Machine learning-based plasma metabolomics for improved cirrhosis risk stratification

Abstract

Background

Cirrhosis is a leading cause of mortality in patients with chronic liver disease (CLD). The rapid development of metabolomic technologies has enabled the capture of metabolic changes related to the progression of cirrhosis.

Methods

This study used proton nuclear magnetic resonance (1 H-NMR) serum metabolomics data from the UK Biobank (UKB) and employed elastic net-regularized Cox proportional hazards models to explore the role of metabolomics in cirrhosis risk stratification in patients with CLD. Metabolomic data were integrated with aspartate aminotransferase to platelet ratio index (APRI) and fibrosis-4 score (FIB-4) to construct predictive models for cirrhosis risk. The model performance was assessed in both the derivation and validation cohorts.

Results

A total of 2,738 eligible patients were included in the analysis. Several metabolites showed an independent association with cirrhosis events (68 out of 168 metabolites after adjustment for age and sex, and 21 out of 168 metabolites after full adjustment). The integration of metabolomics with FIB-4 improved the predictive performance compared to FIB-4 alone (Harrell’s C: 0.717 vs. 0.696, ΔC = 0.021, 95% confidence interval [CI] 0.014–0.028, Net Reclassification Improvement [NRI]: 0.504 [0.488–0.520]). Similarly, the combination of metabolomics with APRI also improved predictive performance compared to APRI alone (Harrell’s C: 0.747 vs. 0.718, ΔC = 0.029, 95% CI 0.022–0.035, NRI: 0.378 [0.366–0.389]). Key metabolites, including branched-chain amino acids (BCAAs), lipids, and markers of oxidative stress, were identified as significant predictors. Pathway enrichment analysis revealed that disruptions in lipid and amino acid metabolism play a central role in the progression of cirrhosis.

Conclusion

1 H-NMR serum metabolomics significantly improves the prediction of cirrhosis risk in patients with CLD. The APRI + Metabolomics model demonstrated strong discriminatory power, with key metabolites involved in fatty acid and amino acid metabolism, providing a promising tool for the early screening of cirrhosis risk.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-03655-y.

Key points

• Integrating metabolomics enhances predictive performance.

• The APRI + Metabolomics model demonstrates strong predictive capability.

• Machine learning effectively selects key metabolites associated with lipid and amino acid metabolism disorders that predict cirrhosis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-03655-y.

Introduction

Cirrhosis is the 11th most common cause of death worldwide, with more than 1 million deaths annually [1]. The primary causes of cirrhosis are chronic liver diseases (CLD), including viral infections, alcoholic liver disease (ALD), metabolic-associated fatty liver disease (MAFLD, also known as non-alcoholic fatty liver disease [NAFLD]), autoimmune liver diseases, chronic cholestasis, and drug-related or toxic liver injuries. The progression of CLD follows the course from liver fibrosis to cirrhosis. While most patients with cirrhosis have a single underlying cause, a minority have multiple contributing factors [2]. CLD often culminates in cirrhosis at the final stage. CLD progresses to irreversible cirrhosis through metabolic dysregulation, leading to fat accumulation, inflammation, and fibrosis, which gradually damages the structure and function [1]. The progression of cirrhosis takes several years, and owing to the heterogeneity among patients with CLD, predicting the advancement of cirrhosis and its complications remains challenging [3].

Current research primarily focuses on the diagnosis of cirrhosis, with few studies addressing the risk stratification for cirrhosis in patients with CLD. Previous studies have explored risk scores for predicting cirrhosis in such patients, but these have often been limited by the type of liver disease studied [4, 5] or by small sample sizes [6]. Although liver biopsy remains the gold standard for diagnosing and staging liver fibrosis, its invasive nature and associated risks limit its suitability for the longitudinal monitoring of fibrosis progression or assessment of treatment response [7]. Serum markers such as the aspartate aminotransferase-to-platelet ratio index (APRI) and fibrosis-4 (FIB-4) index are commonly used to predict cirrhosis, but their sensitivity and specificity remain suboptimal [8]. Managing CLD is challenging and more reliable methods are needed to assess the risk of cirrhosis progression in these patients. Early identification of the risk of progression to cirrhosis in patients with CLD allows for increased monitoring frequency, implementation of preventive measures, and reduction in treatment burden.

The liver is a central regulator of metabolism. Increases in free fatty acids, hyperglycemia, lipotoxicity, and significant alterations in protein synthesis following cell activation can disrupt the liver structure, promoting the development of liver fibrosis and, eventually, cirrhosis [9]. Persistent metabolic dysfunction in the liver can lead to chronic mitochondrial impairment and CLD, ultimately progressing to end-stage liver disease [10].

The rapidly advancing field of metabolomics has become a powerful tool in clinical research, facilitating the identification of biomarkers, phenotyping, disease staging, and uncovering the underlying mechanisms [11]. Metabolites are known to be associated with the progression of CLD to cirrhosis [9]. Proton nuclear magnetic resonance (1 H-NMR) spectroscopy-based metabolomics of serum samples is a quantitative method for studying multi-parameter metabolic changes and responses, and has been widely applied in liver disease research [12, 13]. However, the potential of serum metabolomics to predict the progression of CLD to cirrhosis has yet to be systematically evaluated and benchmarked.

The UK Biobank (UKB) is a large prospective cohort study that recruited more than 500,000 participants between 2006 and 2010 [14]. The UKB participants underwent comprehensive phenotypic characterization and their health records were subsequently recorded. Large-scale metabolomic profiling was conducted on approximately 120,000 baseline serum samples using 1 H-NMR, covering 168 individual metabolites. In this study, we utilized the resources of the UKB and integrated metabolomic data with machine learning to enhance risk stratification for cirrhosis in CLD.

Materials and methods

Study design

The participants underwent extensive baseline assessments, which included the collection of clinical information and biological samples, with regular updates and follow-up [15]. Blood, urine, and saliva samples were collected for analysis at baseline.

Metabolic biomarkers were obtained from 118,019 baseline EDTA non-fasting venous serum samples using the high-throughput nuclear magnetic resonance (NMR) metabolomics platform developed by Nightingale Health Ltd., between June 2019 and April 2020. Detailed information can be found in UKB research documentation (https://biobank.ctsu.ox.ac.uk/ukb/ukb/docs/nmrm_companion_doc). After controlling for quality and batch effects, 249 metabolic biomarkers were available (168 original measurements and 81% ratios).

We selected 168 primary metabolites based on their direct concentrations and biological relevance, as these biomarkers are widely recognized for their utility in predicting disease risk and their strong association with clinical outcomes [16]. The remaining 81 indicators, including metabolite ratios, percentages of individual metabolites within their total category, and measures of unsaturation, were excluded to maintain focus on the absolute concentrations of specific metabolites [17].

We focused on 168 measurements representing the concentrations of various metabolites that were categorized into 17 groups. These included: Amino Acids (n = 10), Apolipoproteins (n = 2), Cholesterol (n = 21), Cholesteryl Esters (n = 18), Fatty Acids (n = 9), Fluid Balance (n = 2), Free Cholesterol (n = 18), Glycolysis-Related Metabolites (n = 4), Inflammation (n = 1), Ketone Bodies (n = 4), Lipoprotein Particle Concentrations (n = 4), Lipoprotein Particle Sizes (n = 3), Lipoprotein Particles (n = 14), Other Lipids (n = 4), Phospholipids (n = 18), Total Lipids (n = 18), and Triglycerides (n = 18).

In this study, we included complete data from all UKB participants who had 168 original serum metabolite measurements at their initial assessment center visit. We further excluded individuals with incomplete parameters, such as missing data on age or liver enzymes, those with a baseline diagnosis of cirrhosis, and those with significantly abnormal metabolomic measurements (defined as values exceeding 5 standard deviations from the mean). Individuals with missing data on key parameters, such as age and liver enzymes, were excluded to ensure the integrity and accuracy of the analysis. Only complete datasets were used in the final model to avoid potential biases and ensure robust, reliable results. To account for the influence of lipid-lowering medications on the metabolomic profiles, participants taking these medications were also excluded. This left 2,738 eligible patients with CLD for analysis, allowing us to explore the association between serum metabolomics and the progression of CLD. The cohort was then divided into derivation (80%) and validation (20%) cohorts. This approach was chosen to preserve the heterogeneity within the sample, ensuring the model’s generalizability to a broader population. Given the relatively limited sample size, stratified sampling could have led to insufficient sample sizes in certain subgroups, potentially compromising the model’s stability. Furthermore, random sampling helps simulate a more diverse, real-world population while minimizing potential biases that could arise from an overly complex stratification process. We used the derivation subset to train the Elastic Net (EN) models to predict cirrhosis risk, which was subsequently validated in the validation subset (Fig. 1).

Fig. 1.

Flow diagram

Definition of CLD and cirrhosis

We defined the starting point as CLD and the endpoint as cirrhosis [15], both determined through clinical diagnoses obtained from electronic hospital health records or in cases of death or surgery related to the disease. Individuals diagnosed with liver fibrosis or cirrhosis at the baseline were excluded from the study. Supplementary Table 1 provides detailed disease codes and definitions. Detailed definitions of the diseases and predictors used in this study can be found in Supplementary Table 2.

Cirrhosis risk models and predictor extraction

We extracted cirrhosis-related predictors from the UKB dataset [15]. Supplementary Table 2 provides detailed descriptions of all relevant data fields, diseases, and associated information. Independent predictors of liver fibrosis severity include age, male sex, obesity, hypertension, diabetes, elevated alanine aminotransferase (ALT), elevated aspartate aminotransferase (AST), reduced platelet count (PLT), decreased albumin (Alb), and the presence of fatty liver and hepatic steatosis on ultrasound [18].

Upon enrollment, we first tested the association between each metabolite and cirrhosis events and identified a series of significantly correlated metabolites. For risk prediction, we incorporated various factors into the risk assessments, including sociodemographic factors (age, sex), patient history (smoking, alcohol consumption, sleep patterns), physical measurements (body mass index [BMI], systolic blood pressure, blood pressure, blood glucose), and clinical chemistry markers (liver enzymes, albumin, and additional relevant metrics). Potential confounders, including alcohol consumption, medication use, and dietary habits, were accounted for in the analysis framework. Alcohol consumption was adjusted using self-reported data on drinking frequency and intensity from the UK Biobank questionnaire. Medication use was addressed by excluding individuals taking lipid-lowering drugs with known effects on metabolic profiles. Although detailed dietary information was not available, the metabolites included in the analysis indirectly reflect nutritional status and dietary patterns. Systolic blood pressure was recorded twice, and a lower reading was used. If there was an error in the automated measurements, manual readings were recorded.

We utilized a total of five models, with the base model being the “metabolomics " model, which focused on the specificity of metabolomics. To predict cirrhosis risk, we employed the FIB-4 and APRI models, both of which are widely used for assessing the risk of cirrhosis development in patients with CLD. The FIB-4 model incorporates four parameters: age, PLT, ALT, and AST levels. The APRI model is based on the ALT to AST ratio and platelet counts. We further developed APRI + metabolomics and FIB-4 + metabolomics models by integrating metabolomics data with commonly used APRI and FIB-4 models to assess whether adding metabolomics data could improve model performance.

Survival analysis of individual metabolite associations

All metabolites were log-normalized and standardized to a mean of 0 and standard deviation of 1 to reduce skewness, ensure comparability across metabolites, and eliminate biases introduced by differences in their original scales, thereby enhancing the robustness and interpretability of the model. The Cox-PH model was used to assess the risk of cirrhosis associated with each metabolite. The model was adjusted for age, sex, and complete set of CLD risk factors. The Benjamini-Hochberg method was applied to correct for multiple comparisons of P-values.

Elastic net model development and evaluation

To identify the most predictive metabolites, we applied a Cox-PH model with EN regularization to the training set, optimizing model discrimination using 10-fold cross-validation with performance assessed by Harrell’s C-index. To identify the most predictive metabolites, we applied a Cox proportional hazards model with EN regularization to the training set. EN regularization combines the benefits of both L1 and L2 penalties, balancing sparsity and model complexity by adjusting the α parameter (0 ≤ α ≤ 1), which controls the weight ratio between L1 (lasso) and L2 (ridge) regularization. The L1 component promotes variable selection, while the L2 component enhances model stability. We optimized the model’s performance using 10-fold cross-validation to determine the optimal α and λ (regularization strength) parameters, maximizing the Cox model’s discriminatory power (e.g., Harrell’s C-index) and minimizing overfitting. The predictive accuracy of the final model was evaluated in the validation set using metrics such as Harrell’s C-index, sensitivity, Youden’s index specificity, and net reclassification improvement (NRI). Additionally, we performed receiver operating characteristic (ROC) curve and decision curve analyses, stratifying individuals into quintiles based on the predicted cirrhosis risk. The performance of the model was further validated through calibration and network visualization.

We calculated and visualized the Spearman’s correlation of the model features using the corrplot package. Network visualization followed the standard weighted gene co-expression network analysis (WGCNA) workflow, where metabolite correlations were calculated and converted into an adjacency matrix with a soft threshold of β = 30. The resulting topological overlap matrix was processed using a hard threshold (threshold = 0.2) to construct an unweighted network graph. Metabolite nodes included in the final APRI + metabolomics model are highlighted with saturated colors, and node sizes were adjusted according to the degree of connectivity. The entire network was generated and visualized using WGCNA and igraph packages.

Identification of key metabolites and pathway enrichment analysis

Key metabolites were identified from the model using EN Cox regression. Metabolites with non-zero coefficients in the final model were considered critical, and their importance was ranked based on the magnitude of their coefficients. To further elucidate the biological pathways these metabolites are involved in, we performed pathway enrichment analysis using the WebGestalt tool (https://www.webgestalt.org). The key metabolites were mapped to their corresponding pathways using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.

Significance, software, and data availability

All analyses were performed using R (v.4.3.3) software. Statistical significance was controlled using the Benjamini-Hochberg method, with significance defined as an adjusted P-value of less than 0.05. These data are available on the UKB website.

Results

Baseline characteristics

Following strict selection criteria, the final study cohort included 2,738 patients with CLD (Table 1). The median age of the eligible participants was 56 years (interquartile range: 50–62 years), and 51.4% were male. During the study period, individuals who developed cirrhosis (n = 142 [5.2%]) tended to be older. Significant differences were observed in baseline characteristics and event occurrence rates, particularly in various sociodemographic factors, such as the Townsend deprivation index (TDI) and alcohol consumption frequency. Participants with higher TDI scores were more likely to progress to cirrhosis. Lower socioeconomic status and greater social disadvantages were associated with cirrhosis (P = 0.002). Frequency of alcohol consumption was also associated with cirrhosis, with more frequent drinking being linked to a higher incidence of cirrhosis (P < 0.001). Clinical chemistry indicators, such as ALT, AST, alkaline phosphatase (ALP), PLT, Alb, total bilirubin (TBil), and direct bilirubin (DBil), as well as commonly used scores, such as APRI and FIB-4, were significantly associated with cirrhosis. These findings are consistent with expectations based on the literature [19–21].

Table 1.

Baseline characteristics and event outcomes of the study population

Characteristic	Whole cohort(n = 2738)	Incident cirrhosis
		Yes(n = 142)	No(n = 2596)
Sociodemographics
Age at recruitment (years)	56(50–62)	58(51–64)	56(50–62)
Sex
Female	1491(54.5%)	69(48.6%)	1422(54.7%)
Male	1247(45.5%)	73(51.4%)	1174(45.3%)
Ethnicity
Black	36(1.3%)	2(1.4%)	34(1.3%)
White	2702(98.7%)	140(89.6%)	2562(98.7%)
TDI *	-1.405(-3.300-1.863)	-0.105(-2.798-3.165)	-1.490(-3.320-1.790)
Higher education	637(23.4%)	28(19.7%)	609(23.5%)
Lifestyle factors
Current smoker	452(16.5%)	28(19.7%)	424(16.3%)
Drink frequency *
Never	274(10.0%)	25(17.6%)	249(9.6%)
Special occasions only	380(13.9%)	24(1.9%)	356(13.7%)
One to three times a month	310(11.3%)	10(7.1%)	300(11.6%)
Once or twice a week	642(23.4%)	22(15.5%)	620(23.9%)
Three or four times a week	439(16.0%)	12(8.4%)	427(16.4%)
Daily or almost daily	447(16.3%)	40(28.2%)	407(15.7%)
Sleep duration (hours/day)	7(6–8)	7(6–8)	7(6–8)
Clinical chemistry
ALT(U/L) *	26.320(18.593–39.620)	35.240(22.585–55.706)	25.885(18.480-38.693)
AST(U/L)*	26.7000(22.100–34.400)	36.450(26.250-53.225)	26.500(22.000-33.525)
ALP(U/L) *	87.150(72.100-105.700)	98.000(82.275–130.300)	86.400(71.900-104.800)
PLT(109 cells/Litre) *	250.150(215.000-295.000)	223.400(182.500-266.975)	251.450(217.000-296.000)
Alb(g/L) *	45.020(43.035–46.755)	44.550(41.84–46.27)	45.060(43.060–46.770)
TBil(umol/L) *	7.870(6.220-10.293)	9.390(7.160–13.910)	7.830(6.170–10.140)
DBil(umol/L) *	1.600(1.290–2.110)	2.050(1.545-3.100)	1.580(1.290–2.070)
Glu(mmol/L)	4.954(4.615–5.405)	4.987(4.674–5.4575)	4.952(4.610–5.402)
HbA1c(mmol/mol)	35.600(33.100–38.400)	36.300(33.325–39.75)	35.600(33.100–38.300)
Physical measurements
Systolic blood pressure	134(123–147)	137(125–151)	134(123–147)
BMI (kg/m2)	29.498(26.432–33.305)	29.481(26.177–33.519)	29.503(26.462–33.287)
Waist/hip ratio	0.905(0.843–0.963)	0.920(0.861–0.978)	0.904(0.842–0.963)
Body fat percentage (%)	34.8(27.8–42.1)	34.4(27.2–40.9)	34.9(27.9–42.1)
Basal metabolic rate (KJ)	6732(5807–7916)	7017(5963–8001)	6711(5800–7908)
Cirrhosis score
APRI*	0.270(0.206–0.372)	0.419(0.274–0.706)	0.266(0.204–0.362)
FIB-4*	1.181(0.915–1.550)	1.580(1.122–2.350)	1.170(0.909–1.516)

Values are presented as median (interquartile range) or n (%). Statistical tests used include the Mann–Whitney U test for continuous variables and the chi-square (𝜒2) test for categorical variables. Yes: Indicates the presence of the condition or characteristic in the respective category. Benjamini–Hochberg correction was applied for multiple testing.* indicates statistically significant differences P < 0.05

Abbreviations: ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALP, alkaline phosphatase; PLT, platelet count; Alb, albumin; TBil, total bilirubin; DBil, direct bilirubin; Glu, glucose; HbA1c, glycated hemoglobin; BMI, body mass index; APRI, aspartate aminotransferase to platelet ratio index; FIB-4, fibrosis-4 score; TDI, Townsend deprivation index

Relationship between individual metabolites and cirrhosis

To examine the relationship between individual metabolites and the risk of cirrhosis, we used the Cox proportional hazards model to evaluate the association between these metabolites and the progression of CLD to cirrhosis. After adjusting for age and sex, 68 of 168 metabolites (40.5%) were found to be significantly associated with cirrhosis events. However, after further adjustment for all characteristics, the number of significantly associated metabolites decreased to 21 (12.5%). Notably, only 10 metabolites (5.9%) showed consistent significance in both models (detailed statistical results for all metabolites are provided in Supplementary Tables S3 and S4).

In the age- and sex-adjusted models (Figs. 2A and 3A), we observed that most lipoprotein particles, triglycerides (very low-density lipoprotein [VLDL]), phospholipids (VLDL, low-density lipoprotein [LDL], and high-density lipoprotein [HDL]), and total lipids (VLDL and HDL particles) were positively associated with cirrhosis events, whereas other metabolites, such as certain amino acids and free cholesterol, showed a negative association.

Fig. 2.

Cox proportional hazards model results for cirrhosis risk stratification based on serum metabolomics and clinical data. (A) Cox proportional hazards model adjusted for age and gender for each metabolite; (B) Cox proportional hazards model including all features. Several metabolites were significantly associated with cirrhosis events, with false discovery rate-controlled p-values < 0.05 (indicated by dark gray labels). The color of the bars represents the direction of the association: positive (red) or negative (blue)

Abbreviations: Apo, apolipoprotein; DHA, docosahexaenoic acid; FA, fatty acid; HDL, high-density lipoprotein; HDLC, high-density lipoprotein cholesterol; LA, linoleic acid; LDL, low-density lipoprotein; L, large; LP, lipoprotein; M, medium; MUFA, monounsaturated fatty acid; PUFA, polyunsaturated fatty acid; S, small; VLDL, very low-density lipoprotein; VS, very small; XL, very large

Fig. 3.

Top 20 metabolite associations with cirrhosis risk. (A) Represents hazard ratios adjusted for age and sex, while (B) shows the results after adjusting for all variables. The green bars represent the model adjusted for age and sex, and the orange bars represent the model adjusted for all clinical and metabolomic variables

Abbreviations: HDL, high-density lipoprotein; LDL, low-density lipoprotein; VLDL, very low-density lipoprotein; IDL, intermediate-density lipoprotein; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol

After further adjustment for age, sex, lifestyle factors, and biochemical measurements (Figs. 2B and 3B), the strength of the associations between phospholipids, amino acids, free cholesterol, and total lipids decreased. However, the association with lipoprotein particles became stronger, indicating that these metabolites remained significantly correlated with cirrhosis, even after adjusting for factors such as BMI, liver function tests, smoking, and alcohol consumption.

In both Cox proportional hazards model analyses, 10 overlapping metabolites were identified. These metabolites are primarily distributed across the following categories: lipoprotein particle size, lipoprotein particle concentration, cholesteryl esters, phospholipids, and total lipids. Notably, the hazard ratios (HRs) for these metabolites showed consistent associations in both models, indicating that they may be significantly linked to the development of cirrhosis. These findings suggest that certain metabolites may serve as potential biomarkers for cirrhosis risk in patients with CLD. Supplementary Tables S3 and S4 provide detailed results from individual metabolite analyses, highlighting the complexity and specificity of these associations.

Application of serum metabolomics in cirrhosis risk stratification

Our dataset was divided into derivation (80%) and validation (20%) cohorts, with well-balanced baseline characteristics and cirrhosis outcomes between the two groups (Supplementary Table S5). In the derivation cohort, we fitted elastic net regularized risk models. The five models included a metabolomics-only model (36 out of 168 metabolites), FIB-4 model, APRI model, FIB-4 + metabolomics model (18 out of 168 metabolites), and APRI + metabolomics model (22 out of 168 metabolites) (Supplementary Tables S6 and S7 for feature coefficients and optimized model hyperparameters).

Figure 4A shows the ROC curves of the five models. The metabolomics-only, FIB-4 model had an AUC of 0.696, and APRI models had AUC of 0.712, 0.696, and 0.718, respectively. The FIB-4 + metabolomics model, which combines FIB-4 and metabolomics data, achieved an AUC of 0.717, demonstrating a better discriminative ability than using FIB-4 or metabolomics alone. The APRI + metabolomics model had the highest AUC of 0.747, making it the most effective in terms of discriminative power among all models.

Fig. 4.

Internal model validation: discriminative performance of different models. (A) ROC curves illustrating the performance of different models in discriminating cirrhosis risk. (B) DCA curves showing the clinical utility of each model in predicting cirrhosis events

Abbreviations: APRI, aspartate aminotransferase to platelet ratio index; DCA, decision curve analysis; ROC, receiver operating characteristic; FIB-4, fibrosis-4 score

Figure 4B shows the net clinical benefit of the models at different threshold probabilities. Net benefit measures the potential clinical value of the model across various thresholds. Models that incorporated metabolomics data outperformed those that used only traditional parameters across most thresholds, showing a higher net benefit. The APRI + metabolomics model demonstrated the highest net benefit at nearly all thresholds, indicating its potential for clinical application.

Table 2 provides the internal model validation statistics to assess absolute discriminative ability. The APRI + metabolomics model had the highest C-index (0.747), indicating its strongest ability to predict cirrhosis in patients. The delta C-statistic (ΔC) represents the difference in the C-index between models and is used to evaluate the performance improvement after incorporating metabolomics data. Based on our results.

Table 2.

Internal model validation: discriminative ability of different models for cirrhosis risk prediction

Model	Harrell’s C	Sensitivity	Specificity
Metabolomics	0.712[0.707–0.716]	0.599[0.588–0.610]	0.734[0.733–0.736]
FIB-4	0.696[0.692-0.700]	0.563[0.524–0.603]	0.749[0.714–0.784]
FIB-4 + Metabolomics	0.717[0.714–0.720]	0.445[0.436–0.454]	0.878[0.876–0.880]
APRI	0.718[0.715–0.721]	0.584[0.575–0.594]	0.780[0.778–0.783]
APRI + Metabolomics	0.747[0.743–0.750]	0.635[0.621–0.649]	0.754[0.742–0.766]

Sensitivity and specificity were calculated using the Youden index. The 95% confidence intervals (in brackets) were obtained through 1,000 bootstrap resamplings of the percentile distribution of the validation cohort

Abbreviations: APRI, aspartate aminotransferase to platelet ratio index; FIB-4, fibrosis-4 score

To further assess the degree of classification improvement, we compared the models in pairs (Table 3), including FIB-4 versus APRI, metabolomics versus FIB-4 + metabolomics, metabolomics versus APRI + metabolomics, FIB-4 versus FIB-4 + metabolomics, APRI versus APRI + metabolomics, and FIB-4 + metabolomics versus APRI + metabolomics. Subsequent performance evaluations indicated that incorporating metabolomic data improved the discriminative ability of the models. For instance, the FIB-4 + metabolomics model showed an increase in the C-index of 0.021 [0.014–0.028] compared to FIB-4 alone, with an NRI of 0.504 [0.488–0.520] (P < 0.001). Similarly, the APRI + metabolomics model showed an improvement of 0.029 [0.022–0.035] in the C-index compared to APRI alone, with an NRI of 0.378 [0.366–0.389] (P < 0.001). These findings suggest that the addition of metabolomics data enhances classification. However, the NRI of the APRI + metabolomics model indicates a significant improvement in case classification but a negative NRI for non-cases, meaning that some low-risk individuals may be misclassified as high-risk. The results of internal validation metrics are summarized in Supplementary Table 8, with relative performance metrics detailed in Supplementary Table 9.

Table 3.

Internal model validation: relative performance metrics of different models for cirrhosis risk stratification

Metric	FIB-4 vs. APRI	Metabolomics vs. FIB − 4 + Metabolomics	Metabolomics vs. APRI + Metabolomics	FIB-4 vs. FIB-4 + Metabolomics	APRI vs. APRI + Metabolomics	FIB-4 + Metabolomics vs. APRI + Metabolomics
ΔC	0.022 [0.015–0.029]	0.005 [-0.002-0.013]	0.035 [0.027–0.043]	0.021 [0.014–0.028]	0.029 [0.022–0.035]	0.030 [0.023–0.036]
NRI
Cases	-0.963[-0.967–0.959]	0.981[0.981–0.981]	0.779[0.770–0.787]	0.536[0.519–0.552]	0.779[0.768–0.789]	-0.734[-0.747–0.721]
Non-cases	0.9923[0.993–0.993]	-0.973[-0.974–0.973]	-0.887[-0.888–0.885]	-0.032[-0.035–0.028]	-0.401[-0.405–0.397]	0.933[0.932–0.935]
Overall	0.030[0.026–0.034]	0.007[0.006–0.008]	-0.108[-0.116–0.099]	0.504[0.488–0.520]	0.378[0.366–0.389]	0.199[0.186–0.213]

This table compares the relative performance of different models for cirrhosis risk stratification. Model performance was evaluated in the validation cohort. The 95% confidence intervals (in parentheses) were obtained from the percentile distribution of 1000 bootstrap resamples within the validation cohort

Abbreviations: ΔC, delta c-statistic; NRI, net reclassification improvement; APRI: aspartate aminotransferase to platelet ratio index; FIB-4, fibrosis-4 score

Overall, the APRI + metabolomics model demonstrated superior performance in terms of both discriminative ability (C-index) and clinical net benefit, making it more effective for risk classification and the prediction of cirrhosis progression.

Metabolomics and cirrhosis risk stratification

To further explore the translational potential of the risk models, we assessed the cumulative cirrhosis incidence, model calibration, and cirrhosis-free survival based on the predicted risk. The inclusion of serum metabolomics in the FIB-4 and APRI models improved risk stratification across the cirrhosis incidence quintiles (Fig. 5A). Model calibration was evaluated for all individuals who completed the 10-year follow-up, and we found that all models were fairly and similarly calibrated, except for the metabolomics-only model, which was less accurate (Fig. 5B). Finally, Kaplan-Meier curves highlighted superior survival stratification over the entire follow-up period in models that included metabolomics, particularly FIB-4 + and APRI + metabolomics (Fig. 5C), with P < 0.05.

Fig. 5.

Metabolomics enhances risk and survival stratification in cirrhosis. (A) Cumulative cirrhosis incidence by risk quintiles, with risk stratified by blue shading indicating different levels of risk. (B) Model calibration plot showing observed versus predicted cirrhosis event risk, providing insight into the accuracy of the model in predicting cirrhosis events. (C) Cirrhosis-free survival by risk quintiles, illustrating the relationship between risk categories and survival without cirrhosis progression

Abbreviations: FIB-4, fibrosis-4 score; APRI, aspartate aminotransferase to platelet ratio index

Model selection and feature retention

Reducing the number of measured features is crucial for a cost-effective clinical implementation and preventive screening. In our study, we used an EN-regularized Cox regression model to select the features that carried the most predictive information. Figure 6A illustrates this approach, in which the network visualization of metabolites highlights the effective representation of highly connected clusters by individual metabolites. The final model retained several unrelated metabolite features (Fig. 6B). A total of 17 metabolites were retained, including 6 amino acids, 2 fatty acids, and 9 lipoprotein subclasses (for all clinical and metabolite feature coefficients, see Supplementary Table S6). Network visualization of the measured metabolites demonstrated that individual metabolites effectively represented highly correlated clusters.

Fig. 6.

Key features representing metabolic risk in cirrhosis. (A) Correlation heatmap for the APRI + metabolomics model, showing the relationship between clinical and metabolite features. The heatmap highlights significant correlations between selected clinical parameters and metabolites, providing insight into how these features contribute to cirrhosis risk prediction. (B) Network visualization of measured metabolomics, illustrating key retained clinical and metabolite features. This network diagram visualizes how clinical and metabolite features are interconnected, offering a clearer understanding of their combined role in metabolic risk

Abbreviations: Apo, apolipoprotein; LP, lipoprotein; LDL, low-density lipoprotein; HDL, high-density lipoprotein; VLDL, very low-density lipoprotein; TBil, total bilirubin; DBil, direct bilirubin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; PLT, platelet count; Alb, albumin; BMI, body mass index; APRI, aspartate aminotransferase to platelet ratio index; FIB-4, fibrosis-4 score

Key metabolites identified and pathway enrichment analysis

The APRI + Metabolomics model identified 21 key metabolites across several functional categories as significant predictors of cirrhosis risk. Amino acids, including branched-chain amino acids (valine, leucine, isoleucine), glutamine, and glycine, were negatively associated with cirrhosis, suggesting their roles in maintaining protein synthesis and nitrogen balance. In contrast, phenylalanine and tyrosine were positively associated, reflecting hepatic dysfunction and metabolic stress. Fatty acids, such as docosahexaenoic acid (DHA) and monounsaturated fatty acids (MUFAs), indicated profound disturbances in lipid metabolism and inflammation, which are key drivers of cirrhosis progression. Cholesterol and lipoprotein markers, including LDL cholesterol, VLDL cholesterol, and total cholesterol minus HDL-C, revealed lipid overload and hepatocyte injury, while alterations in phospholipids and total lipids in large VLDL and very large HDL pointed to membrane instability and metabolic dysregulation. Free cholesterol in small VLDL further implicated oxidative stress as a contributing factor. Among these metabolites, DHA, MUFAs, and total lipids in large VLDL emerged as the most influential, highlighting the critical role of lipid metabolism disturbances in cirrhosis pathogenesis.

The metabolic pathway enrichment analysis (Fig. 7) identified multiple pathways disrupted during cirrhosis progression, underscoring their biological relevance in the disease. The most enriched pathway was phenylalanine and tyrosine metabolism, highlighting its central role in nitrogen imbalance and impaired hepatic clearance, both hallmark features of advanced liver disease. Other significantly enriched pathways included porphyrin metabolism, linked to disruptions in heme biosynthesis and oxidative stress, and nitrogen metabolism, reflecting urea cycle dysfunction and impaired ammonia detoxification. The degradation of valine, leucine, and isoleucine, representing branched-chain amino acid catabolism, was also notably enriched, suggesting impaired protein metabolism and reduced energy production. Additional enrichments were observed in pathways such as retinol metabolism, phosphatidylinositol phosphate metabolism, and amino sugar metabolism, indicating broader alterations in lipid signaling, vitamin metabolism, and carbohydrate utilization. Pathways like arginine and proline metabolism, along with purine metabolism, further emphasized the extensive metabolic reprogramming characteristic of cirrhosis.

Fig. 7.

Pathway enrichment analysis of key metabolites associated with cirrhosis risk

Discussion

Cirrhosis is a major cause of morbidity and mortality in patients with CLD worldwide, and the number of cirrhosis-related deaths is expected to increase over the next decade [22]. Therefore, greater efforts are needed to promote primary prevention, early detection of cirrhosis, and improved access to treatment. Allocating more resources for primary prevention, early diagnosis of cirrhosis, and better integration with healthcare services is crucial to reduce the global burden of cirrhosis [20]. Identifying individuals at high risk of developing cirrhosis is essential for lowering liver disease-related mortality.

The etiology of cirrhosis is changing owing to the rising prevalence of obesity, increased alcohol consumption, and improved management of hepatitis B and C infections [20]. These factors shift the epidemiology and burden of cirrhosis. Previous models often focused on a single disease etiology and were not well equipped to address the evolving spectrum of CLD. When multiple liver diseases co-occur, these models tend to perform poorly. Therefore, there is an urgent need for a tool that is applicable to all types of CLD and offers higher diagnostic accuracy to identify individuals at risk of progressive disease. Early intervention in high-risk populations before clinical symptoms emerge could potentially reverse early stage liver fibrosis.

In this study, we described the association between individual metabolites and cirrhosis events, which is consistent with previously reported findings and established pathological mechanisms. The accumulation of metabolic dysfunctions involving amino acids, lipids, carbohydrates, and fatty acids leads to oxidative stress and damage to hepatocytes, driving CLD to more severe pathological stages [23]. Because the liver is the primary organ for amino acid metabolism, impaired liver function leads to altered amino acid metabolism. Previous studies have shown that early cirrhosis causes an imbalance in peripheral blood amino acids, characterized by a decrease in branched-chain amino acids (BCAAs) and an increase in aromatic amino acids (AAAs) [24], such as phenylalanine and tyrosine, due to impaired hepatic clearance. The enrichment of phenylalanine and tyrosine metabolism pathways highlights the accumulation of AAAs, which has been associated with nitrogen imbalance and complications such as hepatic encephalopathy [25]. BCAA deficiency can occur as early as the chronic hepatitis stage prior to cirrhosis development, contributing to reduced Alb synthesis and diminished antioxidant capacity, further exacerbating disease progression in cirrhosis [26]. BCAAs play an important role in improving immune function and reducing oxidative stress [27], suggesting that targeted BCAAs supplementation could help address these metabolic derangements. At this stage, the ability of the liver to repair tissues is overwhelmed, leading to the progression of fibrosis [28]. These findings were consistent with previously reported associations and established pathological mechanisms. Adjusting amino acid metabolism, such as supplementing BCAAs or targeting AAAs imbalances, could help mitigate oxidative stress, improve immune function, and slow the progression of liver fibrosis.

Metabolic abnormalities reduce antioxidant capacity and increase lipotoxicity, thereby increasing the risk of cirrhosis [29]. Hepatic lipotoxicity, characterized by the ectopic accumulation of triglycerides and their intermediates, leads to hepatocyte injury and structural changes within the liver [30, 31]. Lipid overload triggers apoptotic cascades with subsequent caspase activation, potentially promoting inflammation and fibrosis [32]. Cholesterol crystals can activate NLRP3 inflammasome, leading to hepatocyte inflammation. At the tissue level, repair and remodeling processes occur, resulting in fibrosis. Over time, sustained lipid overload induces oxidative stress, ultimately leading to fibrosis formation [33]. Hepatocytes are particularly vulnerable to the “multiple hits” characterized by oxidative stress, which further promotes the synthesis of pro-inflammatory cytokines, driving disease progression [34].

The enrichment of pathways such as retinol metabolism [35], phosphatidylinositol phosphate metabolism [36], and porphyrin metabolism [37] underscores the systemic nature of lipid metabolic dysfunction in cirrhosis. Elevated DHA levels, although typically anti-inflammatory, may reflect maladaptive responses to chronic inflammation in advanced liver disease. Similarly, reductions in MUFAs and alterations in total lipids in large VLDL particles highlight lipid imbalances that exacerbate hepatocyte injury [38]. Lipid peroxides and products from damaged hepatocytes further activate hepatic stellate cells, driving their transition into myofibroblast-like cells and promoting fibrotic remodeling [39]. These findings highlight the critical role of lipid metabolism in cirrhosis progression and suggest potential therapeutic targets. Anti-inflammatory interventions and strategies to restore lipid balance could mitigate hepatocyte injury and fibrosis, offering promising avenues for cirrhosis management.

In the context of our study, we accounted for several important factors to mitigate potential biases and improve the reliability of our results. First, we addressed variable selection bias by employing Elastic Net regularization combined with 10-fold cross-validation, ensuring that the metabolites selected for the models were robust predictors of cirrhosis risk. Regarding the healthy volunteer bias inherent in the UK Biobank data, which may overrepresent healthier and older individuals, we acknowledge that this could limit the generalizability of our findings to younger or more diverse populations. However, the results still provide valuable insights, and future studies should aim to validate the models in broader cohorts with varying disease severity and comorbidities to better understand their applicability in real-world clinical settings.

To address data imbalance, we used random sampling to divide the dataset into a derivation cohort (80%) and a validation cohort (20%) without stratified sampling. This approach aimed to preserve heterogeneity within the sample, improving the model’s generalizability. Additionally, batch effects were minimized through rigorous data processing and standardization to ensure comparability of metabolite concentrations across different batches. For multiple comparison correction, we applied the Benjamini-Hochberg method to adjust p-values, reducing the likelihood of false positives. These strategies enhanced the robustness of our model while providing a clear framework for future improvements in clinical validation.

We developed five models, and the results showed that the APRI + metabolomics model performed best in cirrhosis risk stratification, demonstrating a strong discriminative ability. Serum metabolomic analysis based on 1 H-NMR is not only reliable but also cost-effective, providing comprehensive systemic metabolic information from a single blood sample. The strength of this study lies in its large sample size and the exclusion of patients with pre-existing cirrhosis, thereby minimizing bias related to treatments, such as lipid-lowering therapies. While this study demonstrates the utility of 1 H-NMR in identifying predictive metabolites for cirrhosis, it is important to acknowledge the platform’s limitations. Compared to mass spectrometry (MS), 1 H-NMR has lower sensitivity and may not capture low-abundance metabolites or complex chemical structures [40]. Particularly, the non-destructive nature, low sample preparation requirements, and reproducibility of the platform make it well-suited for large-scale studies [41]. Furthermore, combining 1 H-NMR data with MS in future analyses could provide a broader metabolic profile, improving the accuracy and generalizability of predictive models. Despite this, we acknowledge that excluding individuals using lipid-lowering medications may have introduced potential biases. Lipid-lowering medications can significantly impact lipid metabolism and, consequently, the metabolomic profiles. While excluding this group minimizes confounding effects, it may limit the generalizability of the findings to broader populations where such medications are commonly used. Future studies could include these individuals and adjust for the effects of lipid-lowering therapies to better assess their impact on cirrhosis risk prediction. This is the first study to provide a detailed description of cirrhosis risk stratification using metabolomics. We validated clinical risk scores based on disease specificity using a well-established metabolomics platform that has received regulatory approval. However, several challenges remain in their clinical application. First, it is well known that the UK population does not fully represent the global population, as participants in this study tend to be older and healthier than the general population (the “healthy volunteer” bias). Another limitation of this study was the absence of liver elastography data. While we validated the internal effectiveness of the models, the results require external validation in other populations to enhance their generalizability. External validation in wider and more heterogeneous populations is necessary to assess the robustness and generalizability of the model. Real-world clinical settings could benefit from integrating data across different ethnicities, age groups, comorbidities, and disease etiologies, providing valuable insights into the model’s utility. Additionally, integrating metabolomics data from platforms like mass spectrometry could offer a more comprehensive metabolic profile, likely improving the model’s performance in diverse populations. Moreover, the blood samples were collected in a non-fasting state and stored for extended periods before analysis, which may have introduced some variation, potentially underestimating the performance of the metabolomics models. The fasting state can be a significant confounding factor in metabolomics research, as it can alter levels of certain metabolites, particularly those related to short-term energy metabolism, such as glucose and ketone bodies [42]. However, previous studies have shown that many key metabolites, including amino acids, certain lipids, and free fatty acids, are minimally affected by fasting [43, 44] This supports the robustness of our findings, as the primary metabolites identified in our study-free fatty acids and amino acids-are less sensitive to fluctuations due to fasting status. Future studies should aim to minimize variability by standardizing sampling conditions, including documenting fasting duration during sample collection. Comparing metabolite levels between fasting and non-fasting states would help assess how variable influences model performance. Furthermore, validating models in independent cohorts with strict fasting controls could ensure the robustness and generalizability of the findings. Addressing these considerations will enable metabolomics studies to better capture disease-specific metabolic features and improve the reliability of predictive models. Potential confounders such as alcohol consumption, medication use, and dietary habits may have also influenced the metabolomic profiles and progression of cirrhosis. While self-reported data on alcohol consumption and medication use were adjusted for in the analysis, detailed dietary information was unavailable. This limitation highlights the need for future studies to collect comprehensive lifestyle and dietary data and conduct sensitivity analyses to evaluate their impact on predictive models. Finally, the metabolomics platform used in this study covered a wide range of lipids and lipid subtypes, offering rich data for further exploration of the relationship between metabolites and cirrhosis risk.

Conclusion

In conclusion, we have demonstrated that 1 H-NMR serum metabolomics can effectively serve as a standalone screening tool for assessing the risk of cirrhosis. We also showed that machine learning can be successfully applied to reduce the number of features used for risk prediction while maintaining strong predictive performance. In the context of cirrhosis, several large metabolite clusters can be efficiently represented by single key metabolites with high predictive value. This approach presents a significant potential for cost-effective implementation, which could facilitate its adoption in clinical practice.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

1H-NMR

Proton nuclear magnetic resonance

AAAs

Aromatic amino acids

ALB

Albumin

ALP

Alkaline phosphatase

ALT

Alanine aminotransferase

APRI

Aspartate aminotransferase to platelet ratio index

AST

Aspartate aminotransferase

BCAA

Branched-chain amino acids

BMI

Body mass index

CI

Confidence interval

CLD

Chronic liver disease

Cox-PH

Cox proportional hazards

DBil

Direct bilirubin

DCA

Decision curve analysis

EN

Elastic net

FA

Fatty acid

FIB-4

Fibrosis-4 score

HbA1c

Glycated hemoglobin

HDL

High-density lipoprotein

HR

Hazard ratio

IQR

Interquartile range

LDL

Low-density lipoprotein

NMR

Nuclear magnetic resonance

NRI

Net reclassification improvement

PLT

Platelet count

ROC

Receiver operating characteristic

TBil

Total bilirubin

TDI

Townsend deprivation index

VLDL

Very low-density lipoprotein

Author contributions

Jingru Song: Software, Methodology, Writing– original draft. Ziwei Gao: Software. Liqun Lai: Investigation, Supervision. Jie Zhang: Visualization. Binbin Liu: Formal analysis. Yi Sang: Investigation. Siqi Chen: Supervision. Jiachen Qi: Formal analysis. Yujun Zhang: Visualization, Validation. Huang Kai: Data curation and project administration. Wei Ye: Funding acquisition, project administration, writing– review, and editing.

Funding

This research was funded by Hangzhou Medical Key Cultivation Discipline (grant number 2020SJZDXK13), Hangzhou Health and Health Commission (grant number A20230658), Provincial Central Management Bureau Heritage Innovation Talent Project (grant number 2024ZR121), Zhejiang Provincial Traditional Chinese Medicine Science and Technology Project (grant number 2021ZA107), and Clinical Research Project of the China Association of Chinese Medicine (grant number ZA_CACM_2024002).

Data availability

UKB data are publicly available to approved researchers via the following link: https://www.ukbiobank.ac.uk. We accessed the UKB data under application ID 100739.

Declarations

Ethics approval and consent to participate

Ethical approval was provided by the North West Multi-Center Research Ethics Committee (REC reference number: 21/NW/0157). All participants provided informed consent prior to inclusion in the study. The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki, and the use of secondary data from the UK Biobank was approved by the committee.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Clinical trial registration

Clinical trial number: not applicable.