Metabolic phenotyping of sepsis in large multicenter cohorts: identification of a reproducible high-risk subgroup

Abstract

Background

Sepsis represents a profound metabolic crisis, yet the prognostic significance of glycolipid biomarkers remains unclear. Most prior studies examined single metabolic indicators, overlooking their combined effects on organ dysfunction and survival. This study aimed to identify glycolipid-related biomarkers predicting outcomes in sepsis and to define a reproducible metabolic phenotype linked to multi-organ dysfunction and mortality.

Methods

We conducted a multicenter retrospective cohort study of 2,970 adults with sepsis admitted to three tertiary hospitals in China (2015–2023), with external validation using the international MIMIC-IV database (2008–2019). A comprehensive panel of glycolipid markers was evaluated through univariable Cox and LASSO regression, supported by tree-based machine learning and logistic regression for sepsis-associated acute kidney injury. Low-density lipoprotein cholesterol (LDL-C) and the triglyceride–glucose (TyG) index emerged as the most consistent prognostic biomarkers and were used to define metabolic phenotypes.

Results

Among 2,970 patients (median age 68 years [IQR 60–80]; 62.5% male), LDL-C and TyG independently predicted mortality across statistical and machine-learning analyses. Patients with a high TyG–low LDL phenotype exhibited the most severe multi-organ dysfunction—renal, hepatic, coagulative, and inflammatory—and the highest mortality risk (HR 1.31, 95% CI 1.04–1.65). Validation in MIMIC-IV confirmed the reproducibility of this phenotype across populations and healthcare systems.

Conclusions

Integrating LDL-C and TyG identifies a reproducible metabolic phenotype marking sepsis patients at high risk of multi-organ failure and death, offering a simple framework for early risk stratification and personalized management in critical care.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12944-026-02888-4.

Background

Sepsis remains one of the most formidable challenges in critical care—a life-threatening syndrome marked by a dysregulated host response to infection and a major cause of global mortality [1, 2]. Among its complications, multi-organ dysfunction—especially acute kidney injury (AKI)—is both common and devastating, worsening short-term outcomes and leaving survivors with long-term health burdens [3]. Despite significant progress in antimicrobial therapy and organ support, our ability to recognize high-risk patients early and stratify them effectively remains limited [4]. Current scoring systems such as SOFA and KDIGO capture physiological derangements but largely ignore metabolic disturbances, which may represent critical yet overlooked dimensions of sepsis biology.

Recent research has brought renewed attention to metabolism as a central player in sepsis pathophysiology [5–7]. Severe infection disrupts glucose and lipid homeostasis, driving insulin resistance, hyperglycemia, and disordered lipoprotein metabolism [8]. These changes are not passive reflections of illness severity—they actively fuel inflammation, oxidative stress, mitochondrial dysfunction, and endothelial injury, ultimately amplifying organ damage [9]. Yet, how these metabolic abnormalities might inform clinical risk prediction remains insufficiently explored.

Individual metabolic markers such as triglycerides, glucose indices, and cholesterol fractions have each been linked to outcomes in chronic disease and critical illness [10, 11]. However, most studies have taken a reductionist view, assessing single pathways rather than their intricate interplay [12]. This approach misses the fact that glucose and lipid metabolism are tightly interconnected, forming a metabolic network that profoundly shapes immune and organ responses during sepsis. To date, no study has systematically defined metabolic phenotypes that integrate both glycolipid axes in sepsis or examined how such phenotypes relate to organ dysfunction and mortality. Identifying reproducible metabolic subtypes grounded in glucose–lipid interactions could transform our understanding of disease heterogeneity and provide new tools for clinical decision-making.

To fill this gap, we conducted a comprehensive screening of glycolipid-related biomarkers in a large, multicenter sepsis cohort. By combining traditional regression with penalized and machine learning models, we sought to identify biomarkers that robustly predict adverse outcomes. Low-density lipoprotein cholesterol (LDL-C) and the triglyceride–glucose (TyG) index consistently emerged as the most stable and clinically relevant markers. Using these two indices, we defined metabolic phenotypes and tested their prognostic significance. Finally, to ensure generalizability, we validated our findings in the international MIMIC-IV database. Our study had two overarching goals: (1) to identify key metabolic predictors associated with organ dysfunction and mortality in sepsis; and (2) to establish and externally validate reproducible metabolic phenotypes that capture glycolipid dysregulation in critically ill patients.

Methods

This study followed the TRIPOD-AI guideline for developing and validating prediction models using regression and machine learning methods. Because of its retrospective and non-interventional design, informed consent was waived. All procedures complied with the Declaration of Helsinki (1964) and later revisions and were approved by the institutional review boards of all participating hospitals.

Data source and study population

The internal cohort included consecutive adults (≥ 18 years) diagnosed with sepsis and admitted to three tertiary academic hospitals in China—Xinhua Hospital and Ruijin Hospital (Shanghai Jiao Tong University School of Medicine) and Huashan Hospital (Fudan University)—between September 2015 and May 2023. Sepsis was defined according to Sepsis-3 criteria. Patients were excluded if they (1) died within 24 h of admission or declined aggressive care, (2) lacked essential demographic or laboratory data for severity scoring, or (3) had more than 30% missing data among key variables. For external validation, we used the MIMIC-IV database, which contains de-identified intensive care unit (ICU) data from Beth Israel Deaconess Medical Center (Boston, MA, USA; 2008–2019). One author (Huailong Shang) obtained access credentials (Certification ID: 65871591). Based on Sepsis-3 definitions, 41,295 admissions were identified, of which 26,651 unique patients remained after excluding duplicates. Among these, 2,750 individuals with available LDL-C, triglycerides, HbA1c, and plasma glucose data were included in the analysis. In the MIMIC-IV cohort, sepsis was defined per Sepsis-3 criteria. A patient was included if both suspected infection (concurrent antibiotics and cultures) and a ≥ 2-point increase in SOFA score occurred within a time window of 48 h before to 24 h after the infection time.

Candidate biomarkers and variable definitions

We evaluated a comprehensive panel of glycolipid-related biomarkers, including triglyceride–glucose index (TyG), LDL-C, HDL-C, total cholesterol (TC), triglycerides (TG), non-HDL/HDL ratio, remnant cholesterol (RC), atherogenic index of plasma (AIP), arterial stiffness index (ASI), HbA1c, HbA1c/HDL ratio, glucose, and stress hyperglycemia ratio (SHR). For both the TyG index and LDL-C, the values used in all primary analyses were the first available measurement obtained within the first 24 h following ICU admission. This ‘first-value’ approach was TCsen to reflect the patient’s initial metabolic state upon presentation for critical care, maximizing clinical utility for early risk stratification.

The TyG index was calculated as ln [triglycerides (mmol/L) × glucose (mmol/L) / 2]. Demographic variables included age and sex, while clinical variables encompassed comorbidities, Acute Chronic Comorbidity Index (ACCI), and Sequential Organ Failure Assessment (SOFA) scores.

Screening of glycolipid biomarkers

We used a multi-step strategy to identify prognostic glycolipid biomarkers. Each candidate was first tested using univariable Cox regression to assess its association with mortality. LASSO regression with 10-fold cross-validation was then applied to identify predictors with non-zero coefficients while minimizing overfitting. In parallel, a Light Gradient Boosting Machine (LightGBM) with SHAP (Shapley Additive Explanations) interpretation was trained to rank variable importance, and Random Forest models were used as complementary confirmation. Associations with sepsis-associated acute kidney injury (AKI) were further examined using logistic regression adjusted for prespecified confounders. Candidate markers were reviewed against previous literature and expert clinical judgment to ensure biological plausibility. Full analytical details are provided in the Supplement (Additional file 1). Through this integrative process, LDL-C and TyG consistently demonstrated strong and stable prognostic value and were selected for phenotype construction.

Definition and validation of phenotype cut-offs

Patients were stratified into tertiles (T1, T2, T3) based on their TyG index and LDL-C levels, respectively. In the primary discovery cohort, the numerical cut points defining these tertiles were empirically determined as follows: for the TyG index, the T1/T2 boundary was 1.347 and the T2/T3 boundary was 1.934; for LDL-C (mmol/L), the T1/T2 boundary was 1.99 and the T2/T3 boundary was 1.26. Patients concurrently located in the highest TyG tertile (TyG ≥ 1.934) and the lowest LDL-C tertile (LDL-C < 1.26) were classified as having the “high TyG–low LDL” phenotype. To assess the transportability and generalizability of this phenotype definition, a fixed-cutoff validation strategy was employed for the external cohort. Specifically, the identical absolute cut points derived from the discovery cohort were directly applied to assign phenotypic groups to patients in the MIMIC-IV cohort. This approach evaluates the performance of a stable, pre-defined criterion across distinct populations, which is essential for potential future clinical application.

Metabolic phenotype construction

Patients were categorized into four phenotypes based on tertiles of TyG and LDL-C: T1: low TyG–high LDL; T2: high TyG–high LDL; T3: low TyG–low LDL; T4: high TyG–low LDL. Clinical characteristics, laboratory indices, and outcomes were compared across groups. The primary outcomes were in-hospital mortality and severity of organ dysfunction; secondary outcomes included 30-, 60-, and 90-day and ICU mortality.

External validation and model interpretability

To verify the reproducibility of the “high TyG–low LDL” phenotype, we trained and compared seven machine learning algorithms: CatBoost, decision tree (DT), multilayer perceptron (MLP), LightGBM, logistic regression (LR), random forest (RF), and XGBoost. Ten-fold cross-validation was performed, and model performance was assessed using the area under the receiver operating characteristic curve (AUC). SHAP values were used to interpret both global and case-level feature importance.

Data processing and visualization

Continuous missing data were imputed using the median. Yeo–Johnson transformation followed by robust scaling (10th–90th percentile) was applied to reduce skewness. Class imbalance was corrected by random oversampling with a fixed seed for reproducibility. For unsupervised analysis, t-distributed stochastic neighbor embedding (t-SNE) was used for dimensionality reduction, followed by Birch clustering. A “true” metabolic class based on LDL and TyG was predefined to assess concordance between hypothesis-driven and data-driven classifications. Cluster-level differences were expressed as mean Z-scores standardized against the overall sepsis population. Diverging heatmaps illustrated coordinated derangements in hepatic, coagulation, and oxygenation markers.

Statistical analysis

Continuous variables were expressed as mean ± SD or median (IQR) and compared using Student’s t-test or Wilcoxon rank-sum test. Categorical variables were presented as counts (%) and compared with chi-square or Fisher’s exact test. In the internal cohort, missing values were imputed using the missForest algorithm in R; in MIMIC-IV, missingness was handled using median imputation. Cox proportional hazards models estimated hazard ratios (HRs) for mortality, with proportionality tested using STCenfeld residuals. To assess the independent association between metabolic phenotypes and mortality, we performed multivariable Cox proportional-hazards regression. Covariates were selected a priori based on their established clinical relevance to sepsis outcomes and their potential as confounders of the relationship between metabolic state and survival. We employed a sequential adjustment strategy with two pre-defined models to incrementally address different layers of confounding: Model 1: Adjusted for demographic factors and chronic comorbidities, specifically age, sex, hypertension, sepsis shock, diabetes mellitus. Model 2: Built upon Model 1 by further adjusting for key indicators, including PLT, WBC, TBIL, PCT, and HCT. The proportional hazards assumption for the Cox models was formally tested using Schoenfeld residuals; no significant violation was detected for the primary exposure variable (the metabolic phenotype) in either model (all p > 0.05). Kaplan–Meier survival curves were compared using log-rank tests. Restricted cubic splines modeled nonlinear associations of TyG and LDL-C with mortality. Prespecified subgroup analyses were conducted by age, sex, shock, hypertension, and diabetes status. All analyses were performed in SPSS 27.0, R 3.6.0, and Python 3.9, with p < 0.05 considered significant.

Results

Study population and baseline characteristics

We included 2,970 adult patients with sepsis across the three Chinese tertiary hospitals (Fig. 1). The median age was 68 years (IQR 60–80), and 62.5% were male. Hypertension (48.9%) and diabetes (35.3%) were the most common comorbidities. The in-hospital mortality rate was 24.8%. Baseline laboratory results covered a broad range of glucose- and lipid-related parameters, alongside renal, hepatic, and coagulation indices (Additional File 1-Table A.1).

Fig. 1.

Study flowchart showing patient enrollment, exclusion criteria, and cohort assignment for analysis

Systematic screening of glycolipid biomarkers

We examined multiple glycolipid indices, including TC, TG, HDL-C, LDL-C, TyG, HbA1c, HbA1c/HDL ratio, non-HDL/HDL ratio, AIP, ASI, SHR, and RC. Univariable Cox regression identified several biomarkers—TC, HDL, LDL, AIP, ASI, HbA1c/HDL, non-HDL/HDL, and TyG—as significantly associated with mortality (all p < 0.05; Table 1). LASSO regression retained LDL, AIP, and HbA1c/HDL as key predictors (Fig. 2A–B). Tree-based machine learning models ranked SHR, TC, TyG, RC, and LDL among the top five contributors to prognosis (Fig. 2C). Logistic regression confirmed that both LDL-C and TyG were independently associated with sepsis-associated AKI (adjusted OR 1.18 [1.02–1.37] for TyG; OR 0.74 [0.64–0.85] for LDL; Additional File 3-Table A.2). Across all approaches, LDL-C and TyG consistently emerged as the most reliable and clinically meaningful biomarkers.

Table 1.

Univariable cox regression of candidate glycolipid biomarkers for overall survival in sepsis

Factor	HR	95%CI	P value
TC	0.9	0.85–0.95	< 0.001
HDL	0.64	0.52–0.79	< 0.001
LDL	0.83	0.77–0.9	< 0.001
AIP	1.51	1.25–1.83	< 0.001
ASI	1.02	1.01–1.03	< 0.001
HbA1C/HDL	1.02	1.01–1.02	< 0.001
Non-HDL/HDL	1.02	1.01–1.04	0.002
TyG	1.13	1.03–1.25	0.014
HbA1C	1.02	1-1.04	0.058
SHR	1.11	0.99–1.24	0.076
RC	1.03	0.97–1.09	0.293

TC total cholesterol, HDL high-density lipoprotein cholesterol, LDL low-density lipoprotein cholesterol, AIP atherogenic index of plasma, ASI arterial stiffness index, HbA1c glycated hemoglobin, SHR stress hyperglycemia ratio, RC remnant cholesterol, HR hazard ratio, CI confidence interval. P < 0.05 was considered statistically significant

Fig. 2.

Feature selection and model interpretation using LASSO regression and SHAP analysis. A Coefficient trajectories of variables across log-transformed penalty values. B Partial deviance curve for optimal model selection. C SHAP summary plot showing the distribution of feature importance, with colors representing low (blue) to high (red) feature values

Prognostic value of TyG and LDL-C

Higher TyG values were modestly associated with greater in-hospital mortality (HR 1.13 [1.02–1.25]; Table 2), showing linear relationships across 30-, 60-, and 90-day endpoints (p < 0.01; Fig. 3A). LDL-C, conversely, displayed a robust protective effect, inversely associated with mortality (HR 0.83 [0.77–0.90]; Table 2). Kaplan–Meier curves demonstrated significantly higher mortality in patients within the lowest LDL tertile (Fig. 4B). These trends remained consistent across age, sex, and comorbidity subgroups (Fig. 3D).

Table 2.

Multivariable Cox regression analysis of TyG index and LDL-C for in-hospital mortality in sepsis patients

Groups	Non-adjusted HR (95% CI) P-value	Model 1 HR (95% CI) P-value	Model 2 HR (95% CI) P-value
TyG index
Continuous	1.129 (1.024–1.245)	1.144 (1.034–1.265)	1.118 (1.009–1.238)
	0.015	0.009	0.032
T1	ref	ref	ref
(≤ 1.347033, N = 990)	ref	ref	ref
T2 (> 1.347033, ≤ 1.933929; N = 990)	0.969 (0.810–1.160)	1.006 (0.840–1.206)	0.993 (0.828–1.191)
	0.736	0.947	0.94
T3 (> 1.933929; N = 990)	1.171 (0.984–1.394)	1.189 (0.994–1.424)	1.140 (0.950–1.368)
	0.075	0.059	0.158
P for trend	0.066	0.084	0.204
LDL
Continuous	0.833 (0.768–0.904)	0.890 (0.819–0.967)	0.922 (0.847–1.002)
	< 0.001	0.006	0.057
T1	ref	ref	ref
(> 1.99; N = 998)	ref	ref	ref
T2	1.164 (0.967–1.401)	0.969 (0.810–1.160)	1.051 (0.871–1.268)
(> 1.26, ≤ 1.99; N = 989)	0.109	0.736	0.603
T3	1.461 (1.222–1.747)	1.171 (0.984–1.394)	1.173 (0.972–1.417)
(≤ 1.26; N = 993)	< 0.001	0.075	0.097
P for trend combined	< 0.001	0.016	0.106
T4	ref	ref	ref
(Low LDL and high TyG; N = 375)	ref	ref	ref
T3	0.772 (0.636–0.937)	0.765 (0.627–0.932)	0.800 (0.655–0.977)
(Low LDL and low TyG; N = 618)	0.009	0.008	0.028
T2	0.658 (0.534–0.810)	0.747 (0.605–0.922)	0.794 (0.641–0.983)
(High LDL and high TyG; N = 615)	< 0.001	0.007	0.034
T1 (High LDL and low TyG; N = 1362)	0.599 (0.488–0.736) < 0.001	0.645 (0.521–0.797) < 0.001	0.701 (0.564–0.872) 0.001

Model 1 adjusted for age, sex, hypertension, sepsis shock, and diabetes mellitus

Model 2 additionally adjusted for PLT, WBC, TBIL, PCT, and HCT

Fig. 3.

Restricted cubic spline regression and subgroup interaction analyses for TyG index and LDL-C as predictors of in-hospital mortality. A Nonlinear dose–response association between TyG index and mortality risk. B Nonlinear dose–response association between LDL-C and mortality risk. C Subgroup analysis of TyG index stratified by sex, age, hypertension, diabetes, and SOFA score. D Subgroup analysis of LDL-C stratified by sex, age, hypertension, diabetes, and SOFA score

Fig. 4.

Kaplan–Meier survival curves for in-hospital mortality stratified by metabolic markers. A TyG index; B LDL-C; C combined TyG–LDL phenotypes

Identification of metabolic phenotypes

Based on LDL-C and TyG levels, patients were grouped into four phenotypes (T1-T4). The distribution was as follows: T1: n = 1362 (45.86%); T2: n = 615 (20.71%); T3: n = 618 (20.81%); T4: n = 375 (12.62%). When analyzed individually, neither LDL nor TyG alone fully captured outcome heterogeneity. Combined, however, they revealed a distinct “high TyG–low LDL” phenotype (T4), characterized by the worst survival (HR 1.31 [1.04–1.65]; Table 2; Fig. 4C).

Clinical characteristics of metabolic phenotypes

Patients with the T4 phenotype exhibited the most severe multi-organ dysfunction (Table 3). Specifically, they showed more pronounced renal injury (higher SOFA scores, creatinine, and BUN), greater hepatic impairment (elevated TBIL, ALT, and AST), marked coagulation derangements (lower platelet counts, prolonged PT, INR, and APTT), and exaggerated systemic inflammation (increased CRP and PCT; all p < 0.01; Figs. 5 and 6). Together, these abnormalities delineate a profoundly dysmetabolic state characterized by widespread organ injury and the highest mortality across all time points (Fig. 7).

Table 3.

Comparison of clinical and laboratory parameters between T4 (high TyG–low LDL) and non-T4 phenotypes

SOFA-score	Non-T4	T4	P
	5.00	6.38	< 0.001
Cr	104.51	129.90	< 0.001
BUN	9.74	13.52	< 0.001
TBIL	17.06	21.90	< 0.001
ALT	37.22	51.72	< 0.001
AST	45.70	65.14	< 0.001
PLT	181.85	141.60	< 0.001
PT	14.00	14.40	< 0.001
INR	1.20	1.24	< 0.001
APTT	33.65	35.58	< 0.001
WBC	10.81	11.74	0.03
CRP	106.30	133.27	< 0.001
PCT	7.06	13.12	< 0.001

Values are presented as mean or median as appropriate

SOFA Sequential Organ Failure Assessment score, Cr creatinine, BUN blood urea nitrogen, TBIL total bilirubin, ALT alanine aminotransferase, AST aspartate aminotransferase, PLT platelet count, PT prothrombin time, INR international normalized ratio, APTT activated partial thromboplastin time, WBC white blood cell count, CRP C-reactive protein, PCT procalcitonin

P-values < 0.05 were considered statistically significant

Fig. 5.

Comparison of renal and hepatic function between T4 (high TyG–low LDL) and non-T4 phenotypes. A SOFA score; B serum creatinine (Cr); C blood urea nitrogen (BUN); D total bilirubin (TBIL); E alanine aminotransferase (ALT); F aspartate aminotransferase (AST)

Fig. 6.

Comparison of coagulation and inflammatory markers between T4 (high TyG–low LDL) and non-T4 phenotypes. A. Platelet count (PLT); B prothrombin time (PT); C. international normalized ratio (INR); D. activated partial thromboplastin time (APTT); E. C-reactive protein (CRP); F. procalcitonin (PCT)

Fig. 7.

Kaplan–Meier survival curves comparing T4 (high TyG–low LDL) and non-T4 phenotypes. A. In-hospital mortality; B. 30-day mortality; C. 60-day mortality; D. 90-day mortality

External validation in MIMIC-IV

In the international validation cohort, the TyG–LDL phenotypes were reproducible. CatBoost achieved the best discrimination (AUC = 0.726, F1 = 0.69, Precision = 0.70, Recall (the overlap between Cluster 1 and the predefined high-risk phenotype T4) = 0.69; Fig. 8A). SHAP analysis identified PT, bilirubin, lymphocyte count, hematocrit, PaO₂, hemoglobin, calcium, BUN, and APTT as the most influential features for phenotype classification (Fig. 8B–C). Cluster analysis revealed that the unfavorable cluster—characterized by higher bilirubin and PT but lower PaO₂—aligned closely with the high TyG–low LDL group (Fig. 8D–E). This phenotype showed higher SOFA scores, longer vasopressor use, and greater in-hospital mortality (Fig. 8F–H), confirming its reproducibility and prognostic relevance across independent datasets.

Fig. 8.

Machine learning validation and characterization of the low LDL–high TyG phenotype in the MIMIC-IV cohort. A Performance of seven machine learning models. B SHAP beeswarm plot from the CatBoost model. C Global feature importance ranking. D Overlap of the predefined low LDL–high TyG phenotype with data-driven clusters. E Heatmap of mean Z-score divergence across clinical features. F–H. Differences in SOFA scores, vasopressor use, and in-hospital mortality between phenotypes

Discussion

In this multicenter study, we systematically examined the prognostic relevance of glycolipid-related biomarkers in sepsis and identified low-density lipoprotein cholesterol (LDL-C) and the triglyceride–glucose (TyG) index as the most robust predictors of adverse outcomes. By integrating these two parameters, we defined a reproducible metabolic phenotype characterized by high TyG and low LDL-C levels, which was consistently associated with greater organ dysfunction and increased mortality. Importantly, this association was independently validated in the external MIMIC-IV cohort, supporting the generalizability of this metabolic subtype across different populations and healthcare settings.

Metabolic dysregulation is increasingly recognized as a defining feature of sepsis pathophysiology [13, 14]. Severe infection induces a cascade of insulin resistance, lipolysis, and mitochondrial dysfunction, fueling systemic inflammation and energy imbalance [15]. In parallel, altered lipoprotein metabolism impairs endotoxin clearance and modulates immune signaling [16, 17]. Although numerous glucose- and lipid-related markers have been individually linked to outcomes, few studies have systematically compared their prognostic contributions. In our analysis, among multiple candidates—including TG, HDL-C, non-HDL/HDL ratio, AIP, ASI, HbA1c, SHR, and RC—only LDL-C and TyG consistently emerged as independent predictors across both traditional and machine learning models. This convergence suggests that insulin resistance and lipid availability represent two interrelated metabolic axes that jointly shape sepsis progression.

The apparent protective association of LDL-C merits particular attention. While LDL is traditionally viewed as deleterious in chronic atherosclerotic disease, it may exert beneficial effects under acute inflammatory stress [18, 19]. LDL particles and their apolipoprotein B-100 component can neutralize circulating endotoxins, modulate macrophage polarization by suppressing IL-6 and enhancing IL-10, and provide cholesterol substrates for membrane repair and immune cell function [3, 16, 17, 20, 21]. Notably, recent clinical evidence further supports this concept: in patients with acute respiratory distress syndrome (ARDS), non-survivors had significantly lower total cholesterol, LDL-C, and apoA-I levels than survivors [22], reinforcing the notion that inadequate lipid availability may impair host defense. In chronic inflammatory settings such as atherosclerosis, macrophages are continuously exposed to oxidative stress, which leads to sustained activation of lipid uptake pathways, including scavenger receptors such as CD36 and SR-A1. This persistent stimulation promotes excessive internalization of oxidatively modified low-density lipoproteins, driving foam cell formation and the progression of atherosclerotic plaques. These processes are regulated through complex intracellular signaling cascades, including the cAMP–PKA–CREB pathway [23]. Collectively, these findings highlight that the same lipid-sensing receptors can support markedly different immunometabolic functions depending on the inflammatory milieu. A similar principle may be relevant in the context of sepsis. During the early phase of acute infection, profound systemic inflammation and metabolic stress are often accompanied by a rapid decline in circulating native LDL levels. Such depletion may limit the availability of lipid substrates required to sustain energy-demanding macrophage functions, including phagocytosis and bactericidal activity. Under these conditions, macrophages may transition from a metabolic state that supports effective innate immune responses toward a dysfunctional or immunosuppressive phenotype due to inadequate metabolic fuel. From this perspective, the “low LDL-C” observed in our study may reflect exhaustion of a critical metabolic resource rather than merely serving as a surrogate marker of inflammatory severity.

In contrast, the TyG index—a well-established surrogate for insulin resistance [24, 25]—reflects systemic metabolic imbalance that amplifies organ injury. Insulin resistance promotes hyperglycemia, free fatty acid release, mitochondrial stress, and endothelial dysfunction, ultimately leading to tissue hypoxia and inflammatory damage [26, 27]. The combined use of TyG and LDL-C provided superior risk discrimination compared with either marker alone, capturing the interplay between impaired insulin signaling, oxidative stress, and diminished lipid-mediated detoxification. This metabolic interaction likely drives the inflammatory–metabolic amplification that underlies organ injury in severe sepsis [3, 28–32].

The principal clinical implication of our findings is that the “high TyG–low LDL-C” phenotype provides prognostic information that complements, rather than duplicates, established severity scoring systems. While the SOFA score reflects the extent of overt organ dysfunction and therefore represents a downstream manifestation of critical illness, the identified metabolic phenotype appears to capture upstream metabolic dysregulation that precedes and potentially contributes to organ injury. Integrating metabolic phenotyping with conventional severity assessment may thus enable a more nuanced risk stratification strategy, in which current disease severity and future deterioration risk are jointly considered. In particular, patients presenting with moderate SOFA scores but exhibiting a high-risk metabolic phenotype may warrant closer monitoring and earlier intervention. Although the “high TyG–low LDL-C” phenotype remained independently associated with mortality after adjustment for acute illness severity and chronic comorbidities, the observational nature of this study precludes causal inference. Accordingly, this phenotype should be interpreted as a robust integrative marker of risk rather than a direct etiological factor. It likely reflects a convergent metabolic–immune dysfunction induced by severe infection, representing a final common pathway of systemic derangement rather than a singular initiating mechanism.

For practical implementation, the TyG index and LDL-C should ideally be measured within the first 24 h of ICU admission to capture the early metabolic derangement and inform initial risk stratification. Beyond prognosis, this framework suggests phenotype-tailored therapeutic considerations. For instance, patients identified with the “high-TyG” component might be prioritized for more stringent glycemic control and monitoring for insulin resistance-associated complications, while those with “low-LDL” may highlight a population in whom the risks of aggressive lipid-lowering therapy outweigh potential benefits during acute infection. Clinically, patients with the high TyG–low LDL phenotype exhibited the most pronounced multi-organ dysfunction—renal, hepatic, coagulation, and inflammatory abnormalities. This constellation suggests that systemic metabolic imbalance may act as a unifying mechanism of sepsis-related deterioration. Mechanistically, insulin resistance leads to excessive fatty acid oxidation and mitochondrial overload, promoting tubular and hepatocellular injury [15, 33–38]. while reduced LDL availability limits endotoxin clearance and dampens anti-inflammatory responses [3, 28]. In the coagulation system, insulin resistance and hypertriglyceridemia enhance tissue factor expression and thrombin generation, whereas low LDL levels may exacerbate LPS-driven coagulopathy by facilitating fibrinolytic inhibition [39–41]. Together, these processes contribute to prolonged PT, thrombocytopenia, and microvascular dysfunction [42, 43]. The dual disruption of glucose and lipid homeostasis therefore provides a mechanistic basis for the profound multi-organ injury and high mortality observed in this subgroup.

The external validation in MIMIC-IV further reinforced the reproducibility and interpretability of this phenotype. SHAP-based model explanations identified coagulation (PT), hepatic (bilirubin), and respiratory (PaO₂) parameters as key discriminators, aligning closely with the observed pathophysiological pattern. The strong concordance between hypothesis-driven classification and data-driven clustering supports the biological authenticity of this metabolic subtype rather than a statistical artifact.

From a clinical standpoint, this metabolic phenotype offers a simple yet powerful framework for early recognition of high-risk patients. It highlights the limitations of assessing glucose or lipid indices in isolation and underscores the value of an integrated metabolic approach. Beyond risk prediction, the findings suggest potential therapeutic implications: modulating insulin resistance, supporting mitochondrial function, and avoiding indiscriminate lipid lowering may all represent rational strategies. By bridging metabolic mechanisms with clinical stratification, this approach could help move sepsis care toward precision phenotyping and personalized management in critical illness. In summary, the “high TyG–low LDL” phenotype consolidates disparate metabolic signals into a unified risk axis that is reproducible, mechanistically grounded, and clinically informative. Its performance alongside SOFA score warrants prospective evaluation to determine if its integration into early clinical assessment can improve patient triage or guide targeted metabolic support.

Several limitations should be noted. First, the retrospective design may introduce residual confounding despite multivariable and machine-learning adjustments. Second, the primary cohort included only Chinese adults, and although findings were validated in an international database, caution is warranted in extrapolating to pediatric or geriatric populations. Third, both TyG and LDL-C are dynamic variables that fluctuate during sepsis in response to inflammation and treatment; longitudinal studies are needed to characterize their temporal trajectories. Fourth, in the MIMIC-IV validation, we restricted our analysis to patients who had LDL-C and TyG index measurements available. This inevitably selected a subpopulation from the overall sepsis cohort. Although a direct comparison revealed no substantial differences in baseline severity between the included and excluded groups (Additional File 3-Table A.3), the need for these specific laboratory tests could affect how widely our findings apply—especially in clinical environments where such metabolic panels are not part of routine practice. Whether the phenotype retains its prognostic value in those settings awaits future investigation. Finally, mechanistic studies integrating metabolomic, lipidomic, or transcriptomic data could help clarify causal pathways linking metabolic dysregulation to organ failure.

Conclusion

In this multicenter study, low LDL-C and high TyG were identified as key metabolic predictors of poor outcomes in sepsis. Their combination defined a reproducible “high TyG–low LDL” phenotype characterized by severe multi-organ dysfunction and elevated mortality. This glycolipid-derived phenotype provides a clinically accessible and cost-effective tool for early risk stratification and may serve as a foundation for individualized metabolic interventions in critically ill patients.

Abbreviations

ACCI

Acute chronic comorbidity index

AKI

Acute kidney injury

AIP

Atherogenic stiffness index of plasma

ALB

Albumin

ALT

Alanine aminotransferase

APTT

Activated partial thromboplastin time

ASI

Arterial stiffness index

AUC

Area under the curve

BUN

Blood urea nitrogen

CKD

Chronic kidney disease

DAG

Directed acyclic graph

HbA1c

Glycated hemoglobin

HCT

Hematocrit

HDL

High density lipoprotein

HR

Hazard ratio

ICU

Intensive care unit

IR

Insulin resistance

LDL-C

Low density lipoprotein cholesterol

MIMIC-IV

Medical Information Mart for Intensive Care IV

OR

Odds ratio

PaO₂

Partial Pressure of oxygen in arterial blood

PCT

Procalcitonin

PLT

Platelet

PT

Prothrombin time

RC

Remnant Cholesterol

ROC

Receiver operating characteristic

SA-AKI

Sepsis-associated acute kidney injury

SHAP

SHapley Additive exPlanations

SHR

Stress Hyperglycemia Ratio

SOFA

Sequential organ failure assessment

TBIL

Total bilirubin

TC

Total Cholesterol

TG

Triglycerides

TLR4/NF-κB

Toll-like Receptor 4 / Nuclear Factor Kappa B. TyG

TyG

Triglyceride-glucose index

WBC

White blood cell

Authors’ contributions

Zhang Yining, Guoxiang Liu, Xianxian Yu, Zhaoming Shang, and Huailong Shang analyzed data and wrote the paper. Yining Zhang, Xianxian Yu, Xiao Cui, and Yidan Zhai conducted the statistical analysis. Chenjin Gao, Guoxiang Liu, Jiameng Chen, and Huailong Shang designed and supervised the research. Guoxiang Liu, Zhaoming Shang, Xiao Cui, Jiameng Chen, Jiawei Ye, Jiyuan Zhang, Junwei Qian, Wenjie Liu, Chaoping Ma contributed to data collection and correction. Chengjin Gao, Guoxiang Liu, Bing Zhao, and Mingquan Chen served as the academic consultants. All authors contributed to the process and the editing of the manuscript. All authors provided critical revisions of the manuscript and approved the final manuscript.

Funding

This work was supported by Key Supporting Subject Researching Project of Shanghai Municipal Health Commission (No. 2023ZDFC0106); the Science and Technology of Shanghai Committee (23Y31900100 and 23Y31900102); the National Natural Science Foundation of China (No. 82172138); and the Innovation Research Project of Shanghai Science and Technology Commission (No. 21Y11902400).

Data availability

Some or all datasets generated during and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study has received approval from the Ethics Committee of Xinhua Hospital, Shanghai Jiaotong University School of Medicine.

Consent for publication

The manuscript is original, has not been published elsewhere, and is not under consideration by another journal. All authors have approved the final version and declare no conflicts of interest.

Competing interests

The authors declare no competing interests.