and GDM 27 .Although Cer(d18:1/24:0) was positively related to GDM, it did not improve the prediction of GDM when clinical risk factors and triglycerides were considered.
Possible explanations for discordant results may be differences in sample size and settings of studies, diagnostic criteria of GDM, varying methods of determining ceramides and differences in study populations and analysing methods, including adjustments for covariates, especially where pre-pregnancy BMI plays an important role.
This study has several strengths.Four ceramides, previously validated in nonpregnant populations 19,24 , were measured during early pregnancy in a large group of pregnant women in this well-defined case-control setting.This was the first study assessing the early pregnancy levels of these ceramides and the Cer(d18:1/18:0)/ Cer(d18:1/16:0) ratio together with traditional lipids and evaluating their roles as predictors for subsequent GDM.The LASSO regression, also previously applied in several lipidomic studies assessing circulating lipids in early pregnancy with subsequent GDM 26,39,42 , was selected as an efficient method to create a parsimonious predictive model in the presence of multicollinear predictors.The GDM status of each participant was confirmed from the medical records, and several potential confounders were considered in the analyses.The study provides reference data for ceramide lipids among pregnant women in relation to GDM status.
The study also has some limitations.Firstly, serum samples were taken at non-fasting state, which may have a minor effect on triglyceride levels 43 .Secondly, the majority of the participants were of Finnish ancestry, which may limit the generalisability of the findings.Thirdly, the quantifications of other ceramides, diacyl-or triacylglycerols, or lipidomic analyses could have brought a broader perspective to this subject but they were not possible to realise within this study.Finally, the results could not be validated in an external cohort; to control this limitation, the LASSO regression and cross-validation were performed.
Future studies with independent external validation are needed to confirm the findings of this study.Further, it would be important to study whether these early pregnancy alterations in lipid metabolism are related to the long-term metabolic health and development of type 2 diabetes, and whether the ceramide profile varies depending on the stage of the diabetic cascade.
The early pregnancy levels of ceramides and traditional lipids were higher among women who developed GDM compared to those who did not.Cer(d18:1/24:0), triglycerides and LDL were found to be independent predictors of GDM.Clinical risk factors played a dominant role in predicting GDM and after combined with triglycerides, ceramides did not markedly improve the predictive performance for GDM.However, adverse alterations in lipids reflects the clustering of metabolic risk factors related to GDM.

Figure 1 .
Figure 1.Flowchart of the study population.FinnGeDi Finnish Gestational Diabetes study, GDM gestational diabetes.

Figure 2 .
Figure 2. Odds ratios (ORs) per 4th quartile (Q4) for GDM.Model 1 was unadjusted.Model 2 was adjusted with pre-pregnancy BMI, age, parity (dichotomous variable) and gestational weeks at sampling.Model 3 was adjusted for Model 2 and educational attainment, history of GDM, parental type 2 diabetes and delivery unit.Cer ceramide, GDM gestational diabetes, HDL high-density lipoprotein, LDL low-density lipoprotein.

Table 1 .
Baseline characteristics of study participants (n = 1998).GDM gestational diabetes.a p values based on the Student's t test or χ 2 test.

Table 2 .
Means (SD) and mean differences (95% CI) of ceramides and traditional lipids in early pregnancy in women with subsequent gestational diabetes (GDM) compared with non-diabetic women (n = 1998).Cer ceramide, HDL high-density lipoprotein, LDL low-density lipoprotein.Model 1: Linear regression, unadjusted.Model 2: Linear regression adjusted for pre-pregnancy BMI, maternal age, parity and gestational weeks at sampling.Model 3: Linear regression adjusted for Model 2 and education, history of GDM, parental type 2 diabetes and delivery unit.