Serum metabolomic signatures integrating sphingosine-1-phosphate and tetrahydrocortisone improve prognostic assessment in non-ischemic cardiomyopathy

Abstract

Introduction

Dilated cardiomyopathy (DCM) and left ventricular non-compaction (LVNC) are major non-ischemic cardiomyopathy (NICM) subtypes with heterogeneous outcomes. Conventional clinical and echocardiographic markers remain insufficient for long-term risk stratification.

Objectives

This study aimed to identify serum metabolites associated with adverse cardiovascular outcomes and evaluate their incremental prognostic value in NICM.

Methods

Thirty-two patients with DCM or LVNC and left ventricular ejection fraction < 50% were enrolled and followed for a median of 45.5 months. The primary endpoint was a composite of cardiovascular death, heart failure-related hospitalization, or clinically indicated cardiovascular device implantation. Baseline serum samples underwent liquid chromatography–mass spectrometry-based untargeted metabolomic profiling. Differential metabolites were identified using OPLS-DA and KEGG enrichment analysis. Prognostic metabolites were screened using Cox regression, Kaplan–Meier analysis, correlation filtering, and ROC analysis. An integrated Cox model combining clinical and metabolic markers was evaluated using bootstrapping, calibration, and time-dependent ROC analysis.

Results

A total of 299 differential metabolites were identified and enriched in bile secretion, steroid hormone biosynthesis, and neuroactive ligand-receptor interaction pathways. Sphingosine-1-phosphate (S1P) and tetrahydrocortisone (THE) were selected as final prognostic metabolite biomarkers, with ROC AUCs of 0.777 and 0.793, respectively. The integrated model incorporating S1P, THE, tricuspid annular plane systolic excursion, and total protein achieved a bootstrap-corrected C-index of 0.772, with time-dependent AUCs of 0.92 and 0.86 at 3 and 5 years.

Conclusions

Serum metabolomics may provide complementary prognostic information in NICM. S1P and THE are exploratory biomarkers linked to remodeling and stress, supporting risk stratification in NICM with reduced ejection fraction.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11306-026-02497-3.

Introduction

Non-ischemic cardiomyopathy (NICM) comprises a heterogeneous group of myocardial disorders characterized by structural and functional abnormalities of the myocardium in the absence of coronary artery disease or ischemic injury (Elliott et al., 2008; Wang et al., 2023). NICM has a reported prevalence of 2% to 15% in community and hospital-based populations, increasing to nearly 50% in large-scale clinical trials (Wu, 2007). Although NICM encompasses diverse etiologies, it typically results in progressive myocardial injury, ventricular dysfunction, and ultimately clinical heart failure (HF). Among its subtypes, dilated cardiomyopathy (DCM) is the most common, characterized by ventricular dilatation and impaired contractility, and represents a leading cause of heart transplantation worldwide (Weintraub et al., 2017). In contrast, left ventricular non-compaction (LVNC) arises from disrupted myocardial embryogenesis and is considered a rare congenital cardiomyopathy, with disease severity ranging from asymptomatic cases to severe cardiac dysfunction and sudden cardiac death, leading to substantial prognostic heterogeneity (Paluszkiewicz et al., 2022). These features highlight the urgent need for accurate prognostic assessment of adverse cardiovascular outcomes to support risk stratification and inform clinical management in patients with NICM (Chow et al., 2017).

Conventional prognostic evaluation in NICM largely depends on established clinical indicators such as New York Heart Association (NYHA) functional class, natriuretic peptide levels, and left ventricular ejection fraction (LVEF) (Feng et al., 2022; Li et al., 2024, 2025). However, these indicators primarily reflect the clinical and functional consequences of established myocardial dysfunction and may not fully capture the systemic biological alterations associated with disease progression (Chen et al., 2023). Emerging evidence indicates that NICM is not merely a localized cardiac disorder but a systemic syndrome involving maladaptive neurohumoral activation, chronic inflammation, metabolic remodeling, and adverse inter-organ crosstalk, all of which may contribute to myocardial injury and remodeling (Borovac et al., 2020; McDonagh et al., 2022; Zhou et al., 2024). Therefore, relying solely on cardiac-centric functional metrics may provide an incomplete assessment of long-term risk(Vancheri et al., 2024). Approaches that capture circulating metabolic alterations may offer complementary prognostic information and help characterize systemic pathophysiological processes associated with adverse cardiovascular outcomes in NICM.

Metabolomics, by comprehensively profiling circulating metabolites, offers a powerful window into these systemic pathological processes and has emerged as a promising tool in cardiovascular research for biomarker discovery and mechanistic insight (Kozhevnikova et al., 2025). Previous metabolomics studies have demonstrated significant associations between metabolic signatures and HF severity or outcomes (Lanfear et al., 2017). For example, Cheng et al. (2015) identified significant alterations in histidine, phenylalanine, phosphatidylcholine, and taurine across HF stages and demonstrated that metabolite panels outperformed the conventional biomarker BNP in prognostic prediction. However, most of these studies were conducted in mixed HF cohorts dominated by ischemic cardiomyopathy, with limited follow-up duration and minimal integration of key clinical variables such as echocardiographic measures, thereby restricting their applicability to risk prediction in NICM.

Given these gaps in prior research, we established a longitudinal cohort of patients with DCM and LVNC with long-term follow-up and performed untargeted liquid chromatography–mass spectrometry (LC–MS) analysis. Patients were stratified according to the occurrence of adverse cardiovascular outcomes during follow-up. We aimed to: (1) identify serum metabolites and metabolic pathways associated with adverse cardiovascular outcomes; (2) evaluate their prognostic value alone and in combination with conventional clinical indicators; and (3) develop an integrated risk prediction model that incorporates systemic metabolic information and may support individualized risk assessment in NICM.

Methods

Study population

This retrospective study enrolled 32 adult patients diagnosed with LVNC or DCM between May 2012 and October 2022 at Peking Union Medical College Hospital (PUMCH), a tertiary referral hospital. All patients had reduced left ventricular ejection fraction (LVEF < 50%) confirmed by transthoracic echocardiography (TTE) at baseline and were confirmed to have a non-ischemic etiology based on coronary computed tomography angiography (CTA). Patients were excluded if they had coexisting cardiovascular conditions such as hypertension, primary valvular heart disease, congenital heart disease, or if they had cancer, diabetes mellitus, systemic metabolic disorders, or severe multi-organ dysfunction. Patients without available baseline serum samples, complete medical records, essential laboratory or echocardiographic data, or sufficient follow-up information for endpoint adjudication were also excluded.

LVNC was diagnosed primarily according to the echocardiographic criteria proposed by Jenni et al. (2001), including a non-compacted to compacted myocardial ratio > 2.0 at end-systole, prominent trabeculations, and deep intertrabecular recesses. Cardiac magnetic resonance (CMR) findings, when available, were used as supportive evidence. DCM was diagnosed according to the 2016 European Society of Cardiology (ESC) position statement (Pinto et al., 2016). The study protocol was approved by the Ethics Committee of PUMCH (Approval No. I-25PJ0792), conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants. The overall study workflow, including patient enrollment, serum sample collection, metabolomic profiling, follow-up and outcome assessment, and prognostic model development, is illustrated in Fig. 1.

Fig. 1.

Schematic of the study workflow. CTA, computed tomography angiography; DCM, dilated cardiomyopathy; LC-MS, liquid chromatography-mass spectrometry; LVNC, left ventricular non-compaction; OPLS-DA, orthogonal partial least squares discriminant analysis; ROC, receiver operating characteristic; S1P, sphingosine-1-phosphate; TAPSE, tricuspid annular plane systolic excursion; THE, tetrahydrocortisone; TP, total protein; TTE, transthoracic echocardiography

Echocardiographic assessment

TTE was performed at baseline using a commercially available ultrasound system (Vivid E95, GE Healthcare, USA) equipped with a phased-array transducer. Multiplanar images were acquired at a frame rate of approximately 60–90 frames per second, and at least three cardiac cycles were digitally stored for each view. All examinations were performed by experienced sonographers and independently interpreted by two cardiologists following guidelines from the American Society of Echocardiography (ASE). LVEF was calculated using the biplane Simpson’s method, with values below 50% defined as indicative of impaired systolic function. Available echocardiographic parameters, including global longitudinal strain (GLS), left ventricular mass index (LVMI), relative wall thickness (RWT), tricuspid annular plane systolic excursion (TAPSE), and lateral mitral annular systolic velocity (S’), were collected according to standard echocardiographic protocols. All measurements were averaged over three cardiac cycles. The interpreting cardiologists were blinded to patients’ clinical outcomes and metabolomic data.

Follow-up and outcome assessment

Patients were followed annually through outpatient visits at the Department of Cardiology, PUMCH, or through structured telephone interviews conducted with patients or their family members. Follow-up data included current clinical symptoms, medication use, HF-related rehospitalization, clinically indicated cardiovascular device implantation, and survival status. For deceased patients, the date and cause of death were verified through review of inpatient and outpatient medical records, electronic health system entries, official death certificates when available, and structured telephone interviews with patients’ family members. Cardiovascular death was defined as death attributable to HF, sudden cardiac death, fatal arrhythmia, or other clearly documented cardiovascular causes, in accordance with standardized cardiovascular endpoint definitions (Hicks et al., 2018).

The primary endpoint was a composite adverse cardiovascular endpoint, defined as cardiovascular death, HF-related hospitalization, or clinically indicated cardiovascular device implantation. For time-to-event analyses, the event time was defined as the interval from baseline serum sample collection to the first occurrence of any component of the composite endpoint. Patients who did not experience any component of the composite endpoint were censored at the date of their last confirmed follow-up contact before October 28, 2024. Patients were categorized into event and non-event groups according to the occurrence of the composite endpoint during follow-up for subsequent statistical and metabolomic analyses. All follow-up procedures and endpoint adjudication were conducted by trained research personnel who were blinded to patients’ metabolomic data to minimize observer bias.

Sample preparation and untargeted LC-MS analyses

Fasting serum samples were collected from all participants at baseline, centrifuged at 3,000 × g for 10 min at 4 °C, and stored at − 80 °C until analysis. For metabolite extraction, 100 µL of serum was mixed with 400 µL of a pre-chilled methanol: acetonitrile solution (1:1, v/v) containing 2-chloro-L-phenylalanine as an internal standard. The mixture was vortexed for 1 min, incubated at − 20 °C for 1 h, and centrifuged at 14,000 rpm for 15 min at 4 °C. The supernatant was collected for LC-MS analysis. Quality control (QC) samples were prepared by pooling equal volumes from all serum samples. QC samples were injected at regular intervals throughout the analytical sequence to monitor instrument stability and analytical reproducibility. Untargeted metabolomic profiling was performed using an ultra-performance liquid chromatography (UPLC) system (UltiMate 3000, Thermo Scientific) coupled with a high-resolution Orbitrap Exploris 480 mass spectrometer (Thermo Scientific). Chromatographic separation was achieved using an ACQUITY UPLC HSS T3 column (2.1 × 100 mm, 1.8 μm; Waters) with a flow rate of 0.3 mL/min and an injection volume of 1 µL. For both positive and negative ionization modes, the mobile phases consisted of 0.1% formic acid in water (phase A) and 0.1% formic acid in acetonitrile (phase B). Mass spectrometric data were acquired in data-dependent acquisition (DDA) mode with a scan range of m/z 67–1,000. The electrospray ionization (ESI) source parameters were as follows: spray voltage, + 3500 V/−2500 V; sheath gas, 50 arb; auxiliary gas, 10 arb; sweep gas, 1 arb; ion transfer tube temperature, 325 °C; and vaporizer temperature, 350 °C.

Metabolomics data analysis

Raw LC-MS data files (.raw) were imported into Compound Discoverer software (Thermo Fisher Scientific) for data preprocessing, including peak detection, retention time alignment, m/z correction, peak extraction, and peak area quantification. Peak alignment was performed across all samples using a retention time tolerance of 0.2 min and a mass tolerance of 5 ppm. Features were retained according to the following criteria: mass deviation ≤ 5 ppm, signal intensity deviation ≤ 30%, signal-to-noise ratio ≥ 3, and minimum intensity ≥ 1 × 10⁵. Adduct ions were integrated during peak processing. MS/MS spectra were matched against the mzCloud, mzVault, and ChemSpider databases for metabolite identification, and annotations were further verified by theoretical fragment analysis. Identified metabolites were annotated using KEGG, HMDB, and LIPID MAPS databases for compound classification and pathway information.

For downstream statistical analysis, metabolite abundance data were normalized to the maximum value, log10-transformed, and scaled. Age was adjusted as a covariate using a linear regression model, and age-adjusted metabolite abundances were used for subsequent multivariate and differential analyses. Principal component analysis (PCA) was performed to evaluate sample distribution and QC reproducibility. Orthogonal partial least squares discriminant analysis (OPLS-DA) was conducted using the R package ropls to assess metabolic differences between patients with and without adverse cardiovascular outcomes. The robustness of the OPLS-DA model was evaluated using 1,000 permutation tests. Differential metabolites were identified based on a variable importance in projection (VIP) score ≥ 1 and a false discovery rate (FDR)-adjusted P value < 0.05 derived from Student’s t-test comparing patients with and without adverse cardiovascular outcomes (Zhou et al., 2025). Significantly altered metabolites were then subjected to KEGG pathway enrichment analysis. Pathways with a q-value < 0.05, calculated using a hypergeometric test, were considered significantly enriched.

Analysis of clinical indicators

To address missing values in clinical indicators, multiple imputation was performed using the R package mice. Five imputed datasets (m = 5) were generated using the predictive mean matching (PMM) method. Variables were selected as predictors for imputation if their absolute correlation coefficient with the variable being imputed was greater than 0.2. Subsequent analyses involving imputed clinical indicators were performed across the five imputed datasets, and pooled estimates were obtained according to Rubin’s rules.

After imputation, correlation analyses were performed between clinical indicators and the metabolites identified as prognostically significant. Pearson correlation was used for continuous clinical variables when appropriate, whereas Spearman’s rank correlation was used for non-normally distributed, ordinal, or categorical variables.

Screening of candidate clinical and metabolic prognostic biomarkers

All metabolites involved in the significantly enriched pathways were selected for further prognostic evaluation. The associations between these metabolites and adverse cardiovascular outcomes were assessed using univariable Cox proportional hazards regression and Kaplan–Meier survival analysis. For Kaplan–Meier analysis, metabolite levels were stratified according to the median value. Metabolites were retained as candidates only if they achieved statistical significance (P < 0.05) in both univariable Cox regression and Kaplan–Meier survival analysis.

To refine the candidate metabolic biomarker panel, two complementary analytical approaches were applied. Multivariable Cox proportional hazards regression was performed on all candidate metabolites simultaneously to assess their independent prognostic contributions after mutual adjustment. Pearson correlation analysis was conducted among the candidate metabolites to quantify inter-marker redundancy and identify markers conveying distinct biological signals. Final biomarker selection was determined by jointly considering evidence of independent prognostic contribution and low inter-marker redundancy. Receiver operating characteristic (ROC) curve analysis was subsequently performed for the selected biomarkers to corroborate their individual predictive utility for adverse cardiovascular outcomes.

To identify clinical indicators with potential prognostic significance, univariable Cox proportional hazards regression analyses were performed for each candidate clinical variable. Variables demonstrating a statistically suggestive association with adverse cardiovascular outcomes (P < 0.1) were considered for further analysis and subsequently incorporated into the multivariable prognostic model (Chen et al., 2022).

Development and validation of the integrated prognostic prediction model

An integrated prognostic prediction model was constructed using multivariable Cox proportional hazards regression based on the pre-screened candidate clinical predictors and metabolite prognostic biomarkers. Independent prognostic factors were identified, and their corresponding hazard ratios (HRs) with 95% confidence intervals (CIs) were reported.

To evaluate model discrimination and internal validity, bootstrap resampling with 1,000 iterations was performed to calculate the bootstrap-corrected concordance index (C-index). Calibration curves at 3 and 5 years were plotted to assess the agreement between model-predicted probabilities and observed outcomes, with the ideal diagonal used as a reference. Time-dependent ROC curves and corresponding AUC values at 3 and 5 years were generated to evaluate time-specific predictive accuracy. All statistical analyses were implemented in R (version 4.5.0) using the survival, rms, and timeROC packages.

Results

Baseline characteristics of study participants

A total of 32 patients diagnosed with either LVNC or DCM, all with LVEF < 50%, were included in the final analysis. The median follow-up duration was 45.5 months (interquartile range [IQR]:

15.0–80.5 months). During follow-up, 16 patients (50.0%) experienced the composite adverse cardiovascular endpoint and were classified into the event group. The first qualifying endpoint events included 5 cardiovascular deaths, 9 HF-related hospitalizations, 1 clinically indicated pacemaker implantation, and 1 ventricular assist device implantation. Among the 16 patients in the event group, 15 subsequently died during long-term follow-up. The remaining 16 patients (50.0%) did not experience any component of the composite endpoint or death and were classified into the non-event group.

Baseline demographic, laboratory, and echocardiographic characteristics are summarized in Table 1. Notably, TAPSE was significantly lower in the event group than in the non-event group (15 ± 4 mm vs. 19 ± 4 mm, P = 0.025), suggesting impaired right ventricular systolic function in patients who experienced adverse cardiovascular outcomes. No other baseline clinical or echocardiographic variables differed significantly between the two groups.

Table 1.

Baseline characteristics of the study population

	Event N = 16	Non-event N = 16	P-value
Demographics
Age, years	50 ± 16	43 ± 16	0.145
Males, n (%)	11 (68.8%)	9 (56.3%)	0.465
BMI, kg/m2	24.1 ± 4.3	25.0 ± 2.8	0.604
Risk factors
Smoking, n (%)	5 (31.3%)	3 (18.8%)	0.685
Drinking alcohol, n (%)	5 (31.3%)	3 (18.8%)	0.685
SBP (mmHg)	108 ± 19	105 ± 15	0.743
DBP (mmHg)	70 ± 12	71 ± 10	0.531
HR (bpm)	84 ± 13	79 ± 7	0.469
eGFR, ml/min/1.73m2	82.82 ± 21.00	99.44 ± 22.22	0.077
Echocardiographic parameters
LVEF (%)	26 ± 7	32 ± 10	0.217
TAPSE (mm)	15 ± 4	19 ± 4	0.025 *
S′ (cm/s)	8 ± 3	8 ± 3	0.604
LVMI (g/m2)	140.8 ± 36.5	129.2 ± 35.8	0.463
RWT	0.21 ± 0.05	0.19 ± 0.03	0.404
GLS (%)	−7.0 ± 3.0	−9.7 ± 4.4	0.248
Biochemical analyses
HGB (g/L)	150 ± 14	154 ± 15	0.411
TP (g/L)	68 ± 7	73 ± 6	0.072
Alb (g/L)	41 ± 7	45 ± 3	0.075
A/G	1.5 ± 0.3	1.6 ± 0.1	0.225
NT-proBNP (ng/L)	6733 [1094, 8287]	599 [214, 2313]	0.082
Time to endpoint or censoring (months)	26 [10, 56]	54 [30, 82]	0.046 *

Data are presented as mean ± SD, median [IQR], or n (%), as appropriate.

^* p < 0.05 was considered statistically significant.

Event was defined as the first occurrence of the composite adverse cardiovascular endpoint, including cardiovascular death, HF-related hospitalization, or clinically indicated cardiovascular device implantation.

Alb, albumin; A/G, albumin to globulin ratio; BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; GLS, global longitudinal strain; HGB, hemoglobin; HR, heart rate; LVEF, left ventricular ejection fraction; LVMI, left ventricular mass index; NT-proBNP, N-terminal pro B-type natriuretic peptide; RWT, relative wall thickness; S’, systolic velocity at the mitral annulus; SBP, systolic blood pressure; TAPSE, tricuspid annular plane systolic excursion; TP, total protein.

Follow-up echocardiographic data were available in 24 of 32 patients and are summarized in Additional file 1: Table S1. Compared with the non-event group, patients in the event group showed significantly lower follow-up LVEF and greater left ventricular end-systolic diameter (LVESD), left atrial diameter (LAD), and tricuspid regurgitation velocity (TRV), whereas the interval from baseline to follow-up TTE was not significantly different between the two groups. Baseline treatment status is summarized in Additional file 1: Table S2. Baseline medical and device therapies were descriptively compared between the event and non-event groups, and no statistically significant between-group differences were observed.

Metabolomic profiling and differential analysis

Untargeted serum metabolomic profiling was performed on all 32 patients using LC-MS. PCA and OPLS-DA analyses demonstrated clear separation between the event and non-event groups, indicating distinct metabolic phenotypes associated with the occurrence of adverse cardiovascular outcomes (Fig. 2A, Additional file 1: Fig. S1A). In contrast, PCA stratified by diagnosis showed substantial overlap between DCM and LVNC samples, indicating no clear diagnosis-driven separation in the overall metabolomic profiles (Additional file 1: Fig. S1B). In exploratory subgroup comparisons, S1P and THE showed similar trends across the DCM and LVNC subgroups (Additional file 1: Fig. S2). The OPLS-DA model showed good fit (R²Y = 0.793, P = 0.028) and moderate predictive ability (Q² = 0.436, P = 0.001), and its robustness was supported by 1,000 permutation tests (Fig. 2B). Using predefined thresholds of VIP ≥ 1 and FDR-adjusted P value < 0.05, a total of 299 differential metabolites were identified between the two groups (Fig. 2C). KEGG pathway enrichment analysis showed that these metabolites were mainly enriched in bile secretion, steroid hormone biosynthesis, and neuroactive ligand-receptor interaction pathways (Fig. 2D, Additional file 1: Table S3). These findings suggest that dysregulation of neurohumoral and endocrine-related metabolic pathways is associated with adverse cardiovascular outcomes in patients with NICM.

Fig. 2.

Serum metabolomic profiling and pathway enrichment analysis in NICM. A OPLS-DA score plot showing distinct separation between the event and non-event groups. B Permutation test validating the robustness of the OPLS-DA model with 1,000 permutations. C Volcano plot of differential metabolites identified based on VIP ≥ 1 and FDR-adjusted P value < 0.05. D KEGG pathway enrichment analysis of significantly altered metabolites

Identification of prognostic metabolite biomarkers

To identify candidate metabolic biomarkers associated with adverse cardiovascular outcomes, differential metabolites from significantly enriched pathways were further evaluated using univariable Cox proportional hazards regression and Kaplan–Meier survival analyses. Patients were stratified according to the median level of each metabolite for survival analysis. Four metabolites—sphingosine-1-phosphate (S1P), tetrahydrocortisone (THE), uric acid, and desoxycortone—were consistently associated with adverse cardiovascular outcomes in both analyses (P < 0.05) (Fig. 3A, B, Additional file 1: Fig. S3A, B). Notably, all four metabolites exhibited hazard ratios (HRs) greater than 1, suggesting that higher levels of these metabolites were associated with increased risk of adverse cardiovascular outcomes. Consistent with this finding, the relative abundances of these four metabolites were significantly higher in the Event group than in the Non-event group (Fig. 3C, D, Additional file 1: Fig. S3C, D). This suggests that higher levels of these metabolites were associated with increased risk of adverse cardiovascular outcomes.

Fig. 3.

Prognostic value and clinical correlations of key serum metabolites in NICM. A–B Cox regression analysis and Kaplan–Meier survival curves for sphingosine-1-phosphate (S1P) (A) and tetrahydrocortisone (THE) (B), stratified by median concentration, 1 indicates expression levels above the median, and 0 indicates expression levels below the median. C-D Violin plots showing serum levels of S1P (C) and THE (D). E Receiver operating characteristic (ROC) curve for S1P. F ROC curve for THE

To evaluate the independent prognostic contribution of each metabolite after mutual adjustment, all four candidate metabolites were simultaneously entered into a multivariable Cox proportional hazards model. In this model, THE (HR = 3.742, P = 0.085) and S1P (HR = 2.494, P = 0.025) retained consistent risk directionality and comparatively larger effect sizes, whereas the prognostic contributions of uric acid (HR = 0.523, P = 0.280) and desoxycortone (HR = 1.344, P = 0.639) were attenuated (Additional file 1: Table S4). These findings suggest that S1P and THE may provide more independent prognostic information than uric acid and desoxycortone. This interpretation was further supported by Pearson correlation analysis, which revealed strong inter-correlations among uric acid, desoxycortone, and S1P (r = 0.60–0.73), indicating considerable overlap in the prognostic information conveyed by these metabolites. In contrast, S1P and THE showed a weaker inter-correlation (r = 0.36), suggesting that they may capture partially distinct and complementary prognostic signals (Additional file 1: Fig. S4).

Integrating these two lines of evidence—consistent risk directionality with larger effect sizes in multivariable Cox regression and lower inter-marker redundancy in correlation analysis—S1P and THE were selected as the final prognostic metabolite biomarkers for subsequent integrated model development. Their individual discriminative utility was further supported by ROC curve analysis, yielding AUCs of 0.777 for S1P and 0.793 for THE, both indicating good predictive performance for adverse cardiovascular outcomes (Fig. 3E, F).

Selection of candidate clinical variables for integrated model development

To identify candidate clinical variables for integrated model development, 14 baseline clinical indicators were screened using univariable Cox proportional hazards regression analysis with a threshold of P < 0.1. As shown in Fig. 4A, total protein (TP) and TAPSE were the only two variables meeting the predefined threshold. TP was significantly associated with a lower risk of adverse cardiovascular outcomes (HR = 0.923, 95% CI: 0.852–0.999, P = 0.049), suggesting that lower serum total protein levels may reflect impaired nutritional or synthetic reserve and confer higher prognostic risk (Additional file 1: Table S5). Similarly, TAPSE showed a suggestive inverse association with adverse cardiovascular outcomes (HR = 0.856, 95% CI: 0.731–1.003, P = 0.054), indicating that lower right ventricular systolic function was associated with increased risk (Additional file 1: Table S5). Both variables were therefore retained for subsequent multivariable modeling.

Fig. 4.

Univariable Cox regression analysis of clinical variables and correlations between selected metabolites and clinical parameters. A Forest plot showing hazard ratios (HRs) with 95% confidence intervals from univariable Cox proportional hazards regression analysis for clinical variables. Red squares denote variables meeting the predefined screening threshold (P < 0.1), and blue squares indicate variables not meeting this threshold (P ≥ 0.1). The x-axis is presented on a logarithmic scale. B Heatmap of Spearman correlation coefficients between selected metabolites and clinical parameters. Significant correlations are indicated with asterisks (^*P < 0.05)

To further explore the relationships between the identified candidate metabolites and prognostic clinical variables, Spearman correlation analyses were performed between the four screened metabolites (S1P, uric acid, desoxycortone, and THE) and baseline clinical parameters (Fig. 4B). S1P was significantly negatively correlated with TAPSE (P < 0.05), suggesting a potential association between sphingolipid metabolism and right ventricular systolic dysfunction. In contrast, THE showed a significant inverse correlation with TP (P < 0.05), which may reflect a relationship between glucocorticoid metabolism and nutritional or hepatic synthetic status. The corresponding correlation coefficients and P-values are provided in Additional file 1: Tables S6 and S7. These correlation patterns provide supportive biological context for the clinical relevance of S1P and THE as prognostically informative metabolites in patients with NICM.

Prognostic performance of clinical, metabolic, and integrated models

An integrated prognostic model was developed using multivariable Cox proportional hazards regression incorporating four candidate predictors: TAPSE, TP, S1P, and THE. Internal validation using 1,000 bootstrap resampling iterations yielded a bias-corrected concordance index (C-index) of 0.772 (95% CI: 0.732–0.812), indicating satisfactory discriminative performance.

In the multivariable Cox model, higher S1P levels (HR = 1.814, CI: 1.026–3.206, P = 0.067) and higher THE levels (HR = 1.816, CI: 0.810–4.071, P = 0.178) were associated with increased risk of adverse cardiovascular outcomes, whereas higher TAPSE (HR = 0.580, CI: 0.294–1.144, P = 0.153) and higher TP (HR = 0.739, CI: 0.412–1.328, P = 0.337) were associated with reduced risk (Additional file 1: Table S8). Given the limited sample size and low events-per-variable ratio, with four model degrees of freedom and approximately 4 events per variable, none of the four predictors reached conventional statistical significance (P < 0.05). Nevertheless, the direction of their effect estimates was clinically consistent, and all four variables were retained in the final integrated prognostic model.

Calibration curves were constructed based on bootstrap-corrected estimates averaged across five multiply imputed datasets. At the 3-year time point, the bias-corrected calibration curve fell below the ideal diagonal across most of the prediction range (predicted risk: 0.33–0.97), indicating a tendency toward overestimation of event risk; additionally, visible separation between the apparent and bias-corrected curves suggested a degree of optimism (Additional file 1: Fig. S5). At the 5-year time point, the apparent and bias-corrected curves were nearly superimposable throughout the prediction range (predicted risk: 0.02–0.85), suggesting minimal optimism and better internal stability; however, both curves lay above the ideal diagonal across much of the prediction range, indicating possible underestimation of long-term event risk (Fig. 5A). These findings suggest that the integrated model showed stronger internal consistency at the 5-year horizon, whereas calibration at 3 years appeared less stable. Time-dependent ROC analysis further demonstrated strong discriminative performance, with AUC values of 0.92 at 3 years and 0.86 at 5 years, supporting the potential prognostic utility of the integrated model across short- and medium-term follow-up horizons (Fig. 5B).

Fig. 5.

Prognostic performance of the integrated model. A Calibration curve for the integrated model at 5 years, showing the agreement between predicted and observed risks of adverse cardiovascular outcomes. The dashed line represents ideal prediction, and the bootstrap-corrected curve represents the internally validated calibration performance. B Time-dependent receiver operating characteristic (ROC) curves of the integrated model at 3 and 5 years. The areas under the curve were 0.92 at 3 years and 0.86 at 5 years

Discussion

In this study, we applied untargeted serum metabolomic profiling to baseline samples from a longitudinal cohort of patients with DCM or LVNC and reduced LVEF. By stratifying patients according to the occurrence of a composite adverse cardiovascular endpoint, we identified distinct metabolic signatures associated with adverse outcomes. Among the candidate metabolites, S1P and THE were selected as prognostically informative circulating biomarkers after multivariable Cox regression and correlation-based filtering. We further developed an integrated Cox prognostic model combining these two metabolites with TAPSE and TP, which showed favorable discrimination, calibration, and time-dependent predictive performance. These findings suggest that circulating metabolomic profiling may provide complementary prognostic information beyond conventional clinical and echocardiographic indicators in NICM.

Our study has several notable strengths. First, the cohort consisted exclusively of patients with non-ischemic myocardial disease, including DCM and LVNC, allowing us to focus on metabolic alterations associated with intrinsic myocardial remodeling rather than ischemic injury. Second, patients with major metabolic comorbidities, such as diabetes mellitus and hypertension, were excluded, thereby reducing potential metabolic confounding. Third, the primary endpoint was defined as a clinically meaningful composite adverse cardiovascular endpoint, including cardiovascular death, HF-related hospitalization, and clinically indicated cardiovascular device implantation. This endpoint captures not only fatal events but also clinically important disease progression requiring hospitalization or device-based intervention. Together, these design features allowed us to explore a serum metabolomic signature associated with adverse cardiovascular outcomes in a relatively well-defined NICM population.

S1P emerged as one of the final prognostic metabolite biomarkers in our analysis. S1P is a bioactive sphingolipid involved in vascular homeostasis, immune regulation, inflammation, endothelial function, and cardiac remodeling. Previous studies have suggested that S1P exerts context-dependent effects in cardiovascular disease (Kovilakath et al., 2023). Under physiological or early compensatory conditions, S1P signaling, particularly through S1PR, may support endothelial barrier integrity and cardioprotection (Phan et al., 2024; Zhao et al., 2025). In contrast, persistent or dysregulated S1P signaling in advanced disease may promote inflammatory activation, fibroblast-to-myofibroblast differentiation, extracellular matrix remodeling, and adverse ventricular remodeling, partly through S1PR3- and TGF-β-related pathways (Ohkura et al., 2017; Pérez-Carrillo et al., 2022). In clinical HF populations, Xue et al. (Xue et al., 2020) reported a U-shaped association between plasma S1P levels and mortality, with abnormally elevated S1P associated with increased risk of death (Xue et al., 2020). However, the upstream mechanisms driving elevated circulating S1P in human HF, including whether fibrotic remodeling is a major contributor, remain incompletely characterized.

In the present study, elevated serum S1P was associated with adverse cardiovascular outcomes and was negatively correlated with TAPSE, suggesting a potential link between sphingolipid-related signaling and right ventricular systolic dysfunction in NICM (Wang et al., 2022). These findings support the hypothesis that elevated S1P may reflect adverse remodeling, inflammatory activation, or hemodynamic stress in NICM. However, given the incomplete availability of CMR-based fibrosis assessment and the absence of histological validation or validated circulating fibrosis biomarkers, elevated S1P should be interpreted cautiously as a circulating metabolic signal potentially associated with adverse remodeling or inflammatory–hemodynamic stress, rather than as a definitive marker of myocardial fibrosis. Future studies integrating serum S1P measurements with CMR-based tissue characterization, circulating fibrosis biomarkers, and experimental validation are warranted to clarify the pathophysiological role of S1P in NICM progression.

THE was also selected as a final prognostic metabolite biomarker in our analysis. As a downstream product of cortisol metabolism, THE may reflect activation of the neuroendocrine–chronic stress axis and participate in stress- and inflammation-related regulation (Qin et al., 2003). In patients with NICM, persistent hemodynamic stress and HF progression may activate the hypothalamic–pituitary–adrenal (HPA) axis, leading to increased cortisol secretion and altered downstream cortisol metabolism. Under conditions of myocardial injury and oxidative stress, increased reactive oxygen species production may further modify glucocorticoid–mineralocorticoid receptor signaling, thereby contributing to adverse cardiovascular remodeling, ventricular stiffness, and progressive cardiac dysfunction (Funder, 2005; Li et al., 2002; Nagata et al., 2006). In this context, elevated baseline serum THE levels in patients with adverse outcomes may reflect enhanced cortisol turnover and sustained neuroendocrine stress activation. Notably, THE was negatively correlated with TP, a marker of nutritional and synthetic reserve, suggesting that increased THE may also be linked to systemic catabolic stress and depletion of metabolic reserves. Together, these findings support the hypothesis that THE captures a neuroendocrine–catabolic stress signal associated with adverse cardiovascular outcomes in NICM, although this interpretation requires further validation.

The integrated prognostic model combining S1P, THE, TAPSE, and TP demonstrated satisfactory discriminative performance, with a bootstrap-corrected C-index of 0.772 and time-dependent AUCs of 0.92 at 3 years and 0.86 at 5 years. These findings support the potential value of integrating circulating metabolic biomarkers with conventional clinical indicators for risk stratification in NICM. The prognostic value of the integrated model may be explained by the complementary information provided by its four retained variables: TAPSE reflects cardiac mechanical function, TP indicates nutritional and synthetic reserve, S1P may capture remodeling- or inflammation-related metabolic stress, and THE may reflect neuroendocrine and catabolic stress.

Because the integrated model was developed in a cohort including both DCM and LVNC, the potential influence of diagnostic subtype should be considered. Although DCM and LVNC represent distinct phenotypic entities within the NICM spectrum, they may overlap clinically and morphologically, particularly in patients with reduced LVEF, as excessive left ventricular trabeculation can also be observed in some patients with DCM (Gregor et al., 2022; Hänselmann et al., 2020). In the present study, all patients had NICM with reduced LVEF after exclusion of major secondary cardiovascular and metabolic conditions, providing a shared high-risk clinical context for metabolomic prognostic assessment. In exploratory subgroup analyses, DCM and LVNC patients showed substantial overlap in global metabolomic profiles, and S1P and THE showed similar trends across the two subgroups. These findings suggest that the identified metabolic signatures were not solely driven by diagnostic subtype. Nevertheless, given the limited sample size, residual subtype-related heterogeneity cannot be fully excluded.

From a clinical translational perspective, our findings support the potential value of combining circulating metabolomic biomarkers with conventional clinical indicators for risk stratification in patients with NICM. Recent studies have shown that systemic inflammation-, nutrition-, and stress-related scores may provide complementary prognostic information in HF populations. For example, the modified Glasgow Prognostic Score (mGPS), which integrates inflammatory and nutritional markers, has been reported to predict long-term mortality in patients hospitalized for acute decompensated heart failure with reduced ejection fraction (HFrEF) (Tunca et al., 2025). The Scottish Inflammatory Prognostic Score (SIPS), based on neutrophil and albumin levels, and the Endothelial Activation and Stress Index (EASIX) have also been shown to predict in-hospital mortality in acute heart failure or acute decompensated HFrEF(Özlek et al., 2026; Taş et al., 2025). Together, these studies highlight the prognostic relevance of systemic biological processes beyond traditional cardiac functional indicators. In this context, our metabolomics-based model may capture broader circulating metabolic alterations related to neuroendocrine activation, inflammation, remodeling, and nutritional reserve. Moreover, given the variable or limited benefit of current anti-fibrotic strategies in advanced HF and the lack of established therapies directly targeting oxidative stress (Carlino et al., 2025; Graziani et al., 2021; Lewis et al., 2021), S1P- and THE-related metabolic signals may help refine prognostic assessment and provide clues for future mechanistic and translational studies in high-risk NICM.

Several limitations merit consideration. First, the relatively small sample size (n = 32), limited number of events (n = 16), and large number of metabolites identified through untargeted profiling may increase the risk of model overfitting and optimistic performance estimates. Although the integrated Cox model was internally validated by bootstrap resampling and evaluated using time-dependent ROC and calibration analyses, external validation in larger independent multicenter cohorts is required to confirm its generalizability. Second, the limited number of events restricted robust adjustment for longitudinal cardiac remodeling, NICM subtype, and detailed longitudinal medical treatment exposure, including guideline-directed medical therapy (GDMT) titration, medication dose changes, and treatment adherence; therefore, residual confounding related to disease progression and treatment exposure cannot be fully excluded. Third, excluding patients with diabetes mellitus and hypertension reduced metabolic confounding but may limit generalizability to broader HF populations, where these comorbidities are common and may modulate the S1P and THE metabolic axes. Fourth, genetic testing data, including pathogenic variants in cardiomyopathy-related genes such as TTN, LMNA, and DSP, were not available, and residual genetic confounding may remain. Finally, future targeted quantitative assays, external validation cohorts, and experimental models are needed to confirm the robustness and biological relevance of S1P and THE and to clarify their potential translational implications in NICM.

Conclusions

In this study, untargeted serum metabolomic profiling identified metabolic alterations associated with adverse cardiovascular outcomes in patients with NICM and reduced LVEF. S1P and THE were selected as prognostically informative circulating biomarkers, potentially reflecting adverse remodeling and neuroendocrine stress. Their integration with TAPSE and TP provided complementary value for risk stratification. These findings support the potential utility of combining serum metabolomic biomarkers with conventional clinical indicators for individualized prognostic assessment in NICM, thereby providing a rationale for further validation and mechanistic studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under the curve

BMI

Body mass index

CTA

Computed tomography angiography

DCM

Dilated cardiomyopathy

DDA

Data-dependent acquisition

DE analysis

Differential expression analysis

GLS

Global longitudinal strain

HF

Heart failure

HPA

Hypothalamic-pituitary-adrenal

HR

Hazard ratio

LC-MS

Liquid chromatography-mass spectrometry

LVEF

Left ventricular ejection fraction

LVMI

Left ventricular mass index

LVNC

Left ventricular non-compaction

NICM

Non-ischemic cardiomyopathy

OPLS-DA

Orthogonal partial least squares discriminant analysis

PCA

Principal component analysis

PMM

Predictive mean matching

QC

Quality control

RAAS

Renin-angiotensin-aldosterone system

ROC

Receiver operating characteristic

RWT

Relative wall thickness

S’

Lateral mitral annular systolic velocity

S1P

Sphingosine-1-phosphate

TAPSE

Tricuspid annular plane systolic excursion

THE

Tetrahydrocortisone

TP

Total protein

TTE

Transthoracic echocardiography

UPLC

Ultra-performance liquid chromatography

VIP

Variable importance in projection

Author contributions

Conceptualization, FW, BY, XL, and NS, data curation, FW and XL, formal analysis, FW, BY, NS, and XL, methodology, FW, BY, NS, and XL, software, BY, supervision, LF, WC, NS, XY, and XL, writing—original draft, FW, writing—review and editing, FW, BY, NS, and XL, visualization, FW and BY, project administration, LF, WC, and XY. All authors participated in revising the initial draft and approving the final version.

Funding

This research was supported by the CAMS Innovation Fund for Medical Sciences (CIFMS) (2022-I2M-1-023), the National High-Level Hospital Clinical Research Funding of China (2022-PUMCH-B-098), and the National Natural Science Foundation of China (62471273).

Data availability

All data generated or analysed during this study are included in this published article and its supplementary information files.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethical approval

The study protocol was approved by the Ethics Committee of Peking Union Medical College Hospital (Approval No. I-25PJ0792) and was conducted in accordance with the principles of the Declaration of Helsinki.

Consent to participate

Written informed consent was obtained from all participants prior to enrollment.

Consent for publication

Not applicable.