Development and validation of clinical phenotypes in sepsis-associated liver injury using multi-algorithm consensus classification: a retrospective cohort study.

Development and validation of clinical phenotypes in sepsis-associated liver injury using multi-algorithm consensus classification: a retrospective cohort study - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Infect Dis . 2026 Apr 18;26:1050. doi: 10.1186/s12879-026-13131-x Development and validation of clinical phenotypes in sepsis-associated liver injury using multi-algorithm consensus classification: a retrospective cohort study Xiaoman Xie 1, 2, 3 , ZhiXian Zhu 1, 2, 3 , XinYu Cao 1, 2, 3 , Qiong Pan 1, 2, 3 , Dong Zhao 1, 2, 3, ✉ , Jin Chai 1, 2, 3, ✉ Author information Article notes Copyright and License information 1 Department of Gastroenterology, Institute of Digestive Diseases of PLA, The First Affiliated Hospital (Southwest Hospital) of Third Military Medical University (Army Medical University), Chongqing, 400038 China 2 Cholestatic Liver Diseases Center, The First Affiliated Hospital (Southwest Hospital) of Third Military Medical University (Army Medical University), Chongqing, 400038 China 3 Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD) Medical Research Center, The First Affiliated Hospital (Southwest Hospital) of Third Military Medical University (Army Medical University, Chongqing, 400038 China ✉ Corresponding author. Received 2025 Nov 6; Accepted 2026 Mar 16; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13224431  PMID: 42001043 Abstract Background Sepsis-associated liver injury (SALI) is a heterogeneous syndrome with high mortality and pathophysiological heterogeneity. Current prognostic models are limited in their ability to capture the heterogeneity in pathophysiology, highlighting the demand for improved stratification of patient subgroups and the development of personalized therapeutic strategies. Methods We performed a retrospective analysis of ICU patients diagnosed with SALI from 2001 to 2022. The research included 5,755 adult patients obtained from four databases, utilizing the eICU-CRD cohort ( n = 3,893) for initial exploration and three independent cohorts (NWICU-1.0, MIMIC-III, MIMIC-IV; n = 1,862) for subsequent validation. We developed and utilized a novel Multi-algorithm Consensus-based Sepsis-Associated Liver Injury Classification (MCSALIC) method that integrates ten clustering algorithms. This technique identified four distinct SALI phenotypes. phenotypes were subsequently validated and characterized through XGBoost-based SHAP analysis and survival analysis, which elucidated essential distinguishing features and variability in treatment response. Results phenotypes A (47.9%, liver preservation), B (22.6%, hyperbilirubinemia), C (16.7%, hepatocellular dysfunction), and D (12.8%, cholestasis) were delineated. In the development cohort, phenotypes C and D had higher 30-day mortality rates than phenotype A (HR = 1.44 and 1.32, respectively). There was no significant difference between phenotypes A and B. In SALI, therapeutic responses were specific to phenotypes. Patients with phenotype A experienced advantages from early antibiotic treatment alongside moderate fluid resuscitation. Patients with phenotype B benefited from immediate antibiotics, limited fluids, and low-dose norepinephrine. Patients with phenotype C benefited from a moderate fluid range, low-dose norepinephrine, and relatively lenient antibiotic initiation. Patients with phenotype D benefited from high-volume fluid resuscitation, high-dose norepinephrine administration, and relatively lenient antibiotic timing. Conclusions SALI consists of four distinct phenotypes, each characterized by unique pathophysiology, prognostic outcomes, and treatment responses. This stratification, strongly validated across international cohorts, sets the stage for precision medicine and calls into question the current one-size-fits-all approach to treatment. Clinical tiral number Not applicable. Supplementary Information The online version contains supplementary material available at 10.1186/s12879-026-13131-x. Keywords: Sepsis, SALI, Phenotype, Phenotype, Multi-algorithm consensus clustering Introduction Sepsis is a life-threatening condition defined by a dysregulated host response, manifesting as systemic inflammation, immune and metabolic dysfunction, and consequent multi-organ failure [ 1 – 2 ]. The pathogenesis of sepsis, initiated by microbial invasion and immune dysfunction, triggers a systemic cascade involving cytokine release, increased vascular permeability, and microcirculatory impairment, which can progress to multi-organ failure [ 3 – 5 ]. The liver, owing to its central role in metabolism, detoxification, and immunity, is particularly vulnerable to these insults [ 6 ]. This vulnerability is reflected epidemiologically, as sepsis-associated liver injury (SALI) develops in 34–46% of septic patients [ 7 ]. Moreover, SALI carries a grave prognosis, with a mortality rate of 54–68% that surpasses even that of frequent failures such as pulmonary dysfunction [ 8 ]. Significant molecular heterogeneity exists in SALI, encompassing variations in inflammation, microcirculation, and metabolism, yet is encompassed under a common clinical definition [ 9 – 11 ]. The spectrum of SALI severity, driven by mechanistic heterogeneity, ranges from localized inflammation to multiorgan failure. This heterogeneity in pathophysiology and clinical manifestation mandates improved subphenotyping to enable individualized therapeutic strategies [ 12 – 14 ]. Unsupervised machine learning (UML) is progressively utilized to characterize clinical heterogeneity, demonstrated by its efficacy in identifying illness phenotypes, including acute-on-chronic liver failure (ACLF) and acute respiratory distress syndrome (ARDS) [ 15 – 17 ]. Nevertheless, their application in SALI continues to be limited. Although Yang et al. [ 16 ] identified three distinct subtypes of acute liver failure (ALF) using a multiple-algorithm consensus clustering approach—revealing heterogeneity in prognosis and treatment responses and thereby providing an important framework for liver disease subtyping—this classification system may not be directly applicable to sepsis-associated liver injury (SALI). This is due to fundamental differences in pathophysiology: SALI arises secondary to sepsis, a generalized infectious process involving more complex immune-inflammatory responses, microcirculatory dysfunction, and gut–liver axis disturbances, whereas ALF typically results from primary hepatocyte necrosis or toxic injury [ 16 ]. Consequently, the direct application of the ALF subtyping framework to SALI is likely inappropriate. Moreover, the heterogeneity of SALI has not yet been systematically investigated, and there is currently a lack of precision stratification studies focused specifically on SALI. In this study, we used a computational phenotyping method that used data from several electronic health records (EHRs) to find different SALI phenotypes, which solved the problem. These subgroups exhibit varied clinical trajectories and distinct responses to conventional ICU care. Our results offer a foundation for the management of SALI according to distinct phenotypes. This study makes three important contributions: We created and tested the Multi-algorithm Consensus-based Sepsis-Associated Liver Injury Classification (MCSALIC) system. This system used ten different clustering algorithms to reliably and consistently find SALI phenotypes in clinical data. Second, we identified four SALI phenotypes with distinct pathophysiological profiles and outcomes, and we confirmed their generalizability through external validation in three independent ICU databases (NWICU-1.0, MIMIC-III-CareVue, and MIMIC-IV-3.1). Third, we examined the variability of treatment effects, with a specific emphasis on the most effective approaches for antibiotic timing, fluid management, and vasopressor administration among various phenotypes. Methods Data sources We analyzed data from four critical care databases. The discovery cohort comprised patients from the eICU Collaborative Research Database (eICU-CRD) [ 18 ], a multicenter US repository of over 200,000 ICU admissions. For external validation, we used three independent databases: (1) The Medical Information Mart for Intensive Care (MIMIC)-III CareVue subset [ 19 ] (patient records from 2001 to 2008). (2) MIMIC-IV version 3.1 [ 20 ], a comprehensive, modular update of MIMIC-III. (3) The Northwestern ICU (NWICU) database, version 1.0 [ 21 ], which contains de-identified data from over 25,000 admissions (2020–2022), including extensive COVID-19 information, and was built using a MIMIC-compatible architecture. SALI identification criteria To identify Sepsis-Associated Liver Injury (SALI) based on pathophysiological changes, we applied the modified Drug-Induced Liver Injury (DILI) criteria within the first seven days following sepsis diagnosis (defined as the initial fulfillment of Sepsis-3 criteria) [ 21 ]. SALI was defined by meeting any one of the following conditions: (1) Alanine aminotransferase (ALT) ≥ 5 times the upper limit of normal (ULN). (2) Alkaline phosphatase (ALP) ≥ 2 times ULN. (3) Total bilirubin ≥ 2.0 mg/dL concurrent with ALT ≥ 3 times ULN. (4) Total bilirubin ≥ 2.0 mg/dL concurrent with ALP ≥ ULN. Data preprocessing Sepsis was diagnosed at ICU admission for all included patients. All baseline characteristics were assessed at sepsis diagnosis. Clinical variables including demographics, vital signs, laboratory values, interventions, and comorbidities were extracted. We didn’t include variables that had more than 20% of their data missing. Using Multiple Imputation by Chained Equations (MICE) with five imputations, we filled in the rest of the missing data. For continuous variables, we used predictive mean matching, and for categorical variables, we used logistic or multinomial regression [ 16 , 22 ]. In this study, patients with pre-existing liver disease were excluded to minimize cohort heterogeneity. This exclusion encompassed cirrhosis, hepatic neoplasms, chronic medical liver diseases (e.g., chronic hepatitis, autoimmune liver diseases), and active hepatobiliary conditions (e.g., cholecystitis, liver abscess, cholangitis), irrespective of surgical status. We also excluded post-operative hepatobiliary patients (including those returning from theatre or ward after hepatobiliary surgery), patients with a history of liver transplant (especially in the immediate post-transplant period), and those with prior cholecystectomy. These criteria aimed to define a SALI cohort primarily comprising sepsis patients without significant pre-existing or concurrent hepatobiliary disorders, thereby focusing on new-onset liver injury attributable to sepsis. Clustering analysis This integrative methodology employed ten distinct clustering algorithms [ 16 , 23 , 24 ] to formulate a robust consensus classification, including K-Means, Gaussian Mixture, Agglomerative Clustering, MiniBatch K-Means, BIRCH, DBSCAN, COCA, NEMO, Consensus K-Means, and PINSPlus. All 10 algorithms performed clustering based on the uniformly determined K = 4. Validation of cluster assignments We evaluated the stability of the identified clusters by comparing the original MCSALIC phenotypes with those generated by two alternative machine learning algorithms: Least Absolute Shrinkage and Selection Operator (LASSO) and Support Vector Machine (SVM). We used Cohen’s kappa coefficient to see how the methods agreed with each other, and heatmaps to show this. We tested the SVM and LASSO models made for the discovery cohort (eICU-CRD) on three different critical care databases: MIMIC-III-CareVue, MIMIC-IV-3.1, and NWICU-1.0. In the validation cohorts, analyses involved evaluating the concordance between SVM and LASSO predictions and comparing the clinical profiles and outcomes of the identified phenotypes. Feature importance and SHAP analysis We utilized extreme gradient boosting (XGBoost) within a one-versus-rest classification framework to discern the differentiating features of each phenotype. Employing SHAP, which clarifies the influence of features on prediction results, enhanced the model’s interpretability. We utilized tailored waterfall charts, beeswarm distributions, and relevance rankings to depict the performance of each sub-phenotype. This technique allowed us to systematically assess the degree and manner in which each feature distinguished the phenotypes. Analysis of heterogeneous treatment response The discovery and validation cohorts were amalgamated to examine the differential responses of phenotypes to conventional therapies. These strategies encompassed the timing of administering antibiotics, fluids, and norepinephrine. After that, multi-predictor partial dependence plots (PDPs) were used to show how the amount and length of treatment affected the risk of death. We noted how long it took to start antibiotics after being admitted to the ICU, how much norepinephrine was given, and how much fluid was taken in during the first six hours of ICU admission. A random forest classifier was created for each phenotype to predict death in the ICU based on common clinical features and therapies. Statistical analysis and software Descriptive statistics for continuous variables are presented as the median [IQR] and analyzed using the Kruskal-Wallis test. The Chi-square or Fisher’s exact test is used to compare counts of categorical variables. We used R 4.2.0 for all of the statistical tests, and a p-value of less than 0.05 was considered statistically significant. Results Baseline characteristics This research encompassed 5,755 patients with SALI from U.S. databases between 2001 and 2022 (Fig. 1 ). The eICU-CRD cohort ( n = 3893) served as the discovery cohort, accompanied by three external validation cohorts: NWICU-1.0 ( n = 532), MIMIC-III CareVue ( n = 52), and MIMIC-IV-3.1 ( n = 1,278). To maintain data integrity, the subsequent exclusion criteria were established: pediatric patients, ICU stays shorter than 24 h, non-first ICU admissions, and individuals with pre-existing hepatic disorders. The aggregated cohort ( n = 5,755) had varied clinical features reflective of many data sources. The median age was 64.3 years, with 55.2% of the patients being male. Significant inter-database disparities were seen in critical physiological and laboratory parameters, such as temperature, blood pressure, and electrolyte levels ( p < 0.05 for the majority of comparisons), indicating variations in patient demographics and data sources. Hospital mortality rates exhibited considerable variation among the datasets, from 29.9% in the eICU-CRD cohort to 69.3% in the MIMIC-IV-3.1 subset ( p = 0.002). Laboratory trends exhibited variability, with indications like salt and liver enzymes showing significant variation ( p < 0.001)(Supplementary Table S1 ). Fig. 1. Open in a new tab Inclusion exclusion flow chart for the study Determination of the optimal number of SALI phenotypes We performed an extensive cluster analysis using various validation metrics to determine the ideal number of phenotypes in SALI. The elbow method was used to find the best number of clusters, and it clearly showed that K = 4 was the best (Fig. 2 a). The Calinski-Harabasz index confirmed the best number of clusters by reaching a clear maximum at K = 4 (1782) (Fig. 2 b). The Davies-Bouldin index reached its minimum at K = 4 (0.891), signifying an ideal cluster configuration (Fig. 2 c). The multi-algorithm consensus methodology exhibited significant resilience, as pairwise Adjusted Rand Index (ARI) comparisons indicated varied, yet predominantly robust, concordance among the 10 algorithms (Fig. 2 d-e). All 10 algorithms performed clustering based on the uniformly determined K = 4, and the robust agreement among ten clustering algorithms through ARI analysis highlights the dependability of the multi-algorithm consensus methodology. Fig. 2. Open in a new tab Validation of SALI phenotypes by clustering analysis. ( a ) The elbow method plot identified K = 4 as the optimal number of clusters based on the inflection point in the within-cluster sum of squares. ( b ) The Calinski-Harabasz index peaked at K = 4, indicating maximal inter-cluster separation. ( c ) The Davies-Bouldin index was minimised at K = 4, reflecting the least inter-cluster similarity. ( d ) Performance of the ten clustering algorithms was plotted as a mean pairwise Adjusted Rand Index (ARI) against consensus agreement. ( e ) Concordance of patients’ cluster assignments generated by the ten algorithms. All analyses were performed using data from the development cohort Clinical phenotypes of four SALI phenotypes A comparative examination of phenotype-specific clinical features and outcomes was performed across the discovery and validation cohorts (Supplementary Tables S2 and S3 ). We also demonstrated associations between baseline variables and phenotypes (Fig. 3 a), with chord diagrams illustrating abnormal clinical variables stratified by phenotype (Fig. 3 b-f). Each sub-phenotype is depicted in an individual column, with variables organized vertically to enhance comparison. We conducted a SHAP analysis using gradient boosting models to identify the clinical variables that differentiated the four SALI phenotypes and predicted sub-phenotype membership (Fig. 4 ). This method uncovered unique feature significance patterns that corresponded to the clinical phenotypes identified via unsupervised clustering. Examination of the clinical characteristics of the four SALI phenotypes demonstrated unique phenotype profiles, highlighting considerable disparities in hemodynamic state, organ dysfunction, and disease severity. In the results, all metrics are reported as two values in the order of the development cohort (eICU-CRD) and pooled validation cohort (NWICU-1.0, MIMIC-III, MIMIC-IV), separated by commas. Phenotype A (47.9%, 40.4%): Mild liver injury. This phenotype is designated as mild liver injury owing to mild alterations in all hepatic function indicators. It was characterized by low levels of bilirubin (TBil: 1.36 mg/dL, 0.94 mg/dL) and transaminases (ALT: 31.9 U/L, 25.2 U/L; AST: 42.0 U/L, 36.7 U/L) in development and validation cohorts. Moderate renal impairment was noted (creatinine: 1.98 mg/dL, 2.06 mg/dL), alongside stable haemodynamics (SBP: 109.0 mmHg, 110.6 mmHg; DBP: 60.6 mmHg, 56.8 mmHg) and a comparatively low hospital mortality rate (27.4%, 59.5%). Phenotype B (22.6%, 19.6%): Mixed liver injury. This phenotype is defined as mixed liver injury, with the validation cohort and the development cohort showing a prominent mixed injury trend with concurrent elevation of hepatocellular and cholangiolar injury markers. It was characterized by elevated transaminases (ALT: 91.1 U/L, 141.0 U/L; AST: 118.8 U/L, 190.2 U/L), hyperbilirubinemia (TBil: 1.10 mg/dL, 2.26 mg/dL) and raised alkaline phosphatase (ALP: 252.2 U/L, 167.5 U/L) across both cohorts. Severe renal impairment was observed (creatinine: 2.24 mg/dL, 2.20 mg/dL), accompanied by haemodynamic instability (SBP: 108.8 mmHg, 106.4 mmHg; DBP: 60.8 mmHg, 58.3 mmHg). Phenotype C (16.7%, 14.4%): Hepatocellular liver injury. This phenotype meets the clinical diagnostic criterion for hepatocellular liver injury, defined as alanine aminotransferase ≥ 5 times the upper limit of normal. It was characterised by markedly elevated transaminase levels (ALT: 442.3 U/L, 402.3 U/L; AST: 572.9 U/L, 569.6 U/L) in development and validation cohorts. Additional notable features included renal impairment (creatinine: 2.36 mg/dL, 2.57 mg/dL), hyperglycemia (glucose: 151.0 mg/dL, 154.9 mg/dL), and metabolic acidosis (pH: 7.31, 7.29), with notable elevated hospital mortality rates (33.7%, 69.8%). Phenotype D (12.8%, 25.6%): Cholestatic liver injury. This phenotype aligns with the clinical diagnostic criteria for cholestatic liver injury, defined as total bilirubin ≥ 2.0 mg/dL and alkaline phosphatase ≥ the upper limit of normal. It was characterized by significantly elevated total bilirubin (TBil: 5.26 mg/dL, 2.21 mg/dL) and cholestatic features with elevated alkaline phosphatase (ALP: 330.1 U/L, 268.2 U/L) in development and validation cohorts. Renal impairment (creatinine: 2.21 mg/dL, 2.14 mg/dL), hyperglycemia (glucose: 134.4 mg/dL, 152.3 mg/dL), and stable haemodynamics (SBP: 108.1 mmHg, 106.9 mmHg) were also observed in this phenotype. Fig. 3. Open in a new tab Abnormal Clinical Variables by Phenotype. ( a ) Pearson correlations are shown between variables and phenotypes. The right triangle displays inter-variable associations (purple: positive correlation; yellow: negative correlation), with square size indicating the magnitude of association strength. The lower-left section shows correlations between phenotypes and clinical features, where colour denotes the strength of association and marker width corresponds to statistical significance. ( b )-( f ) Chord diagrams showing abnormal clinical variables by phenotype. In ( b ), the ribbons connect from an phenotype to the clinical variables if the group mean is greater or lesser than the overall mean for the cohort. In ( c )-( f ), each phenotype is highlighted separately and the ribbons connect to the different patterns of clinical variables on the top of the circle. All analyses were performed using data from the development cohort Fig. 4. Open in a new tab Interpretation of phenotype-defining features by machine learning. ( a - d ) Beeswarm plots for phenotypes A-D, illustrating the top 20 most significant features arranged by mean absolute SHAP value. Each point is a distinct patient; its location on the x-axis denotes the feature’s influence on the forecast, while its color reflects the feature value (red: high, blue: low). ( e - h ) Waterfall plots illustrate the decision-making process for a typical patient per phenotype, identified by median core clinical features. The plots show how distinct features influence the model output from baseline to final prediction (red bars push toward the phenotype; grey bars push away) External validation of SALI subtypes across multiple databases To evaluate the generalizability of the SALI phenotypes, we externally validated them using LASSO and SVM classifiers across three critical care databases (NWICU-1.0, MIMIC-III-CareVue, and MIMIC-IV-3.1). In the discovery cohort (eICU-CRD), both algorithms showed strong agreement with the original MCSALIC classification (LASSO κ = 0.966; SVM κ = 1.0) (Fig. 5 a-b). High inter-algorithm concordance was maintained across all datasets: eICU-CRD (κ = 0.966), NWICU-1.0 (κ = 0.872), MIMIC-IV-3.1 (κ = 0.791), and MIMIC-III-CareVue (κ = 0.879) (Fig. 5 c-f). The non-significant p-value ( p = 0.513) in MIMIC-III-CareVue likely reflects a limited sample size rather than poor agreement given the high κ value. These results confirmed the reproducibility of the phenotype across diverse clinical settings, and the clinical significance of the phenotypes was further validated by survival analysis. LASSO-predicted phenotypes reliably differentiated hospital mortality across all external cohorts ( p < 0.001). Phenotype C demonstrated the highest mortality followed by phenotype D, whereas phenotype A displayed the most advantageous outcomes (Fig. 6 ). Fig. 5. Open in a new tab Algorithm concordance in phenotype identification. Concordance between the primary MCSALIC classification and two alternative algorithms, LASSO and SVM, was assessed using Cohen’s kappa. ( a ), ( b ) Agreement in the eICU-CRD discovery cohort between MCSALIC and (a) LASSO or ( b ) SVM. ( c )-( f ) Concordance between LASSO and SVM predictions across the four independent validation cohorts: ( c ) eICU-CRD, ( d ) NWICU-1.0, ( e ) MIMIC-IV-3.1, and ( f ) MIMIC-III-CareVue Fig. 6. Open in a new tab Kaplan-Meier survival curves stratified by LASSO-predicted phenotypes in the development cohorts. Log-rank p-values indicate statistical significance of survival Clinical outcomes across phenotypes The adjusted Kaplan-Meier analysis revealed consistent stratification of 30-day mortality risk among SALI phenotypes in both the development and validation cohorts. Phenotypes C and D had markedly elevated mortality rates compared with phenotype A, but no significant difference was noted between phenotypes A and B in the development cohort (Fig. 7 a, b, Table 1 ). Cox regression analysis in the development cohort validated these findings. Phenotypes C (crude HR 1.44, 95% CI: 1.23–1.69, p < 0.001) and D (crude HR 1.32, 95% CI: 1.10–1.58, p = 0.0024) exhibited considerably increased hazards. Conversely, phenotype B had no significant association with heightened mortality (crude HR 1.12, 95% CI: 0.96–1.30, p = 0.1508). These associations remained significant after demographic adjustment: Phenotype C (adjusted HR 1.41, 95% CI: 1.23–1.69, p < 0.001) and phenotype D (adjusted HR 1.32, 95% CI: 1.11–1.58, p = 0.0022) were significant, whereas phenotype B exhibited no difference (adjusted HR: 1.11, 95% CI: 0.96–1.30, p = 0.1631). External validation corroborated these trends, with phenotypes C and D consistently linked to increased mortality both before and after adjustment. Phenotype B had a distinct risk profile in the external cohort (Table 1 ). Fig. 7. Open in a new tab Kaplan-Meier survival analysis for the four SALI phenotypes. The figure shows 30-day survival curves in the ( a ) development and ( b ) validation cohorts. The analysis was adjusted for age, gender, and race Table 1. Multivariate Cox regression analyses for 30-day mortality, adjusted for demographic characteristics (age, gender, and race), were performed across phenotypes in the development and validation cohorts Phenotype A Phenotype B Phenotype C Phenotype D HR(95%CI) P value HR(95%CI) P value HR(95%CI) P value Development cohort 30-day mortality Crude Reference 1.12 (0.96, 1.30) 0.1508 1.44 (1.23, 1.69) <0.001 1.32 (1.10, 1.58) 0.0024 Adjusted Reference 1.11 (0.96, 1.30) 0.1631 1.41 (1.21, 1.66) <0.001 1.32 (1.11, 1.58) 0.0022 Validation cohort 30-day mortality Crude Reference 1.23 (1.01, 1.50) 0.0399 1.97 (1.62, 2.40) <0.001 1.48 (1.25, 1.76) <0.001 Adjusted Reference 1.25 (1.03, 1.53) 0.0242 1.93 (1.59, 2.35) <0.001 1.53 (1.28, 1.82) <0.001 Open in a new tab Heterogeneity of treatment responses of ICU admission We utilized Partial Dependence Plots (PDP) to analyze the relationship between ICU mortality and early interventions, specifically fluid resuscitation, norepinephrine administration, and antibiotic therapy within the first 0–6 h. All intervention data analyzed in this study represent the implementation within the first 0–6 h following sepsis diagnosis. This showed that the phenotypes had different risk patterns. The PDP results indicate that each SALI phenotype requires distinct treatment strategies (Fig. 8 ). In phenotype A (mild liver injury), initiation of antibiotics within 50 h was associated with lower ICU mortality. Additionally, a combination of moderate fluid resuscitation (2000–4000 mL/6 hours) and moderate-dose norepinephrine (500–1200 mg/6 hours) was also associated with reduced mortality risk in this phenotype. In phenotype B (mixed liver injury), initiation of antibiotics within 50 h was associated with lower ICU mortality. Additionally, a combination of moderate fluid resuscitation (2000–4000 mL/6 hours) and moderate-dose norepinephrine (500–1000 mg within the first 0–6 h) was also associated with reduced mortality risk in this phenotype. In phenotype C (hepatocellular dysfunction), initiation of antibiotics within 50 h was associated with a more favorable ICU mortality profile. Additionally, a combination of substantial fluid resuscitation (5000–8000 mL/6 hours) and high-dose norepinephrine (> 1000 mg/6 hours) was also associated with reduced mortality risk in this phenotype. In phenotype D (cholestasis), initiation of antibiotics within 0–50 h was associated with lower ICU mortality, creating an adaptable therapeutic window for timing adjustments. Furthermore, a combination of substantial fluid resuscitation (4000–6000 mL/6 hours) and high-dose norepinephrine (1500–2000 mg/6 hours) was also associated with reduced mortality risk in this phenotype. These findings provide exploratory insights for personalized treatment strategies in SALI; however, as this is an observational study, the observed associations require validation in prospective trials before clinical recommendations can be made. Fig. 8. Open in a new tab Partial dependence plots illustrate the joint effects of variable pairs on ICU mortality, with subfigures arranged from left to right corresponding to phenotypes A to D. ( a )-( d ) show the joint effects of time to first antibiotic administration and 0–6 h fluid administration; ( e )-( h ) show the joint effects of 0–6 h norepinephrine dose and 0–6 h fluid administration Discussion SALI is a common type of organ dysfunction that occurs during the advancement of sepsis. The pathophysiology is complex and remains incompletely elucidated, involving circulatory abnormalities, dysregulated inflammation, immunological dysfunction, and disruption of the gut-liver axis [ 7 ]. At the hepatic level, microcirculatory dysfunction and systemic inflammation increase vascular permeability, resulting in ischemic hepatocyte necrosis and potentially liver failure [ 25 , 26 ]. Concurrently, sepsis-induced gut dysbiosis and intestinal barrier dysfunction promote bacterial translocation, exacerbating a systemic inflammatory response that intensifies liver damage [ 8 , 27 – 29 ]. The intricate etiology and clinical variability result in disparate treatment outcomes, whereas existing uniform therapeutic approaches yield contradictory clinical results, rendering SALI a significant determinant of prognosis in septic patients. This study performed a retrospective review of 5,755 patients diagnosed with SALI, resulting in the development of MCSALIC, which revealed four distinct SALI phenotypes characterized by diverse clinical profiles, prognosis trajectories, and treatment responses. The reliability and applicability of this classification were confirmed across various independent cohorts. These findings offer novel clinical and data-driven insights into the variability of SALI and contest the traditional uniform-treatment paradigm, establishing a crucial basis for the advancement of precision medicine in SALI. The multi-algorithm consensus clustering method used in this study is similar to the strategy Yang et al. [ 16 ] applied in ALF research. Both aim to understand disease heterogeneity and guide precision medicine. However, the key features of these two diseases are different. First, ALF starts with hepatocellular necrosis [ 16 ], where liver injury causes problems throughout the body. In contrast, SALI is caused by systemic infection and inflammatory responses, and is one of the organ damages linked to sepsis. Because their key underlying mechanisms are different, specific classification methods are required, and it is not suitable to directly use one system for the other. Secondly, the ALF classification by Yang et al. focuses on hemodynamic status and organ failure severity, using mean arterial pressure, heart rate, and other parameters as key criteria to identify three subtypes. In contrast, the MCSALIC in this study prioritizes liver injury, with a focus on its type and severity to delineate four subtypes that align with distinct clinical manifestations of liver injury. Finally, therapeutic differences in ALF primarily lie in the selection of vasoactive agents and sedatives; whereas the core of SALI treatment is the individualized adjustment of standard sepsis therapy, with varying requirements for fluid resuscitation volume, norepinephrine dosage, and antibiotic initiation timing across subtypes. In conclusion, despite both utilizing multi-algorithm consensus clustering to enhance robustness, the inherent differences between ALF and SALI in pathophysiological mechanisms and therapeutic needs determine the uniqueness of their respective classification systems. The MCSALIC classification developed in this study provides a more targeted stratification tool for SALI. We used a consensus clustering strategy to combine the strengths of ten different clustering algorithms, such as K-Means, Gaussian Mixture Model, and hierarchical clustering, and reduce the differences between them. The appropriate number of subtypes (K = 4) was established using threefold validation: the elbow technique, a peak Calinski-Harabasz index (1782 at K = 4), and a minimal Davies-Bouldin index (0.891 at K = 4), ensuring statistical rigor. Our methodological strategy provides superior comprehensiveness compared to current SALI subtyping investigations. The De Ritis ratio classification by Palmowski et al. [ 21 ] concentrated exclusively on liver function indices, neglecting essential systemic factors such as inflammation, metabolism, and hemodynamics. Conversely, our model integrated multidimensional clinical factors (demographics, vital signs, laboratory results, and treatments) and utilized SHAP analysis to identify the most distinguishing traits for each phenotype. This guarantees that the phenotypes represent both statistical coherence and clinically significant pathophysiological differences. The clustering architecture exhibited strong external validity, as evidenced by substantial concordance between the SVM and LASSO classifiers across all cohorts (Cohen’s kappa > 0.791). This reproducibility verifies that phenotypic patterns are inherent and consistent across healthcare systems and across temporal contexts. In the MIMIC-III-CareVue cohort, a non-significant P-value ( P = 0.513) indicated a near-perfect distribution that diminished statistical power rather than reflecting poor agreement. Consequently, the kappa statistic serves as the principal measure of robust inter-algorithm consistency. Cross-database validation is crucial for ensuring generalizability in accordance with precision medicine guidelines [ 30 , 31 ]. The Kaplan-Meier survival analysis and Cox regression confirmed a significant prognostic gradient among the four phenotypes. phenotypes C and D were classified as high-risk (adjusted HR 1.41 and 1.32, respectively); phenotype A was designated as low-risk, while phenotype B demonstrated intermediate results without significant distinction from phenotype A. This prognostic classification provides an accurate instrument for clinical risk evaluation, exceeding the constraints of conventional methods dependent on isolated liver function markers. Currently, SALI prognosis is frequently assessed based on the extent of liver enzyme elevation or bilirubin concentrations. Our findings indicate that single-parameter assessment inadequately represents the severity of illness. For example, phenotype A had liver enzyme elevations, along with stable hemodynamics and mild renal impairment, and consequently exhibited a low 30-day mortality rate. In contrast, phenotype D was characterized by moderately elevated liver enzymes but severe cholestasis, resulting in a higher 30-day mortality rate than phenotype A.These results confirm that the prognosis of SALI is not determined by a single liver function index alone, but by the combined effects of liver injury severity, injury type, systemic hemodynamic status, metabolic conditions and multi-organ involvement. The present care of SALI primarily follows conventional sepsis guidelines, highlighting the importance of early administration of broad-spectrum antibiotics, sufficient fluid resuscitation, and organ support. This uniform approach neglects illness heterogeneity, which may result in inadequate or excessive treatment. The Surviving Sepsis Campaign guidelines advocate for an initial fluid resuscitation target of 30 mL/kg [ 32 – 35 ]; however, our study revealed that mortality escalated in phenotype B patients receiving over 4000 mL in 6 h, while phenotype D patients necessitated greater fluid volumes for improved outcomes. This aligns with the findings of Zhang et al. [ 36 ] in septic shock, where subtype-specific fluid strategies were associated with improved outcomes, further supporting the value of phenotype-directed therapy for SALI. Our findings regarding the timing of antibiotics also have important clinical consequences. Conventional wisdom emphasizes the administration of antibiotics within the initial hour of sepsis recognition; however, our data indicate that initiation may be justifiably postponed in phenotypes C and D. This could be due to differences in the main causes of liver damage in different subtypes. In phenotypes A and B, the damage is directly caused by an infection, so source control needs to be done quickly. In phenotypes C and D, the damage is mostly caused by hepatocellular necrosis or cholestasis, where antibiotics don’t help much and may even make gut dysbiosis worse. These insights justify customized antibiotic administration, aiding in the reduction of unwarranted exposure. We recognize these constraints while evaluating our findings. This retrospective design limits causal inference, as unmeasured variables may influence both phenotype classification and treatment associations. Therefore, while these findings offer exploratory insights for clinical management, the observed associations require validation in prospective trials before clinical recommendations can be made. Secondly, this study’s retrospective definition of SALI may have included a minor proportion of patients with liver injury attributable to medications (e.g., antivirals) or other etiologies. While strict exclusion criteria were applied and sensitivity analyses confirmed the robustness of the primary findings, the generalizability of the phenotype classification and associated treatment implications to sepsis patients without SALI requires prospective validation. Lastly, we did not characterize the inflammatory profiles underlying SALI phenotypes or validate their correspondence with established inflammatory subtyping systems. As highlighted by van Amstel et al. [ 37 ], the agreement of phenotype definitions across independent studies is only moderate at best. This inconsistency warrants caution when translating phenotype-based observations into therapeutic implications or comparisons with prior literature. Despite these limitations, this study developed a comprehensive, multi-algorithm consensus framework for SALI subphenotyping using high-quality longitudinal intervention data from international ICU databases. This method mitigates the bias of an individual algorithm and ensures the stability of phenotypes. We identified four distinct phenotypes that had varying responses to therapy upon ICU admission. This provided us with valuable insights for urgent resuscitation that surpassed the capabilities of standard protocols. External validation across geographically diverse cohorts enhances the generalizability of the results. Future research should focus on the potential evaluation of phenotype-guided management, including the customization of hemodynamic support, sedation techniques, and organ-specific assistance, relative to standard care. Incorporating proteomic or metabolomic markers could enhance the precision of stratification. The development of real-time phenotyping technologies compatible with EHRs and the capacity to monitor alterations in a patient’s phenotype during their ICU stay are crucial for clinical translation and early detection of patient deterioration. Conlusion Our research revealed four distinct phenotypes of SALI using EHR data, each exhibiting unique clinical profiles, outcomes, and reactions to critical ICU interventions. These phenotypes were rigorously confirmed across global datasets, offering enhanced pathophysiological insights. We utilized machine learning to enhance individualized clinical decision-making. Our results contest the dominant universal therapy paradigm and propose a data-driven framework for phenotype-specific management, necessitating further prospective validation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (2MB, docx) Acknowledgements We thank the teams behind the Medical Information Mart for Intensive Care (MIMIC)-IV version 3.1, MIMIC-III CareVue subset, eICU Collaborative Research Database version 2.0, and Northwestern ICU database (NWICU) version 0.1.0 for providing access to these valuable datasets. We are also grateful to our colleagues from he First Affiliated Hospital (Southwest Hospital) of Third Military Medical University (Army Medical University). Abbreviations SALI Sepsis-associated liver injury MCSALIC Multi-algorithm Consensus-based Sepsis-Associated Liver Injury Classification ALF Acute liver failure ACLF Acute-on-chronic liver failure MASLD Metabolic Dysfunction-Associated Steatotic Liver Disease ARDS Acute respiratory distress syndrome ALT Alanine aminotransferase ULN Upper limit of normal ALP Alkaline phosphatase TBil Total bilirubin AST Aspartate aminotransferase SBP Systolic blood pressure DBP Diastolic blood pressure BUN Blood urea nitrogen CR Creatinine GLU Glucose WBC White blood cell HR Heart rate RR Respiratory rate SpO2 Pulse oxygen saturation CRP C-reactive protein UML Unsupervised machine learning MICE Multiple Imputation by Chained Equations LASSO Least ASupport Vector Machine XGBoost Extreme Gradient Boosting SHAP SHapley Additive exPlanations PDPs Partial dependence plots ARI Adjusted Rand Index IQR Interquartile range HR Hazard ratio CI Confidence interval SVM Support Vector Machine IRB Institutional Review Board MIT Massachusetts Institute of Technology BIDMC Beth Israel Deaconess Medical Center NU Northwestern University GDPR General Data Protection Regulation ICU Intensive Care Unit DILI Drug-Induced Liver Injury EHRs Electronic health records Author contributions Conception and design: Xiaoman Xie, ZhiXian Zhu; Data analysis and interpretation: Xiaoman Xie, ZhiXian Zhu; Xinyu Cao; Administrative support: Jin Chai, Dong Zhao, and Qiong Pan; Manuscript writing: All authors; (Xiaoman Xie, and ZhiXian Zhu contributed equally to this work and share first authorship. Jin Chai, and Dong Zhao contributed equally to this work and share corresponding authorship.) Funding This work was supported by the National Natural Science Foundation of China [82450104], Open Project of Chongqing Engineering Research Center of Organ Intelligent Bio-Manufacturing (CQERC-OIBM20241010), and the Natural Science Foundation of Chongqing, China(cstc2025jcyj-jqX008). Data availability The data for this study are available from the corresponding author upon reasonable request. Declarations Ethical approval and consent to participate All data utilized in this study were anonymized, and informed consent was waived in accordance with the General Data Protection Regulation (GDPR) and U.S. federal regulations. Each database received ethical approval from its respective institutional review board (IRB) prior to data collection. The eICU Collaborative Research Database (eICU-CRD) was approved by theMassachusetts Institute of Technology Institutional Review Board (MIT IRB). The Medical Information Mart for Intensive Care III (MIMIC-III) and MIMIC-IV databases were approved by theBeth Israel Deaconess Medical Center Committee on Clinical Investigations (BIDMC IRB). The Northwestern ICU (NWICU-1.0) database was approved by theNorthwestern University Institutional Review Board (NU IRB). The requirement for individual patient consent was waived due to the use of retrospectively collected, de-identified data. All researchers completed appropriate ethics and security training (Credential ID: 70883411) prior to data access. This study adhered to all relevant data use agreements and was conducted in accordance with the principles of the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Dong Zhao, Email: dong.zhao@cldcsw.org. Jin Chai, Email: jin.chai@cldcsw.org. References 1. Meyer NJ, Prescott HC. Sepsis and septic shock. N Engl J Med. 2024;391:2133–46. 10.1056/NEJMra2403213IF. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Chen L, Zhang X, Shi P. Recent advances in biomarkers for detection and diagnosis of sepsis and organ dysfunction: a comprehensive review. Eur J Med Res. 2025;30:1081. 10.1186/s40001-025-03324-6IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Holmes CL, Albin OR, Mobley HLT, Bachman MA. Bloodstream infections: mechanisms of pathogenesis and opportunities for intervention. Nat Rev Microbiol. 2025;23:210 – 24. 10.1038/s41579-024-01105-2IF. [ DOI ] [ PMC free article ] [ PubMed ] 4. Girardis M, David S, Ferrer R, et al. Understanding, assessing and treating immune, endothelial and haemostasis dysfunctions in bacterial sepsis. Intensive Care Med. 2024;50:1580–92. 10.1007/s00134-024-07586-2IF. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Giamarellos-Bourboulis EJ, Aschenbrenner AC, Bauer M, et al. The pathophysiology of sepsis and precision-medicine-based immunotherapy. Nat Immunol. 2024;25:19–28. 10.1038/s41590-023-01660-5IF. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Maiwall R, Kulkarni AV, Arab JP, Piano S. Acute liver failure. Lancet. 2024;404:789–802. 10.1016/S0140-6736(24)00693-7IF. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Chen JW, Liu CY, Li S, et al. Sepsis-associated liver injury: mechanisms and potential therapeutic targets. World J Gastroenterol. 2024;30:4518–22. 10.3748/wjg.v30.i42.4518IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Sun J, Zhang J, Wang X, et al. Gut-liver crosstalk in sepsis-induced liver injury. Crit Care. 2020;24:614. 10.1186/s13054-020-03327-1IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Abraham E, Singer M. Mechanisms of sepsis-induced organ dysfunction. Crit Care Med. 2007;35(10):2408–16. 10.1097/01.ccm.0000282072.56245.91IF. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Saini K, Bolia R, Bhat NK. Incidence, predictors and outcome of sepsis-associated liver injury in children: a prospective observational study. Eur J Pediatr. 2022;181(4):1699–. 10.1007/s00431-022-04374-2IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. McCann MR, Wang Y, Liu X et al. Early sepsis metabolic changes in kidney and liver precede clinical evidence of organ dysfunction. Am J Respir Cell Mol Biol. 2025;73(2):299–309. 10.1165/rcmb.2024-0391OCIF. [ DOI ] [ PMC free article ] [ PubMed ] 12. Hu D, Li Y, Zhang H, et al. Mitochondrial dysfunction in sepsis: mechanisms and therapeutic perspectives. Crit Care. 2024;28(1):292. 10.1186/s13054-024-05069-wIF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Sahoo DK, Mohanty S, Patra S, et al. Exploring the role of antioxidants in sepsis-associated oxidative stress: a comprehensive review. Front Cell Infect Microbiol. 2024;14:1348713. 10.3389/fcimb.2024.1348713IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Guo Y, Guo W, Chen H, et al. Mechanisms of sepsis-induced acute liver injury: a comprehensive review. Front Cell Infect Microbiol. 2025;15:1504223. 10.3389/fcimb.2025.1504223IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Verma N, Sharma A, Singh R, et al. Identification of four novel acute-on-chronic liver failure clusters with distinct clinical trajectories and mortality using machine learning methods. Aliment Pharmacol Ther. 2024;60(11–12):1534–48. 10.1111/apt.17415IF. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Yang Y, Yang W, Tang B, et al. Multi-algorithm consensus classification identifies three distinct acute liver failure subtypes with differential treatment responses: a multi-database cohort study. J Adv Res. 2025;39:103–12. 10.1016/j.jare.2025.06.019IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Zheng Z, Chen J, Li Y, et al. Subtyping CKD patients by consensus clustering: the Chronic Renal Insufficiency Cohort (CRIC) study. J Am Soc Nephrol. 2021;32(3):639–53. 10.1681/ASN.2020050568. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Pollard TJ, Raffel D, Celi LA, et al. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci Data. 2018;5:180178. 10.1038/sdata.2018.178IF. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Johnson AEW, Pollard TJ, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1. 10.1038/s41597-022-01899-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Hyland SL, Combes A, Guttmann C, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med. 2020;26(3):364–73. 10.1038/s41591-020-0786-0. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Palmowski L, Müller M, Schmidt M, et al. phenotypes and the De Ritis ratio for mortality risk stratification in sepsis-associated acute liver injury: a retrospective cohort study. EClinicalMedicine. 2025;82:103173. 10.1016/j.eclinm.2025.103173. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Liu P, Wang Q, Zhang L, et al. Subphenotyping heterogeneous patients with chronic critical illness to guide individualised fluid balance treatment using machine learning: a retrospective cohort study. EClinicalMedicine. 2023;59:101970. 10.1016/j.eclinm.2023.101970. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Hoadley KA, Yau C, Wolf DM, et al. Multiplatform analysis of 12 cancer types reveals molecular classification within and across tissues of origin. Cell. 2014;158(4):929–44. 10.1016/j.cell.2014.07.011. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Ramazzotti D, Carvalho BS, Bontempi G. Multi-omic tumor data reveal diversity of molecular mechanisms that correlate with survival. Nat Commun. 2018;9(1):4453. 10.1038/s41467-018-06862-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Lu Y, Shi Y, Wu Q, et al. An overview of drug delivery nanosystems for sepsis-related liver injury treatment. Int J Nanomed. 2023;18:765–79. 10.2147/IJN.S394802. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Srdic T, Durasevic S, Lakic I, et al. From molecular mechanisms to clinical therapy: understanding sepsis-induced multiple organ dysfunction. Int J Mol Sci. 2024;25:7770. 10.3390/ijms25147770. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Zhang Z, Tan XJ, Shi HQ, et al. Bibliometric study of sepsis-associated liver injury from 2000 to 2023. World J Gastroenterol. 2024;30:3609–24. 10.3748/wjg.v30.i30.3609. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Adelman MW, Woodworth MH, Langelier C, et al. The gut microbiome’s role in the development, maintenance, and outcomes of sepsis. Crit Care. 2020;24:278. 10.1186/s13054-020-02989-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Liang H, Song H, Zhang X, et al. Metformin attenuated sepsis-related liver injury by modulating gut microbiota. Emerg Microbes Infect. 2022;11:815–28. 10.1080/22221751.2022.2045876. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Rappoport N, Shamir R. NEMO: cancer subtyping by integration of partial multi-omic data. Bioinformatics. 2019;35(18):3348–56. 10.1093/bioinformatics/btz297. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Reel PS, Singh A, Pandey A, et al. Using machine learning approaches for multi-omics data analysis: a review. Biotechnol Adv. 2021;49:107739. 10.1016/j.biotechadv.2021.107739. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Xu X, Yang T, An J, et al. Liver injury in sepsis: manifestations, mechanisms and emerging therapeutic strategies. Front Immunol. 2025;16:1575554. 10.3389/fimmu.2025.1575554. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Dellinger RP, Levy MM, Rhodes A, et al. Surviving sepsis campaign: international guidelines for management of severe sepsis and septic shock: 2012. Crit Care Med. 2013;41(2):580–637. 10.1097/CCM.0b013e31827e83af. [ DOI ] [ PubMed ] [ Google Scholar ] 34. He XL, Zhang Y, Li J, et al. Single-cell RNA sequencing deciphers the mechanism of sepsis-induced liver injury and the therapeutic effects of artesunate. Acta Pharmacol Sin. 2023;44:1801–14. 10.1038/s41401-023-01065-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Durand F, Kellum JA, Nadim MK. Fluid resuscitation in patients with cirrhosis and sepsis: a multidisciplinary perspective. J Hepatol. 2023;79:240–6. 10.1016/j.jhep.2023.02.024. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Zhang Z, Li Y, Wang H, et al. Identifying septic shock subgroups to tailor fluid strategies through multi-omics integration. Nat Commun. 2024;15:9028. 10.1038/s41467-024-49028-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. van Amstel RBE, Kennedy JN, Scicluna BP, et al. Uncovering heterogeneity in sepsis: a comparative analysis of phenotypes. Intensive Care Med. 2023;49:1360–9. 10.1007/s00134-023-07239-w. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (2MB, docx) Data Availability Statement The data for this study are available from the corresponding author upon reasonable request. Articles from BMC Infectious Diseases are provided here courtesy of BMC ACTIONS View on publisher site PDF (3.1 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Biomarkers to Predict Acute Respiratory Distress Syndrome(ARDS) in Patients With Sepsis#332843
• Identification and multi-layered experimental validation of feature genes related to butyrate metabolism and oxidative stress in sepsis.#302085
• Microbiome and Enteric Signatures in Sepsis-associated Hepatorenal Injury#1031904
• Identification of common diagnostic biomarkers and immune landscapes in sepsis and acute kidney injury: a transcriptomic study integrating machine learning and single-cell analysis.#430723
• TRanscriptomic Analysis of Circulating Endothelial Cells During Sepsis - TRACES Study#1010145