Triglyceride-glucose index trajectories and in-hospital mortality in heart failure: a retrospective cohort study Zongjun Hu, Xiangjun Kong, Quanwei Liu, Jianguo Chen, Xia Huang Abstract BackgroundHigh triglyceride-glucose (TyG) index levels are linked to poor prognosis in the heart failure (HF) population. However, existing studies have predominantly relied on static baseline measurements, and the association of dynamic TyG trajectories with outcomes in critically ill HF patients remains unclear. This study aimed to identify TyG index trajectories in critically ill patients with HF and evaluate their association with in-hospital mortality.MethodsThis retrospective study used the MIMIC-IV (v3.1) as the derivation cohort and the eICU-CRD (v2.0) as the external validation cohort. Patients with HF admitted to the intensive care unit (ICU) were included. We applied group-based trajectory modeling (GBTM) to identify distinct TyG index trajectories during the first 7 days of ICU admission. Multivariable logistic regression was used to calculate odds ratios (ORs) and 95% confidence intervals (CIs) for the association between TyG trajectories and in-hospital mortality. Complete-case analysis, Cox regression, and E-value calculation were conducted as sensitivity analyses.ResultsThis analysis included 2,180 patients from MIMIC-IV and 1,433 from eICU-CRD. GBTM identified three distinct TyG trajectories: Class 1 (low-level), Class 2 (moderate-level), and Class 3 (high-level). The high-level trajectory (Class 3) exhibited the highest in-hospital mortality in both cohorts. After full adjustment, Class 3 was associated with significantly increased mortality compared with Class 1 (MIMIC-IV: OR = 1.62, 95% CI: 1.13–2.31; eICU-CRD: OR = 2.73, 95% CI: 1.59–4.67). Class 2 showed a weaker and non-significant association (MIMIC-IV: OR = 1.17, 95% CI: 0.87–1.58; eICU-CRD: OR = 1.45, 95% CI: 0.97–2.17). Sensitivity analyses showed generally consistent trends.ConclusionIn critically ill patients with HF, a high-level TyG trajectory was independently associated with increased in-hospital mortality, showing a level-dependent gradient across two independent cohorts. Introduction Heart failure (HF) remains a major global health challenge, characterized by its rising prevalence and substantial morbidity (, ). Patients with HF requiring admission to the intensive care unit (ICU) represent a particularly vulnerable population, marked by clinical complexity and high mortality rates (). While hemodynamic instability has traditionally been the primary focus in this setting, HF is increasingly reconceptualized as a complex cardiometabolic syndrome (, ). Central to this metabolic dysfunction is insulin resistance (IR); in the critically ill, the acute stress response exacerbates IR, precipitating a vicious cycle that accelerates systemic organ failure and worsens prognosis (, ). However, accurate quantification of IR in the critical care setting remains elusive. Traditional gold standards, such as the hyperinsulinemic-euglycemic clamp, are invasive and labor-intensive. Similarly, insulin-based surrogates such as the homeostasis model assessment of insulin resistance (HOMA-IR) are frequently rendered unreliable by acute stress hyperglycemia and exogenous insulin use (). In this context, the triglyceride–glucose (TyG) index is increasingly recognized as a validated surrogate for insulin resistance (). Although observational studies and meta-analyses have linked elevated TyG levels to adverse outcomes in patients with heart failure, these investigations have predominantly focused on static, baseline measurements (–). Reliance on a single baseline measurement, however, is inherently limited. A static measurement may not capture the sustained metabolic burden experienced during the early ICU phase. Longitudinal trajectory modeling, which characterizes patient-level patterns over time, can identify distinct TyG trajectory patterns that static metrics may overlook. Although TyG trajectories have been associated with adverse outcomes in sepsis and atrial fibrillation (, ), their association with mortality in critically ill HF patients remains poorly characterized. To ensure generalizability and overcome the limitations inherent to single-center studies, this large-scale retrospective study adopted a derivation–validation design using two independent critical care databases (MIMIC-IV for derivation and eICU-CRD for external validation). By applying group-based trajectory modeling (GBTM), the analysis sought to delineate distinct longitudinal phenotypes of the TyG index during the early ICU phase and evaluate their association with in-hospital mortality. Methods Database This retrospective study utilized a derivation-validation design using data from two distinct critical care databases. The Medical Information Mart for Intensive Care-IV (MIMIC-IV, v3.1) was employed as the derivation cohort, while the eICU Collaborative Research Database (eICU-CRD, v2.0) served as the external validation cohort. MIMIC-IV is a comprehensive single-center registry from Beth Israel Deaconess Medical Center (Boston, MA), encompassing 94,458 ICU admissions from 2008 to 2022 (). In contrast, the eICU-CRD is a multicenter repository containing over 200,000 admissions from 208 hospitals across the United States (2014–2015) (). Both datasets are hosted on PhysioNet (https://physionet.org/). Access was granted to author Zongjun Hu (Credential ID: 60604021) following completion of the Collaborative Institutional Training Initiative (CITI) curriculum. Both databases are publicly available de-identified datasets; as a secondary analysis of these data, no additional ethics approval or informed consent was required. The study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Study participants This study included patients with HF admitted to the ICU, identified from the MIMIC-IV (n = 2,180) and eICU-CRD (n = 1,433) databases. HF diagnoses were ascertained based on International Classification of Diseases, Ninth and Tenth Revision (ICD-9/ICD-10) codes () (, ). Exclusion criteria were as follows: (1) re-admissions (only the first ICU stay was analyzed); (2) insufficient longitudinal glucose data, defined as fewer than three daily mean values within the first 7 days; and (3) absence of triglyceride measurements within the same 7-day window. Data retrieval and variable definition Data retrieval was conducted via Structured Query Language (SQL) scripts using Navicat Premium. Baseline laboratory parameters and vital signs were retrieved from records within the initial 24 h of ICU admission; for variables with repeated measurements during this window, arithmetic means were calculated. The retrieved variables were classified into seven principal domains: (1) Demographics (age, sex, weight); (2) Vital signs (temperature, heart rate, respiratory rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), peripheral oxygen saturation (SpO2)); (3) Comorbidities (hypertension, diabetes, chronic kidney disease, cerebrovascular disease, atrial fibrillation, chronic pulmonary disease, liver disease, coronary artery disease, malignancy); (4) Laboratory data (white blood cells (WBC), platelets, hemoglobin, creatinine, blood urea nitrogen (BUN), sodium, potassium, chloride, bicarbonate); (5) Severity scores (Acute Physiology and Chronic Health Evaluation (APACHE) II for MIMIC-IV, APACHE IV for eICU-CRD); (6) Medications (angiotensin-converting enzyme inhibitors/angiotensin receptor blockers (ACEIs/ARBs), β-blockers, diuretics, statins, glucocorticoids, insulin); and (7) Interventions (vasoactive agents, invasive mechanical ventilation (IMV), renal replacement therapy (RRT)). Longitudinal TyG calculation and outcomes Consistent with established methodologies (, ), consecutive 24 h windows were established commencing at ICU admission, with mean glucose values computed for each interval. Given the intermittent nature of lipid profiling in the ICU, a temporal matching protocol was implemented whereby each triglyceride measurement was paired with the mean glucose value from the corresponding 24 h window. The TyG index was derived using the formula: ln[triglyceride (mg/dL) × glucose (mg/dL)/2] (). For patients with an ICU length of stay ≥7 days, the analysis covered the first 7 days. For those with shorter stays, all available data were utilized. The primary endpoint was all-cause in-hospital mortality during the index admission. Group-based trajectory modeling Group-based trajectory modeling (GBTM) was applied to characterize distinct longitudinal TyG patterns among patients with a minimum of three recorded values (). Analyses were performed using the lcmm package in R (), which implements group-based trajectory modeling and inherently accommodates unbalanced longitudinal data, such that patients with different numbers of observations contribute information. Trajectories were modeled as second-order polynomials (incorporating linear and quadratic terms), with parameters estimated via maximum likelihood. Model selection was guided by the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), sample-size-adjusted BIC (SABIC), and the average posterior probability of assignment (AvePP) (). Final model determination balanced statistical adequacy (requiring AvePP >0.70 and a minimum class proportion of 5%) with clinical interpretability. Statistical analysis Multiple imputation was applied to manage incomplete data, following the exclusion of variables with > 20% missingness (). To describe the study population, baseline characteristics were compared across TyG trajectory groups. For continuous variables, data were presented as mean ± SD or median (IQR), contingent on the distribution; differences among groups were evaluated using one-way ANOVA or the Kruskal–Wallis test. Categorical data were expressed as frequencies (percentages) and analyzed via the chi-square test. To evaluate potential selection bias, baseline characteristics were compared between included and excluded patients using standardized mean differences (SMDs), with SMD >0.1 considered indicative of a meaningful between-group difference. Excluded patients comprised those with fewer than three glucose measurements or missing triglyceride measurements within the first 7 days of ICU admission. Multivariable logistic regression was employed to evaluate the association between TyG trajectory groups and in-hospital mortality, reporting odds ratios (ORs) with 95% confidence intervals (CIs). Covariates were selected based on clinical relevance, prior literature, and the change-in-estimate criterion (>10% change in OR), resulting in the inclusion of 22 variables. Collinearity diagnostics confirmed that variance inflation factors (VIFs) for all covariates were < 5 (). Four hierarchical models were constructed: the Crude Model (unadjusted); Model 1 (adjusted for age, sex, and weight); Model 2 (additionally adjusted for temperature, respiratory rate, WBC, BUN, APACHE scores, hypertension, cerebrovascular disease, chronic pulmonary disease, liver disease, diabetes, chronic kidney disease, coronary artery disease, and atrial fibrillation); and Model 3 (further adjusted for β-blockers, diuretics, statins, glucocorticoids, vasoactive agents, and IMV). Subgroup analyses were conducted by age (<65 vs. ≥65 years), sex, and key comorbidities (hypertension, diabetes, chronic kidney disease, and atrial fibrillation) to investigate potential effect modification in the MIMIC-IV cohort, with interactions assessed via likelihood ratio tests. To verify the robustness of the primary results, three sensitivity analyses were undertaken: (1) complete-case analysis (excluding participants with missing covariates); (2) calculation of E-values to quantify the potential impact of unmeasured confounding (); and (3) Cox regression analysis using hospital length of stay as the time variable. All analyses were performed using R 4.6.1. A two-sided P value less than 0.05 was considered statistically significant. Results Study population and trajectory modeling The initial screening identified 24,942 critically ill patients with HF (14,984 from MIMIC-IV and 9,958 from eICU-CRD). After applying exclusion criteria, 2,180 patients from the MIMIC-IV derivation cohort and 1,433 from the eICU-CRD validation cohort were retained for final analysis (). To characterize the excluded population, baseline characteristics of included (n = 2,180) and excluded (n = 12,804) patients were compared (). Excluded patients were older (median 75.0 vs. 71.2 years) and exhibited a profile of greater physiological derangement, including lower temperature (36.6 vs. 36.8 °C) and diastolic blood pressure (64.0 vs. 64.4 mmHg), higher heart rate (85.0 vs. 83.8 beats/min), and slightly higher APACHE II scores (19.0 vs. 18.0), whereas other baseline characteristics were largely comparable (SMD < 0.1). In-hospital mortality was slightly higher in the excluded group (21.0% vs. 17.8%; SMD = 0.081). Using GBTM, we identified three distinct longitudinal TyG phenotypes. The three-trajectory model was selected as the optimal solution based on model fit indices (lower AIC, BIC, and SABIC), assignment quality (AvePP >0.70), a minimum class proportion exceeding 5%, and clinical interpretability (). The identified trajectories were characterized as follows: Class 1 (low-level), comprising 40.1% of the MIMIC-IV cohort and 39.7% of the eICU-CRD cohort; Class 2 (moderate-level), comprising 39.5% (MIMIC-IV) and 49.2% (eICU-CRD); Class 3 (high-level), comprising 20.3% (MIMIC-IV) and 11.1% (eICU-CRD) (). Baseline characteristics In the MIMIC-IV cohort, compared with Class 1, Class 3 patients were younger (66.7 vs. 73.8 years) (), had higher body weight (87.7 vs. 76.7 kg), and presented with markedly greater prevalence of diabetes (64.1% vs. 21.3%) and chronic kidney disease (37.7% vs. 28.3%). Disease severity scores were highest in Class 3, with APACHE II substantially elevated (median 22.0 vs. 17.0). Therapeutic intensity was also greatest in this group: insulin use (73.4% vs. 42.9%), vasoactive agents (51.9% vs. 33.4%), invasive mechanical ventilation (64.1% vs. 31.7%), and renal replacement therapy (12.2% vs. 4.6%) were all markedly higher. Biochemically, Class 3 patients showed elevated white blood cell count (12.2 vs. 10.1 × 109/L), creatinine (1.5 vs. 1.1 mg/dL), and blood urea nitrogen (28.5 vs. 22.0 mg/dL). Unadjusted in-hospital mortality was highest in Class 3 across both cohorts (27.5% in MIMIC-IV and 25.2% in eICU-CRD, both P < 0.001) (; ). Association between TyG trajectories and in-hospital mortality In the MIMIC-IV derivation cohort (), with Class 1 (low-level) as the reference, Class 3 exhibited a significantly elevated risk of in-hospital mortality in the crude model (OR = 2.49, 95% CI: 1.87–3.31, P < 0.001), and this association remained significant after full adjustment (OR = 1.62, 95% CI: 1.13–2.31, P = 0.009). Class 2 (moderate-level) showed a weaker and non-significant association after full adjustment (OR = 1.17, 95% CI: 0.87–1.58, P = 0.302). In the eICU-CRD validation cohort, Class 3 remained significantly associated with mortality after full adjustment (OR = 2.73, 95% CI: 1.59–4.67, P < 0.001), whereas Class 2 did not reach significance (OR = 1.45, 95% CI: 0.97–2.17, P = 0.071). Subgroup analysis In the MIMIC-IV cohort (), the association between Class 3 and in-hospital mortality remained stable across most subgroups, with no significant interaction observed for age, sex, hypertension, chronic kidney disease, or atrial fibrillation (all P for interaction > 0.05). Sensitivity analyses In the complete-case analysis excluding participants with missing covariates (), Class 3 remained associated with increased mortality in both cohorts (MIMIC-IV: OR = 1.68, 95% CI: 1.14–2.48, P = 0.008; eICU-CRD: OR = 3.02, 95% CI: 1.51–6.01, P = 0.001). Cox regression analysis using hospital length of stay as the time variable yielded directionally consistent results (): with Class 1 as the reference, eICU-CRD Class 3 hazard ratio (HR) = 2.09 (95% CI: 1.30–3.36, P = 0.003), with MIMIC-IV showing a similar direction albeit not reaching statistical significance (HR = 1.23, 95% CI: 0.92–1.65, P = 0.166). E-values were 2.62 (MIMIC-IV) and 4.90 (eICU-CRD) (). Discussion In this retrospective study using a dual database derivation-validation design, we characterized the longitudinal trajectories of the TyG index among critically ill patients with heart failure. Our analysis identified three distinct TyG trajectory patterns: a low-level (Class 1), a moderate-level (Class 2), and a high-level (Class 3) trajectory. The key finding is that the high-level trajectory (Class 3) was independently associated with the highest risk of in-hospital mortality in both cohorts, whereas the moderate-level trajectory (Class 2) showed a weaker and non-significant association. These results suggest a level-dependent gradient between TyG trajectory and mortality risk. Subgroup analyses further showed that this association remained stable across most predefined subgroups. In critically ill patients, the TyG index has been consistently associated with adverse outcomes. Liao et al. reported that the TyG index was linked to all-cause mortality in a cohort of 3,026 critically ill patients (), and Zheng et al. demonstrated its association with in-hospital mortality in sepsis using the MIMIC-IV database (); Wang et al. further confirmed its prognostic value in critically ill patients with coronary artery disease (). The prognostic relevance of the TyG index also extends to heart failure. Zhou et al. showed that in patients with heart failure with preserved ejection fraction, a high TyG index was associated with increased mortality and rehospitalization (), and Xiao et al. demonstrated an association between the TyG index at ICU admission and short-term all-cause mortality in critically ill patients with heart failure (). However, these studies predominantly relied on static, baseline measurements, which may not fully capture the dynamic metabolic burden experienced during critical illness. In the ICU setting, dynamic changes in the TyG index during hospital stay have been associated with in-hospital and 1-year all-cause mortality (), and trajectory-based approaches have linked distinct TyG trajectories to mortality in sepsis ()and atrial fibrillation (). Our study extends this line of evidence by applying GBTM-based trajectory classification specifically to critically ill patients with heart failure, and by confirming the findings in an independent validation cohort (eICU-CRD), thereby addressing the generalizability concern of single-database studies. Patients assigned to the high-level trajectory had the highest prevalence of diabetes (64.1% vs. 21.3% in the low-level group) and the highest APACHE II scores (22.0 vs. 17.0) at baseline, suggesting a more pronounced state of insulin resistance and metabolic burden in this group. Insulin resistance impairs myocardial glucose utilization and shifts substrate preference toward fatty acid oxidation, leading to intramyocardial lipid accumulation and lipotoxicity, which promotes myocardial dysfunction and progression of heart failure (, ). In the ICU setting, the acute stress response—characterized by elevated catecholamines, cortisol, and inflammatory cytokines—induces stress hyperglycemia and further aggravates insulin resistance (, ), and a persistently high TyG trajectory may capture this sustained state of stress-related metabolic dysregulation. Accordingly, these metabolic perturbations may contribute to the higher mortality observed in the high-level group. Subgroup analyses showed that the association between the high-level trajectory and in-hospital mortality remained stable across most predefined strata (all P for interaction > 0.05). Notably, the association appeared stronger in non-diabetic patients than in diabetic patients, although the interaction with diabetes did not reach statistical significance (P for interaction = 0.06). This differential pattern may reflect the distinct insulin resistance states between diabetic and non-diabetic patients: diabetic patients exhibit chronic insulin resistance with higher baseline TyG levels (), which may attenuate the additional prognostic information conveyed by dynamic metabolic changes during critical illness. These findings warrant cautious interpretation and confirmation in larger cohorts. The TyG index can be readily calculated from routinely measured glucose and triglyceride levels at no additional cost, without requiring additional blood draws or specialized assays. Trajectory analysis may therefore help identify patients with a persistently high-level TyG trajectory during the early ICU phase, who face a substantially elevated risk of in-hospital mortality, and may facilitate the characterization of dynamic metabolic status in critically ill patients with heart failure. Several limitations warrant consideration. First, as a retrospective observational study, causal inference cannot be established, and the observed associations between TyG trajectories and in-hospital mortality may be influenced by unmeasured confounders and reverse causality. Moreover, trajectories were constructed from TyG measurements during the first 7 days of ICU admission (at least three measurements), and this exposure window partially overlaps with the outcome assessment period. Second, NT-proBNP and lactate were excluded due to high missingness and could act as unmeasured confounders; however, E-values (2.62 and 4.90) suggest that residual confounding of sufficient magnitude is unlikely. Third, excluding patients with fewer than three measurements may introduce survivor bias, as excluded patients were older with greater physiological derangement (); nevertheless, in-hospital mortality was comparable between included and excluded patients (SMD = 0.081 < 0.1), indicating that the selection process had limited impact on the primary outcome. Given that the inclusion criteria required at least three measurements, our findings apply to patients who survive the initial acute period. In addition, ICU length of stay is correlated with disease severity and may influence trajectory estimation. Fourth, triglyceride concentrations are influenced by ICU factors (insulin, nutrition, catecholamines), and TyG dynamics are predominantly driven by glucose fluctuations (); accordingly, our analysis focused on sustained TyG levels rather than day-to-day variability. Fifth, glucose was not measured under standardized fasting conditions, which is consistent with the real-world ICU setting, and prior studies have shown that the TyG index remains predictive under non-fasting conditions (–). Finally, both cohorts were derived from US critical care settings, which may limit the generalizability of our findings to other populations; future prospective multicenter studies are warranted to validate these results and to assess whether trajectory-guided metabolic management improves clinical outcomes. Conclusion In critically ill patients with HF, a high-level TyG index trajectory during the early ICU period is independently associated with increased in-hospital mortality, showing a level-dependent gradient across two independent cohorts. These findings suggest that sustained high TyG levels represent a distinct high-risk metabolic profile. Prospective studies are warranted to determine whether trajectory-guided metabolic management can improve outcomes.