Metabolic signatures and machine learning identify gut-liver-heart axis dysfunction as a potential link to major adverse cardiovascular events in coronary artery disease

Abstract

Background

Traditional risk factors do not fully account for the residual cardiometabolic risk of major adverse cardiovascular events (MACE) in coronary artery disease (CAD). We aimed to identify circulating metabolic signatures associated with MACE susceptibility and uncover potential pathobiological mechanisms underlying the gut-liver-heart axis.

Methods

In this retrospective case-control study, untargeted high-performance liquid chromatography-mass spectrometry (HPLC-MS) was performed on fasting serum from 200 patients with CAD and MACE, 200 with CAD without MACE, and 400 matched non-CAD controls. Metabolomics data were processed using univariate analysis, multivariate analysis, and eXtreme Gradient Boosting (XGBoost) machine learning. Pathway enrichment was conducted using metabolite set enrichment analysis. Circulating fibroblast growth factor 19 (FGF19) was quantified via enzyme-linked immunosorbent assay to validate enterohepatic endocrine disruption.

Results

The MACE cohort exhibited a pronounced cardiometabolic phenotype, characterized by significantly highest rates of diabetes, hypertension, and dyslipidemia (p < 0.01). The XGBoost model robustly discriminated patients with CAD and MACE from non-CAD controls (area under the curve [AUC] = 0.984) and from patients with CAD without MACE (AUC = 0.932). Pathway analysis revealed marked dysregulation of linoleic acid metabolism and peroxisome proliferator-activated receptor (PPAR) signaling (p < 0.05). Specifically, pro-inflammatory oxidized linoleic acid metabolites, including 9- and 13-hydroxyoctadecadienoic acid (HODE)—which drive plaque instability—were significantly elevated in the MACE cohort. Furthermore, atheroprotective primary bile acids were significantly depleted in patients with CAD (p < 0.001). This depletion was accompanied by an elevated serum FGF19 level (p = 0.003), reflecting a potential disruption of the gut-liver-heart endocrine axis.

Conclusions

In conclusion, dysregulated linoleic acid oxidation, altered PPAR signaling, and disturbed primary bile acid-FGF19 metabolism may represent key metabolic pathways associated with MACE susceptibility. Integrating these gut-liver-heart axis signatures into machine learning models holds significant promise for refining cardiovascular risk stratification and guiding targeted preventive interventions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40842-026-00329-w.

Background

Coronary artery disease (CAD) remains a leading cause of global mortality [1]. Although established cardiovascular risk prediction algorithms offer moderate discrimination in the general population, they frequently fail to accurately identify individuals at the highest risk of adverse outcomes, particularly among patients with complex cardiometabolic comorbidities like diabetes and metabolic syndromes [2–4]. Furthermore, these algorithms rely heavily on traditional clinical variables (e.g., blood pressure, static lipid panels) that do not capture dynamic metabolic disturbances or novel pathobiological pathways. This reliance limits our understanding of the intricate biological interactions driving residual risk and hinders precise prognostic stratification.

CAD progression is driven not only by canonical lipid abnormalities but also by complex, systemic metabolic disturbances, including metabolic inflammation, oxidative stress, insulin resistance, and altered bile acid metabolism [5–7]. Consequently, atherosclerosis is increasingly recognized not merely as a localized vascular pathology, but as a systemic metabolic disorder governed by inter-organ communication. At the forefront of this systemic perspective is the gut-liver-heart axis, a dynamic network where circulating metabolic products act as endocrine messengers to drive cardiovascular pathogenesis [8, 9]. The liver synthesizes primary bile acids and regulates circulating lipids; these molecules are subsequently modified by the gut microbiota before entering the systemic circulation [10]. Once in circulation, gut- and liver-derived metabolites, including specific bile acids and fatty acids like linoleic acid, profoundly modulate vascular inflammation, endothelial function, and lipid homeostasis via peroxisome proliferator-activated receptor (PPAR) signaling [11]. Because routine clinical assays fail to capture these intricate metabolic shifts, vulnerable individuals may be misclassified as low or intermediate risk, obscuring key mechanistic pathways and limiting opportunities for targeted prevention.

Untargeted metabolomics provides a comprehensive landscape of circulating metabolites, enabling the discovery of novel biomarkers and the elucidation of disease mechanisms. The integration of metabolic signatures with advanced machine learning algorithms significantly improves prognostic accuracy for major adverse cardiovascular events (MACE) [12]. Despite this, comprehensive metabolic profiling of high-risk patients with CAD—particularly those exhibiting advanced cardiometabolic dysfunction who subsequently experience MACE—remains limited. It remains unclear whether specific metabolic pathways can reliably distinguish patients with CAD destined for MACE from those who remain event-free.

In this study, we utilized untargeted metabolomics coupled with XGBoost machine learning to characterize the circulating metabolic profile of healthy controls and patients with CAD, stratified by the occurrence of MACE. By mapping these distinct signatures, we aimed to identify culprit metabolic pathways, specifically within the gut-liver-heart axis, to provide novel prognostic biomarkers and mechanistic insights into CAD progression and cardiometabolic vulnerability.

Methods

Study design and study population

This retrospective case-control study included 200 patients with CAD who experienced MACE (Group A), 200 patients with CAD without MACE (Group B), and 400 controls without CAD (Group C). The study groups were matched for age and sex. Participants were recruited from three established cohorts: the Hong Kong Chinese CAD Cohort (comprising patients with angiographically confirmed CAD), the Hong Kong Cardiovascular Risk Factors Prevalence Study (CRISPS), and the Hong Kong West Diabetes Registry. This multi-cohort selection strategy ensured a robust representation of diverse cardiometabolic phenotypes. Detailed descriptions of these cohorts have been published previously [13].

The study protocol was approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster. Written informed consent was obtained from all participants at enrollment. For patients who experienced MACE, fasting blood samples were collected after the occurrence of the event, following an overnight fast of at least 8 hours. Serum was separated and stored at −80 °C until metabolomic analysis. This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines [14].

Untargeted metabolomics

Untargeted metabolomics was performed using high-performance liquid chromatography-mass spectrometry (HPLC-MS), adhering to established protocols for robust large-scale metabolic profiling [15]. Detailed sample preparation procedures have been described previously [16]. Briefly, 10 μL of serum was mixed with 240 μL extraction buffer (methanol:acetonitrile:water) for 30 minutes at 4 °C, followed by centrifugation at 12,000×g for 10 minutes at 4 °C. The supernatant was injected into a Vanquish HPLC system (Thermo Fisher Scientific, Germering, Germany) coupled to a Q-Exactive Plus mass spectrometer (Thermo Fisher Scientific, Bremen, Germany).

The mobile phases comprised 0.1% formic acid in water (phase A) and 0.1% formic acid in acetonitrile (phase B). A constant flow rate of 0.05 mL/min was maintained. Auto-sampler and column oven temperatures were kept at 4 °C and 25°C, respectively. Quality control (QC) samples, generated by pooling study serum, were analyzed every 10 sample injections to monitor signal stability. Analytes were ionized via electrospray ionization (spray voltage: 4 kV; capillary temperature: 325°C) in both positive and negative ionization modes to acquire full-scan mass-to-charge (m/z) data. Method setup and raw data acquisition were performed using Xcalibur version 4.1.50 software (Thermo Fisher Scientific).

Metabolomics data processing and pathway enrichment analysis

Raw HPLC-MS data were processed using the mProbe Cloud Analytical Platform. The computational workflow encompassed m/z peak extraction, peak identification, retention time correction, and peak alignment to generate a structured feature matrix. QC was assessed using pooled QC samples injected at regular intervals throughout the analytical sequence to monitor retention time, mass accuracy, signal stability, and feature reproducibility. Features demonstrating acceptable reproducibility in the QC samples, defined as a relative standard deviation of less than 30%, were retained for downstream analysis. Signal intensities underwent drift correction and were normalized to the median intensity of the pooled QC samples. Features with a missing value ratio of greater than 20% were excluded from further analysis. Metabolites were initially annotated based on their m/z values (mass tolerance: ±5 ppm) using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. For metabolite hits within key enriched pathways, targeted tandem mass spectrometry (MS/MS) fragmentation was performed to resolve specific chemical structures. Targeted structural annotation was achieved by matching m/z values and MS/MS spectra against the Human Metabolome Database (HMDB). This corresponds to level 2 confidence in metabolite identification according to the Metabolomics Standards Initiative (MSI) guidelines [17].

Pathway enrichment was conducted via metabolite set enrichment analysis. In each pathway, the total number of metabolites, observed hits, expected hits (under a null model), enrichment p-value, and enrichment ratio (observed hits divided by expected hits) were calculated.

Clinical parameters and fibroblast growth factor 19 assay

Routine clinical and biochemical parameters were measured by the Department of Biochemistry at Queen Mary Hospital. Serum level of fibroblast growth factor 19 (FGF19) was quantified using an in-house enzyme-linked immunosorbent assay (ELISA) developed by the Antibody and Immunoassay Services at the University of Hong Kong. The intra- and inter-assay coefficients of variation were 4.5% and 5.6%, respectively [18].

Statistical analysis and machine learning

For univariate analyses, group-level comparisons of individual metabolite features were performed by calculating fold changes, p values (using the Mann-Whitney U test), odds ratios, false discovery rate (FDR)-adjusted p values, and areas under the receiver operating characteristic curve (AUCs). Differentially abundant features were considered statistically significant if they met all of the following criteria: (1) a fold change of less than 0.83 or greater than 1.20; (2) an FDR-adjusted p value of less than 0.05; (3) an AUC of greater than 0.60; and (4) a missing value ratio of less than 20%. Unsupervised multivariate analyses, including principal component analysis (PCA) and hierarchical clustering, were used to visualize sample variances and global profiling similarities.

For supervised classification, XGBoost machine learning models were constructed to distinguish cases from controls based on key metabolite features. To prevent data leakage and overfitting, feature selection was performed independently within each training fold using feature importance scores, without accessing the corresponding test set. Model hyperparameters were optimized using a randomized search across more than 20 combinations, evaluating boosting rounds from 5 to 20 and maximum tree depths from 2 to 5. Model performance was evaluated using 5-fold cross-validation. Predicted probabilities from all five test folds were concatenated to construct a combined receiver operating characteristic curve and calculate the overall AUC. No data normalization or missing value imputation was performed.

Continuous variables with normal distributions are expressed as means with standard deviations and compared using analysis of variance (ANOVA) or Welch’s ANOVA. Skewed variables (e.g., serum FGF19) were logarithmically transformed to achieve normality prior to parametric testing. Categorical variables are presented as frequencies and percentages and compared using the chi-square test. Statistical analyses were performed using R (version 4.4.1; R Foundation for Statistical Computing), SPSS (version 28.0; IBM Corp), and GraphPad Prism (version 9.5.0; GraphPad Software Inc). All statistical tests were 2-sided, and a p value of less than 0.05 was considered statistically significant.

Results

Characteristics of the study population

The baseline clinical characteristics of the cohorts are summarized in Table 1 (workflow illustrated in Fig. 1). The mean age of the overall population was 62.0 years, and 623 participants (77.9%) were male. We observed significant differences in the prevalence of hypertension, diabetes, and dyslipidemia across the study groups (all p < 0.01). Post-hoc pairwise comparisons revealed that CAD patients with MACE had a significantly higher prevalence of hypertension and diabetes than those without MACE (all p < 0.01). Furthermore, body mass index, waist circumference, and waist-to-hip ratio were highest among CAD patients with MACE (all p < 0.001). Conversely, circulating levels of total cholesterol, high-density lipoprotein cholesterol, and low-density lipoprotein cholesterol were significantly higher in the non-CAD control group than in both CAD groups (all p < 0.001). This difference is likely attributable to the lower prevalence of dyslipidemia and the consequently lower rate of lipid-lowering therapy prescriptions in the control cohort. Notably, these standard lipid parameters did not differ significantly between CAD patients with versus without MACE (p > 0.05), underscoring the limitations of traditional lipid profiles in capturing residual cardiovascular risk.

Table 1.

Baseline characteristics of the study population

	CAD with MACE (N = 200)	CAD without MACE (N = 200)	Non-CAD (N = 400)	p-value
Age, y	62.3 (9.1)	62.5 (8.5)	61.6 (8.8)	0.39‡
Male, n (%)	158 (79.0)	162 (81.0)	303 (75.8)	0.31†
Smoker, n (%)	86 (43.0)	95 (47.5)	154 (38.5)	0.54†
Diabetes, n (%)	127 (63.5)	115 (57.5)	200 (50.0)	0.006†
Hypertension, n (%)	161 (80.5)	155 (77.5)	237 (59.3)	<0.001†
Dyslipidemia, n (%)	191 (95.5)	191 (95.5)	225 (56.3)	<0.001†
SBP, mmHg	137 (20.0)	135 (17.8)	130 (17.4)	<0.001§
DBP, mmHg	75 (10.7)	76 (10.9)	76 (9.1)	0.62§
Body mass index, kg/m2	25.9 (4.0)	25.5 (3.6)	24.4 (3.6)	<0.001‡
Waist circumference, cm	91.1 (9.5)	90.5 (9.3)	86.6 (10.0)	<0.001‡
Hip circumference, cm	94.8 (6.6)	95.0 (7.2)	93.6 (6.7)	0.047‡
Waist-to-hip ratio	0.96 (0.07)	0.95 (0.06)	0.93 (0.08)	<0.001§
Fasting glucose, mmol/L	6.8 (2.1)	6.5 (1.6)	6.3 (2.0)	0.023§
Triglyceride, mmol/L	1.4 (0.7)	1.4 (0.8)	1.3 (0.7)	0.39‡
Total cholesterol, mmol/L	3.9 (0.9)	3.9 (0.8)	4.7 (1.0)	<0.001§
HDL-cholesterol, mmol/L	1.2 (0.3)	1.2 (0.3)	1.3 (0.4)	<0.001§
LDL-cholesterol, mmol/L	2.1 (0.8)	2.1 (0.6)	2.8 (0.8)	<0.001§
Ln_FGF19, pg/mL	4.9 (1.3)	4.9 (1.1)	4.5 (1.2)	<0.001‡

†: chi-square test; ‡: analysis of variance (ANOVA); §: Welch’s ANOVA

Data are presented as mean (standard deviation) for continuous variables and frequency (percentage) for categorical variables. CAD: coronary artery disease; DBP: diastolic blood pressure; HDL-cholesterol: high-density lipoprotein cholesterol; LDL-cholesterol: low-density lipoprotein cholesterol; Ln_FGF19: natural logarithm of fibroblast growth factor 19; MACE: major adverse cardiovascular event; SBP: systolic blood pressure

Fig. 1.

Workflow of the study. CAD: coronary artery disease; HPLC-MS: high-performance liquid chromatography-mass spectroscopy; MACE: major adverse cardiovascular event; XGBoost: eXtreme Gradient Boosting

Metabolomic profiling and pathway enrichment

Untargeted HPLC-MS initially detected 4,333 metabolite features across both ionization modes. After filtering, 2,466 features (56.9%) were retained, yielding 1,295 annotated metabolite features for pathway enrichment. These were categorized into 19 chemical superclasses, predominantly organic acids (19.4%) and fatty acyls (18.2%) (Table 2).

Table 2.

Chemical superclasses of annotated metabolites

Chemical superclasses	Proportion (%)
Organic acids	19.4
Fatty Acyls	18.2
Benzenoids	9.23
Organoheterocyclic compounds	8.02
Alkaloids	7.87
Prenol Lipids	7.72
Nucleic acids	5.75
Carbohydrates	5.45
Sterol Lipids	4.39
Organic nitrogen compounds	3.33
Organic oxygen compounds	3.33
Polyketides	2.87
Organohalogen compounds	1.36
Sphingolipids	1.36
Organosulfur compounds	0.61
Glycerolipids	0.45
Lignans	0.45
Inorganic compounds	0.15
Organometallic compounds	0.15

In comparing patients with CAD and MACE (Group A) with non-CAD controls (Group C), the top 20 significantly differential metabolite features, ranked by absolute fold change, were all upregulated (Supplementary Table S1). PCA scatterplots showed substantial overlap between the groups, with the first principal component explaining 14.6% of the total variance, indicating that the two groups were well matched (Supplementary Figure S1a). Within this internal dataset of 600 samples, an XGBoost machine learning model identified the 10 most important features, demonstrating excellent performance in distinguishing patients with CAD and MACE from non-CAD controls (AUC, 0.984; Fig. 2a, Supplementary Table S2).

Fig. 2.

Metabolic signatures and machine learning results comparing patients with CAD and MACE with non-CAD controls (n = 600). (a) The eXtreme Gradient Boosting machine learning model demonstrates discriminative performance with an area under the curve of 0.984, based on combined 5-fold cross-validation. (b) The bar chart of pathway enrichment analysis illustrates enrichment ratio and raw p-value of each pathway, using metabolite set enrichment analysis. (c) The bubble plot of pathway enrichment analysis illustrates total number of hits and raw p-value of each pathway, using metabolite set enrichment analysis. ABC: ATP-binding cassette; CAD: coronary artery disease; MACE: major adverse cardiovascular event; PPAR: peroxisome proliferator-activated receptor

Pathway enrichment analysis revealed significant alterations in linoleic acid metabolism (14 hits; enrichment ratio = 54.8; p < 0.001) and the PPAR signaling pathway (2 hits; enrichment ratio = 245.4; p = 0.029) (Fig. 2b–c). Metabolites mapping to these significantly enriched pathways were putatively annotated with level 2 confidence using targeted MS/MS fragmentation. Elevated levels of 9-hydroxyoctadecadienoic acid (9-HODE) and 13-HODE in the CAD with MACE group were identified in both pathways (Supplementary Table S3). Although overall enrichment in the primary bile acid biosynthesis pathway did not reach statistical significance (p > 0.05), its high enrichment ratio (4.2) suggested potential pathobiological involvement in MACE. Specifically, three primary bile acid metabolites were significantly depleted in the CAD with MACE group (p < 0.001 for all).

When comparing patients with CAD and MACE (Group A) to those without MACE (Group B), 11 features remained significantly upregulated, consistent with the findings from the comparison between Group A and Group C (Supplementary Table S4). Using data from these 400 patients, the XGBoost model selected the 10 most important features and maintained robust predictive capacity for CAD with MACE (AUC, 0.932; Fig. 3a, Supplementary Table S5). Pairwise PCA also demonstrated high concordance between the groups, with the first principal component accounting for 15.9% of the total variance (Supplementary Figure S1b). Linoleic acid metabolism and the PPAR signaling pathway, along with the annotated metabolites mapped to these pathways, exhibited enrichment patterns consistent with those observed in the comparison of Group A vs Group C (Fig. 3b–c, Supplementary Table S6).

Fig. 3.

Metabolic signatures and machine learning results comparing patients with CAD and MACE with those with CAD without MACE (n = 400). (a) The eXtreme Gradient Boosting machine learning model demonstrates discriminative performance with an area under the curve of 0.932, based on combined 5-fold cross-validation. (b) The bar chart of pathway enrichment analysis illustrates enrichment ratio and raw p-value of each pathway, using metabolite set enrichment analysis. (c) The bubble plot of pathway enrichment analysis illustrates total number of hits and raw p-value of each pathway, using metabolite set enrichment analysis. CAD: coronary artery disease; MACE: major adverse cardiovascular event; PPAR: peroxisome proliferator-activated receptor

In a pooled analysis comparing all patients with CAD (Groups A and B) to non-CAD controls (Group C), the 20 most differential metabolite features were identified, of which 8 were among the key features characterizing patients with CAD and MACE (Supplementary Table S7). The first five principal components explained approximately 37.3% of the total variance, with PCA showing no distinct group-wise separation (Supplementary Figure S1c). The XGBoost model trained on the combined sample of 800 individuals identified the 10 most important features and achieved strong discriminative performance for overall CAD (AUC, 0.944; Fig. 4a, Supplementary Table S8). Within the overall CAD cohort, linoleic acid metabolism and PPAR signaling remained significantly enriched (p = 0.003 and p = 0.004, respectively) (Fig. 4b–c). Although fewer metabolites were putatively annotated in this pooled comparison, 9-HODE and 13-HODE were retained in both pathways (Supplementary Table S9).

Fig. 4.

Metabolic signatures and machine learning results comparing all patients with CAD with non-CAD controls (n = 800). (a) The eXtreme Gradient Boosting machine learning model demonstrates discriminative performance with an area under the curve of 0.944, based on combined 5-fold cross-validation. (b) The bar chart of pathway enrichment analysis illustrates enrichment ratio and raw p-value of each pathway, using metabolite set enrichment analysis. (c) The bubble plot of pathway enrichment analysis illustrates total number of hits and raw p-value of each pathway, using metabolite set enrichment analysis. CAD: coronary artery disease; PPAR: peroxisome proliferator-activated receptor

Serum FGF19 level

To further investigate the disruption of the enterohepatic endocrine axis indicated by the metabolomic analyses, we evaluated circulating FGF19 levels. Circulating FGF19 was significantly elevated in patients with CAD and MACE compared with non-CAD controls (median, 134.3 pg/mL vs 90.0 pg/mL; p = 0.003). However, FGF19 levels did not differ significantly between patients with CAD and MACE and those with CAD without MACE (median, 134.3 pg/mL vs 121.5 pg/mL; p = 0.99) (Table 1, Fig. 5). In receiver operating characteristic analysis, circulating FGF19 demonstrated modest diagnostic performance for distinguishing patients with CAD from controls (AUC, 0.597; Supplementary Figure S2).

Fig. 5.

Naturally log-transformed circulating fibroblast growth factor 19 (Ln_FGF19) in the study cohort (n = 800). The study cohort is stratified by patients with CAD and MACE (red), CAD without MACE (blue), and non-CAD controls (green). Mean with standard deviation is plotted. Analysis of variance and Tukey’s multiple comparisons are used to compare Ln_ FGF19 between groups. **: p-value < 0.01. CAD: coronary artery disease; MACE: major adverse cardiovascular event

Discussion

Using robust machine learning, this study provides a comprehensive characterization of the circulating metabolome in CAD patients stratified by their vulnerability to MACE. Our findings reveal that dysregulated linoleic acid metabolism and altered PPAR signaling may represent the dominant metabolic signatures associated with high-risk cardiometabolic profiles predisposed to MACE. Additionally, the depletion of primary bile acids concurrent with elevated FGF19 levels suggests a profound disruption in enterohepatic signaling. Collectively, these findings emphasize that residual cardiovascular risk is potentially related to systemic cardiometabolic and gut-liver-heart axis dysfunction.

Across all comparative analyses, linoleic acid metabolism demonstrated a robust association with CAD progression and MACE occurrence. Linoleic acid, an essential omega-6 polyunsaturated fatty acid, plays a dual role in cardiovascular homeostasis [19]. While historically viewed as atheroprotective [20], linoleic acid is highly susceptible to lipid peroxidation. The resulting oxidized linoleic acid metabolites (OXLAMs) actively promote atherogenesis by inducing endothelial cell activation, upregulating adhesion molecules, and driving macrophage foam cell formation [19, 21]. Crucially, our targeted profiling revealed that individuals who developed MACE exhibited marked accumulation of OXLAMs, including HODE, dihydroxyoctadecadienoic acid (DiHODE), hydroperoxyoctadecadienoic acid (HPODE), and epoxyoctadecenoic acid (EpOME). In the context of atherogenesis, 9-HODE drives plaque inflammation and instability; 13-HODE exacerbates advanced atherosclerosis by amplifying macrophage inflammatory responses [22].

Intertwined with these lipid disturbances is the PPAR signaling pathway, which emerged as a defining signature of MACE vulnerability. PPARs are ligand-activated transcription factors that govern the intersection of systemic metabolism, insulin sensitivity, and vascular inflammation [23, 24]. Clinically, the atheroprotective potential of targeting this pathway is highly relevant to cardiometabolic diseases; for example, the PROactive trial demonstrated that the PPARγ agonist pioglitazone significantly reduced cardiovascular events in high-risk diabetic cohorts [25]. In our study, 9- and 13-HODE, both endogenous ligands for PPARγ, were significantly elevated in the MACE cohort. We hypothesize that the localized accumulation of 9- and 13-HODE acts as a critical messenger between dysregulated lipid metabolism and vascular inflammation. The overactivation of this pathway in advanced atherosclerosis may override normal anti-inflammatory feedback mechanisms, driving maladaptive foam cell formation. Tracking circulating biomarkers linked to PPARγ activity may therefore refine MACE prediction and help identify patients who would benefit most from insulin-sensitizing or PPAR-targeted pharmacotherapies.

Beyond lipid peroxidation, the observed depletion of primary bile acids and concurrent elevation of FGF19 provide suggestive evidence that CAD is a systemic metabolic disorder governed by the gut-liver-heart axis [8, 9, 26]. The significant elevation of FGF19 in our CAD cohorts is highly consistent with our previous prospective observation that elevated FGF19 was associated with incident MACE in CAD [27], as well as external reports demonstrating the predictive value of FGF19 for subclinical atherosclerosis in men with diabetes [28]. In our cohorts, FGF19 levels were higher in individuals with CAD compared to non-CAD controls, suggesting an association with CAD presence. However, there was no significant difference in FGF19 levels between CAD patients with MACE and those without MACE, indicating that FGF19 may be more closely related to CAD status rather than serving as a specific predictor of MACE risk in this dataset.

The concurrent elevation of FGF19 and depletion of primary bile acids likely represent a linked pathobiological cascade in the gut-liver-heart axis. The increase in FGF19 may occur as a result of enhanced expression of the Farnesoid X receptor (FXR), which has been shown to act as a compensatory mechanism in response to linoleic acid-induced inflammation [29]. FXR activation subsequently reduces hepatic bile acid synthesis, either directly or indirectly increasing the expression of FGF19, a known FXR target enterokine [30]. Once in circulation, FGF19 impacts lipid metabolism via dual mechanisms: it increases serum cholesterol and triglycerides through fibroblast growth factor receptor (FGFR) 4 signaling, while reducing them via FGFR1-induced effects on energy metabolism [31]. The pro-atherogenic potential of the FGFR4-mediated pathway is highly relevant in clinical settings. Notably, in human clinical trials, administration of FGF19 analogs for steatohepatitis treatment was associated with raised low-density lipoprotein cholesterol levels and reduced bile acid synthesis, raising concerns about potential cardiovascular risks [32].

The linoleic acid-driven upregulation of FXR and subsequent overproduction of FGF19 may contribute to a pro-atherogenic lipid profile, which could influence overall CAD pathology and MACE risk. To better understand these relationships, correlation analyses between FGF19 levels and specific bile acid species would be valuable, and should be pursued in future studies.

Limitations

Our study has several limitations. First, the retrospective case-control design precludes causal inference and introduces imbalances in baseline characteristics, including differences in the types and doses of various medications, such as lipid-lowering therapies. While multivariable adjustment for these imbalanced baseline characteristics could strengthen the results, we did not perform this analysis due to the lack of detailed medication data—particularly in the “healthy” non-CAD cohort—and because of the limited sample size relative to the complexity of metabolomics, which could impact statistical power. Second, in our predictive model using XGBoost, we employed the full metabolomics dataset as input. Future studies could explore the trade-off between stepwise reduction of input variables and the impact on predictive accuracy. However, this was beyond the scope of the current study. Third, the absence of an external validation dataset limits the evaluation of the XGBoost model’s performance. Forth, we did not validate the levels of downstream metabolites (e.g., linoleic acid and bile acids) identified in the pathway analysis by targeted metabolomics. Future external or prospective validation, including assessment of both the machine learning model and metabolite levels, may further support our findings. Finally, although we implicate the gut-liver-heart axis, the lack of matched fecal microbiome sequencing prevents us from linking circulating bile acid shifts to specific microbial disturbances. Future prospective cohorts and mechanistic studies are needed to validate these metabolite panels and clarify their role in vascular pathobiology.

Conclusions

By integrating untargeted metabolomics with XGBoost machine learning, we demonstrate that dysregulated linoleic acid oxidation and altered PPAR signaling are potential metabolic pathways associated with MACE susceptibility in patients with CAD. Concurrently, the possible link on depletion of primary bile acids and the elevation of FGF19 underscore the critical role of gut-liver-heart axis dysfunction in atherogenesis. Translating these distinct metabolic and endocrine signatures into clinical algorithms may offer a promising strategy to capture residual cardiovascular risk, refine risk stratification, and guide targeted therapeutic interventions for high-risk patients.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

ANOVA

Analysis of variance

AUC

Area under the receiver operating characteristic curve

CAD

Coronary artery disease

DiHODE

Dihydroxyoctadecadienoic acid

EpOME

Epoxyoctadecenoic acid

FGF19

Fibroblast growth factor 19

FGFR

Fibroblast growth factor receptor

FXR

Farnesoid X receptor

HODE

Hydroxyoctadecadienoic acid

HPLC-MS

High-performance liquid chromatography-mass spectrometry

HPODE

Hydroperoxyoctadecadienoic acid

KEGG

Kyoto Encyclopedia of Genes and Genomes

MACE

Major adverse cardiovascular event

m/z

Mass-to-charge ratio

OXLAM

Oxidized linoleic acid metabolite

PCA

Principal component analysis

PPAR

Peroxisome proliferator-activated receptor

QC

Quality control

XGBoost

EXtreme Gradient Boosting

Author contributions

AX and HFT conceived the study design. MQL, CKW, KWA, CYYC, YML, SYL, KSLL, AX, and HFT acquired data. MQL, CKW, and HFT performed data analysis and interpretation. MQL and CKW wrote the first draft of the manuscript. MQL, CKW, and HFT revised the manuscript. All authors read the final manuscript and authorized its submission.

Funding

This work was supported in part by the InnoHK initiative of the Innovation and Technology Commission of the Hong Kong Special Administrative Region Government (H.F.T); Hong Kong the Impact Research Fund (R7036-21), Lee Wan Keung Charity Foundation Fund; National Key Research and Development Program of China (2022YFA0806102), and Hong Kong Research Grants Council/Area of Excellence (AoE/M/707-18).

Data availability

All data supporting the findings of this study are available within the paper and its Supplementary Information.

Declarations

Ethics approval and consent to participate

The study was approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster. Written informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.