Introduction

Heart failure (HF) is the leading cause of hospitalization and mortality worldwide (Savarese and Lund 2017; Rajadurai et al. 2017). Although advanced new pharmacological therapies and devices have improved outcomes, functional capacity and quality of life for HF patients remain poor, especially in the advanced stages (Lavie et al. 2013; Sperry et al. 2015). B-type natriuretic peptide (BNP) is a robust biomarker for diagnosing acute HF (Maisel et al. 2004). However, BNP represents cardiac wall stress rather than global functioning status. HF care involves the function of multiple end organs, such as the liver and skeletal muscles, and patients’ nutritional state and metabolic energy systems. New assessment tools that compensate for the weakness of BNP are needed to improve HF care and prognostication (Chioncel et al. 2016).

Circulating amino acid profiles are regarded as systemic metabolic biosignatures that represent the metabolic balance among disease-related pathophysiology, food intake, absorption, and tissue synthesis and breakdown (Hakuno et al. 2015; Griffin et al. 2011). Our previous metabolomics study demonstrated several amino acids as the best metabolites for identifying HF. We also found that the amount of total essential amino acids indicated nutritional status and provided significant prognostic value (Cheng et al. 2015). Exploring the dynamics of essential amino acids in plasma may help physicians better interpret HF-related metabolism disturbances, guide future nutritional interventions, and lead to development of new medications and devices.

Although metabolomics provides copious information about metabolism status, the limited availability of mass spectrometry and the complexity of multi-metabolite measurement present obstacles for its clinical application. Tools such as ultra-performance liquid chromatography (UPLC) make quantification of essential amino acids feasible. Aiming for clinical application, this study sought to develop a simplified plasma essential amino acid-based profile by measuring only two amino acids to risk-stratify patients after acute/decompensated HF. We tested whether the metabolism information the profiles provide has prognostic value additive to traditional risk factors and BNP.

Methods

Patients and study design

From January 2011 to May 2014, patients hospitalized due to acute or decompensated chronic HF (the American College of Cardiology and the American Heart Association HF classification system stage C) with a left ventricular ejection fraction (LVEF) of  < 50%, and aged 20–85 years, were consecutively enrolled. Exclusion criteria included (i) the presence of disorders other than HF that might compromise survival within 6 months; (ii) patients being bed-ridden for > 3 months and/or unable to stand alone; (iii) the presence of systemic diseases such as hypothyroidism, decompensated liver cirrhosis, and systemic lupus erythematosus; (iv) patients with severe coronary artery disease without complete revascularization therapy; and (v) patients with a serum creatinine of > 3 mg/dl. Normal controls were aged 20–85 years, and had no significant systemic disease, such as hypertension, diabetes mellitus, or coronary artery disease. They were not on any medications, and had an LVEF of  > 60%. Informed consent was obtained from all patients. The study was designed and carried out in accordance with the principles of the Declaration of Helsinki and with approval from the Ethics Review Board of Chang Gung Memorial Hospital.

To validate the prognostic value of the simplified metabolic profile, a second independent cohort, including 391 HF patients as defined above, was recruited from July 2013 to March 2016. (Study flow diagram is provided in Supplemental Fig. 1).

Fig. 1
figure 1

Prognostic value of simplified essential amino acid profile. a Blood leucine levels in patients with and without events. Dotted line boxes indicate leucine level at 145 μM; b Association between leucine level and risk of the composite events in the additive Cox regression models; c The profile for defining high-risk types 1 and 2 based on leucine (Leu) and phenylalanine (Phe) levels; d and e The Kaplan–Meier curves for high-risk types 1 and 2, and for low-risk patients in the initial and the validation cohorts, respectively (for composite events). Amino acids were measured by UPLC

Full size image

Blood sampling and examination

For HF patients, blood samples for metabolomics were collected in the early morning after fasting for 8 h before discharge from hospital in EDTA-containing tubes. Plasma was analyzed by UPLC workflow. BNP was measured in triplicate with the Triage BNP Test (Biosite, San Diego, CA), which was a fluorescence immunoassay for quantitative determination of plasma BNP. Precision, analytical sensitivity and stability characteristics of the assay were previously described (Maisel et al. 2004). For normal controls, blood samples for UPLC analysis and BNP were collected at enrollment. Measurement of other parameters, including estimated glomerular filtration rate (eGFR), hemoglobin, and albumin, was conducted in the central core laboratory.

Follow-up program

Follow-up data were prospectively obtained every month from hospital records, personal communication with the patients’ physicians, telephone interviews with patients, and patients’ regular visits to staff physician outpatient clinics (followed for up to 3 years until they died, the study ended, or they were lost to follow-up). “Re-hospitalization” was defined as HF-related re-hospitalizations. A committee of three cardiologists adjudicated all hospitalizations without knowledge of patients’ clinical variables to determine whether the events were related to worsening HF. “All-cause death” was chosen as an endpoint because of the interrelationship of HF with other comorbidities in the patient cohort. The composite event of HF-related re-hospitalization and all-cause death (time to the first event) was analyzed for prognostic purposes. We also performed a competing risks analysis. For validation, patients in the second independent population were followed up for one year.

Ultra-performance liquid chromatography-based amino acid measurement

EDTA plasma samples were collected and stored at − 80 °C until assayed. The plasma samples (100 μl) were precipitated by adding an equal volume (100 μl) of 10% sulfosalicylic acid containing an internal standard (norvaline 200 µM). After protein precipitation, the samples were vortexed and centrifuged at 12,000 g for 10 min at room temp. After the samples were centrifuged, 20 μl of the supernatant was mixed with 60 μl working buffer (borate buffer, pH 8.8). The derivatization was initiated by the addition of 20 μl of 10 mM AQC in acetonitrile. After 10 min incubation, the mixture was added with an equal volume Eluent A (20 mM ammonium formate/0.6% Formic acid/1% acetonitrile) and analyzed using the ACQUITY UPLC System (Frank and Powers 2007; Pappa-Louisi et al. 2007). AQC derivatization reagent was obtained from Waters Corporation (Milford, MA, USA). An aqueous amino acid standard mixture was prepared at different concentrations (0, 25, 50, 100, 250, 500 μM) for each amino acid and was done by the same procedure. The Waters ACQUITY UPLC® System consisted of a Binary Solvent Manager (BSM), a Sample Manager fitted with a 10 μl loop, and a Tunable UV (TUV) detector. The system was controlled and data collected using Empower™ 2 Software. Separations were performed on a 2.1 × 100 mm ACQUITY BEH C18 column at a flow rate of 0.70 mL/min. The average intra-assay coefficient of variation was 4.5% for leucine and 4.6% for phenylalanine. A total coefficient of variation was 4.1% for leucine and 3.7% for phenylalanine. The limit of detection was 0.9 μM for leucine and 3.3 μM for phenylalanine. The linear range was 25–500 μM for these amino acids. For other amino acids, the average intra-assay coefficient of variation was 4.3–4.6% and the limit of detection was 0.1–4.7 μM. Leucine and isoleucine were separated by UPLC.

Quantification of plasma acylcarnitines by tandem mass spectrometry

The acylcarnitine analyses were carried out with the AbsoluteIDQ® p180 Kit (Biocrates Life Science AG, Innsbruck, Austria) (Cheng et al. 2015). The kit enabled us to identify and quantify 40 acylcarnitines. 10 μL aliquot of each plasma sample was mixed with isotopically labeled internal standards in a multititer plate, and dried under nitrogen. The samples were derivatized with 5% phenylisothiocyanate (PITC) for the amino acids and biogenic amines, and dried again. 300 μL of extraction solvent (5 mM ammonium acetate in methanol) was added, and after 30 min incubation, it was centrifuged for 2 min at 100 × g. The filtrate of sample extracts was diluted with kit running solvent (Biocrates Life Sciences AG) for flow injection analysis coupled with tandem mass spectrometric analysis (FIA-MS/MS), Xevo TQ mass spectrometer (Waters Corporation). The analysis was performed in positive electrospray ionization mode. Identification and quantification were achieved by multiple reactions monitoring (MRM). The FIA-MS/MS method was used (100% organic running solvent) with varying flow conditions (0 min, 30 μL/min; 1.6 min 30 μL/min; 2.4 min, 200 μL/min; 2.8 min, 200 μL/min; 3 min 30 μL/min). The corresponding MS settings were as follows: dwell time of 0.019–0.025 s; capillary voltage at 3.92 kV for positive mode; capillary voltage at 1.5 kV for negative mode; nitrogen as collision gas medium; source temperature at 150 °C. It was standardized by spiking in of isotopically labeled standards. For example, deuterium-labeled acylcarnitine standards, including D3-C2, D3-C3, D3-C4, D9-C5, D3-C6, D3-C8, D3-C10, D3-C16, and D3-C18 carnitines, were used for C2, C3, C4, C5, C6, C8, C10, C16, and C18, respectively. For a subset of acylcarnitines, measurements were semi-quantitative since specific standards were not commercially available. A detailed overview of the measured acylcarnitines is provided in Supplemental Table I.

Statistical analyses

Results are expressed as the mean ± SD for variables with normal distribution, and median (interquartile range) for variables with skewed distribution, and as the number (percentage) for categorical variables. Data were compared by two-sample t tests, ANOVA (subgroup analysis was conducted by Bonferroni correction) and Chi square (multiple comparison with Bonferroni-adjusted p values) when appropriate. Receiver operating characteristic (ROC) curves were estimated, and we used Youden’s index to identify the cutoff value of variables. Follow-up data were collected as scheduled or at the last available visit. We used Cox proportional hazards models with stepwise analysis to determine independent predictors of the first defined events (death, or HF-related re-hospitalization). Variables with p value < 0.05 in the univariate analysis were selected for the multivariable analysis. Hazard ratios (HRs) and 95% confidence intervals (CIs) were calculated. To compare the time-dependent outcomes, we performed Kaplan–Meier analyses with a log-rank test. Furthermore, since all-cause death and HF-related re-hospitalizations were also competing risks, we used a competing risks analysis model to estimate the risk of each of these events independent of each other. Since we have finished the validation study, we used R software to calculate the observed power. Based on the data of the control group (n = 257, event rate = 0.202 at 1 year) and the high-risk type 2 group (n = 45, event rate = 0.556 at 1 year, with a hazard ratio of 3.62) in the validation cohort, the observed power was found to be 99.9%, indicating adequate sample size. All statistical analyses were two-sided and performed using SPSS software (version 22.0, SPSS, Chicago, IL, USA) and R software (version 3.5.1). A p value of < 0.05 was considered significant.

Results

Baseline characteristics

The study enrolled a total of 599 HF patients and 94 normal controls. We observed 164 (27.4%) composite events among the HF patients over a follow-up period of 3 years, including 72 (11.9%) deaths and 132 (22.7%) HF-related re-hospitalizations. Comparisons of baseline characteristics and laboratory data between patients with and without composite events are shown in Table 1. Patients who experienced events were older, more often female, had higher New York Heart Association functional classes, incidence of diabetes mellitus, chronic kidney disease, hypertension and use of diuretics, and higher BNP levels compared to the patients with no events. They also had lower systolic blood pressure and incidence of using beta-blockers, as well as lower total cholesterol, triglyceride, hemoglobin, albumin, and eGFR. No significant difference in left ventricular ejection fraction was noted between the two groups.

Table 1 Demographic and laboratory data in heart failure patients (Initial cohort)
Full size table

Identification of amino acids with potential for assessing prognosis

Comparisons of all amino acids between patients with and without composite events are shown in Table 2 (measured by UPLC). Plasma levels of all essential amino acids were lower in patients who experienced events compared to those who did not, except for phenylalanine, which increased in patients with events. To identify amino acids that might be useful for prognosis, we performed Cox univariate and multivariable analysis. Results showed that four essential amino acids—leucine, methionine, phenylalanine and tryptophan—have strongly significant prognostic value, independent of others. In addition, leucine had the best correlation with the total amount of essential amino acids (r = 0.83, p < 0.001), followed by methionine, phenylalanine and tryptophan (r = 0.42, 0.35 and 0.10, respectively). We, therefore, used leucine as the main essential amino acid for risk stratification. To keep the quality of measuring amino acids stable, we used UPLC to measure these amino acids for statistical analysis in the initial and the validation studies.

Leucine levels were poorly correlated with albumin levels (r = 0.26, p < 0.001). Although lower leucine levels were significantly associated with higher event rates, a significant portion of events also happened in patients with high leucine levels (Fig. 1a), indicating a reverse J shaped relationship (Fig. 1b). We performed the ROC curve analyses using a cutoff value based on Youden’s index (supplemental Fig. II). Based on the cutoff, patients were separated into two subgroups (patients with leucine levels ≥145 μM and < 145 μM). In these two subgroups, we performed multivariable analyses to identify the amino acids that were independently significant for prognosis. In patients with leucine levels ≥ 145 μM and < 145 μM, phenylalanine (HR 1.87, 95% CI 1.26–2.78) and leucine (HR 2.80, 95% CI 2.25–3.48) were found to be the best prognostic amino acids, respectively; including other amino acids did not provide additional prognostic value. Based on ROC curves, the cutoff points for phenylalanine and leucine were set at 88.9 and 81.2 μM, respectively. On this basis, we created a profiling method and categorized patients into low risk or high risk types 1 and 2 (Fig. 1c).

Confirming the value of the profile for stratifying risk, Cox univariate analysis showed significant prognostic value for high-risk types 1 and 2 (Table 3). Further Cox multivariable analysis showed that both high-risk types 1 and 2 retained significant prognostic value after adjusting for age, sex, diabetes mellitus, chronic kidney disease, hypertension, hemoglobin, albumin, and BNP (model 1) and after cumulative adjustment for New York Heart Association functional class, diastolic blood pressure, body mass index, medications and cholesterol (model 2). Kaplan–Meier curves showed that both type 1 and 2 high-risk patients had lower three-year event-free survival rates compared to patients classified as low risk (Log rank = 9.98 and 118.7, p = 0.002 and < 0.001, respectively) (Fig. 1d). The competing risks analysis also demonstrated that the high-risk types had value for predicting each of these events independent of each other (p < 0.001; Supplemental Fig. III and Table II). The prognostic value of the metabolic profile was further validated in 391 stage C patients. During the 1-year follow-up period, 111 (28.4%) composite events were noted. Kaplan–Meier curves show that both type 1 and 2 high-risk patients had lower 1-year event-free survival rates compared to low-risk patients (Fig. 1e). The demographic data for the validation cohort are shown in Supplemental Table III.

Table 2 Comparisons and analysis of plasma amino acid concentrations between heart failure patients with and without composite events (Initial cohort)
Full size table
Table 3 COX univariate and multivariable analysis for the prognostic value of different high-risk types in patients with heart failure (Initial cohort)
Full size table

Since patients who suffered an event had a significantly higher rate of diabetes mellitus and hypertension, we conducted subgroup analysis for patients with diabetes mellitus and hypertension (supplemental Fig. IV). In the different subgroups, the cutoff values for leucine and phenylalanine were slightly different. However, the prognostic value of high-risk types 1 and 2 remained significant in these subgroups.

Characteristics of patients at high risk

Compared to the low-risk group, both types of high-risk patients were older, had lower blood pressure and eGFR, but higher New York Heart Association functional classes, and BNP. In addition, high-risk type 2 patients were more often female, but had lower body mass indexes; they also had a lower percentage of using guideline-based medications, and lower levels of cholesterol, hemoglobin, and albumin (Table 4).

Table 4 Demographic and laboratory data for patients with different risk types (initial cohort)
Full size table

Lipidomics in patients of different risk types

In this study, in addition to our amino acid-based definition of high risk, we would like to provide additional supportive data based on other metabolites as well. Since the value of acylcarnitines in assessing the severity of HF has been demonstrated (Ahmad et al. 2016), we investigated lipidomics by measuring a variety of acylcarnitines in normal controls and patients with different risk types (Table 5) (measured by tandem mass spectrometry). Compared to the normal controls, HF patients at low risk of events had higher levels of long-chain (C18:1, C18:2, C16, and C10:2) and short-chain acylcarnitines (C5 and C4). It is noteworthy that compared to the low-risk group, high-risk type 1 was associated with higher levels of long- (C18, C18:1, C14, C14:2, C12, C12:1, C10, and C10:1), medium- (C9, C8, and C6:1) and short-chain acylcarnitines (C5, C5:1, C4, C3 and C2). High-risk type 2 patients had similar levels of long-, medium- and short-chain acylcarnitines, but they had lower levels of C3 compared to the low risk group.

Table 5 Comparisons of acylcarnitines among patients with different risk types and normal controls (Initial cohort)
Full size table

Discussion

This study developed a novel risk stratification strategy for patients with acute/decompensated HF based on a simplified essential amino acid-based profile. Our profile defined two different types of patients at high risk for HF-related events, namely, high risk types 1 and 2. The different risk types demonstrated different metabolic and nutritional statuses associated with HF. Both high-risk types provided prognostic value additive to BNP and traditional risk factors.

BNP is a powerful diagnostic biomarker for HF, with ST2 and galectin-3 providing additional prognostic information (Maisel et al. 2004; Wang et al. 2016; Rehman et al. 2008). However, all these biomarkers offer limited information beyond wall stress and activation of global fibrosis. Previous studies have shown that metabolic assessment provided superior prognostic value for HF patients compared to BNP and galectin-3 (Cheng et al. 2015; Wang et al. 2017). This study revealed that our simplified amino acid profile had prognostic value independent of albumin, BNP and clinical risk factors. The profile provides a quantifiable readout with information that is often not obvious from assessing cardiac wall stress and traditional nutritional biomarkers.

Based on the amino acid profile, high-risk type 1 was defined by high circulating leucine and phenylalanine levels. Of all essential amino acids, our data showed that leucine, phenylalanine, tryptophan and methionine were independently significant for identifying HF patients at high risk of composite events. Additionally, leucine had the best correlation with the total amount of essential amino acids that have been shown to predict HF-related events (Cheng et al. 2015). Thus, we used leucine levels as the main biomarker for risk stratification.

Interestingly, we noted a reverse J shaped relationship between leucine levels and event rates. Although lower leucine levels predicted higher event rates, event rates were also increased in patients with high leucine and phenylalanine levels. Phenylalanine was the only essential amino acid that was elevated in HF patients compared to the normal controls, suggesting a pathophysiological link. Nitric oxide synthesis dysregulation and tetrahydrobiopterin depletion were probably related to impaired conversion of phenylalanine to tyrosine, followed by accumulation of phenylalanine (Nishijima et al. 2011; Ziolo et al. 2004). Furthermore, the increase of phenylalanine levels may also be associated with increased muscular protein breakdown (Aquilani et al. 2012) and impaired liver function caused by HF. Accordingly, the increased leucine and phenylalanine levels in type 1 high-risk patients indicate substantial tissue breakdown, which is probably related to insufficient tissue perfusion, increased insulin resistance, and dysfunctional energy production machinery associated with HF. Accumulated circulating long-chain and medium-chain acylcarnitines point to mitochondrial β-oxidation dysfunction in processing lipids for energy production, which may further cause insulin resistance (Hunter et al. 2016; Schooneman et al. 2013). Short-chain acylcarnitines are partially derived from the metabolism of amino acids; elevated levels are thus compatible with the suggested phenomena of tissue protein breakdown for energy production. However, the formal evaluation of muscle protein dynamics involves complex calculation based on isotope infusion studies (Nagabhushan and Narasinga Rao 1978; Liu and Barrett 2002), which are not applicable to patients with HF.

Intriguingly, high-risk type 2 is associated with remarkable malnutrition as indicated by extremely low amounts of circulating amino acids; such low levels could shut down a variety of protein and enzyme functions, causing HF deterioration. The type 2 high-risk population presented circulating acylcarnitine levels similar to those of the low-risk population—even lower for some short-chain acylcarnitines. It is worth noting that type 2 high-risk patients were characterized by a variety of clinical risk factors indicating poor prognosis. One advantage of amino acid-based risk stratification is its potential for providing information about nutritional deficiency as part of risk assessment, as well as enabling clinicians to propose plans for intervention. It remains to be elucidated whether patient outcomes can be improved if the type 2 risk pattern is identified and corrected earlier.

Although researchers search enthusiastically for a nutritional regimen specific for HF, our findings do not point to a universal solution. Nutritional supplementation as recommended by current guidelines may not be adequate for managing the metabolic changes associated with HF from moment to moment. Both HF patients and patients in critical condition are commonly subjected to over- and under-feeding (Reid 2006). Both practices are harmful, yet poorly defined. A strategic approach to providing appropriate calorie and protein supply would attempt to switch the catabolic state of amino acids to an anabolic state, benefitting patients by shifting them away from the high-risk profiles, while maintaining appropriate metabolic condition.

Study limitations

This study has a few limitations. Although we have identified essential amino acids critical for prognosis, the causal mechanisms to explain why they work are speculative and not well established. To clarify the value of amino acid profiling, serial measurement is needed to understand the dynamics of the amino acid profile over time and to track changes associated with nutrition and medications.

Conclusions

Our study shows that a simplified amino acid-based risk stratification profile provides information on amino acid and lipid metabolism in HF patients with robust prognostic value additive to BNP and traditional risk factors. High-risk type 1 represents patients with substantial metabolism disturbance and accumulated circulating waste of incompletely metabolized lipids. High-risk type 2 represents patients characterized by severe malnutrition and multiple clinical factors signaling poor prognosis. Future research is needed to investigate whether multidisciplinary interventions based on such profiles can improve outcomes among patients who are already receiving guideline-based medications.