Ceramide stearic to palmitic acid ratio predicts incident diabetes

The FINRISK survey has been performed every 5 years since 1972 mainly to monitor trends in cardiovascular risk factors in the Finnish population [9]. The FINRISK 2002 study is a stratified random sample of the population aged 25–74 years from specific geographical areas of Finland. The survey included participants from North Karelia and Northern Savo in eastern Finland, the Turku and Loimaa regions in south-western Finland, the cities of Helsinki and Vantaa in the capital region, the provinces of Northern Ostrobothnia and Kainuu in north-western Finland and the province of Lapland in northern Finland. Sampling was stratified by sex, region and 10 year age groups so that each stratum contained 250 participants. In North Karelia, Lapland and the cities of Helsinki and Vantaa, the strata of 65–74-year-old men and women were also sampled, each with 250 participants. The original population sample was thus 13,500 (minus 64 who had died or moved away between sampling and the survey); the overall participation rate was 65.5%, including a questionnaire and a health examination during which blood samples were drawn. The study protocol was approved by the Coordinating Ethics Committee of the Helsinki and Uusimaa Hospital District and all participants gave written informed consent.

Serum samples from 8121 participants were available for the present study, and pregnant women (n = 46) as well as those with missing clinical variables (n = 30) were excluded from the data. This left 8045 participants, of whom 445 had prevalent type 2 diabetes and 593 developed incident type 2 diabetes during the follow-up. The participants were advised to fast for 4 h or longer and to avoid heavy meals earlier in the day. Serum was separated and stored at −70°C. Some analytes, such as conventional lipids, were measured immediately. The detailed methods have previously been described [9]. The authors had full access to the data used in this study. The participants were followed up until 31 December 2014, i.e. for 13 years. Only participants who had permanently moved abroad (<1%) before the disease event or 31 December 2014 were lost to follow-up. Prevalent diseases were defined as those that had developed prior to or at the baseline examination date, and incident events were those occurring thereafter. Individuals with incident diabetes during follow-up were identified by record linkage of the FINRISK 2002 data with the countrywide electronic health registers on the basis of the personal ID-code, unique to every permanent resident of Finland. Registers included the Causes-of-Death Register, Hospital Discharge Register and Drug Reimbursement Registers. This procedure has previously been described in detail [10].

All 45–74-year-old individuals were invited further to participate in a 2 h glucose tolerance test. In addition, HbA_1c was measured in participants of all ages in North Karelia, the capital region and the Turku/Loimaa regions. Both fasting glucose and HbA_1c measurements were available for 2119 participants; this population is referred to here as the FINRISK 2002 substudy.

Citrate plasma samples were obtained from the Western Norway Coronary Angiography Cohort (WECAC), which consists of 5209 adults who underwent coronary angiography at Haukeland University Hospital, Bergen, or Stavanger University Hospital, Stavanger, in Norway between January 2000 and April 2004 [11]. The study was carried out according to the Declaration of Helsinki and approved by the Regional Committee for Medical and Health Research Ethics and the Norwegian Data Inspectorate. The primary aim of WECAC was to study various prognostic biomarkers of cardiovascular endpoints. WECAC includes 3090 participants who were randomised to participate in the Western Norway B vitamin Intervention Trial (WENBIT; ClinicalTrials.​gov number NCT00354081) to investigate the effect of B vitamin supplementation on mortality and cardiovascular events [12]. The trial did not find an effect of treatment with folic acid/vitamin B₁₂ or B₆ on cardiovascular events or total mortality [12].

Altogether, 4634 baseline samples were analysed, and only participants examined because of suspected stable angina pectoris were considered for the current investigation (n = 3927). Those individuals with missing ceramide or clinical data were excluded (n = 69). Moreover, we excluded participants with any of the following: a self-reported diagnosis of diabetes mellitus (type 1 or type 2) at baseline (n = 442), fasting plasma glucose ≥7.0 mmol/l (n = 45) or non-fasting plasma glucose ≥11.1 mmol/l (n = 27). Using these criteria, a total of 3344 eligible individuals were included in the final analyses. Individuals with incident diabetes during the follow-up were identified from the Norwegian prescription database, hospital discharge diagnosis and WENBIT trial follow-up, as previously described in detail [13]. Because information on incident diabetes was collected through 31 December 2009, i.e. before the HbA_1c criterion was implemented in the diagnosis of diabetes [14], the diagnosis of type 2 diabetes was based only on glucose levels. We performed additional analyses after excluding participants with HbA_1c ≥ 47.5 mmol/mol (6.5%) at baseline (n = 2352). The median follow-up time for incident diabetes in the WECAC cohort was 7.6 years.

The Prevent Metabolic Syndrome (PrevMetSyn) trial is a population-based randomised study that consists of either overweight or obese individuals [15]. All procedures of the study comply with the Declaration of Helsinki (World Medical Association 2013). All study participants gave written informed consent, and the study was approved by the Ethics Committee of the Hospital District of Northern Ostrobothnia, Oulu, Finland (decision number 29/2012). The study is registered at ClinicalTrials.​gov with the identifier NCT01959763.

The sample was collected using the address and information system of the Finnish Population Register Center. Invitation letters were sent to people aged 20–60 years living in the city of Oulu. Eligibility criteria were: (1) a BMI of 27–35 kg/m²; (2) access to the Internet; (3) no health-related restrictions for losing weight (such as pregnancy); and (4) no other ongoing treatment for obesity. Volunteers were invited to a screening visit at the enrolment site at the Oulu University Hospital. After the screening, the participants were randomised and started the lifestyle intervention; they attended follow-up visits after the first and second years of the intervention.

Blood samples were drawn during the screening and study visits at the research laboratory of the Oulu University Hospital. Blood pressure, waist circumference, height and weight were also measured. All the procedures have been described in detail and approved by the Ethics Committee of the Hospital District of Northern Ostrobothnia, Oulu, Finland. Samples from 375 participants were analysed, and four were excluded due to missing ceramide or clinical data, leaving 371 individuals for the analyses.

The analysis of ceramides from serum samples has previously been described in detail [8, 16]. In brief, ceramide quantification was performed using a validated high-throughput mass spectrometry approach [16]. Serum or plasma samples (10 μl) were spiked with ²H (deuterium [D])-labelled internal standards, D7-Cer(d18:1/16:0), D7-Cer(d18:1/18:0), D7-Cer(d18:1/24:0) and D7-Cer(d18:1/24:1), and extracted in isopropanol:ethyl acetate (8:2, vol./vol.) using a Hamilton MICROLAB STAR robot (Hamilton Robotics, Kista, Sweden). The levels of Cer(d18:1/16:0), Cer(d18:1/18:0), Cer(d18:1/24:0) and Cer(d18:1/24:1) were quantified on a QTRAP6500 (AB SCIEX, Concord, ON, Canada) mass spectrometer equipped with an Eksigent 100-XL ultra-HPLC system (AB SCIEX). The individual ceramides were quantified in multiple-reaction-monitoring mode. Quantification was assessed through calibration line samples comprising of known amounts of synthetic Cer(d18:1/16:0), Cer(d18:1/18:0), Cer(d18:1/24:0) and Cer(d18:1/24:1) and corresponding ²H-labelled standards. The peak area ratios of each ceramide to its corresponding ²H-labelled form were calculated, plotted against the added ceramide concentration and fitted by linear regression analysis. Final ceramide concentrations were presented in μmol/l.

The baseline characteristics are described using medians (interquartile range) for continuous variables, and numbers (percentages) for categorical variables. Cox proportional hazard regression analyses were used to determine associations of the ceramide ratios with incident diabetes, after excluding individuals with prevalent diabetes. Full models were adjusted for the following risk factors: age, BMI, HDL-cholesterol (HDL-C), C-reactive protein (CRP), current smoking status, lipid-lowering treatment and systolic blood pressure (in FINRISK, 15 mmHg was added for participants on antihypertensive medication). In addition, WECAC was adjusted by sex, and FINRISK 2002 by waist circumference, eastern location in Finland, family history of diabetes (first-degree relatives), physical activity (binary variable for physical activity during free time) and socioeconomic status, defined as educational level (low, medium, high). Some models were additionally adjusted for glucose and HbA_1c. In WECAC, the vitamin B intervention arms, alone or in combination with the investigated ceramide ratio, did not show an interaction with the development of diabetes in Cox regression models, but the intervention arms were used as a stratifying variable. In FINRISK 2002, the models were stratified by sex. All continuous covariates, including baseline age and ceramide level, were log _e -transformed, and their effects are expressed per SD of the transformed variable. Individuals were excluded from the analyses if data relating to any of the variables was missing.

For prediction, Cox regression models were constructed based on FINRISK 2002 to predict the 10 year absolute risk of incident diabetes, using tenfold crossvalidation to prevent overoptimism. The increase in C-statistic resulting from the addition of the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio into the model was evaluated using the method of Antolini et al [17] for correlated C-statistics. Calibration of the models was estimated using the Hosmer–Lemeshow test. For validation, the regression coefficients from FINRISK 2002 were used to predict the 7 year risk in the WECAC study. In this validation, the baseline hazard was recalibrated in the WECAC study. The diabetes risk scores were derived directly from the models developed in FINRISK 2002.

In the PrevMetSyn study, the data were log₂-transformed and paired t tests were performed to calculate the significance of changes at 1 or 2 years relative to baseline. Software package R version 3.4.3 [18] was used for all statistical analyses. A value of p < 0.05 was considered statistically significant, and all tests were two-sided.

Four ceramides, Cer(d18:1/16:0), Cer(d18:1/18:0), Cer(d18:1/24:0) and Cer(d18:1/24:1), were analysed in the FINRISK 2002 population study (n = 8045; see electronic supplementary material [ESM] Table 1). In unadjusted Cox regression models, all ceramides and ceramide ratios were significantly associated with future type 2 diabetes, and several of these remained significant after adjusting for other diabetes-associated variables (ESM Table 2). However, in an analysis of the substudy with fasting glucose and HbA_1c data available (ESM Table 3), adjusting additionally for these variables the highest significant HRs were observed for Cer(d18:1/18:0) and Cer(d18:1/18:0)/Cer(d18:1/24:0) (ESM Table 2).

As distinct ceramides have specific roles in different cardiometabolic disorders, ceramide ratios may appear to be better predictors than individual ceramide concentrations [7]. To this end, we investigated the ratios of Cer(d18:1/18:0) to the other ceramides, and observed that the most significant predictor in univariate and multivariable analyses was the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio (Table 1). The HR for this ceramide ratio was more significant than for several traditional risk markers including cholesterol measurements (total cholesterol, LDL-cholesterol and HDL-C), triacylglycerols (TGs), CRP and systolic blood pressure, while glucose and HbA_1c were stronger predictors of incident type 2 diabetes, which is expected as they form the diagnostic definition of the disease (ESM Table 4).

The Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio was elevated in individuals with prevalent diabetes or with incident diabetes compared with individuals who remained free of type 2 diabetes during the follow-up (Fig. 1a). Furthermore, this ratio was elevated by lipid-lowering treatment (Fig. 1b). This ratio, followed by Cer(d18:1/18:0), was the ceramide variable that correlated most significantly with BMI, waist circumference, insulin, glucose and HbA_1c (Fig. 1c). Of note, Cer(d18:1/16:0) was not statistically significantly associated to any of these variables. There was no difference in the ratio between men and women (p = 0.55).

The predictive value of the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio was validated in the WECAC study (n = 3344; ESM Table 5). In WECAC as well, the Cer(d18:1/18:0)/Cer(d18:1/16:0) ratio showed the strongest association with incident type 2 diabetes (Table 2). The Cox regression results for other clinical measurements were similar to those for the FINRISK 2002 study, i.e. CRP was significant primarily in unadjusted models, HDL-C and TGs were the only significant measurements from the lipid panel, and glucose showed the highest univariate and multivariable HRs (ESM Table 6). In addition to excluding the participants with self-reported diabetes and high glucose levels, we also analysed the WECAC study after excluding all individuals with HbA_1c ≥ 47.5 mmol/mol (6.5%) at baseline. The results remained consistent, although the HRs decreased for all the investigated variables (Table 2, ESM Table 6).

Cox regression models were constructed to predict the 10 year risk of incident diabetes in the FINRISK 2002 study and validated in the WECAC study for 7 year risk. For validation, regression coefficients from the FINRISK 2002 study were used, but due to the different nature and follow-up time of the study, baseline hazard was recalibrated for the WECAC study. In addition to age, sex and BMI, the variables selected for these models were those that showed significant HRs (ESM Table 4) and were available for all participants in the FINRISK 2002 study. Furthermore, in FINRISK 2002 we compared the results with risk factors currently used in diabetes risk scores. In all models, the addition of Cer(d18:1/18:0)/Cer(d18:1/16:0) statistically significantly increased the C-statistic in both the FINRISK 2002 and WECAC studies (Table 3). Most of the models also showed adequate calibration in both studies. In the WECAC study with an HbA_1c cut-off, the C-statistic also showed an increase but were not statistically significant (ESM Table 7). The highest C-statistics were obtained when the ceramide ratio was added on top of age, BMI, sex and TGs (C-statistic 0.824 in FINRISK 2002 and 0.736 in WECAC) or traditional risk factors (C-statistic 0.828 in FINRISK) (Table 3). These two models had equally good predictive value in FINRISK 2002 (p = 0.40 for the difference in C-statistics).

Two diabetes scores (dScores) were constructed for Cer(d18:1/18:0)/Cer(d18:1/16:0) that could be used to predict the onset of diabetes, using also BMI and age with or without TG. For scoring, the regression models developed in the FINRISK 2002 for the 10 year risk were used, and individuals were divided into three groups based on the predicted risk. For both scores in FINRISK 2002, approximately 60% of the population had a low risk of 1.5% for developing incident diabetes in 10 years, whereas for the individuals in the high risk group the risk was on average 25%, i.e. 17-fold the low-risk group (Table 4). The scores were validated in the WECAC cohort, where roughly half of the population belonged to the low-risk group (2.4% and 1.7% developed incident diabetes, for dScore1 and dScore2, respectively), while the highest risk individuals had an average 17% risk of developing incident diabetes (Table 4).

The effect of nutritional intervention and resulting weight loss on the ceramide ratio was investigated in the PrevMetSyn dietary intervention study (n = 371; ESM Table 8). The ceramide ratio was significantly reduced in individuals with weight loss of 5% or more, and the result was more significant after 2 years (Fig. 2, ESM Fig. 1) than 1 year. Glucose showed a similar trend, whereas for insulin the changes between participants with either weight loss or weight gain were stronger. Interestingly, HbA_1c and CRP did not show trends according to change in weight. The diabetes scores were significantly reduced in individuals with a weight loss of 2% or more.