Introduction
Acute-on-chronic liver failure (ACLF) is a clinical syndrome that develops in patients with chronic liver disease who present with acute hepatic dysfunction and is characterized by hepatic and/or extrahepatic organ failure and systemic inflammation.1 The 90-day mortality for ACLF ranges from 40% to 58%.2,3 ACLF is associated with concomitant immune dysfunction, which increases susceptibility to bacterial infection.4,5 Bacterial infection is the most common extrahepatic precipitant and complication of ACLF, affecting nearly 50% of cases.4 It can trigger a cytokine storm, leading to multiple organ failure and driving a vicious cycle of immunoparalysis and sepsis that substantially elevates mortality.6,7 Early identification of bacterial infection and antimicrobial intervention are crucial for improving clinical outcomes in this population.
Early and accurate diagnosis of bacterial infection in patients with ACLF remains challenging. Inaccurate diagnosis can lead to overuse of antimicrobials or delays in treatment.8 Current diagnostic approaches for bacterial infection predominantly rely on conventional biomarkers such as C-reactive protein (CRP), white blood cell count (WBC), and procalcitonin (PCT).8,9 These indicators can be altered by both infectious and non-infectious causes, and their responses exhibit a temporal lag.10,11 CRP and PCT are acute-phase reactants. A study showed that bacterial infection does not significantly influence the concentrations of acute-phase proteins in patients with fulminant hepatic failure.10 Due to the wide range of diagnostic thresholds and the limited time window of acute-phase reactants, their diagnostic accuracy is limited. CRP has low specificity for infection, as it frequently exhibits elevated levels due to systemic inflammation itself.11,12 PCT has higher specificity but is not suitable for distinguishing infectious from non-infectious inflammation.11 In the context of liver injury, WBC counts in infected patients are often within the clinically normal range, rendering it inadequate as a standalone diagnostic marker.12,13 Although pathogen culture remains the diagnostic gold standard for bacterial infection, it has low positivity rates and is unsuitable for routine screening in patients with cirrhosis who have atypical clinical symptoms.14 None of these methods can capture earlier or more precise signatures of bacterial infection. Previous studies have identified protein signatures associated with infection-related acute decompensation of cirrhosis.15 However, the metabolomic characteristics of patients with bacterial infection-related ACLF remain unclear. Since our previous work indicated the prognostic value of metabolomics in ACLF,16 we hypothesized that metabolomics could also identify biomarkers for bacterial infection.
To identify novel biomarkers, we used untargeted metabolomics to analyze a large cohort of plasma samples from patients with acute-on-chronic liver disease (AoCLD). Our objectives were to develop diagnostic models for bacterial infection in ACLF and to assess the models’ risk stratification performance for 90-day outcomes. Given the high risk of ACLF progression in non-ACLF (defined in the Materials and methods section) patients with bacterial infection, we analyzed this subgroup separately to complement the primary analysis.
Methods
Patient cohort, data collection, and study design
Patients were recruited from 14 medical centers in the Chinese Acute-on-Chronic Liver Failure study between January 2015 and March 2019. The inclusion criteria for AoCLD were: (1) diagnosis of chronic liver disease; (2) hospitalization for acute liver injury (total bilirubin (TB) > 2 mg/dL, alanine aminotransferase (ALT) or aspartate aminotransferase (AST) > 3× the upper limit of normal); (3) hospitalization for acute decompensation (e.g., gastrointestinal bleeding). The exclusion criteria were: (1) age < 15 or > 80 years; (2) pregnancy; (3) hepatocellular carcinoma or other malignant liver tumors; (4) malignant tumors of other organs; (5) severe extrahepatic diseases; and (6) immunosuppressive therapy.16 This study was registered at ClinicalTrials.gov (NCT02457637, NCT03641872).
AoCLD patients were classified as ACLF or non-ACLF after enrollment based on clinical follow-up data. ACLF was diagnosed according to the COSSH-ACLF criteria.14 In patients with cirrhosis, ACLF was defined by the presence of at least one organ failure. In non-cirrhotic patients with advanced fibrosis, ACLF was diagnosed based on the co-occurrence of liver failure (TB > 12 mg/dL) and coagulation failure (international normalized ratio (INR) > 1.5).8,16 Organ failure was defined as liver, renal, coagulation, circulatory, respiratory, or cerebral failure. The EASL-CLIF ACLF criteria were used to reclassify the samples for further validation of the models.17 Non-ACLF patients were those with AoCLD who did not meet the corresponding ACLF criteria.
The criteria for infection were based on evidence of bacterial infection within 48 h of admission.8 The inclusion criteria were: (1) pulmonary infection: an infectious focus shown on X-ray or CT, or a positive bacterial culture obtained from sputum, tracheal aspirate, or bronchoalveolar lavage fluid; (2) urinary tract infection: urine leukocyte count > 10/high-power field, positive bacterial Gram stain, or positive midstream urine culture; (3) spontaneous bacterial peritonitis (SBP): absolute neutrophil count in ascites ≥ 250/mm3; (4) spontaneous bacteremia: positive blood culture; (5) skin and soft tissue infection: clinical manifestations of infection with swelling, erythema, fever, and skin tenderness; (6) other bacterial infections (e.g., cholangitis). Patients with fungal or non-hepatotropic viral infections were excluded from this study.18,19
Demographic information, etiological data, clinical symptoms, complications, laboratory tests, and treatments after admission were recorded. EDTA plasma samples collected within 48 h of admission and before empirical antibiotic therapy were obtained and stored at −80 °C. The study design is summarized in Figure 1. After excluding 28 patients who underwent liver transplantation and 44 patients with fungal or non-hepatotropic viral infections and missing infection data, the final analysis included 1,314 of the initial 1,386 patients. We applied machine learning (ML) and statistical analyses to identify candidate biomarkers in the discovery set. Model validation and prognostic risk stratification assessment were conducted in the validation set.
Metabolomics profiling
Untargeted metabolomics analysis was based on the Discovery HD4 metabolomics platform from Metabolon (Durham, North Carolina, USA). The detailed methods were previously reported and are provided in the Supplementary File 1.16 In brief, samples were extracted with methanol to precipitate proteins. The supernatant was divided into four aliquots following centrifugation, and each aliquot was analyzed using a distinct ultra-high-performance liquid chromatography/tandem mass spectrometry (UHPLC-MS/MS) method. Two aliquots were analyzed using separate reverse-phase (RP) UPLC-MS/MS methods with positive-ion mode electrospray ionization (ESI), one aliquot using RP/UPLC-MS/MS with negative-ion mode ESI, and one aliquot using hydrophilic interaction liquid chromatography/UPLC-MS/MS with negative-ion mode ESI. Metabolite identification, quantification, curation, normalization, and quality control were performed using the platform’s software.
Statistical analysis
Statistical analyses were performed using R (version 4.4.3, http://www.r-project.org ) and DecisionLinnc software. Normally distributed continuous variables were compared using Student’s t-test, while the Mann-Whitney U test was used for non-normally distributed variables. Categorical variables were compared using the chi-square test or Fisher’s exact test. P-values were adjusted using the Benjamini-Hochberg method to control the false discovery rate (FDR). For untargeted metabolomics data, metabolites with >80% missing values were excluded. Remaining missing values were imputed using half of the minimum detected value for the corresponding metabolite. The metabolomics data were log2-transformed, and differential metabolite expression was evaluated by calculating fold change, P-values, and estimated FDR. Spearman correlation analysis was used to evaluate relationships between variables. The composite endpoint of infection-related prognostic events was assessed as having hierarchical clinical importance using the win ratio method described in the Supplementary File 1.20 All statistical tests were two-tailed, and statistical significance was defined as P < 0.05.
Model development, evaluation, and interpretation
The cohort was randomly split into discovery and validation sets in a 7:3 ratio using stratified random sampling according to infection status. Feature selection was performed using the Boruta algorithm. Within the confirmed variable set identified by Boruta, recursive feature elimination (RFE) was applied to determine the optimal number of variables in the selected feature subset. To ensure translational applicability, features were further filtered based on the availability of corresponding chemical standards, yielding the final feature subset.
Eight ML algorithms were used to predict the risk of bacterial infection: categorical boosting (CatBoost), multilayer perceptron (MLP), support vector machine (SVM), decision tree (DT), random forest (RF), naive Bayes (NB), logistic regression (LR), and k-nearest neighbors (KNN). The prevalence of bacterial infection was relatively balanced in the ACLF group (0.439) but imbalanced in the non-ACLF group (0.153). Therefore, the synthetic minority over-sampling technique (SMOTE) was applied to the discovery set of the non-ACLF group to balance the minority class, while the original distribution was retained in the ACLF dataset. The optimal hyperparameters for each model were determined through grid search using accuracy as the metric, supplemented by manual tuning (Supplementary Table 1). Each model was developed using the final feature subset and optimal hyperparameters. The final algorithm was selected based on evaluation metrics including discrimination, calibration, and overfitting assessment.
Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), precision-recall curves, accuracy, specificity, recall, precision, and F1 score. Calibration was assessed using calibration curves. Clinical utility was estimated using decision curve analysis (DCA). AUROC differences were compared using DeLong’s test, while net reclassification improvement (NRI) quantified the improvement in discrimination over previous benchmarks. The Youden index was used to determine optimal model thresholds. The SHapley Additive exPlanations (SHAP) method was used to interpret model predictions and quantify the contribution of key features.
Results
Patient characteristics
Among the 1,314 enrolled patients with AoCLD, 451 met the COSSH-ACLF criteria. Bacterial infection occurred in 198 (43.90%) ACLF patients and 132 (15.30%) non-ACLF patients. Baseline characteristics were compared between the infected and non-infected groups (Table 1). In ACLF patients, no significant differences in sex, age (median age: 49.20 vs. 48.19 years, P = 0.735), or cirrhosis (84.34% vs. 80.63%, P = 0.306) were observed between the infected and uninfected groups. Among non-ACLF patients, those with infection were older (median age: 53.0 vs. 49.0 years, P < 0.001) and had a higher prevalence of cirrhosis (84.85% vs. 65.12%, P < 0.001). HBV was the primary etiology of ACLF, accounting for 355 (78.71%) cases. The proportion of alcohol-related cases was higher in ACLF patients with bacterial infection than in those without (54 (27.27%) vs. 46 (18.18%), P = 0.021). In the non-ACLF population, the proportion of HBV infection was lower in patients with bacterial infection than in those without (79 (59.85%) vs. 552 (75.51%), P < 0.001), whereas the proportion of alcohol-related etiology was higher in the infected group compared with the non-infected group (42 (31.8%) vs. 149 (20.38%), P = 0.004). In both the ACLF and non-ACLF cohorts, infected patients exhibited higher rates of jaundice and ascites as acute decompensation events than uninfected patients (all P < 0.001). Laboratory parameters also differed significantly: infected patients exhibited elevated levels of CRP, WBC, PCT, neutrophils, neutrophil-to-lymphocyte ratio (NLR), prothrombin time, INR, TB, AST/ALT ratio, blood urea nitrogen, and creatinine, whereas albumin and sodium were significantly decreased (all P < 0.05).
Table 1Comparison of baseline clinical parameters between infected and uninfected in the ACLF and non-ACLF groups
| Characteristics | ACLF (n = 451) | ACLF | P-value | Non-ACLF (n = 863) | Non-ACLF | P-value |
|---|
| Infection (n = 198) | Non-infection (n = 253) | Infection (n = 132) | Non-infection (n = 731) |
|---|
| Age, Median (Q1, Q3) | 48.67 (41.00–57.00) | 49.20 (42.00–57.22) | 48.19 (40.91–56.00) | 0.735 | 49.00 (40.00–58.22) | 53.00 (43.46–62.69) | 49.00 (39.00–57.42) | <0.001 |
| Male, n (%) | 365 (80.93) | 161 (81.31) | 204 (80.63) | 0.855 | 615 (71.26) | 93 (70.45) | 522 (71.41) | 0.824 |
| Cirrhosis, n (%) | 371 (82.26) | 167 (84.34) | 204 (80.63) | 0.306 | 588 (68.13) | 112 (84.85) | 476 (65.12) | <0.001 |
| Etiology, n (%) |
| HBV | 355 (78.71) | 156 (78.79) | 199 (78.66) | 0.973 | 631 (73.12) | 79 (59.85) | 552 (75.51) | <0.001 |
| HCV | 3 (0.67) | 1 (0.51) | 2 (0.79) | 0.711 | 9 (1.04) | 3 (2.27) | 6 (0.82) | 0.131 |
| HEV | 16 (3.55) | 3 (1.52) | 13 (5.14) | 0.039 | 8 (0.93) | 1 (0.76) | 7 (0.96) | 0.825 |
| Alcohol | 100 (22.17) | 54 (27.27) | 46 (18.18) | 0.021 | 191 (22.13) | 42 (31.82) | 149 (20.38) | 0.004 |
| Autoimmune | 33 (7.32) | 13 (6.57) | 20 (7.91) | 0.588 | 96 (11.12) | 10 (7.58) | 86 (11.76) | 0.159 |
| NAFLD | 16 (3.55) | 3 (1.52) | 13 (5.14) | 0.039 | 38 (4.40) | 5 (3.79) | 33 (4.51) | 0.708 |
| Schistosomiasis | 4 (0.89) | 2 (1.01) | 2 (0.79) | 0.805 | 8 (0.93) | 3 (2.27) | 5 (0.68) | 0.080 |
| Cryptogenic | 18 (3.99) | 11 (5.56) | 7 (2.77) | 0.133 | 14 (1.62) | 4 (3.03) | 10 (1.37) | 0.164 |
| Type of AD, n (%) |
| HE | | 0.775 | | 0.759 |
| 0 | 371 (82.26) | 162 (81.82) | 209 (82.61) | | 810 (93.86) | 126 (95.45) | 684 (93.57) | |
| 1 | 22 (4.88) | 8 (4.04) | 14 (5.53) | | 29 (3.36) | 4 (3.03) | 25 (3.42) | |
| 2 | 30 (6.65) | 15 (7.58) | 15 (5.93) | | 21 (2.43) | 2 (1.52) | 19 (2.60) | |
| 3 | 17 (3.77) | 9 (4.55) | 8 (3.16) | | 3 (0.35) | 0 (0.00) | 3 (0.41) | |
| 4 | 11 (2.44) | 4 (2.02) | 7 (2.77) | | 0 (0.00) | 0 (0.00) | 0 (0.00) | |
| Jaundice | 404 (89.58) | 188 (94.95) | 216 (85.38) | <0.001 | 247 (28.62) | 60 (45.45) | 187 (25.58) | <0.001 |
| Ascites | 284 (62.97) | 158 (79.80) | 126 (49.80) | <0.001 | 385 (44.61) | 97 (73.48) | 288 (39.40) | <0.001 |
| GI bleeding | 31 (6.87) | 13 (6.57) | 18 (7.11) | 0.819 | 133 (15.41) | 19 (14.39) | 114 (15.60) | 0.725 |
| Laboratory data, M (Q1, Q3) |
| TB (mg/dL) | 19.57 (13.11–27.37) | 22.18 (14.96–30.06) | 17.77 (12.39–23.78) | <0.001 | 2.53 (1.29–5.72) | 4.48 (2.14–8.91) | 2.27 (1.25–5.15) | <0.001 |
| INR | 2.01 (1.55–2.61) | 2.09 (1.66–2.69) | 1.86 (1.44–2.50) | 0.002 | 1.34 (1.17–1.59) | 1.53 (1.32–1.80) | 1.31 (1.15–1.55) | <0.001 |
| Cr (mg/dL) | 0.85 (0.68–1.13) | 0.88 (0.72–1.27) | 0.82 (0.67–1.04) | 0.002 | 0.77 (0.63–0.89) | 0.81 (0.64–1.04) | 0.76 (0.63–0.88) | 0.028 |
| BUN (mmol/L) | 4.59 (3.27–7.27) | 5.23 (3.44–9.23) | 4.20 (3.11–6.15) | <0.001 | 4.50 (3.50–6.00) | 5.01 (3.73–7.33) | 4.40 (3.41–5.89) | 0.002 |
| ALT (U/L) | 154.60 (54.10–481.00) | 123.00 (53.00–388.00) | 180.60 (55.20–562.00) | 0.151 | 68.00 (30.00–308.00) | 42.20 (26.00–119.55) | 79.10 (30.40–347.00) | <0.001 |
| AST (U/L) | 169.60 (82.40–363.00) | 153.35 (87.00–342.00) | 182.00 (79.00–400.00) | 0.904 | 84.90 (44.00–231.00) | 73.00 (45.60–146.10) | 88.00 (43.80–246.00) | 0.170 |
| AST/ALT | 1.11 (0.71–1.74) | 1.21 (0.79–2.00) | 1.00 (0.69–1.62) | 0.002 | 1.16 (0.72–1.65) | 1.47 (1.05–2.06) | 1.10 (0.67–1.57) | <0.001 |
| ALB (g/L) | 30.80 (27.20–33.80) | 29.70 (25.80–33.10) | 31.11 (28.30–34.50) | 0.001 | 32.30 (28.19–36.80) | 28.20 (24.45–32.10) | 33.10 (29.10–37.70) | <0.001 |
| ALP (U/L) | 142.00 (109.00–189.00) | 143.50 (103.53–194.05) | 141.50 (112.00–186.00) | 0.797 | 120.00 (88.90–163.00) | 132.88 (92.00–166.00) | 119.00 (88.00–163.00) | 0.165 |
| γ-GT (U/L) | 79.15 (47.00–123.00) | 80.10 (52.00–138.95) | 78.45 (44.90–119.00) | 0.387 | 87.00 (35.00–166.00) | 82.00 (34.00–161.00) | 87.00 (36.00–167.00) | 0.457 |
| WBC (×109/L) | 6.66 (4.70–9.46) | 7.88 (5.70–11.37) | 5.80 (4.21–8.15) | <0.001 | 4.60 (3.31–6.07) | 5.95 (3.96–8.49) | 4.49 (3.17–5.89) | <0.001 |
| PLT (×109/L) | 100.00 (65.00–143.00) | 86.00 (60.00–131.00) | 105.00 (68.00–151.00) | 0.010 | 93.00 (58.00–151.00) | 77.00 (51.00–122.50) | 97.00 (60.00–154.00) | 0.003 |
| MAP (mmHg) | 89.00 (82.33–96.00) | 87.67 (81.00–94.33) | 90.33 (83.33–97.00) | 0.002 | 88.33 (83.33–95.33) | 87.00 (80.83–93.67) | 88.67 (83.33–95.67) | 0.040 |
| NLR ratio | 3.84 (2.41–6.39) | 4.92 (2.81–8.18) | 3.27 (2.26–5.16) | <0.001 | 2.11 (1.36–3.32) | 3.46 (2.07–6.20) | 1.98 (1.31–2.94) | <0.001 |
| HGB (g/L) | 119.00 (101.00–134.00) | 116.50 (98.00–129.00) | 123.00 (105.00–137.00) | 0.005 | 117.00 (98.00–135.00) | 109.50 (85.50–123.00) | 119.00 (99.00–136.00) | <0.001 |
| N (×109/L) | 4.45 (3.03–6.69) | 5.59 (3.86–8.93) | 3.85 (2.61–5.33) | <0.001 | 2.60 (1.79–3.81) | 3.77 (2.38–6.46) | 2.45 (1.70–3.55) | <0.001 |
| L (×109/L) | 1.18 (0.78–1.63) | 1.19 (0.74–1.65) | 1.17 (0.78–1.62) | 0.832 | 1.22 (0.81–1.77) | 1.11 (0.76–1.78) | 1.24 (0.83–1.77) | 0.240 |
| K + (mmol/L) | 3.83 (3.50–4.20) | 3.80 (3.39–4.16) | 3.90 (3.58–4.22) | 0.030 | 3.89 (3.57–4.15) | 3.80 (3.40–4.12) | 3.90 (3.60–4.15) | 0.135 |
| Na+ (mmol/L) | 136.55 (133.50–139.00) | 135.00 (131.30–137.90) | 137.50 (135.10–139.70) | <0.001 | 139.00 (136.90–141.05) | 137.10 (133.90–140.00) | 139.10 (137.00–141.40) | <0.001 |
| CRP (mg/L) | 12.60 (6.85–19.90) | 17.55 (11.40–30.40) | 9.64 (5.90–14.99) | <0.001 | 5.90 (2.80–11.61) | 13.35 (7.49–33.25) | 4.80 (2.50–10.30) | <0.001 |
| PCT (ng/mL) | 0.56 (0.28–1.03) | 0.71 (0.45–1.26) | 0.47 (0.23–0.77) | <0.001 | 0.17 (0.08–0.36) | 0.28 (0.15–0.56) | 0.14 (0.07–0.32) | <0.001 |
| PT (s) | 23.30 (17.80–29.80) | 24.75 (20.00–30.60) | 21.40 (16.80–28.00) | <0.001 | 15.60 (13.60–18.20) | 18.05 (15.55–20.70) | 15.20 (13.30–17.80) | <0.001 |
The distribution of bacterial infection types was similar between ACLF and non-ACLF patients. Pulmonary infection was the predominant type (ACLF: 70.71%, non-ACLF: 71.21%), followed by SBP (15.66%, 12.88%), other infections, urinary tract infection, skin and soft tissue infection, and spontaneous bacteremia. Among pathogens isolated via bacterial culture, Gram-negative bacteria (ACLF: 53.66%, non-ACLF: 57.58%) were the most prevalent, followed by Gram-positive bacteria (ACLF: 34.15%, non-ACLF: 24.24%) and mixed infections (Supplementary Table 2).
Plasma metabolic profiles altered by bacterial infection in ACLF and non-ACLF
Untargeted metabolomics identified 1,072 known metabolites, of which 176 detected in fewer than 20% of samples were excluded. Orthogonal partial least squares discriminant analysis (OPLS-DA) of the remaining 896 metabolites distinguished infected from uninfected patients in both the ACLF and non-ACLF groups (Supplementary Fig. 1). Consistent with prior findings, the plasma metabolome differed significantly between ACLF and non-ACLF patients.16 Hence, we analyzed the two subgroups separately in subsequent analyses. A total of 400 metabolites (44.6%) in the ACLF group and 481 (53.7%) in the non-ACLF group showed significant differences between infected and uninfected patients, respectively (Fig. 2A and B).
The metabolites formed four clusters (C1–C4) during disease progression (Fig. 2C). Metabolites in C1 (e.g., thyroxine) showed a sustained decrease with both bacterial infection and ACLF progression, whereas metabolites in C3 (e.g., pipecolate and 2-isopropylmalate) exhibited a sustained increase. Metabolites in C2 and C4 were associated with bacterial infection independent of ACLF status, characterized by elevated levels of C2 (e.g., orotate and ribitol) and reduced levels of C4 (e.g., sphingosine-1-phosphate (S1P)) in infected patients. These clusters were enriched in superpathways including lipids, amino acids, and xenobiotics (Supplementary Fig. 2).
Bacterial infection remodeled the ACLF plasma metabolome, particularly affecting pathways related to lipid biosynthesis, amino acid catabolism, and pyrimidine metabolism (Fig. 2D). Among the top 40 metabolites, 13 were lipids, including lysophospholipids, phospholipids, sphingomyelins, and sphingosines. These lipids were generally downregulated in patients with infection. Given their roles in maintaining membrane integrity and lipid-mediated immune signal transduction, these changes may reflect membrane remodeling or degradation during infection in ACLF. For example, reduced S1P levels have been associated with sepsis-induced liver injury.21 Another 13 metabolites belonged to amino acid metabolism. Among these, thyroxine was significantly downregulated, whereas others (e.g., pipecolate) were upregulated. Thyroxine can enhance leukocyte function during bacterial pneumonia, and hypothyroidism is a known risk factor for infection.22 Elevated levels of pipecolate reflected both infection and ACLF progression and indicated perturbed lysine metabolism.23 Of the remaining 14 metabolites, 6 nucleotides (e.g., orotate) and 2 carbohydrates (e.g., ribitol) were all upregulated.24 The other 6 metabolites consisted of 3 xenobiotics (e.g., 2-isopropylmalate, upregulated) and 3 cofactors and vitamins (e.g., gamma-CEHC, downregulated). KEGG enrichment analysis highlighted disturbed metabolic pathways in ACLF patients with bacterial infection (Supplementary Fig. 3), including pyrimidine metabolism, characterized by increased levels of 3-ureidopropionate, orotate, and thymine (Fig. 2D). Pyrimidine derivatives possess antimicrobial and antiviral properties and are modulated during host–pathogen interactions.24 Other core metabolic pathways were also perturbed, including pantothenate and CoA biosynthesis and beta-alanine metabolism. Pantothenate and its derivative CoA are central coenzymes in cellular metabolism. Bacteria can regulate virulence gene expression by modulating pantothenate synthesis via the beta-alanine pathway.25
We investigated the associations between the top 10 infection-related metabolites and key clinical indicators. In ACLF, metabolites (1-linoleoyl-GPE (18:2), thyroxine, pipecolate, S1P, and 2-isopropylmalate) were significantly correlated with inflammatory markers, including CRP, PCT, WBC, neutrophils, and NLR. In non-ACLF, a distinct set (3-ureidopropionate, orotate, ribitol, and 3-phenylpropionate) showed similar associations (Fig. 2E and F).
In summary, the disturbed plasma metabolome elucidates the link between infection-induced metabolic stress and immune homeostasis imbalance, emphasizing the potential of metabolite biomarkers for monitoring organ function and the interplay between host immunity and pathogen infection.
Model development and validation for bacterial infection in ACLF
Among patients who met the COSSH-ACLF criteria, concurrent bacterial infection was present in 136 of 315 in the discovery set and 62 of 136 in the validation set. Baseline clinical parameters are summarized in Supplementary Table 3.
In the initial exploration (Fig. 3A), the top 40 metabolites exhibited independent diagnostic value (AUCs: 0.650–0.773). Candidate metabolite biomarkers were screened for importance using the Boruta algorithm to identify a confirmed variable set (Fig. 3B). Subsequently, feature reduction was performed using RFE to determine the minimum number of variables required for stable performance (Supplementary Table 4). The final feature subset was further refined based on both the availability of chemical standards and statistical significance (Supplementary Fig. 4). Five metabolites (thyroxine, 2-isopropylmalate, pipecolate, 1-linoleoyl-GPE (18:2), and S1P) were identified as the optimal biomarkers.
Eight ML algorithms were trained using a five-metabolite panel. The CatBoost model demonstrated superior performance in the discovery set (AUC: 0.881 (95% CI: 0.844–0.918)) and validation set (AUC: 0.835 (95% CI: 0.769–0.902)) (Fig. 3C and Supplementary Table 5). The CatBoost model was well calibrated (Brier score: 0.145), confirming the reliability of its probability estimates (Supplementary Figs. 5 and 6). DCA showed that the CatBoost model provided a higher net benefit across a wide range of threshold probabilities than the other models (Fig. 3D).
Model development and validation for the diagnosis of bacterial infection in non-ACLF
Among the 863 non-ACLF patients, concurrent bacterial infection was found in 89 of 606 in the discovery set and 43 of 257 in the validation set. Baseline clinical parameters are summarized in Supplementary Table 6.
Individual metabolites showed moderate diagnostic value (AUCs: 0.689–0.784, Fig. 4A). The above steps were repeated to identify a six-metabolite panel (3-phenylpropionate, ribitol, 1-linoleoyl-GPC (18:2), orotate, S1P, and 1-carboxyethylphenylalanine) for model training (Fig. 4B, Supplementary Fig. 7, and Supplementary Table 7). Given the class imbalance of positive events, we applied the SMOTE method to the discovery set. Model validation was conducted on the unsampled validation set.
Among eight ML algorithms, the six-metabolite CatBoost model demonstrated superior overall performance in the discovery set (AUC: 0.935 (95% CI: 0.920–0.949)) and validation set (AUC: 0.889 (95% CI: 0.859–0.920)) (Fig. 4C and Supplementary Table 8). Calibration curves showed that the model was acceptable at reasonable thresholds, with a Brier score of 0.136 (Supplementary Fig. 8). DCA indicated that the model had good clinical utility (Fig. 4D).
Model comparison and sensitivity analysis
We reclassified patients according to the EASL-CLIF ACLF criteria and compared baseline parameters between infected and uninfected patients with ACLF (Supplementary Table 9). The performance of the ACLF bacterial infection model remained robust under this reclassification, with AUCs of 0.872 (95% CI: 0.828–0.916) in the discovery set (Fig. 5A) and 0.824 (95% CI: 0.743–0.906) in the validation set (Fig. 5B).
We used DeLong’s test and NRI to compare the five-metabolite model with conventional indicators (CRP and WBC) and with the combined model (five metabolites + CRP) (Fig. 5C). The discrimination of the five-metabolite model was better than that of conventional clinical indicators (AUCs: 0.665–0.691, all P < 0.05) and showed no significant difference compared with the combined model (AUC: 0.840 (95% CI: 0.774–0.905), P = 0.855; NRI: 0.008, P = 0.910) in the validation set (Supplementary Tables 10 and 11).
We compared the performance of the six-metabolite model with that of conventional indicators (CRP, neutrophils, or PCT) and the combined model (six metabolites + CRP) (Fig. 5D). The six-metabolite model outperformed conventional indicators (AUCs: 0.624–0.782, all P < 0.001) and showed a significant improvement in AUC but a non-significant NRI compared with the combined model (AUC: 0.929 (95% CI: 0.905–0.953), P < 0.001; NRI: 0.056, P = 0.082) in the validation set (Supplementary Tables 12 and 13). This discrepancy arose because DeLong’s test evaluates overall discrimination across all thresholds, whereas NRI assesses reclassification improvement at specific cutoffs.
To further explore the diagnostic information that metabolites add beyond the clinical model, we used RFE to evaluate all clinical indicators and selected the optimal combination (Supplementary Tables 14 and 15). Metabolomics models outperformed the best composite clinical model in both ACLF (AUC: 0.752 (0.670–0.834), P < 0.05) and non-ACLF (AUC: 0.823 (0.784–0.863), P < 0.05) patients in the validation set (Supplementary Table 16).
We conducted a sensitivity analysis using microbiologically confirmed cases. Pathogens were isolated from 82 of 198 ACLF cases and 66 of 132 non-ACLF cases. Models for bacterial infection in ACLF and non-ACLF remained robust in the discovery (AUCs: 0.948 and 0.962, respectively) and validation sets (AUCs: 0.869 and 0.881, respectively) (Supplementary Table 17).
We analyzed model performance across ACLF grades. Of the 451 patients with ACLF, 311 had ACLF-1, 110 had ACLF-2, and 30 had ACLF-3. Model performance varied slightly across grades and remained generally robust. The AUCs for the discovery and validation sets were 0.880 and 0.824 for ACLF-1, 0.947 and 0.830 for ACLF-2, and 0.961 and 0.859 for ACLF-3, respectively (Supplementary Table 18).
Baseline metabolic risk stratification for 90-day outcomes
During the 90-day follow-up, the bacterial infection group had significantly higher 90-day mortality than the non-infection group in both ACLF (infection: 45.45%, non-infection: 22.53%) and non-ACLF (infection: 12.12%, non-infection: 3.69%) populations (all P < 0.0001, Supplementary Fig. 9). We employed the win ratio method to assess the risk stratification capacity of the diagnostic model for 90-day outcomes. In this approach, for each patient pair, the first step was to determine whether the high-risk patient died before the low-risk patient. Pairs that remained tied were then compared for non-fatal events in a prespecified order of clinical importance.20
Patients with ACLF in the validation set were stratified into high infection-related risk (n = 73) and low infection-related risk (n = 63) groups based on the model’s best threshold of 0.357. This stratification yielded a recall of 0.839, a specificity of 0.716, and a negative predictive value of 0.841. Using the win ratio method to analyze the hierarchical composite endpoint comprising all-cause death, organ failure, sepsis, new-onset acute decompensation, and SIRS, the high-risk group had more wins than the low-risk group (50.4% vs. 24.9%, win ratio: 2.03 (95% CI: 1.21–3.39), P = 0.007, Fig. 6). This result was primarily driven by death, organ failure, and sepsis, indicating that patients in the high-risk group had worse 90-day outcomes than those in the low-risk group. Reordering of the hierarchical components did not affect the robustness of the analysis (Supplementary Table 19).
Non-ACLF patients in the validation set were stratified into high infection-related risk (n = 83) and low infection-related risk (n = 174) groups based on the model’s best threshold of 0.378. This stratification yielded a recall of 0.818, a specificity of 0.772, and a negative predictive value of 0.954. The high-risk group also had worse 90-day outcomes than the low-risk group (win ratio: 7.13 (95% CI: 3.66–13.89), P < 0.001, Supplementary Table 20).
Model interpretation
SHAP analysis identified thyroxine, followed by 2-isopropylmalate, pipecolate, 1-linoleoyl-GPE (18:2), and S1P, as the top predictors in the ACLF infection model (Supplementary Fig. 10A). In the non-ACLF infection model, 3-phenylpropionate, ribitol, and 1-linoleoyl-GPC (18:2) were the top three contributors (Supplementary Fig. 10B).
The SHAP waterfall plots illustrated how individual predictions were generated. For an ACLF patient who was correctly diagnosed with infection, decreased 1-linoleoyl-GPE (18:2), S1P, and thyroxine, along with increased pipecolate, pushed the prediction toward infection, whereas 2-isopropylmalate pushed it away from infection (Supplementary Fig. 10C). In non-ACLF, infection prediction was driven by positive contributions from orotate, 1-carboxyethylphenylalanine, S1P, 1-linoleoyl-GPC (18:2), and 3-phenylpropionate, while ribitol contributed negatively (Supplementary Fig. 10D).
SHAP dependence plots showed how the model output varied with the log2-transformed relative abundance of metabolites. In ACLF, log2 (relative abundance) thresholds of thyroxine < 18.66, S1P < 18.71, 1-linoleoyl-GPE (18:2) < 23.06, pipecolate > 28.05, and 2-isopropylmalate > 19.58 pushed classification toward infection (Supplementary Fig. 10E). In non-ACLF, thresholds associated with infection were S1P < 19.40, 3-phenylpropionate < 19.14, 1-linoleoyl-GPC (18:2) < 26.20, orotate > 21.39, ribitol > 19.33, and 1-carboxyethylphenylalanine > 19.60 (Supplementary Fig. 10F).
Discussion
This study delineated plasma metabolic alterations associated with bacterial infection in patients with AoCLD and identified key diagnostic biomarkers through metabolomics profiling. ACLF and non-ACLF patients have distinct plasma metabolomic profiles.16 To precisely address the different clinical needs of this heterogeneous population, we developed a disease-stage-specific early diagnostic framework. The framework outperformed conventional infection indicators (CRP, WBC, and PCT) and the best composite clinical model. As HBV patients comprised the highest proportion of our cohort, we initially used the COSSH-ACLF criteria to define ACLF and then used the EASL-CLIF ACLF criteria to validate these biomarkers.17
Consistent with previous studies, the five metabolites in our model were associated with bacterial infection in ACLF. Thyroxine regulates innate and adaptive immunity, and its serum concentration decreases during severe infections.22 Low thyroxine levels are predictors of hepatic encephalopathy in patients with cirrhosis.26 2-Isopropylmalate is a key intermediate in bacterial biosynthesis of leucine. We found that its levels correlated with CRP and NLR. Pipecolate, a catabolite of lysine, exhibits elevated levels associated with infection and liver injury.23 1-linoleoyl-GPE (18:2) is a lysophosphatidylethanolamine (LysoPE). LysoPEs can support gut homeostasis and barrier function and possess antimicrobial activity.27 S1P is a lipid signaling molecule linked to inflammation. A previous study found significantly reduced serum S1P levels in sepsis, consistent with our findings.21 In addition, we observed significant correlations between S1P levels and markers of inflammation and immunity, including WBC, CRP, neutrophils, and albumin.
The 6-metabolite non-ACLF model shares S1P with the ACLF model. The other five metabolites reflect the interplay between bacterial infection and host antimicrobial defense. 1-linoleoyl-GPC (18:2) is a lysophosphatidylcholine (LysoPC) produced by phospholipase-mediated cleavage of phosphatidylcholine, which has been shown to prevent microbial infection and treat sepsis.28 Consistently, we observed significant downregulation of various LysoPCs in infected patients. 3-phenylpropionate promotes intestinal epithelial barrier formation and exhibits antimicrobial activity.29 Ribitol is a Gram-positive bacterial cell wall component. Increased orotate reflects pyrimidine pathway perturbation due to bacterial metabolism.24 1-carboxyethylphenylalanine, a phenylalanine derivative, indicates elevated oxidative stress.30
The panel of metabolites used in the diagnostic models for bacterial infection differs between ACLF and non-ACLF patients. This suggests that liver function status may influence the diagnostic performance of these markers. A possible explanation is that severe liver impairment compromises biosynthetic and immune functions. The model performance showed mild differences among ACLF grades but remained robust overall (AUCs: 0.824–0.859 in the validation set). This indicates its potential applicability across ACLF populations of varying severity. Exploring the heterogeneity of host metabolic responses to bacterial infection across liver function statuses and extrahepatic organ failure conditions is a promising research direction. Metabolomic profiles may also differ by infection type or pathogen. In our cohort, metabolic differences between the two most common infection types (pulmonary infection and SBP) were limited (Supplementary Tables 21 and 22). A previous study identified species-specific metabolic signatures of Acinetobacter baumannii and Klebsiella pneumoniae infections in ACLF, indicating that metabolomic profiles may vary by pathogen.31 However, whether pathogens directly affect metabolism remains unclear. In the future, models can be developed for populations with specific infection types or pathogens, and animal models can be used to explore their impact on metabolism. Class imbalance can bias models toward the majority class. To avoid overfitting and noise, we did not apply sampling techniques to the ACLF group, which was relatively balanced (prevalence: 0.439). In contrast, the non-ACLF group was imbalanced (prevalence: 0.153), so we applied SMOTE only to its discovery set.32
When analyzing composite endpoints for prognosis, we usually focus on multiple clinical outcomes and prioritize more severe events. Traditional approaches often treat all clinical events as equally important and focus only on the first-occurring endpoint. These methods are inconsistent with real-world clinical scenarios. For example, subsequent fatal events (death) may be overlooked when patients experience non-fatal prognostic events (e.g., sepsis). We used the win ratio method to emphasize differences in the clinical importance of infection-related outcomes.20 All-cause death was designated as the most important endpoint. To mitigate the impact of tied events, we calculated win odds; the results also supported our conclusions.
This study has several advantages. First, it was based on a large-scale, prospective multicenter cohort that is representative of the Chinese AoCLD population, in which HBV was the predominant etiology. Second, the ACLF model showed consistent diagnostic performance when applied according to the two major ACLF criteria. Third, it accounted for plasma metabolome heterogeneity during disease progression and identified novel metabolic biomarkers for bacterial infection in ACLF. Finally, risk stratification based on baseline metabolic levels was aligned with 90-day outcomes, demonstrating its potential for prognostic management.
This study also has several limitations. First, as a proof-of-concept investigation, it provided potential biomarkers for clinical practice; however, the models were not externally validated in an independent cohort, which may limit the generalizability of our findings across different populations and clinical settings. Second, our model performed well in the HBV-predominant ACLF population but requires further validation in cohorts where other etiologies, such as alcohol-related liver disease, predominate. Third, future studies should validate how metabolic biomarkers relate to the progression of ACLF or infection through interventional animal studies or clinical trials. Finally, as the definition of ACLF is not yet uniform, these findings require further external validation across multiple ACLF definitions in more recent cohorts.
Given a 90-day mortality of 45.45% in ACLF patients with bacterial infection, early and accurate diagnosis is critical. Our model demonstrated moderate to high discriminatory ability and provided a complementary tool that outperformed conventional indicators, including CRP, WBC, and the best composite clinical model. Using a threshold of 0.357, patients classified as high-risk by the model had significantly worse 90-day outcomes than those classified as low-risk, indicating the early warning value of this stratification approach. Early antimicrobial treatment in high-risk patients identified by these biomarkers may improve survival and reduce unnecessary empirical antibiotic use. Building on our prior work in translating metabolic biomarkers into a practical test,16 future efforts will focus on developing rapid diagnostic tests based on targeted quantitative assays and integrating them into a clinical decision-support system for individualized management.
Supporting information
Supplementary File 1
Supplementary Methods.
(PDF)
Supplementary Table S1
Hyperparameters for eight ML models.
(PDF)
Supplementary Table S2
Type distribution of bacterial infection.
(PDF)
Supplementary Table S3
Comparison of baseline clinical parameters between infected and uninfected patients with ACLF in the discovery set and validation set.
(PDF)
Supplementary Table S4
The predictive performance for infection risk in ACLF with incrementally reduced feature subsets evaluated by RF-RFE.
(PDF)
Supplementary Table S5
Performance of eight ML models to predict bacterial infection in ACLF.
(PDF)
Supplementary Table S6
Comparison of baseline clinical parameters between infected and uninfected non-ACLF patients in the discovery set and validation set.
(PDF)
Supplementary Table S7
The predictive performance for infection risk in non-ACLF with incrementally reduced feature subsets evaluated by RF-RFE.
(PDF)
Supplementary Table S8
Performance of eight ML models to predict bacterial infection in non-ACLF.
(PDF)
Supplementary Table S9
Comparison of baseline clinical parameters between infected and uninfected patients with EASL-CLIF ACLF.
(PDF)
Supplementary Table S10
Comparison of AUC and NRI between ML models and conventional indicators for bacterial infection in ACLF.
(PDF)
Supplementary Table S11
Performance of ML models and conventional indicators to predict bacterial infection in ACLF in the validation set.
(PDF)
Supplementary Table S12
Comparison of AUC and NRI between ML models and conventional indicators for bacterial infection in non-ACLF in the validation set.
(PDF)
Supplementary Table S13
Performance of ML models and conventional indicators to predict bacterial infection in non-ACLF in the validation set.
(PDF)
Supplementary Table S14
The performance for composite clinical models to predict infection risk in ACLF with incrementally reduced feature subsets evaluated by RF-RFE.
(PDF)
Supplementary Table S15
The performance for composite clinical models to predict infection risk in Non-ACLF with incrementally reduced feature subsets evaluated by RF-RFE.
(PDF)
Supplementary Table S16
Comparison of AUC and NRI between metabolomics model and the best composite clinical model in the validation set.
(PDF)
Supplementary Table S17
Performance of metabolomics models to predict bacterial infection using microbiologically confirmed cases.
(PDF)
Supplementary Table S18
The diagnostic performance of metabolomics models in patients with different ACLF grades.
(PDF)
Supplementary Table S19
Evaluation of component ranking variations in the hierarchical composite endpoint using win ratio analysis in ACLF.
(PDF)
Supplementary Table S20
Win ratio analysis to evaluate the hierarchical composite endpoint in non-ACLF.
(PDF)
Supplementary Table S21
Top 10 differential metabolites between pulmonary infection and Spontaneous bacterial peritonitis in ACLF patients.
(PDF)
Supplementary Table S22
Top 10 differential metabolites between pulmonary infection and Spontaneous bacterial peritonitis in non-ACLF patients.
(PDF)
Supplementary Fig. S1
Orthogonal partial least squares discriminant analysis (OPLS-DA).
(PDF)
Supplementary Fig. S2
Distribution of metabolites across super pathways within distinct metabolic clusters.
(PDF)
Supplementary Fig. S3
KEGG metabolic pathway enrichment analysis of significant differential metabolites between infected and uninfected ACLF patients.
(PDF)
Supplementary Fig. S4
Relative abundance of five diagnostic variables for bacterial infection in the ACLF cohort.
(PDF)
Supplementary Fig. S5
Calibration curve of the five-metabolite model trained by eight ML models in the validation set.
(PDF)
Supplementary Fig. S6
Calibration curve of the five-metabolite Catboost model.
(PDF)
Supplementary Fig. S7
Relative abundance of six diagnostic variables for bacterial infection in the non-ACLF cohort.
(PDF)
Supplementary Fig. S8
Calibration curve of the six-metabolite model trained by eight ML models in the validation set.
(PDF)
Supplementary Fig. S9
Cumulative incidence of all-cause death in infected and noninfected patients.
(PDF)
Supplementary Fig. S10
Interpretation of bacterial infection models using SHAP method.
(PDF)