Serum Metabolites and Heart Failure Risk: MESA and RS

Abstract

Background

The role of circulating metabolites in heart failure (HF) mechanisms and their clinical utility remain unclear. We aimed to examine the associations between serum metabolites and incident HF and assess their performance in improving HF risk prediction.

Methods

A total of 23 571 serum metabolomic spectral variables were measured using untargeted proton nuclear magnetic resonance spectroscopy. Participants from MESA (Multi‐Ethnic Study of Atherosclerosis; discovery cohort) and RS (Rotterdam Study; replication cohort) with metabolomic data and without prevalent HF at baseline were included. Cause‐specific proportional hazards models were used to estimate the associations between metabolomic features and incident HF, with false discovery rate–adjusted P values for multiple comparisons. We further assessed how adjusting for cardiovascular risk factors influenced these associations. Correlations among metabolites were tested. Discriminative performance of risk prediction was evaluated using Harrell’s C‐statistic.

Results

We included 3942 MESA and 1515 RS participants, with a median of 15 years follow‐up. In MESA, 15 metabolites were significantly associated with incident HF (false discovery rate–adjusted P<0.01), with 5 replicating in RS (false discovery rate–adjusted P<0.05). However, after further adjusting for diabetes or hypertension, observed associations lost statistical significance. Methanol, glucose, proline, and acetoacetate showed significant correlations with fasting glucose (Pearson coefficients, 0.41–0.89). Incorporating the 5 replicated metabolites or all 15 metabolites identified from MESA into PREVENT‐HF (Predicting Risk of Cardiovascular Disease Events–Heart Failure) did not significantly improve C‐statistics in either cohort.

Conclusions

Fifteen proton nuclear magnetic resonance–measured metabolites were associated with incident HF, but these associations were not independent of diabetes or hypertension. These metabolites provided minimal predictive utility beyond PREVENT‐HF equations.

Nonstandard Abbreviations and Acronyms

CPMG

Carr–Purcell–Meiboom–Gill

¹H NMR

proton nuclear magnetic resonance

MESA

Multi‐Ethnic Study of Atherosclerosis

PREVENT‐HF

Predicting Risk of Cardiovascular Disease Events–Heart Failure

RS

Rotterdam Study

What Is New?

• Our research found that 15 identified metabolites measured by proton nuclear magnetic resonance were associated with incident heart failure, but these associations were not independent of diabetes or hypertension.

• More than half of the identified metabolites were moderately to highly related to fasting glucose levels; additionally, the identified metabolites did not significantly improve risk prediction beyond PREVENT‐HF equations.

What Are the Clinical Implications?

• The identified circulating metabolites may be involved in metabolic pathways preceding clinical heart failure and may serve as downstream metabolites of diabetes and hypertension, particularly given that guidelines have proposed diabetes and hypertension as preclinical heart failure conditions.

• Our findings highlight the importance of confounding control of diabetes and hypertension for research on metabolomics and heart failure; future research is needed to explore the intertwined relationships among the identified metabolites and metabolic factors, particularly diabetes.

Heart failure (HF) is a global health concern with a significant disease burden worldwide. According to data from the 2017 to 2020 National Health and Nutrition Examination Survey, an estimated 6.7 million Americans aged >20 years have HF, with a projected increase to 8.5 million Americans by 2030. ¹ Additionally, recent US research indicated a concerning increase in HF‐related death over the past decade. ² These trends underscore the need for new approaches to identify biomarkers that can improve our understanding of risk predictors that could in turn improve early detection and management of HF.

Emerging attention has been gained in omics research, which examines molecules or biological processes, including metabolomics. Metabolomics profiling has been applied to several types of cardiovascular disease (CVD) ³ , ⁴ , ⁵ , ⁶ including HF. ⁵ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ Investigating the association of metabolites linked to HF risk may help elucidate how these factors contribute to HF development and reveal potential pathways involved in its pathogenesis. To date, a few notable metabolites such as branched‐chain amino acid, acylcarnitine, and kynurenine, have emerged as potential HF biomarkers. ⁴ , ¹¹ However, the role of circulating metabolites in HF’s mechanistic pathways and their clinical utility remain poorly understood. ⁷ , ¹¹ , ¹² Besides the limited number of studies, the heterogeneity of methodologies (metabolic profiling measures, model adjustments, etc) and study populations (age, race, health conditions, etc), and the sensitivity characteristics of metabolites, require further research to enhance the current evidence. Moreover, the American College of Cardiology/American Heart Association recently introduced PREVENT‐HF (Predicting Risk of Cardiovascular Disease Events–Heart Failure) equations tailored for CVD risk prediction including HF. ¹³ The PREVENT‐HF equations, which incorporate data on traditional CVD risk factors, have been derived and validated from a vast sample of >6 million individuals, marking an advanced iteration of the pooled cohort equations. ¹⁴ Whether these metabolites can provide added risk prediction value beyond the PREVENT‐HF equations remains unclear.

Using data from 2 community‐based cohorts of middle‐aged to older adults with long follow‐up, and a large set of untargeted proton nuclear magnetic resonance (¹H NMR) metabolomic profiling measured, we aimed to identify novel metabolites related to incident HF and to evaluate their added risk predictive value beyond the PREVENT‐HF equation.

Methods

The data that support the findings of this study are available from the respective Data Coordinating Centers of each study cohort after approval of a manuscript proposal.

Study Population

We included participants with available data on metabolomic profiling and without prevalent HF at the time of metabolomic profiling from 2 population‐based cohort studies, MESA (Multi‐Ethnic Study of Atherosclerosis) and RS (Rotterdam Study). MESA and RS were 2 cohorts participating in the COMBI‐BIO ¹⁵ project, ⁶ which was originally aimed at identifying metabolic biomarkers of subclinical atherosclerosis using novel ways of analyzing human blood serum. Detailed descriptions of the 2 cohorts have been published. ¹⁶ , ¹⁷ Briefly, MESA recruited adults aged 45 to 84 years without known CVD from 6 centers across the United States during 2000 to 2002. The RS cohort is an ongoing, prospective, population‐based cohort in the Netherlands of adults aged ≥45 years from the Ommoord district in the city of Rotterdam, starting in 1990. MESA was approved by institutional review boards at each study site, and the RS was approved by the Medical Ethics Committee of the Erasmus MC (registration number: MEC 02.1015) and by the Dutch Ministry of Health, Welfare, and Sport (Population Screening Act WBO, license number 1071272‐159521‐PG). All participants provided written informed consent.

Serum samples that were collected in 2000 to 2002 in MESA (n=3955) and in 1997 to 1999 in RS (n=1824), respectively, were randomly selected for metabolomic profiling. In MESA, we excluded 13 participants with prevalent HF at baseline. In RS, we excluded those with established cardiovascular disease (coronary heart disease or stroke, n=237) in line with the MESA cohort design, those with prevalent HF at baseline (n=71), and 1 participant without an entry baseline date. Finally, 3942 participants were included in MESA, and 1515 participants were included in RS.

¹H NMR Metabolomic Profiling

Serum samples were collected in a fasting state at baseline, with MESA samples stored at −80 °C and RS samples at −20 °C. Study samples were shipped on dry ice and stored at −80°C upon arrival until metabolomic profiling. ¹⁵ Metabolomic profiling in MESA and RS was conducted following a previously published protocol. ¹⁸ Serum samples were processed in 2 measurement batches over ≈1 year. ¹⁵ Briefly, a standard ¹H NMR 1‐dimensional spectrum with water suppression and gradients (noesygppr1d sequence) and a T₂‐edited spectrum using the Carr–Purcell–Meiboom–Gill (CPMG) sequence with water suppression were obtained for each sample. The free‐induction decays obtained were zero‐filled to 128 000 points and a line broadening of 0.3 Hz was applied before Fourier transformation. The spectra were then phased and baseline corrected. Spectral data were then imported into MATLAB version 8.3 (R2014a; Mathworks Inc., Natick, MA) for further processing. The resulting data sets contained 23 571 1‐dimensional NMR and 23 571 CPMG spectral features. We did not include protein side chains and fatty acyl chains as identified metabolites. Notably, 3942 participants had CPMG data, and 3933 participants had 1‐dimensional NMR data in MESA; 1515 participants had CPMG data, and 1510 participants had 1‐dimensional NMR data in RS.

Assessment of Outcome

HF ascertainments in MESA and RS have been previously described. ¹⁹ , ²⁰ Medical records were reviewed, and diagnoses of HF events while hospitalized were adjudicated by a panel of MESA physicians using standardized criteria. In RS, incident HF was defined as a combination of the presence of typical symptoms or signs of HF based on the criteria of the European Society of Cardiology and was confirmed by medical specialists. Participants were followed until the occurrence of an HF event, death, loss to follow‐up, or end of follow‐up time (December 2018 in MESA, December 2016 in RS), whichever came first.

Assessment of Covariates

Through clinical examination, questionnaires, interviews, and linkage to health records, data were collected in the MESA and RS cohorts, which included age, sex, race and ethnicity, body mass index, systolic blood pressure, high‐density lipoprotein cholesterol, total cholesterol, diabetes, lipid‐lowering medication, blood pressure–lowering medication, physical activity (as defined per cohort), smoking status (never, former, current), and alcohol intake status (current, noncurrent). Physical activity was determined using a 28‐item Typical Week Physical Activity Survey in MESA, and physical activity levels were assessed using an adapted version of the Zutphen Physical Activity Questionnaire in RS. Hypertension was defined by systolic blood pressure ≥140 mm Hg or diastolic blood pressure ≥90 mm Hg or taking antihypertensive medications. The estimated glomerular filtration rate was calculated by the Modification of Diet in Renal Disease study equation ²¹ in MESA, and the Chronic Kidney Disease Epidemiology Collaboration equation of 2012 ²² in RS, by each cohort’s defined protocol. Diabetes was defined as fasting blood glucose ≥7 mmol/L or use of glucose‐lowering medication in both MESA and RS cohorts. In RS, nonfasting blood glucose ≥11.1 mmol/L was used when fasting samples were unavailable. Height and weight were measured, and body mass index was calculated as weight (kg)/height (m) ² .

Statistical Analysis

Baseline characteristics of the study population are presented as mean±SD, number (percentage), and median (interquartile range). Missingness of characteristics was up to 0.2% in MESA and 3.4% in RS. Missing values were singly imputed using the predictive mean matching approach for continuous variables and logistic regression for categorical variables. ²³

All spectral features were standardized to have a mean of 0 and an SD of 1 before analysis. Analyses were performed separately in each cohort following the same analytic protocol, and 1‐dimensional NMR and CPMG spectral features were analyzed separately. Hazard ratios and 95% CIs were estimated with cause‐specific hazard models, treating non‐HF death as a competing risk, to examine the association between per 1‐SD change in each metabolic feature and incident HF. Proportional hazards assumptions were tested using Schoenfeld residuals, and no violations were observed (P>0.05). We performed 5 models to assess the associations of interest. Model 1 adjusted for core characteristic variables including age, sex, race and ethnicity (only in MESA, as 98.8% of the RS cohort were White participants), batch information, estimated glomerular filtration rate, and body mass index. Model 2 built upon model 1 by additionally accounting for lifestyle behaviors including smoking status, alcohol intake, and physical activity. Next, additional models were designed to explore how adjustments for lipids, blood pressure, and diabetes impacted the observed associations. Model 3 built upon model 1 by adding total cholesterol, high‐density lipoprotein, and lipid‐lowering medication use. Model 4 extended model 1 by adding systolic blood pressure and hypertension‐lowering medication use, and model 5 added diabetes to model 1. To account for multiple comparisons, the Benjamini–Hochberg procedure was applied to control the false discovery rate, and P values were adjusted accordingly. A false discovery rate–adjusted P<0.01 for MESA or P<0.05 for RS was considered statistically significant.

Next, we examined the potential predictive value of the identified metabolites for 10‐year HF risk prediction. All participants were censored at follow‐up at 10 years for risk prediction analysis. The base 10‐year PREVENT‐HF equations were selected as the basic model. By adding identified metabolic features into the basic model, changes in discriminative performance (Harrell’s C‐statistic, Δ C‐statistic) were assessed. The 95% CI for Δ C‐statistic was calculated on the basis of the 1000 bootstrap samples, and a 2‐sided P value <0.05 was considered statistically significant for this comparison. Statistical analyses were performed in R software version 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Study Population

Among the 3942 participants from MESA and 1515 participants from RS, the mean ages were 62.9±10.3 years and 70.4±5.6 years, respectively, with little more than half of each sample being women (Table 1). The MESA population included several distinct racial and ethnic groups, with 38.6% White participants, while the RS cohort was predominantly White participants (98.8%). Table S1 presents baseline characteristics of participants who developed incident HF and those who did not. The median follow‐up time was 15.8 (interquartile range, 13.7–16.5) years in MESA and 14.6 (interquartile range, 9.2–16.4) years in RS. During follow‐up, 242 (6.1%) participants in MESA and 248 (16.4%) participants in RS developed HF. The HF incident rate per 1000‐person years was 4.3 (95% CI, 3.8–4.9) in MESA and 13.0 (95% CI, 11.5–14.8) in RS.

Table 1.

Characteristics of the Study Population at Baseline^*

	MESA (N=3942)	RS (N=1515)
Age, y, mean±SD	62.9±10.3	70.4±5.6
Sex, n (%)
Male	1945 (49.3)	644 (42.5)
Female	1997 (50.7)	871 (57.5)
Race and ethnicity, n (%)
White	1523 (38.6)	1378 (98.8)
Black	965 (24.5)	…
Hispanic	923 (23.4)	…
Chinese	531 (13.5)	…
East Asian	…	17 (1.2)
Missing	0	120
Measurement batches, n (%)
First	1967 (49.9)	883 (58.3)
Second	1975 (50.1)	632 (41.7)
Body mass index, kg/m2, mean±SD	28.2±5.4	27.0±4.0
Diabetes, n (%)	516 (13.1)	185 (12.2)
Hypertension, n (%)	1781 (45.2)	1009 (66.7)
Systolic blood pressure, mean±SD	127.1±21.4	143.35±21.16
Diastolic blood pressure, mean±SD	71.9±10.2	76.0±11.0
Blood pressure‐lowering medication use, n (%)	1494 (37.9)	475 (32.4)
High‐density lipoprotein cholesterol, mg/dL, mean±SD)	50.7±14.7	54.7±15.2
Total cholesterol, mg/dL, mean±SD	194.5±36.1	228.3±36.1
Lipid‐lowering medication use, n (%)	665 (16.9)	156 (10.5)
Estimated glomerular filtration rate, mL/min per 1.73 m2, mean±SD	73.8±16.1	73.5±13.1
Smoking status, n (%)
Never	1985 (50.5)	461 (30.6)
Former	1458 (37.1)	782 (51.9)
Current	484 (12.3)	265 (17.6)
Current alcohol drinker, n (%)†	2134 (54.6)	1279 (84.8)
Total moderate and vigorous physical activity, met‐h/wk, median (IQR)	64.4 (31.5–120.2)	83.7 (60.8–121.1)
Heart failure incident rate, per 1000‐person years (95% CI)	4.3 (3.8–4.9)	13.0 (11.5–14.8)

IQR indicates interquartile range; MESA, Multi‐Ethnic Study of Atherosclerosis; and RS, Rotterdam Study.

^*The baseline period was 2000–2002 for MESA and 1997–1999 for RS.

^† In MESA, current drinking was defined as answering yes to the question, “Do you presently drink alcoholic beverages?”; in RS, it was defined as answering yes to the question, “Do you drink beer/white wine/red wine?”

Metabolites Associated With Incident HF

In MESA, 17 metabolites were significantly associated with incident HF in model 1, and 15 metabolites remained significantly associated after further adjustment of lifestyle factors (model 2). Of the 15 metabolites, 10 were positively associated with incident HF (proline betaine, proline, leucine, 1,5‐anhydrosorbitol, myo‐inositol, glucose, acetaminophen glucuronide, glycerol, methanol, and acetoacetate), and 5 were negatively associated with risk of HF (alanine, glutamate, glutamine, histidine, and tyrosine; Figure 1). After further adjustment for lipid measurements beyond model 1 (model 3), a total of 14 metabolites remained significantly associated, with only acetoacetate losing its significant association with HF risk. After further adjustment for blood pressure–related variables (model 4), only 1,5‐anhydrosorbitol and glucose remained significantly associated with incident HF risk. None of the metabolites were significantly associated after adjusting for diabetes (model 5). In the RS cohort, among the 15 significant metabolites derived from model 2 in MESA, only 5 metabolites (1,5‐anhydrosorbitol, methanol, glycerol, glucose, and glutamine) were successfully replicated in RS but only in model 1. No significant metabolites were found to be associated with HF risk in model 2 to model 5 in the RS cohort. Detailed results on model 1 to model 5 in 2 cohorts are reported in Table S2.

Figure 1. Association between metabolites and incident heart failure in the MESA cohort.

HRs and 95% CIs per 1‐SD increment of metabolites were obtained from cause‐specific hazard models, adjusting for age, sex, race and ethnicity, batch information, estimated glomerular filtration rate, body mass index, smoking, alcohol intake, and physical activity. Analyses were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. Colors indicate different types of metabolites. *Metabolites successfully replicated in the RS cohort but only in Model 1. HR indicates hazard ratio; MESA, Multi‐Ethnic Study of Atherosclerosis; and RS, Rotterdam Study.

Correlations Among Metabolites, Hypertension, and Diabetes

Figure 2 shows the heatmap of Pearson correlations among metabolites, hypertension, and diabetes in the MESA cohort. Methanol, glucose, proline, and acetoacetate showed significant correlations with fasting glucose levels, with Pearson coefficients ranging from 0.41 to 0.89. Additionally, methanol, 1,5‐anhydrosorbitol, glutamate, and glucose showed moderate to high correlations with other metabolites.

Figure 2. Heatmap of Pearson correlations among metabolites, hypertension, and diabetes in the MESA cohort.

MESA indicates Multi‐Ethnic Study of Atherosclerosis.

Metabolites for HF Risk Prediction

The C‐statistic of the base model using PREVENT‐HF equations for 10‐year HF risk prediction was 0.778 (95% CI, 0.743–0.813) in MESA and 0.711 (95% CI, 0.681–0.740) in RS (Table 2). Adding the 5 replicated metabolites to the model resulted in a decrease in C‐statistics in both MESA (change in C‐statistic, –0.025 [95% CI, −0.049 to −0.005]) and RS (−0.006 [95% CI, −0.023 to 0.008]), and adding 15 metabolites identified from MESA showed no statistically significant improvements in either cohort (MESA, 0.010 [95% CI, −0.043 to 0.013]; RS, 0.006 [95% CI, −0.022 to 0.023]).

Table 2.

Model Discrimination for 10‐Year Heart Failure Risk Prediction

	MESA		RS
	C‐statistics (95% CI)	Δ C‐statistics (95% CI)	C‐statistics (95% CI)	Δ C‐statistics (95% CI)
Base model (PREVENT‐HF)	0.778 (0.743 to 0.813)	…	0.711 (0.681 to 0.74)	…
Base model +1,5‐anhydrosorbitol, glutamine, glucose, glycerol, methanol	0.753 (0.711 to 0.793)	‐0.025 (−0.049 to −0.005)	0.705 (0.672 to 0.734)	−0.006 (−0.023 to 0.008)
Base model+all 15 metabolites	0.768 (0.726 to 0.8)	0.010 (−0.043 to 0.013)	0.717 (0.681 to 0.741)	0.006 (−0.022 to 0.023)

MESA indicates Multi‐Ethnic Study of Atherosclerosis; PREVENT‐HF, Phase III Study Investigating Heart Failure and Cardiovascular Death With Baxdrostat in Combination With Dapagliflozin; and RS, Rotterdam Study.

Discussion

In this investigation of >5000 participants from 2 well‐characterized prospective population‐based cohorts with approximately 15 years of follow‐up, 15 metabolites measured by ¹H NMR were found to be associated with incident HF, independent of demographic factors, estimated glomerular filtration rate, body mass index, and lifestyle behaviors in the MESA cohort, and 5 of these metabolites were validated in the RS cohort. However, the associations observed were not independent of hypertension and diabetes. More than half of the identified metabolites were moderately to highly related to diabetes, and several metabolites also showed moderate to strong correlations with one another. Incorporating these metabolites into the PREVENT‐HF equations did not improve HF risk prediction in either cohort.

Metabolomic profiling is most often performed using NMR or mass spectrometry. Both techniques enable high‐throughput profiling of large numbers of metabolites. To our knowledge, our study is the largest metabolomics study of incident HF among the general population using untargeted ¹H NMR spectroscopy. This measurement carries advantages including robustness, reproducibility, minimal sample preparation, low cost, and nondestructiveness, but its sensitivity was generally considered to be limited compared with the mass spectrometry method. Most previous research reported mass spectrometry–measured results, ³ , ⁴ , ⁹ which may cause some inconsistencies when compared with our results from NMR measurement, but it also brings insights into how results differed derived from different metabolomic profiling methods.

Metabolic disorders, including obesity, hypertension, and diabetes, are also closely linked to HF development and progression. Notably, these conditions are considered to represent a preclinical HF stage in recent guidelines. ²⁴ , ²⁵ Previous research has demonstrated close relationships between metabolites and diabetes ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ and may be further linked to HF development; for example, aberrant branched‐chain amino acid metabolism in HF development were highly related to insulin resistance and/or type 2 diabetes, with diabetes possibly serving as the dominant contributing factor. ³ Our research also observed high correlation between 15 identified metabolites and diabetes and hypertension, which may partly explain the attenuated associations with HF risk, and their limited added risk predictive value due to overfitting.

In line with previous research, ³ , ⁵ , ⁷ , ¹⁰ , ³¹ among the 15 identified metabolites, 8 metabolites have previously been associated with HF, and interestingly, all of these also showed associations with diabetes in previous research. Regarding amino acids, our study confirmed that leucine and proline were positively related to HF risk, histidine and glutamine were inversely related to HF risk. Leucine, 1 of the branched‐chain amino acids, has been previously linked to various metabolic conditions. ³ Glutamine and glutamate, 2 amino acids central to both nitrogen and carbon cycling and linked with branched‐chain amino acid metabolism, have also been associated with the development of diabetes. ²⁶ Amino acid proline indicates a healthy diet, ³² and its concentrations may reflect metabolic differences underlying diabetes. ³³ Three metabolites linked to carbohydrate metabolism (1,5‐anhydrosorbitol, glucose, and myo‐inositol) were all shown to be positively related to HF risk. 1,5‐Anhydrosorbitol is considered a marker for short‐term glycemic control, reflecting blood glucose levels over few days to 2 weeks, and its clinical use is limited and inferior to hemoglobin A_1c. ²⁷ Previous research also suggested that 1,5‐anhydrosorbitol was closely associated with coronary artery calcium, ⁶ epicardial fat, ³⁴ and CVD risk. ⁶ Recent research also found elevated myo‐inositol levels in patients with HF, and strongly correlated to kidney failure ³¹ ; particularly in patients with preserved ejection fraction, high myo‐inositol levels predicted poor clinical outcomes and were linked to adverse cardiac remodeling, suggesting a role of myo‐inositol and its transporter in its pathophysiology. ³¹ Acetoacetate, a ketone body synthesized from fatty acids as an energy source when glucose is low, may be linked to an increased risk of type 2 diabetes. In a study of 9358 Finnish men, higher acetoacetate levels were positively associated with type 2 diabetes, ²⁹ while in another study of combined 4 Finnish cohorts including 11 896 young participants, the association was positive but not statistically significant. ²⁸ Higher levels of circulating acetoacetate were also found to be associated with higher New York Heart Association classification, higher B‐type natriuretic peptide levels, and worse clinical outcomes in patients with HF. ³⁵

Four amino acids (alanine, tyrosine, glutamate, and proline betaine) have limited evidence linking them to incident HF in previous research. Our research revealed inconsistent relation directions compared with findings from previous CVD research. Alanine, a nonessential amino acid available in protein‐based diets and supplements, was identified as a protective factor for HF risk in our current study. However, previous evidence suggested that alanine was associated with an increased risk of coronary artery disease, ³⁶ major CVD, ³⁷ and may have causal effects on diabetes, lipid profiles, and high blood pressure. ³⁸ Tyrosine has shown inconsistent associations with CVD‐related risks, warranting further research. In our study, higher serum tyrosine levels were associated with reduced HF risk. Xu et al found that higher dietary intake of tyrosine could lower the risk of CVD death. ³⁹ However, Magnusson et al suggested serum tyrosine was not significantly associated with incident CVD but remained a strong predictor of diabetes development. ⁴⁰ In a Finnish male cohort, higher serum tyrosine levels were associated with increased risk of diabetes, coronary artery disease, ischemic stroke, and CVD events. ³⁶ , ⁴¹ Additionally, 1 previous study also reported that HF patients had higher levels of tyrosine compared with controls. ⁴² Although HF development has specific metabolites that differ from other cardiovascular events, there is at least partly shared metabolic disturbance between HF and other cardiovascular events. ⁴³ Whether the metabolic disturbances related to these 4 amino acids are unique to HF requires further investigation.

The remaining 3 metabolites (acetaminophen glucuronide, glycerol, and methanol) that were less reported in previous HF‐related research were observed in our cohorts as potential HF biomarkers. In 1 previous study, Tzoulaki et al used data from the MESA, RS, and London Life Sciences Prospective Population Study cohorts and found that higher levels of acetaminophen glucuronide and glycerol were associated with increased risk of incident CVD, while the relation of glycerol was not independent of CVD risk factors. ⁶ Fasting levels of glycerol were also related to various traits of glucose metabolism. ³⁰ For metabolic methanol, 1 recent MESA study reported that higher levels of methanol were associated with increased left atrial volume, but not with N‐terminal pro‐B‐type natriuretic peptides. ⁴⁴ Metabolic methanol levels also showed association with worse diet health, affected by the intake of dairy, protein, and refined grains. ³²

It is worth noting that the metabolites identified in the MESA cohort were poorly validated in the RS cohort. The successfully validated 5 metabolites (1,5‐anhydrosorbitol, glutamine, glucose, glycerol, and methanol) further highlighted their roles in HF development, mainly driven by or related to glucose metabolism, especially as we observed their high correlation with fasting glucose. Additionally, 3 metabolites (glycine, pyroglutamate, and mannose) were observed to be associated with HF risk only in the RS cohort, suggesting them to be potential biomarkers; particularly, mannose was identified as a possible causal metabolite for diabetes in our earlier research. ²⁶ The inconsistent results between 2 cohorts may be due to differences in the ethnic backgrounds. While the MESA cohort included almost 40% of individuals of European descent, it included greater racial and ethnic diversity than RS, potentially limiting the replicability of the identified metabolites in an overwhelmingly European population. Differences in medical systems and between 2 cohorts from different countries could also contribute to the discrepancies in HF incidence and potential confounding residuals. Further research across diverse populations is needed to assess the utility of metabolomics for precise HF risk stratification. Additionally, the −20 °C storage temperature of serum samples in the RS cohort may have impacted metabolite quality, possibly contributing to the observed discrepancies.

Clinical Implication

The clinical utility of circulating metabolites for HF risk prediction remains questionable. From the disease mechanism side, our research indicates that the identified metabolites are potentially involved in metabolic pathways or drive metabolic changes (diabetes and hypertension) preceding clinical HF. It is also possible that circulating levels of identified metabolites are downstream metabolites of diabetes and hypertension. Our findings also highlight the necessity of confounding control of diabetes and hypertension for research on metabolomics and HF. Future research is needed to explore the intertwined relation among the identified metabolites and metabolic factors, particularly diabetes.

Strength and Limitations

The strengths of this study include long follow‐up, the inclusion of 2 cohorts from different countries, the fact that both were processed identically and analyzed by the same platform, and a comprehensive evaluation of untargeted NMR spectral features beyond known metabolites. There are several limitations. First, the racial and ethnic backgrounds of study participants were different between MESA and RS, and we roughly adjusted for race and ethnicity in analyses in the MESA cohort. Second, the storage temperature in RS was −20 °C, which may have a quality effect on metabolites. Third, although the adjustment was extensive, the possibility of unmeasured confounding remains such as medication use, diet behaviors, and environmental exposures in 2 cohorts. These limitations, along with the absence of genetic admixture markers, restrict our ability to fully compare metabolomic or other omics data between MESA and RS. Another limitation of our study is that potential nonlinear relationships between metabolic features and HF risk were not assessed due to the large number of features analyzed. While our research provided exploratory evidence, a more detailed investigation into nonlinearity is warranted for future studies. Additionally, due to the absence of HF subtype data in RS and the limited sample size in MESA, we did not further explore how involved metabolites differentiated development between HF with preserved ejection fraction and HF with reduced ejection fraction.

Conclusions

In 2 cohorts of middle‐aged to older adults from the United States and the Netherlands, 15 ¹H NMR‐measured metabolites were associated with incident HF risk, but these associations were not independent after adjusting for blood pressure and diabetes status. Incorporating these metabolites into the PREVENT‐HF score did not improve HF risk prediction.

Sources of Funding

The MESA projects are conducted and supported by the National Heart, Lung, and Blood Institute in collaboration with MESA investigators. Support for MESA is provided by contracts 75N92020D00001, HHSN268201500003I, N01‐HC‐95159, 75N92020D00005, N01‐HC‐95160, 75N92020D00002, N01‐HC‐95161, 75N92020D00003, N01‐HC‐95162, 75N92020D00006, N01‐HC‐95163, 75N92020D00004, N01‐HC‐95164, 75N92020D00007, N01‐HC‐95165, N01‐HC‐95166, N01‐HC‐95167, N01‐HC‐95168, N01‐HC‐95169, UL1‐TR‐000040, UL1‐TR‐001079, UL1‐TR‐001420, UL1TR001881, DK063491, R01HL133932, and R01HL105756. Additional support for the metabolomics data was provided by the EU COMBI‐BIO project (FP7, 305 422).

Disclosures

I.N. is a speaker for Bayer and Boehringer Ingelheim/Lilly Alliance; consultant for Lilly and Boehringer Ingelheim; and a prior advisory board member for Novo Nordisk. The remaining authors have no disclosures to report.

Supporting information

Data S1