Plasma Acylcarnitines and Risk of Type 2 Diabetes in a Mediterranean Population at High Cardiovascular Risk

Abstract

Context

The potential associations between acylcarnitine profiles and incidence of type 2 diabetes (T2D) and whether acylcarnitines can be used to improve diabetes prediction remain unclear.

Objective

To evaluate the associations between baseline and 1-year changes in acylcarnitines and their diabetes predictive ability beyond traditional risk factors.

Design, Setting, and Participants

We designed a case-cohort study within the PREDIMED Study including all incident cases of T2D (n = 251) and 694 randomly selected participants at baseline (follow-up, 3.8 years). Plasma acylcarnitines were measured using a targeted approach by liquid chromatography–tandem mass spectrometry. We tested the associations between baseline and 1-year changes in individual acylcarnitines and T2D risk using weighted Cox regression models. We used elastic net regressions to select acylcarnitines for T2D prediction and compute a weighted score using a cross-validation approach.

Results

An acylcarnitine profile, especially including short- and long-chain acylcarnitines, was significantly associated with a higher risk of T2D independent of traditional risk factors. The relative risks of T2D per SD increment of the predictive model scores were 4.03 (95% CI, 3.00 to 5.42; P < 0.001) for the conventional model and 4.85 (95% CI, 3.65 to 6.45; P < 0.001) for the model including acylcarnitines, with a hazard ratio of 1.33 (95% CI, 1.08 to 1.63; P < 0.001) attributed to the acylcarnitines. Including the acylcarnitines into the model did not significantly improve the area under the receiver operator characteristic curve (0.86 to 0.88, P = 0.61). A 1-year increase in C4OH-carnitine was associated with higher risk of T2D [per SD increment, 1.44 (1.03 to 2.01)].

Conclusions

An acylcarnitine profile, mainly including short- and long-chain acylcarnitines, was significantly associated with higher T2D risk in participants at high cardiovascular risk. The inclusion of acylcarnitines into the model did not significantly improve the T2D prediction C-statistics beyond traditional risk factors, including fasting glucose.

A panel of acylcarnitines, especially those with a short chain and long chain, was significantly associated with the incidence of T2D in individuals at high cardiovascular risk.

In 2017, 425 million adults worldwide suffered from type 2 diabetes (T2D), and it is estimated that this number will increase to 629 million by 2045 (1). T2D accounted for 14.5% of deaths worldwide in 2015 and it is a serious burden for health systems of many countries (1). Therefore, early identification of metabolic disturbances that precede clinical manifestations and effective preventive strategies are fundamental to reverse this serious public health problem. Indeed, lifestyle interventions have consistently demonstrated reduced T2D risk in the short and long term. Metabolomics is a rapidly evolving technology that offers a new avenue for identifying novel biomarkers prior to the onset of diabetes beyond the classical risk factors.

The existing literature suggests that plasma acylcarnitines are involved in key metabolic pathways associated with insulin resistance and T2D (2). Particularly, elevated concentrations of short- and long-chain acylcarnitines have been linked to obesity (3), insulin resistance (4), and T2D (5). Carnitine has been shown to play an important role in transporting long-chain fatty acids from the cytosol to the mitochondrial matrix, where β oxidation takes place (6). The accumulation of acylcarnitines may reflect dysregulated fatty acid oxidation, which in turn contributes to metabolic disorders (7). Recent evidence has also suggested that increased concentrations of short-chain acylcarnitines were associated with a Western dietary pattern (8), and some dietary components can modulate acylcarnitine profiles in plasma or other fluids and tissues (9–11).

To date, very few prospective studies have investigated the associations between acylcarnitine profiles and the incidence of T2D (2), and the findings have been inconsistent. Recently, Sun et al. (6) reported that a panel of acylcarnitines, especially with long chain, was significantly associated with increased risk of T2D in a cohort of 2103 Chinese participants followed for 6 years. However, the authors only measured these metabolites at baseline and specifically mentioned that these findings need to be replicated in white populations. Nevertheless, the potential associations between acylcarnitine profiles and incidence of T2D and whether acylcarnitines or their changes can be used to improve diabetes prediction beyond traditional risk factors remain unclear.

Therefore, in the present prospective study based on the PREDIMED Study, we addressed the following hypotheses: (i) Plasma acylcarnitine concentrations at baseline are associated with the incidence of T2D; (ii) 1-year changes in plasma concentrations of acylcarnitines are associated with a subsequent risk of T2D; and (iii) including acylcarnitines in a model for T2D prediction could improve the predictive ability beyond traditional risk factors.

Methods

Study design and participants

The design of the current study was a nested case-cohort study in the framework of the PREDIMED Study (12, 13). All available incident T2D cases diagnosed during follow-up and a random subsample of 20% of participants free of T2D at baseline were included (14). Briefly, the PREDIMED Study (www.predimed.es) was conducted from 2003 through 2010 in Spain and aimed to evaluate the effects of the Mediterranean diet for the primary prevention of cardiovascular disease (CVD). At baseline, 7447 participants aged 55 to 80 years with high cardiovascular risk but initially free from diagnosed CVD were included.

The study population for the present analysis consisted of 892 participants who were free of diabetes at baseline (3541 in the full PREDIMED Study) with available EDTA plasma samples, including 251 incident T2D cases and 641 randomly selected participants (subcohort). The subcohort included 53 overlapping cases of T2D between the subcohort and total cases. Of these, 663 participants out of the 892 had available blood samples after 1-year of follow-up and were included in the 1-year change analyses (15). The Institutional Review Boards at all study locations approved the protocol, and all participants provided written informed consent (13).

Study samples and metabolomics profiling

All analyses used fasting (for ≥8 hours) plasma EDTA samples collected at baseline and at year 1 of intervention. Samples were processed at each recruiting center no later than 2 hours after collection and stored at −80°C. Pairs of samples (baseline and first-year visit) from cases and subcohort participants were randomly distributed before being shipped to the Broad Institute (Boston, MA) for metabolomics assays. Liquid chromatography–tandem mass spectrometry (LC-MS) was used to quantitatively profile acylcarnitines in plasma samples.

Details of the LC-MS platform can be found elsewhere (16–18). Briefly, data were acquired using a system comprised of a Shimadzu Nexera X2 ultra–high-performance liquid chromatograph coupled to a Q Exactive hybrid quadrupole orbitrap mass spectrometer (Thermo Fisher Scientific, Waltham, MA). Metabolite extracts were prepared from plasma samples (10 µL) via protein precipitation with the addition of 9 volumes of 74.9:24.9:0.2 (v/v/v) of acetonitrile/methanol/formic acid–containing stable isotope-labeled internal standards [valine-d8 (Sigma-Aldrich, St. Louis, MO) and phenylalanine-d8 (Cambridge Isotope Laboratories, Cambridge, MA)]. The samples were centrifuged (10 minutes, 9000 × g, 4°C) and the supernatant fluid was injected directly onto a 150-mm × 2-mm, 3-µm Atlantis hydrophilic interaction liquid chromatography column (Waters). The column was eluted isocratically at a flow rate of 250 µL/min with 5% mobile phase A (10 mmol ammonium formate/L and 0.1% formic acid in water) for 0.5 minutes followed by a linear gradient to 40% mobile phase B (acetonitrile with 0.1% formic acid) over 10 minutes. MS analyses carried out using electrospray ionization in the positive-ion and full-scan spectra were acquired over 70 to 800 m/z. Other MS settings were as follows: sheath gas, 40; sweep gas, 2; spray voltage, 3.5 kV; capillary temperature, 350°C; S-lens, 40; heater temperature, 300°C; microscans, 1; automatic gain control target, 1E6; and maximum ion time, 250 ms. To enable assessment of data quality and to facilitate data standardization across the analytical queue and sample batches, pairs of pooled plasma reference samples were analyzed at intervals of 20 study samples. One sample from each pair of pooled references served as a passive quality control sample to evaluate the analytical reproducibility for measurement of each metabolite whereas the other pooled sample was used to standardize data using a “nearest neighbor” approach. Standardized values were calculated using the ratio of the value in each sample over the nearest pooled plasma reference multiplied by the median value measured across the pooled references. Raw data were processed using TraceFinder (Thermo Fisher Scientific) and Progenesis QI (Nonlinear Dynamics). The coefficients of variation measured in the 92 pooled plasma quality control samples are shown in an online repository (15). These data show that the median analytical coefficient of variation for the 28 acylcarnitines measured throughout the study was 4.3%.

Participants’ triglyceride (TG) levels, total cholesterol, low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C) were measured by locally standard measurement methods using fasting plasma samples at baseline. We calculated LDL-C levels with the Friedewald formula (when TGs were <300 mg/dL). We isolated HDL-C particles from the plasma of the participants by density gradient ultracentrifugation. Additionally, fasting plasma glucose and insulin were determined at the central laboratory. Glucose was measured using an enzymatic method to convert glucose to 6-phosphogluconate (ADVIA Chemistry Systems, Tarrytown, NY). The intra-assay and interassay coefficients of variation were 1.2 and 1.6. Insulin concentrations were measured using an immunoenzymometric assay (ADVIA Chemistry Systems) with intra-assay and interassay coefficients of variation equal to 3.7 and 4.4, respectively. Insulin resistance was calculated by using the homeostatic model assessment of insulin resistance (HOMA-IR) method ([insulin resistance = [fasting insulin (μU/mL) × fasting glucose (mmol/L)]/22.5).

Ascertainment of T2D cases

Information was collected from continuous contact with participants and primary health care physicians, annual follow-up visits, yearly ad hoc reviews of medical charts, and annual consultation of the National Death Index. T2D was a prespecified secondary outcome of the PREDIMED Study. At baseline, prevalent T2D was identified by clinical diagnosis and/or use of antidiabetic medication. The diagnosis of incident T2D during follow-up has been described elsewhere (19) and followed the American Diabetes Association criteria (20), namely two confirmations of fasting plasma glucose ≥7.0 mmol/L or 2-hour plasma glucose ≥11.1 mmol/L, after a 75-g oral glucose load. Blinded study physicians collected information on T2D diagnosis. Blinded to the intervention assignment, the Clinical End-Point Committee adjudicated T2D diagnosis according to the standard criteria.

Covariate assessment

Medical conditions, family history of disease, and risk factors were collected through a questionnaire during the first screening visit. At baseline and during annual visits, trained personnel measured participants’ body weight, height, waist circumference, and blood pressure according to the study protocol. Body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared. Physical activity was assessed using the validated Spanish version of the Minnesota Leisure-Time Physical Activity questionnaire (21). Participants were considered to have hypercholesterolemia or hypertension when they had previously been diagnosed and/or they were being treated with cholesterol-lowering or antihypertensive agents, respectively. Participants completed a 14-item questionnaire in a personal interview with a registered dietitian to assess their adherence to the Mediterranean diet (22).

Statistical analysis

We transformed acylcarnitine concentrations to the natural logarithm scale to approximate a normal distribution of metabolites as well as to stabilize the variance. We categorized all participants into quartiles of the individual acylcarnitine concentrations based on the distribution in the subcohort. Person-years of follow-up were calculated from baseline to the earliest T2D event, loss to follow-up, or the end of follow-up. Weighted proportional hazards Cox regression models were applied to estimate hazard ratios (HRs) and their 95% CIs of T2D comparing participants in each quartile to the lowest quartile. Baseline individual acylcarnitine species were analyzed as both continuous variables and using quartiles. We used the weighting scheme suggested by Barlow et al. (23) to account for the overrepresentation of cases in a case-cohort design. To quantify a linear trend, we assigned the median value of acylcarnitine concentration within each quartile and modeled this variable continuously. We also calculated HRs and 95% CIs of T2D associated with a 1-SD increment in the transformed concentrations of acylcarnitines. Model 1 was adjusted for age, sex, recruitment center, smoking, BMI, and stratified by intervention group. Model 2 was additionally adjusted for hypertension, dyslipidemia, physical activity, and baseline plasma glucose (centered on the sample mean and adding a quadratic term). We also examined the associations between 1-year changes in individual acylcarnitine species and subsequent T2D risk (incident cases of T2D occurring after the first year were used as the outcome). We adjusted P values of the multivariable-adjusted associations between a 1-SD increment in individual acylcarnitine concentration and T2D risk using the Benjamini–Hochberg procedure to account for the multiple testing (24).

Additionally, we calculated the Pearson correlations between individual acylcarnitines and used multivariable linear regression models adjusted for age, sex, smoking, and intervention group to test the associations between baseline individual acylcarnitines and fasting glucose, baseline, and 1-year changes in HOMA-IR.

To construct a prediction model for the risk of T2D, we applied elastic net regression, a regularized regression model combining the Lasso and Ridge penalties, to regress the T2D risk on seven traditional risk factors (age, sex, BMI, smoking, baseline hypertension, physical activity, and fasting glucose and its quadratic term) and 27 plasma acylcarnitines. We forced in all the traditional risk factors and let the elastic net engine select acylcarnitines most predictive for T2D risk. The lambda.min parameter was chosen from the 10-fold cross-validation. For the conventional model, we used Cox regression models with robust variance and forced in all the seven risk factors, stratifying by intervention groups and recruiting center. We computed two model scores (one for the conventional model and another for the conventional model plus acylcarnitines) as the weighted sum of the covariates and/or acylcarnitines with weights equal to the regression coefficients from the Cox regressions (model 1) or the elastic net regressions (model 2). An acylcarnitine score was also calculated as the weighted sum of the selected acylcarnitines from elastic net regressions. The leave-one-out cross-validation (LOOCV) approach was used to obtain the unbiased estimates of these model scores and avoid overfitting. In each run, the elastic net or Cox regression was applied to all-but-one samples (i.e., the training dataset), and the model obtained was then applied to the remaining one sample (i.e., the testing dataset) to calculate the model scores. The prediction procedures were carried out in 816 study participants without any missing values in risk factors and acylcarnitines. We calculated the HRs and 95% CI for T2D risk per SD increment in each of the two model scores and the acylcarnitine score using Cox regressions for the whole population and stratifying by intervention group. The likelihood ratio test was used to assess the significance of interaction between the intervention and the acylcarnitine score.

Additionally, we calculated the receiver operator characteristic curves for prediction of T2D to compare the conventional model score with the model score including plasma acylcarnitines. The prediction parameters, including area under the curve (AUC), Akaike information criteria, net reclassification index (NRI), and integrated discriminatory improvement (IDI), were calculated to evaluate the prediction performance. The NRI seeks to quantify whether a new marker provides clinically relevant improvement in prediction summarizing the correctly classified (into disease or nondisease groups) and incorrectly classified individuals comparing the new model with the old model. IDI provides information about the probability differences in the discrimination slopes of events and nonevents by a new marker. A significant positive value suggests that the new model could better classify individuals into the disease groups and nondisease groups (25).

As a sensitivity analysis, we have used robust estimates of the variance to correct for potential intracluster correlation, and we adjusted the models for propensity scores (built with 30 baseline variables to estimate the probability of assignment to each of the intervention groups) (13).

All analyses were performed using SAS software version 9.4 (SAS Institute, Cary, NC) or R 3.4.0, at a two-tailed α of 0.05.

Results

Baseline characteristics

The median follow-up of the analytical population was 3.8 years. Table 1 shows the baseline characteristics of the 892 study participants in those who developed T2D and in controls (251 cases and 641 noncases). Compared with noncases, participants who developed T2D were more likely to be men, current smokers, had higher BMI, higher prevalence of hypertension, and had higher concentrations of fasting plasma glucose and TGs (Table 1). Baseline characteristics of the participants stratified by intervention group are shown in an online repository (15). Baseline acylcarnitine distributions are also shown in an online repository (15).

Table 1.

Baseline Characteristics of the Study Population by Diabetes Case Status

	Cases	Noncases
n	251	641
Age, y	66.4 (5.7)	66.5 (5.7)
Sex, % women	138 (54.9)	408 (63.6)
BMI, kg/m2	30.8 (3.4)	29.7 (3.5)
Hypertension, %	241 (96.0)	577 (90.0)
Dyslipidaemia, %	200 (79.6)	552 (86.1)
Systolic blood pressure, mm Hg	153.8 (20.9)	148.0 (20.1)
Diastolic blood pressure, mm Hg	87.1 (11.3)	84.1 (10.6)
Fasting glucose, mg/dLa	118.8 (21.6)	97.2 (17.5)
HOMA-IR	5.1 (3.1)	3.5 (1.9)
Total cholesterol, mg/dLa	214.7 (52.0)	219.1 (52.8)
LDL-C, mg/dLa	133.8 (45.0)	137.0 (49.0)
HDL-C, mg/dLa	51.0 (15.4)	55.0 (16.02)
TGs, mg/dLa	142.6 (87.0)	112.0 (66.0)
Physical activity, metabolic equivalents, min/d	249.2 (253.5)	237.4 (235.1)
Smoking, %
Never	132 (52.6)	394 (61.4)
Former	56 (22.3)	144 (22.4)
Current	63 (25.1)	103 (16.0)
Mediterranean diet adherenceb	8.4 (2.0)	8.6 (1.9)

Data are expressed as means ± SD or percentage (n).

^aValues are median and interquartile range.

^bThe score is based on the 14-item PREDIMED screener of adherence to the Mediterranean Diet.

Baseline and 1-year changes in acylcarnitine concentrations and T2D risk

The associations between baseline plasma concentrations of each of the individual acylcarnitine species and risk of T2D are presented in Table 2. Several short-chain acylcarnitines, including C2, C3, C4OH, C5, and C6, were associated with a higher risk of T2D, whereas long-chain carnitines C18, C18:1, and C20 were associated with a lower risk of T2D. However, when the models were adjusted for baseline plasma glucose and when the P values were corrected for multiple comparison, most of these associations were no longer significant. The strongest association was observed for C5, in which case the multivariable HR for a 1-SD increment was 1.43 (95% CI, 1.15 to 1.79). As shown by the multivariable linear regression models in an online repository (15), baseline short-chain acylcarnitines tended to be positively associated with fasting glucose, baseline HOMA-IR, and 1-year changes in HOMA-IR. Baseline long-chain acylcarnitines showed inverse associations with fasting glucose, baseline HOMA-IR, and 1-year changes in HOMA-IR (15).

Table 2.

Association of Baseline Acylcarnitine Species With the Risk of T2D (n: Total, 892; Cases, 251; Noncases, 641)

Metabolite	Model 1					Model 2
	Q2	Q4	P for Trend	P for Trenda	HR per SD	Q2	Q4	P for Trend	P for Trenda	HR per SD
Free carnitine	1.55 (1.01–2.38)	0.95 (0.60–1.51)	0.53	0.57	0.94 (0.82–1.08)	1.02 (0.59–1.76)	0.56 (0.32–0.98)	0.04	0.30	0.84 (0.70–1.01)
C2-carnitine	0.96 (0.62–1.51)	1.19 (0.77–1.85)	0.20	0.26	1.16 (1.01–1.32)	0.97 (0.53–1.80)	0.91 (0.51–1.64)	0.76	0.90	1.01 (0.84–1.21)
C3-carnitine	0.89 (0.57–1.38)	1.49 (0.97–2.28)	0.07	0.48	1.20 (1.01–1.43)	1.05 (0.56–1.95)	1.78 (1.00–3.16)	0.05	0.30	1.25 (0.97–1.62)
C3DCH3-carnitine	1.95 (1.25–3.04)	1.68 (1.05–2.67)	0.15	0.80	1.10 (0.96–1.26)	1.76 (0.93–3.33)	1.52 (0.79–2.91)	0.21	0.42	1.08 (0.88–1.33)
C4-carnitine	0.74 (0.48–1.15)	1.05 (0.68–1.62)	0.71	0.01	1.09 (0.93–1.27)	0.71 (0.39–1.31)	0.76 (0.42–1.37)	0.36	0.62	0.92 (0.75–1.12)
C4OH-carnitine	1.53 (0.93–2.50)	1.92 (1.19–3.11)	<0.01	<0.01	1.25 (1.06–1.47)	1.41 (0.72–2.78)	1.64 (0.87–3.09)	0.13	0.30	1.19 (0.95–1.50)
C5-carnitine	1.35 (0.82–2.21)	2.39 (1.47–3.91)	<0.01	0.80	1.37 (1.16–1.60)	1.49 (0.75–2.96)	2.61 (1.39–4.89)	<0.01	0.08	1.43 (1.15–1.79)
C51-carnitine	1.08 (0.69–1.69)	1.14 (0.73–1.78)	0.43	0.03	1.08 (0.93–1.25)	1.45 (0.82–2.56)	1.01 (0.56–1.82)	0.99	0.99	1.05 (0.87–1.26)
C5DC-carnitine	0.75 (0.50–1.13)	0.56 (0.37–0.87)	<0.01	0.04	0.80 (0.68–0.95)	1.11 (0.65–1.89)	0.57 (0.32–1.03)	0.06	0.30	0.86 (0.70–1.07)
C6-carnitine	1.36 (0.86–2.14)	1.92 (1.23–3.01)	<0.01	0.80	1.17 (1.03–1.34)	1.39 (0.75–2.59)	1.69 (0.94–3.02)	0.08	0.30	1.10 (0.90–1.34)
C7-carnitine	0.66 (0.42–1.02)	0.81 (0.53–1.24)	0.54	0.80	0.99 (0.86–1.16)	0.29 (0.15–0.57)	0.56 (0.33–0.97)	0.04	0.30	0.98 (0.80–1.19)
C8-carnitine	1.05 (0.68–1.62)	1.02 (0.64–1.63)	0.70	0.94	1.01 (0.88–1.16)	1.25 (0.66–2.34)	1.19 (0.61–2.30)	0.61	0.82	1.06 (0.87–1.30)
C9-carnitine	1.11 (0.70–1.76)	0.99 (0.61–1.58)	0.91	0.80	0.96 (0.83–1.12)	1.27 (0.71–2.26)	1.24 (0.68–2.27)	0.48	0.70	1.09 (0.89–1.32)
C10-carnitine	1.21 (0.79–1.85)	0.81 (0.51–1.30)	0.36	0.80	0.94 (0.81–1.09)	1.25 (0.70–2.23)	0.87 (0.48–1.58)	0.65	0.82	0.98 (0.81–1.18)
C10:2-carnitine	1.33 (0.87–2.04)	1.20 (0.76–1.89)	0.70	0.80	0.99 (0.86–1.15)	2.03 (1.16–3.52)	1.64 (0.91–2.96)	0.10	0.30	1.05 (0.87–1.25)
C12-carnitine	0.92 (0.59–1.42)	0.93 (0.60–1.44)	0.64	0.80	0.93 (0.80–1.08)	1.08 (0.61–1.91)	0.94 (0.55–1.64)	0.84	0.90	0.89 (0.75–1.07)
C12:1-carnitine	0.81 (0.52–1.26)	1.02 (0.67–1.57)	0.69	0.80	0.98 (0.85–1.13)	0.97 (0.54–1.75)	0.95 (0.56–1.63)	0.85	0.90	0.94 (0.79–1.12)
C14-carnitine	0.85 (0.55–1.33)	1.13 (0.73–1.73)	0.43	0.98	1.05 (0.90–1.22)	0.73 (0.40–1.34)	0.68 (0.38–1.23)	0.20	0.42	0.94 (0.78–1.13)
C14:1-carnitine	1.14 (0.74–1.76)	1.04 (0.66–1.65)	0.98	0.80	0.98 (0.84–1.14)	1.00 (0.57–1.75)	0.94 (0.54–1.66)	0.84	0.90	0.90 (0.74–1.08)
C14:2-carnitine	0.99 (0.64–1.52)	1.06 (0.69–1.65)	0.69	0.90	1.03 (0.89–1.19)	1.19 (0.66–2.12)	1.05 (0.61–1.82)	0.86	0.90	0.98 (0.81–1.17)
C16-carnitine	1.04 (0.68–1.60)	1.02 (0.65–1.59)	0.84	0.04	1.05 (0.91–1.22)	1.14 (0.66–1.97)	0.78 (0.42–1.45)	0.44	0.68	0.97 (0.79–1.18)
C18-carnitine	0.64 (0.42–0.96)	0.57 (0.36–0.89)	<0.01	0.17	0.78 (0.66–0.93)	0.58 (0.35–0.96)	0.71 (0.39–1.29)	0.26	0.49	0.84 (0.67–1.05)
C18:1-carnitine	0.83 (0.54–1.26)	0.57 (0.35–0.90)	0.04	0.80	0.85 (0.73–0.99)	1.03 (0.59–1.80)	0.84 (0.46–1.56)	0.59	0.82	0.94 (0.77–1.14)
C18:1OH-carnitine	1.16 (0.76–1.77)	0.86 (0.55–1.34)	0.48	0.75	0.96 (0.82–1.12)	0.91 (0.52–1.60)	0.62 (0.35–1.09)	0.10	0.30	0.81 (0.66–1.00)
C18:2-carnitine	0.80 (0.53–1.22)	0.78 (0.50–1.21)	0.29	0.11	0.93 (0.80–1.08)	0.93 (0.54–1.61)	1.25 (0.73–2.15)	0.41	0.68	1.10 (0.90–1.33)
C20-carnitine	0.81 (0.53–1.24)	0.58 (0.37–0.90)	0.03	0.80	0.83 (0.71–0.96)	0.65 (0.37–1.16)	0.63 (0.35–1.13)	0.12	0.30	0.93 (0.75–1.15)
C20:4-carnitine	0.67 (0.44–1.03)	0.82 (0.53–1.27)	0.45	0.80	0.90 (0.77–1.05)	0.77 (0.44–1.35)	1.57 (0.91–2.73)	0.11	0.30	1.13 (0.91–1.40)
C26-carnitine	0.84 (0.54–1.29)	0.95 (0.62–1.44)	0.68	0.80	1.03 (0.86–1.23)	0.69 (0.38–1.23)	0.62 (0.35–1.11)	0.11	0.30	0.97 (0.77–1.22)

Natural logarithm scale was applied to raw values of acylcarnitines. Model 1: adjusted for age, sex, center, smoking, and BMI and stratified by intervention group. Model 2: additionally adjusted for hypertension, dyslipidemia, physical activity, and baseline plasma glucose (centered on the sample mean and adding a quadratic term).

^a P value for trend after false discovery rate multiple comparison correction. Q1 was used as the reference.

The associations between 1-year changes in individual acylcarnitines and the risk of T2D are presented in Table 3. Consistent with baseline results, a 1-year increment in short acylcarnitines, including C3, C4OH, and C5, was associated with higher risk of T2D in model 1. Additionally, significant positive associations with a higher risk of T2D for each SD increase in 1-year changes of C12:1, C14, and C14:1 were observed. However, only 1-year changes in C4OH-carnitine remained significant after adjustment for fasting glucose. Per each SD increment in 1-year change in the C4OH-carnitine, the risk of T2D was 44% higher (HR, 1.44; 95% CI, 1.03 to 2.01).

Table 3.

Association of 1-Year Changes in Acylcarnitine Species With the Risk of T2D (n: Total, 663; Cases, 158; Noncases, 505)

	Model 1: HR per SD	Model 2: HR per SD
Free carnitine	1.10 (0.90–1.35)	1.03 (0.81–1.32)
C2-carnitine	1.32 (1.02–1.71)	1.06 (0.77–1.46)
C3-carnitine	1.04 (0.87–1.24)	0.90 (0.73–1.11)
C3DCH3-carnitine	1.08 (0.90–1.30)	1.04 (0.76–1.42)
C4-carnitine	1.05 (0.86–1.29)	1.08 (0.84–1.39)
C4OH-carnitine	1.64 (1.26–2.12)	1.44 (1.03–2.01)
C5-carnitine	1.27 (1.04–1.54)	1.27 (0.95–1.70)
C51-carnitine	1.11 (0.89–1.37)	1.07 (0.83–1.37)
C5DC-carnitine	1.01 (0.82–1.24)	0.87 (0.65–1.16)
C6-carnitine	1.15 (0.87–1.51)	0.93 (0.63–1.37)
C7-carnitine	1.03 (0.84–1.25)	1.10 (0.81–1.48)
C8-carnitine	1.21 (1.00–1.45)	1.14 (0.84–1.53)
C9-carnitine	0.97 (0.79–1.21)	1.05 (0.77–1.44)
C10-carnitine	1.14 (0.95–1.37)	1.02 (0.78–1.34)
C10:2-carnitine	1.19 (0.97–1.47)	1.14 (0.84–1.56)
C12-carnitine	1.14 (0.93–1.39)	1.09 (0.82–1.44)
C12:1-carnitine	1.25 (1.03–1.50)	1.11 (0.85–1.45)
C14-carnitine	1.24 (1.01–1.52)	1.19 (0.94–1.52)
C14:1-carnitine	1.23 (1.00–1.51)	1.09 (0.84–1.42)
C14:2-carnitine	1.18 (0.97–1.43)	1.05 (0.81–1.37)
C16-carnitine	1.18 (0.96–1.45)	1.08 (0.86–1.35)
C18-carnitine	0.99 (0.83–1.19)	1.02 (0.79–1.31)
C18:1-carnitine	1.05 (0.84–1.31)	1.07 (0.81–1.41)
C18:1OH-carnitine	1.25 (1.00–1.56)	1.14 (0.83–1.55)
C18:2-carnitine	1.02 (0.82–1.28)	1.01 (0.77–1.32)
C20-carnitine	0.88 (0.72–1.08)	0.91 (0.70–1.18)
C20:4-carnitine	0.97 (0.79–1.19)	1.14 (0.87–1.49)
C26-carnitine	1.16 (0.99–1.36)	1.00 (0.81–1.23)

A natural logarithmic transformation was applied to raw values of baseline and 1-y changes of acylcarnitines and then the difference between 1-y and baseline levels were calculated. Model 1: adjusted for baseline acylcarnitines, age, sex, center, smoking, and BMI and stratified by intervention group. Model 2: additionally adjusted for hypertension, dyslipidemia, physical activity, and baseline plasma glucose (centered on the sample mean and adding a quadratic term).

Plasma acylcarnitine and prediction of incident T2D

We have constructed two T2D prediction models using the LOOCV method (Table 4). The first model included only conventional risk factors, and the second model included the conventional risk factors plus acylcarnitines (15). The HRs of T2D per SD increment of the predictive score were 4.03 (95% CI, 3.00 to 5.42; P < 0.001) for the conventional model and 4.85 (95% CI, 3.65 to 6.45; P < 0.001) for the model including acylcarnitines, with an HR of 1.33 (95% CI, 1.08 to 1.63; P < 0.001) attributed to the acylcarnitines. When we stratified the models by intervention group, higher risk of T2D was observed in both the Mediterranean diet and the control group for the conventional model and when adding acylcarnitines into the model, with no significant interactions with the intervention (P for interaction = 0.23) (Table 4). In the Mediterranean diet group, when acylcarnitines were added into the model the HR increased to 5.29 (95% CI, 3.69 to 7.57; P < 0.001) whereas in the control group the HR remained consistent with the conventional model (3.95; 95% CI, 2.49 to 6.27; P < 0.001).

Table 4.

Selected Models and Risk of Incident T2D

Analysis Model	Variables	All Participants With Nonmissing Variables		Mediterranean Diet Interventions		Control Group
		HR (95% CI) per SD Increase in Score	P	HR (95% CI) per SD Increase in Score	P	HR (95% CI) per SD Increase in Score	P
Model 1 (conventional model)	Age, sex, BMI, smoking, baseline hypertension, physical activity, fasting glucose	4.03 (3.00–5.42)	<0.001	4.06 (2.85–5.79)	<0.001	3.94 (2.36–6.57)	<0.001
Model 2 (conventional model plus acylcarnitines)	Age, sex, BMI, smoking, baseline hypertension, physical activity, fasting glucose, and plasma acylcarnitines	4.85 (3.65–6.45), among which 1.33 (1.08–1.63) was attributed to acylcarnitines	P < 0.001 for the whole model 2, P = 0.007 for the acylcarnitine score	5.29 (3.69–7.57), among which 1.43 (1.09–1.89) was attributed to acylcarnitines	P < 0.001 for the whole model, P = 0.01 for the acylcarnitine score	3.95 (2.49–6.27), among which 1.17 (0.88–1.56) was attributed to acylcarnitines	P < 0.001 for the whole model, P = 0.28 for the acylcarnitine score

The predictive model scores were computed as the weighted sum of all covariates with weights equal to the regression coefficients from the predictive models built by the Cox regression model using robust variance (model 1) and the elastic net regression model (model 2). In all participants, models were additionally stratified by intervention group and recruitment center. In total, 816 participants (552 in the Mediterranean diet group and 264 in the control group) with nonmissing values for all covariates and acylcarnitines were used to compute the scores. The P value for effect modification of interventions on the association between traditional risk factors and T2D risk was 0.36, and on the association between acylcarnitines and T2D risk was 0.23. Different panels of acylcarnitines were selected for each participant in the leave-one-out cross-validation approach (15).

When adding acylcarnitines into the conventional model, the C-statistics nonsignificantly increased from 0.86 (95% CI, 0.84 to 0.89) to 0.88 (95% CI, 0.85 to 0.90; P = 0.61). However, the comparison of the model including the conventional risk factors and acylcarnitines with the model including only the conventional risk factors yielded a significant improvement in Akaike information criteria (from 713.82 to 687.44, P < 0.001), a continuous NRI of 0.15 (95% CI, 0.00 to 0.31; P = 0.05), and an IDI of 0.03 (95% CI, 0.01 to 0.04; P = 0.002) (Fig. 1).

Figure 1.

Receiver operator characteristic curves for prediction of incident T2D. The black curve indicates the conventional model including age, sex, BMI, smoking, baseline hypertension, physical activity, fasting glucose, and stratification by intervention group and recruitment centers. The red curve indicates the conventional model plus acylcarnitines selected from the elastic net model. Robust standard errors to account for intracluster correlations were used. Statistics of both models were built based on the leave-one-out cross-validation approach. The table below shows the C-statistics, reclassification index (NRI), IDI, and Akaike information criteria for the conventional and the conventional model plus acylcarnitines.

Sensitivity analysis using robust variance estimates to correct for potential intracluster correlations and adjusting for propensity scores yielded consistent results with the ancillary analyses (15). The categorical NDI using a cutoff point with the highest specificity (0.9) and sensitivity (0.3) according to the AUC curve was significant (P = 0.00035) (15).

Discussion

In this prospective nested case-cohort study within the PREDIMED Study, we observed that a panel of acylcarnitines, especially those with short and long chain, was significantly associated with the incidence of T2D in individuals at high cardiovascular risk. Although the inclusion of acylcarnitines into the model did not significantly improve the T2D prediction C-statistics beyond traditional risk factors, including fasting glucose, it slightly improved the model reclassification and discriminatory abilities, suggesting that the new model could better classify individuals into the disease and nondisease groups. To the best of our knowledge, this is the first prospective study in a clinical setting to investigate the associations between plasma acylcarnitines and T2D using repeated measurements of acylcarnitines in individuals at high cardiovascular risk. The present findings suggest that plasma acylcarnitines measured by LC-MS may have the potential to serve as markers of future T2D risk in clinical practice.

Previously, studies in animal models have provided evidence relating acylcarnitines with β-oxidation dysfunction, mitochondrial stress, and insulin resistance (26). Indeed, the accumulation of acylcarnitines could reflect a failed attempt to alleviate reductive and oxidative stress caused by mitochondrial overload (27). Muscle acylcarnitines were substantially higher in Zucker diabetic rats, a model with severe insulin resistance and where incomplete fatty acid oxidation has been observed (28). Alternatively, a 2-week exercise intervention in mice fed a chronic “Western high-fat diet” lowered muscle acylcarnitine concentrations in association with an increase in tricarboxylic acid cycle activity and complete reversal of glucose intolerance (29).

To date, only a few case-control studies have evaluated the associations between acylcarnitines, insulin sensitivity, and T2D. In a case-control study of the Sea Island Genetic African American Family Registry project, acetylcarnitine C2 concentrations were 157% significantly higher in participants with T2D compared with those participants who were free of diabetes (30). In another study, medium-chain acylcarnitines C6-carnitine, C8-carnitine, and C10-carnitine were higher, whereas propionylcarnitine concentrations were lower in participants with T2D compared with nondiabetic control participants (31). In another case-control study conducted in the United States, which targeted 46 acylcarnitines, higher concentrations of plasma long-chain acylcarnitines were observed among participants with T2D (5). Moreover, a recent study, conducted in 570 participants from the Uppsala Longitudinal Study of Adult men, suggested a possible role of C10-carnitine and C12-carnitine in insulin resistant development, assessed by oral glucose tolerance test (OGTT) metabolomics and IV measurement of insulin sensitivity, together with in vitro studies in murine adipocytes (32). The authors concluded that the decline in plasma levels of medium-chain acylcarnitines during a glucose challenge was blunted in insulin resistance, and they demonstrated limited in vitro evidence for a contributing role of medium-chain acylcarnitines in impaired insulin-mediated glucose uptake (32). However, studies in humans are still limited, and controversial findings have been reported in relation to acylcarnitines and glucose metabolism.

Several prospective studies have also indicated that metabolic profiles including acylcarnitines were associated with incident T2D (6, 33–35) and also with CVD and mortality (36–38). For example, findings from the KORA study reported that the combination of higher acetylcarnitine C2 with lower lysophosphatidylcholine 18:2 and glycine were predictors of diabetes (34). Palmer et al. (39) observed that of the 45 acylcarnitines analyzed, higher concentrations of acylcarnitine C4OH and C10:3 and lower concentrations of C5:1 and C20 were associated with insulin resistance. In the present work, we also found that some baseline short-chain and long-chain acylcarnitines were associated with insulin resistance measured by HOMA-IR.

More recently, a prospective study conducted in Chinese individuals followed for 6 years indicated that a panel including several baseline plasma acylcarnitines, especially long chain, were associated with increased risk of T2D (6). The relative risks of T2D per SD increase of the predictive model score were 2.48 (95% CI, 2.20 to 2.78) for the conventional model and 9.41 (95% CI, 7.62 to 11.62) for the full model including traditional risk factors and acylcarnitines (6). These findings are in line with our study, in which we observed that a panel of acylcarnitines, especially those with a short chain and long chain, was associated with a 33% higher risk of incident T2D even after controlling for traditional risk factors. Of note, in our study, the relative risks of T2D per SD increase of the predictive model score were 4.03 (3.00 to 5.42) for the conventional model and 4.85 (3.65 to 6.45) when additionally including acylcarnitines. Because fasting plasma glucose, which is a strong predictor of T2D, was included in the conventional model for T2D risk prediction and the C-statistics was as high as 0.86, adding acylcarnitines has only provided limited improvement. The AUC of the conventional model without fasting glucose was 0.65, and when adding fasting glucose it increased to 0.86. Thus, additional inclusion of other predicting factors, even though the factors are significantly associated with T2D risk, can hardly make additional improvement on the C-statistics (ceiling effect). Despite the significant associations and mechanistic insights provided by the acylcarninitines, the C-statistic is difficult to improve when the conventional model has a very high predictive value (40). Additionally, it is possible that the sample size is small, thus limiting our statistical power to detect significant improvement in C-statistics. Future large cohort studies or pooling projects might provide additional insights.

In contrast, in the reported Chinese population (6), the HR of the conventional model is much lower than ours and the C-statistic is only 0.73; thus, including acylcarnitines could significantly improve the prediction. Despite the limited increase in C-statistics in our data, adding acylcarnitines into the prediction model slightly improved the model discrimination and reclassification abilities, suggesting that acylcarnitine profiles could be promising complementary biomarkers for T2D prediction in early prevention and other clinical settings. Because the AUC may have limitations in terms of clinical relevance and difficulty in interpretation of small magnitude changes (40), concluding the prediction effect solely depending on AUC might be inappropriate; therefore, we have additionally included the NRI and IDI to provide a more comprehensive picture. However, we acknowledge that there are limitations of using the NRI and IDI such as the subjectivity of the cutoff points (25). In the present analysis, the NRI has a 95% bootstrap CI of 0.00 to 0.31 and a P value of 0.053, suggesting that there is a borderline improvement in the new prediction model in terms of more accurate classification of individuals into the disease or nondisease groups. IDI summarizes the sensitivity and specificity of the prediction model, and a significantly positive IDI (in our study, 0.03; P = 0.02) suggested as well a slightly better prediction performance. The additional inclusion of acylcarnitines in the prediction model (adjusted for propensity scores) correctly reclassified 41 more noncases (58 correctly and 17 incorrectly reclassified) and 2 more cases (7 correctly and 5 incorrectly reclassified) compared with the conventional model, when we set the classification cutoff point to achieve an overall maximum sensitivity and specificity (15).

Findings of the current study included the observation that 1-year changes in certain acylcarnitines were associated with the subsequent risk of T2D, although these associations were attenuated when the models were adjusted for plasma glucose. As T2D is a strong risk factor for CVD (41), it is important to note that previous findings from another PREDIMED case-cohort study investigating the association between plasma acylcarnitine profiles and CVD demonstrated that baseline profiles of increased short-chain and medium-chain plasma acylcarnitines were associated with higher risk of CVD independently of established cardiovascular risk factors (42). In our previous study, we also identified elevated concentrations of short-chain acylcarnitines as potential biomarkers of future stroke risk (42). These results are consistent with the previous publication of Sun et al. (6) suggesting different roles of acylcarnitines with different chain length in the physiopathology of T2D. Short-chain acylcarnitines (C3 and C5) can be derived from branched-chain amino acids, which have been related through several potential mechanisms to obesity, insulin resistance, and T2D (3, 16, 43). Acetyl (C2) acylcarnitines derive from carbohydrate catabolism and from the ultimate product of the β-oxidation. Alternatively, the accumulation of long-chain acylcarnitines reflects impaired fatty acid oxidation that has been linked with insulin resistance and oxidative stress (27). The role of medium-chain acylcarnitines is less clear and warrants further investigation, but it is possible that medium-chain acylcarnitines are generated from several precursors and have different functions in the β-oxidation of fatty acids in the mitochondria (44).

Previous studies have suggested that pathways related to acylcarnitine metabolism can be influenced by diet. Bouchard-Mercier et al. (8) found that a metabolite profile composed of branched-chain amino acids and short-chain acylcarnitines was associated with the Western dietary pattern. Interestingly, they also found that a factor composed of medium-chain and long-chain acylcarnitines was inversely associated with the Western dietary pattern (8). Medium-chain and long-chain acylcarnitines were positively associated with butter and negatively associated with margarine and low-fat cheese in the EPIC Study (45). Finally, Schmidt et al. (11) found that meat eaters had higher short-chain acylcarnitines whereas vegans had higher long C18:2-acylcarnitines. In our study, we observed that a model composed of acylcarnitines and conventional risk factors was associated with higher risk of T2D in both the Mediterranean diet intervention and the control group, without apparent evidence of interaction by intervention group.

The present findings should be interpreted in the context of several limitations. First, the generalizability of our findings may be limited to other populations because the analyses were conducted in Mediterranean participants at high cardiovascular risk. Second, although we have adjusted for many potential confounders, residual confounding cannot be completely ruled out. Third, 1-year changes in plasma acylcarnitines might be a short period to observe the effects of changes in these metabolites and the risk of T2D; longer duration of follow-up is needed to confirm the findings. Fourth, although our results were internally validated, we recognize they were not confirmed in a separate prospective cohort. However, the internal validation approach using the LOOCV method allowed us to obtain unbiased estimates without overfitting and rule out other potential conflicting issues such as compatibility between metabolomics platforms. Further larger studies are warranted to replicate these results in other populations and to shed light on the potential underlying mechanisms. Finally, T2D was not the primary endpoint of the PREDIMED Study but was a secondary endpoint. Alternatively, several strengths also deserve mention. The current study was built on a large trial with >4 years of follow up, a well-characterized population, and accurate and blind assessment of incident T2D cases. The case-cohort design maximized the efficiency of the high-throughput metabolomics profiling.

In summary, we have identified a panel composed of acylcarnitines, especially short chain and long chain, that was significantly associated with T2D risk in a Mediterranean population at high cardiovascular risk. The inclusion of acylcarnitines into the model did not significantly improve the T2D prediction C-statistics beyond traditional risk factors, including fasting glucose. However, the identified acylcarnitines may play a role in the early stages of T2D pathogenesis and help to target clinical interventions.

Glossary

Abbreviations:

AUC

area under the curve

BMI

body mass index

CVD

cardiovascular disease

HDL-C

high-density lipoprotein cholesterol

HOMA-IR

homeostatic model assessment of insulin resistance

HR

hazard ratio

IDI

integrated discriminatory improvement

LC-MS

liquid chromatography–tandem mass spectrometry

LDL-C

low-density lipoprotein cholesterol

LOOCV

leave-one-out cross-validation

MS

mass spectrometry

NRI

net reclassification index

TG

triglycerides

T2D

type 2 diabetes