Association of triglyceride-glucose index fluctuation with in-hospital all-cause mortality in critically ill patients: a multidatabase retrospective study

Abstract

This study aims to evaluate the relationship between triglyceride-glucose (TyG) index variability and in-hospital mortality across two intensive care unit (ICU) databases (MIMIC-IV and SWH) and to determine if TyG variability provides more prognostic value than baseline TyG. This retrospective observational study utilized data from the MIMIC-IV (2008–2019) and SWH (2016–2023) ICU cohorts. TyG metrics, including baseline (TyG-BL), median (TyG-Median), mean (TyG-Average), standard deviation (TyG-STD), range (TyG-Range), and coefficient of variation (TyG-CV), were calculated. The association between TyG metrics and all-cause in-hospital mortality was evaluated using multivariable logistic regression. Nonlinear relationships were explored using restricted cubic splines (RCS). Cox proportional hazards models and subgroup analyses were conducted as sensitivity analyses, and post-discharge mortality at 7 days, 1 year, and 2 years was additionally assessed. A total of 2,208 ICU patients were included (MIMIC: n = 1707; SWH: n = 501). In MIMIC, TyG variability metrics (TyG-STD, TyG-Range) were independently associated with mortality, with TyG-STD showing an odds ratio (OR) of 1.66 (95% confidence interval [CI] 1.06–2.61, P = 0.027) and TyG-Range an OR of 1.24 (95% CI 1.05–1.46, P = 0.011). TyG variability metrics in SWH showed similar trends, but associations attenuated after full adjustment. Kaplan–Meier analysis demonstrated clear survival curve separation for TyG variability metrics in MIMIC, while the SWH cohort showed weaker separation. RCS analysis revealed a nonlinear relationship between TyG metrics and mortality risk in MIMIC, with a steeper increase in risk at higher TyG values. In Cox analyses, baseline TyG was associated with the timing of in-hospital death in MIMIC, whereas TyG variability metrics were not. No significant associations were observed between TyG metrics and post-discharge mortality. TyG variability, rather than a single baseline TyG measurement, is associated with in-hospital mortality in critically ill patients, particularly in the MIMIC cohort. These findings suggest that dynamic metabolic instability may provide clinically relevant prognostic information beyond static measurements. Further prospective and multicenter studies are warranted to validate the role of TyG variability in ICU risk stratification.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-42020-1.

Introduction

Despite advances in organ support and infection management, patients in the intensive care unit (ICU) continue to experience high morbidity, mortality, and healthcare utilization¹. Early and accurate risk assessment is therefore essential to guide therapy^2–4. However, biomarkers that are robust, repeatable, and generalizable for mortality prediction in the ICU remain limited^5–8.

Insulin resistance (IR) is a common and clinically important metabolic disturbance in critical illness, driven by stress responses, systemic inflammation, and immune dysregulation, and it is associated with adverse outcomes⁹. Prior work indicates that insulin sensitivity fluctuates dynamically during the course of critical illness and may be 50–70% lower than in healthy individuals¹⁰. Standardized IR assessments-such as the hyperinsulinemic–euglycemic clamp or HOMA-IR, which requires insulin measurements-are impractical for routine ICU use, limiting their bedside applicability^11–15.

The triglyceride-glucose (TyG) index, calculated from routine glucose and triglyceride tests, has emerged as a convenient surrogate of IR and has been linked to cardiometabolic risk and mortality across diverse populations¹⁶. Several studies suggest that TyG performs as well as or better than HOMA-IR for identifying metabolic syndrome (e.g., area under the ROC curve = 0.84 vs. 0.68)¹⁷. Prospective cohorts also indicate that cumulative/long-term TyG exposure correlates with increased cardiovascular risk^16,18. In critical care, admission (baseline) TyG has been associated with mortality. While baseline TyG has shown significant associations with adverse outcomes in various populations, it is primarily a static measure that reflects an individual’s metabolic status at a single point in time. However, metabolic instability—as reflected by fluctuations or variability in biomarkers such as glucose levels—has been increasingly recognized as an important prognostic factor in critically ill patients. A systematic review and meta-analysis found that increased glycemic variability (e.g., standard deviation and coefficient of variation of blood glucose) was consistently associated with higher short-term mortality in ICU populations, independent of mean glucose levels¹⁹. Individual cohort studies also show that high glucose variability is associated with increased ICU mortality, including in large heterogeneous ICU cohorts²⁰, and elevated glycemic variability is linked to prolonged ICU stay and mortality in specific patient subgroups²¹. By extension, fluctuations in TyG levels—which integrate both glucose and lipid metabolism and reflect dynamic changes in IR—may similarly capture metabolic dysregulation and physiological stress responses that are not fully represented by a single baseline measure. In the ICU setting, where patients are under continuous metabolic stress, this variability might reflect the body’s inability to maintain metabolic homeostasis and thus contribute to higher mortality risk. In contrast, baseline TyG levels alone may not fully capture these dynamic metabolic changes and may therefore underestimate the true risk of adverse outcomes in critically ill patients. However, evidence remains limited regarding whether time-series changes and variability in TyG during the ICU/hospitalization period (e.g., standard deviation, coefficient of variation, range) provide incremental prognostic information^12,22,23. Moreover, the consistency and reproducibility of TyG-related metrics across different health systems, patient populations, and care acuity have not been well characterized.

Against this background, we constructed during ICU hospitalization a panel of TyG metrics that capture both level (baseline, median, mean) and temporal variability (standard deviation, coefficient of variation, range). We then evaluated their associations with in-hospital all-cause mortality using multivariable models and explored potential nonlinear exposure–risk relationships via restricted cubic splines. In parallel, we applied an identical analytic workflow to two independent data sources-MIMIC-IV (USA) and a SWH (China)-to compare consistency and reproducibility of findings across clinical contexts.

Methods

Data source and study population

This retrospective observational study used de-identified electronic health record data from two sources: (i) MIMIC-IV (version 2.0) contains hospital and ICU data from Beth Israel Deaconess Medical Center (Boston, USA) for admissions from 2008 to 2019, prior to the COVID-19 pandemic; all personal identifiers in MIMIC-IV are irreversibly de-identified. (ii) SWH cohort: data were obtained from the Clinical Big Data Center of the First Affiliated Hospital of the Army Medical University (Southwest Hospital, Chongqing, China) for hospitalizations between 2016 and 2023; patient identifiers were de-identified prior to analysis.

To enhance comparability, patients with confirmed COVID-19 were excluded from both datasets. The study was approved by the Ethics Committee of the First Affiliated Hospital of the Army Medical University (People’s Liberation Army) (approval No. KY2024116) with a waiver of informed consent due to de-identified data, and it was registered in the China Clinical Trial Registry (ChiCTR2400086782, registration date: July 10, 2024). All procedures adhered to relevant regulations and the Declaration of Helsinki.

Inclusion criteria: age > 18 years; first ICU admission during the index hospitalization; ICU length of stay > 24 h; ≥ 2 paired measurements of blood glucose and triglycerides to compute TyG variability. Exclusion criteria: (1) confirmed COVID-19; (2) missing key demographic or outcome information. Only the first eligible hospitalization for patients with multiple admissions was retained. The unit of analysis was the first ICU admission within the index hospitalization.

Data collection

Data from MIMIC-IV were extracted using structured SQL queries (PostgreSQL 11.0). Data from SWH were exported from the institutional data platform and processed with standardized scripts for statistical analysis. Variable definitions and coding were harmonized a priori across the two datasets. Demographics were obtained at admission. Vital signs were averaged over the first 24 h after ICU admission. For laboratory tests other than blood glucose and triglycerides, the first value after ICU admission was used. Height and weight were taken from measurements within 24 h before ICU admission, when available.

The SOFA score was computable in MIMIC-IV based on components within the first 24 h after ICU admission; SOFA was not uniformly available in SWH. Other covariates included hemoglobin, platelet count, red blood cell count, red cell distribution width, white blood cell count, neutrophil-to-lymphocyte ratio (NLR), albumin, bicarbonate (in MIMIC-IV), creatinine, sodium, calcium, potassium, prothrombin time (PT), activated partial thromboplastin time (PTT), alanine aminotransferase (ALT), alkaline phosphatase (ALP), aspartate aminotransferase (AST), lactate dehydrogenase (LDH), and lipids.

To mitigate potential bias, following Zhang et al.²⁴, variables with > 50% missingness were excluded from multivariable modeling. For remaining missing data, under the missing-at-random assumption we applied multiple imputation by chained equations with m = 10 imputed datasets and 10 iterations each. All covariates were included in the imputation model; the outcome was not imputed. Parameter estimates were combined using Rubin’s rules.

Assessment of the TyG index and related parameters

The TyG index was calculated as ln[TG (mg/dl) × FPG (mg/dl)/2]. To ensure unit consistency, results reported in mmol/L were converted to mg/dL (glucose × 18; triglycerides × 88.57). Because fasting status was not consistently recorded, we used routinely measured serum glucose and triglyceride values. To construct the TyG time series during the ICU stay, each glucose measurement was paired with the temporally closest triglyceride measurement; if the sampling interval between the pair exceeded 6 h, no TyG value was generated for that time point. For each patient’s ICU TyG series, we then derived the following metrics: TyG-BL (first TyG value after ICU admission), TyG-Median (median value of the TyG index sequence), TyG- Average (mean of the TyG index sequence), TyG-STD (standard deviation of the TyG index sequence), TyG-Range (range of the TyG index sequence, maximum—minimum), and TyG-CV (coefficient of variation, SD/mean). TyG-BL, TyG-Median, and TyG-Mean primarily reflect the absolute level, whereas TyG-STD, TyG-Range, and TyG-CV reflect variability/dispersion over time.

Primary and secondary outcomes

The primary outcome was in-hospital all-cause mortality during the index hospitalization; the secondary outcome was ICU mortality. For survival summaries, the time origin was ICU admission. Vital status after discharge was obtained from linked follow-up records, and mortality at 7 days, 1 year, and 2 years after discharge was assessed. Patients were censored at the last known follow-up.

Statistical analysis

Continuous variables were first assessed for normality and homogeneity of variance. Normally distributed variables with equal variances are presented as mean ± standard deviation (Mean ± SD) and compared using the Student’s t-test or one-way ANOVA. Non-normally distributed or heteroscedastic variables are presented as median [Q1, Q3] and compared using the Mann–Whitney U test or Kruskal–Wallis test. Categorical variables are summarized as n (%) and compared using the χ² test or Fisher’s exact test. All tests were two-sided, with P < 0.05 indicating statistical significance.

To examine the association between TyG level/variability and the primary outcome, we classified patients separately within MIMIC-IV and SWH into tertiles (Groups 1–3: low/middle/high) according to the distribution of each TyG metric in the corresponding dataset. With ICU admission as time zero and discharge as censoring, Kaplan–Meier curves were used to estimate the cumulative incidence of in-hospital death across tertiles, and differences were assessed by the log-rank test. For effect estimation with continuous exposures, we fitted multivariable logistic regression models and reported odds ratios (ORs) with 95% confidence intervals (CIs). Pre-specified adjustment sets were: Model 1, unadjusted (TyG metric only); Model 2, adjusted for age, sex, and BMI; and Model 3, further adjusted for clinical and laboratory covariates supported by prior evidence-SOFA (MIMIC-IV only), hemoglobin, platelet count, red blood cell count, RDW (MIMIC-IV), white blood cell count, NLR, albumin, bicarbonate (MIMIC-IV), creatinine, sodium, calcium, potassium, PT, PTT, ALT, ALP, AST, LDH, and lipids (LDL/HDL in SWH). Because covariate availability differed between datasets, models were fit and reported separately for MIMIC-IV and SWH.

To assess potential nonlinear relationships between each continuous TyG metric and the outcome, we incorporated restricted cubic splines (RCS) into the regression framework, placing knots at the 5th, 35th, 65th, and 95th percentiles as recommended by Harrell^25,26; the sample median served as the reference value. Multicollinearity was evaluated using the variance inflation factor (VIF), with VIF < 5 considered acceptable.

In addition, Cox proportional hazards models were fitted as sensitivity analyses with ICU admission as the time origin to account for the time-to-event nature of mortality outcomes, and hazard ratios (HRs) with 95% CIs were reported. Patients were censored at discharge for in-hospital outcomes and at the last available follow-up for post-discharge analyses. Prespecified subgroup analyses were conducted to evaluate the robustness of the associations across clinically relevant subgroups, and interaction terms were tested where appropriate.

Analyses were performed using Python 3.7.5 and R 4.3.3. Two-sided P < 0.05 was considered statistically significant.

Results

Baseline characteristics of study population

As shown in Fig. 1 and Table 1, a total of 2,208 ICU patients were included (MIMIC, n = 1,707; SWH, n = 501). The sex distribution was similar in both cohorts (male 61.2% vs 61.1%). Patients in MIMIC were older (60.6 [47.7–70.5] years) than those in SWH (56.7 [41.1–70.3] years). ICU length of stay was comparable (MIMIC 8.6 [3.4–16.1] vs SWH 8.9 [5.2–15.4] days), as was hospital length of stay (MIMIC 22.7 [13.4–37.1] vs SWH 21.0 [13.0–39.0] days). In-hospital mortality was 23.2% in MIMIC and 14.6% in SWH. Anthropometrics indicated higher weight and BMI in MIMIC (BMI 28.9 [24.5–34.2] vs 23.6 [23.1–23.8] kg/m²). In the MIMIC cohort, 90-day ICU mortality occurred in 57 of 562 patients (10.1%) in the low TyG group, 70 of 563 patients (12.4%) in the middle TyG group, and 59 of 581 patients (10.2%) in the high TyG group, with no statistically significant differences across tertiles (P = 0.363). In the SWH cohort, corresponding mortality rates were 4/165 (2.4%), 3/165 (1.8%), and 7/170 (4.1%), respectively, also showing no significant differences (P = 0.416) (Table 2).

Fig. 1.

The flowchart of study participants.

Table 1.

Demographic and clinical characteristics of patients in the MIMIC and SWH cohorts.

Variable	MIMIC (N = 1707)	SWH (N = 501)
Gender, n (%)
Male	1045 (61.2)	306 (61.1)
Female	662 (38.8)	195 (38.9)
Admission Age (Years)	60.6 [47.7, 70.5]	56.7 [41.1, 70.3]
LOS (ICU) (Days)	8.6 [3.4, 16.1]	8.9 [5.2, 15.4]
LOS (Hospital) (Days)	22.7 [13.4, 37.1]	21.0 [13.0, 39.0]
Height (cm)	170.1 [165.0, 175.0]	162.3 [162.0, 163.0]
Weight (kg)	83.9 [69.5, 101.0]	62.2 [60.0, 62.2]
BMI (kg/m2)	28.9 [24.5, 34.2]	23.6 [23.1, 23.8]
Hospital Expire Flag, n (%)
No	1311 (76.8)	428 (85.4)
Yes	396 (23.2)	73 (14.6)
SOFA Index (Score)	6.0 [3.0, 9.0]	–
Hemoglobin (g/dL)	10.1 [8.5, 11.8]	8.8 [7.9, 10.3]
Platelet (103/µL)	179.0 [113.0, 253.0]	155.0 [83.0, 232.0]
RBC (10⁶/µL)	3.4 [2.8, 4.0]	3.0 [2.6, 3.5]
RDW (%)	15.0 [13.8, 17.0]	–
WBC (103/µL)	11.3 [7.8, 16.3]	9.7 [6.4, 13.8]
Neutrophils (103/µL)	10.9 [6.6, 13.0]	7.9 [5.0, 11.6]
Lymphocytes (103/µL)	1.2 [0.7, 1.4]	0.9 [0.6, 1.3]
NLR	7.9 [5.9, 14.0]	8.4 [5.2, 15.7]
Albumin (g/dL)	2.7 [2.4, 3.1]	3.3 [3.0, 3.6]
Bicarbonate (mEq/L)	22.0 [19.0, 25.0]	–
Creatinine (mg/dL)	1.1 [0.7, 1.8]	0.9 [0.6, 1.9]
Sodium (mEq/L)	138.0 [135.0, 142.0]	140.0 [136.6, 144.0]
Calcium (mg/dL)	8.1 [7.5, 8.6]	1.8 [1.8, 2.1]
Potassium (mEq/L)	4.1 [3.7, 4.6]	4.0 [3.7, 4.3]
PT (Seconds)	14.5 [12.9, 17.2]	12.6 [11.7, 14.2]
PTT (Seconds)	32.0 [27.9, 39.4]	32.5 [28.4, 39.5]
ALT (U/L)	32.0 [16.0, 94.5]	26.1 [12.8, 53.4]
ALP (U/L)	83.0 [59.0, 116.0]	90.0 [67.0, 141.1]
AST (U/L)	47.0 [25.0, 140.0]	43.7 [25.4, 96.1]
LD_LDH (U/L)	381.0 [243.0, 615.2]	518.4 [293.2, 906.3]
LDL (mg/dL)	–	1.7 [1.1, 2.3]
HDL (mg/dL)	–	0.6 [0.4, 0.9]
CKMB (ng/mL)	–	18.8 [10.4, 28.1]
Glucose (mg/dL)	127.0 [105.0, 161.5]	142.2 [111.6, 187.2]
Triglyceride (mg/dL)	170.0 [110.5, 275.5]	159.4 [106.3, 256.9]
Heart Rate (bpm)	91.9 [78.8, 105.4]	92.3 [83.2, 101.0]
SBP (mmHg)	114.9 [104.3, 124.2]	127.0 [117.0, 135.5]
DBP (mmHg)	65.6 [58.4, 71.1]	70.9 [66.0, 77.0]
MBP (mmHg)	81.9 [74.5, 87.8]	89.5 [83.7, 96.0]
Respiratory Rate (breaths/min)	20.4 [17.7, 23.6]	20.0 [19.0, 20.7]
Temperature (°C)	37.0 [36.7, 37.4]	37.0 [36.7, 37.5]
SpO2 (%)	96.9 [95.3, 98.3]	98.3 [98.1, 99.2]

This table presents the demographic and clinical characteristics of patients in the MIMIC and SWH cohorts. Continuous variables are reported as median (Q1, Q3), while categorical variables are expressed as counts (n) and percentages (%). For some variables marked with a "—" (dash), data collection was not performed or was unavailable in that cohort. Abbreviations: LOS, length of stay; ICU, Intensive Care Unit; BMI, body mass index; SOFA, sequential organ failure assessment; RDW, red cell distribution width; WBC, white blood cell count; RBC, red blood cell count; NLR, neutrophil-to-lymphocyte ratio; PT, prothrombin time; PTT, partial thromboplastin time; ALT, alanine aminotransferase; ALP, alkaline phosphatase; AST, aspartate aminotransferase; LD_LDH, lactate dehydrogenase; LDL, low-density lipoprotein; HDL, high-density lipoprotein; CKMB, creatine kinase-MB; SpO₂: Oxygen saturation; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure.

Table 2.

Baseline Characteristics and Clinical Outcomes of ICU Patients Stratified by TyG Index Quantile in the MIMIC and SWH Databases.

Variable	MIMIC			P-Value	SWH			P-Value
TyG-index	Low (N = 562)	Middle (N = 563)	High (N = 581)		Low (N = 165)	Middle (N = 165)	High (N = 170)
Gender, n (%)				0.202				0.309
Male	232 (41.3)	220 (39.1)	210 (36.1)		59 (35.8)	72 (43.6)	64 (37.6)
Female	330 (58.7)	343 (60.9)	371 (63.9)		106 (64.2)	93 (56.4)	106 (62.4)
Diabetes				< 0.001				< 0.001
No	488 (86.8)	438 (77.8)	399 (68.7)		154 (93.3)	133 (80.6)	125 (73.5)
Yes	74 (13.2)	125 (22.2)	182 (31.3)		11 (6.7)	32 (19.4)	45 (26.5)
Sepsis				0.003				0.016
No	129 (23.0)	93 (16.5)	92 (15.8)		75 (45.5)	98 (59.4)	78 (45.9)
Yes	433 (77.0)	470 (83.5)	489 (84.2)		90 (54.5)	67 (40.6)	92 (54.1)
Admission Age (Years)	64.2 [53.5,75.3]	61.0 [49.2,70.7]	56.5 [43.1,65.9]	< 0.001	63.8 [48.7,73.1]	58.7 [41.1,70.3]	50.5 [36.5,63.6]	< 0.001
LOS (ICU) (Days)	6.6 [2.6,13.9]	8.7 [3.5,16.6]	10.4 [4.7,18.0]	< 0.001	9.9 [5.6,15.8]	8.2 [4.9,14.0]	9.6 [5.5,15.7]	0.23
LOS (Hospital) (Days)	23.9 [14.2,38.1]	24.3 [14.3,38.1]	20.4 [11.6,34.4]	0.001	23.0 [14.0,40.0]	19.0 [13.0,34.0]	22.0 [14.0,39.8]	0.242
Height (cm)	170.1 [165.0,173.0]	170.1 [165.0,175.0]	170.1 [165.0,175.0]	0.195	162.3 [161.0,162.3]	162.3 [162.0,163.0]	162.3 [161.2,163.4]	0.925
Weight (kg)	76.3 [63.0,91.5]	83.9 [70.9,100.9]	91.5 [75.3,107.0]	< 0.001	62.2 [58.0,62.2]	62.2 [61.0,62.2]	62.2 [60.2,63.5]	0.224
BMI (kg/m2)	26.5 [22.3,30.9]	28.9 [24.8,33.9]	31.1 [26.3,36.9]	< 0.001	23.6 [22.3,23.6]	23.6 [23.2,24.3]	23.6 [23.6,24.2]	0.004
Hospital Expire Flag				0.912				0.484
No	435 (77.4)	430 (76.4)	445 (76.6)		139 (84.2)	139 (84.2)	150 (88.2)
Yes	127 (22.6)	133 (23.6)	136 (23.4)		26 (15.8)	26 (15.8)	20 (11.8)
SOFA Index (Score)	6.0 [3.0,8.0]	6.0 [3.0,9.0]	7.0 [4.0,10.0]	< 0.001	-	-	-	-
Hemoglobin (g/dL)	10.0 [8.5,11.7]	10.0 [8.5,11.7]	10.3 [8.5,12.1]	0.18	9.0 [8.1,10.5]	8.7 [7.9,10.2]	8.7 [7.7,10.1]	0.347
Platelet (103/µL)	180.5 [110.0,257.0]	185.0 [120.0,261.0]	173.0 [110.0,247.0]	0.338	160.0 [83.0,232.0]	172.0 [86.0,246.0]	141.0 [80.0,217.0]	0.232
RBC (10⁶/µL)	3.3 [2.8,3.9]	3.4 [2.8,4.0]	3.5 [2.9,4.0]	0.078	3.0 [2.7,3.6]	3.0 [2.6,3.5]	2.9 [2.5,3.4]	0.128
RDW (%)	15.4 [13.9,17.2]	15.0 [13.9,16.8]	14.8 [13.7,16.8]	0.013	-	-	-	-
WBC (103/µL)	10.6 [7.5,15.3]	11.9 [7.8,16.4]	11.6 [7.9,17.3]	0.028	9.2 [6.2,13.4]	9.6 [6.9,13.1]	10.1 [6.4,15.2]	0.507
Neutrophils (103/µL)	10.6 [5.9,11.9]	10.9 [6.7,12.9]	10.9 [7.1,13.8]	0.003	7.6 [4.6,11.5]	7.7 [5.0,11.0]	8.1 [5.1,12.6]	0.518
Lymphocytes (103/µL)	1.2 [0.7,1.4]	1.3 [0.7,1.5]	1.3 [0.6,1.5]	0.352	0.8 [0.5,1.3]	0.9 [0.6,1.3]	0.9 [0.6,1.4]	0.465
NLR	7.9 [5.6,13.7]	7.9 [5.9,13.4]	7.9 [6.3,14.8]	0.255	8.4 [5.1,16.2]	8.4 [5.0,15.3]	8.5 [5.7,16.0]	0.692
Albumin (g/dL)	2.7 [2.4,3.1]	2.7 [2.3,3.1]	2.7 [2.4,3.0]	0.748	3.3 [3.0,3.6]	3.3 [3.0,3.6]	3.3 [3.0,3.6]	0.987
Bicarbonate (mEq/L)	22.0 [20.0,25.0]	22.0 [19.0,25.0]	21.0 [17.0,24.0]	< 0.001	-	-	-	-
Creatinine (mg/dL)	1.0 [0.7,1.5]	1.1 [0.7,1.8]	1.2 [0.8,2.1]	< 0.001	0.9 [0.6,1.5]	0.9 [0.6,1.9]	1.0 [0.6,2.2]	0.341
Sodium (mEq/L)	138.5 [135.0,141.0]	139.0 [136.0,142.0]	138.0 [135.0,141.0]	0.006	139.7 [136.8,143.4]	141.0 [136.8,144.8]	139.4 [136.5,143.4]	0.217
Calcium (mg/dL)	8.1 [7.6,8.6]	8.1 [7.6,8.6]	8.0 [7.5,8.5]	0.002	1.8 [1.8,2.1]	1.8 [1.8,2.1]	1.8 [1.8,2.1]	0.419
Potassium (mEq/L)	4.0 [3.7,4.5]	4.1 [3.8,4.6]	4.2 [3.8,4.7]	0.017	4.0 [3.7,4.3]	4.1 [3.8,4.4]	4.1 [3.7,4.3]	0.477
PT (Seconds)	15.0 [13.1,18.3]	14.3 [12.8,17.0]	14.2 [12.7,16.7]	< 0.001	12.8 [11.8,14.8]	12.7 [11.7,14.1]	12.5 [11.6,13.9]	0.21
PTT (Seconds)	33.3 [28.8,42.7]	31.6 [27.6,39.2]	31.2 [27.6,39.2]	< 0.001	32.7 [28.2,43.8]	32.5 [28.8,37.6]	32.3 [28.4,39.4]	0.947
ALT (U/L)	26.0 [14.0,70.0]	30.0 [15.0,82.0]	39.0 [21.0,130.0]	< 0.001	27.2 [12.5,62.7]	27.3 [14.3,49.9]	24.0 [11.5,49.4]	0.411
ALP (U/L)	81.5 [59.0,109.8]	83.0 [58.0,115.0]	85.0 [58.0,125.0]	0.657	88.0 [62.0,142.0]	86.0 [68.0,121.0]	95.5 [71.0,150.9]	0.106
AST (U/L)	41.0 [23.0,111.5]	42.0 [22.0,108.0]	64.0 [31.0,197.0]	< 0.001	38.2 [25.1,91.8]	44.8 [26.3,95.8]	48.0 [25.2,101.6]	0.439
LD_LDH (U/L)	314.5 [213.0,615.2]	354.0 [238.5,615.2]	469.0 [292.0,615.2]	< 0.001	425.0 [263.0,765.1]	501.7 [283.8,906.3]	615.1 [351.1,906.3]	0.002
LDL (mg/dL)	–	–	–	–	1.5 [1.0,2.0]	1.7 [1.1,2.3]	1.9 [1.2,2.6]	0.015
HDL (mg/dL)	–	–	–	–	0.6 [0.4,0.9]	0.6 [0.4,0.9]	0.6 [0.4,0.9]	0.797
CKMB (ng/mL)	–	–	–	–	18.6 [9.7,28.1]	17.6 [10.5,28.1]	23.1 [10.8,28.1]	0.278
Glucose (mg/dL)	119.0 [101.0,145.8]	127.0 [107.0,157.0]	139.0 [110.0,180.0]	< 0.001	123.1[102.6, 156.6]	141.7[111.1, 186.7]	166[126.1, 213.6]	< 0.001
Triglyceride (mg/dL)	99.0 [76.0,133.0]	172.0 [133.0,225.5]	315.0 [227.0,470.0]	< 0.001	101.8[69.9, 131.0]	153.1[116.8, 208.0]	288.5[195.6, 461.8]	< 0.001
Heart Rate (bpm)	88.1 [76.8,101.2]	91.5 [78.8,103.8]	96.7 [83.0,109.0]	< 0.001	91.7 [80.5,100.5]	91.3 [82.4,100.4]	94.2 [86.8,103.7]	0.016
SBP (mmHg)	113.1 [102.1,124.3]	115.4 [104.5,124.0]	115.5 [105.5,124.5]	0.036	124.4 [114.6,135.0]	127.1 [118.3,138.0]	127.1 [119.0,134.5]	0.11
DBP (mmHg)	64.7 [57.0,69.9]	65.6 [59.4,70.5]	65.6 [59.0,72.8]	0.011	69.3 [64.7,75.7]	72.0 [66.6,78.0]	71.0 [66.8,77.1]	0.051
MBP (mmHg)	80.2 [73.5,87.1]	82.2 [75.0,87.7]	82.2 [75.6,88.8]	0.008	87.1 [81.5,95.7]	90.5 [84.9,97.0]	89.6 [84.8,95.2]	0.025
Respiratory Rate (breaths/min)	19.6 [17.2,22.4]	20.2 [17.6,23.5]	21.6 [18.7,25.1]	< 0.001	20.0 [18.9,20.8]	20.0 [18.9,20.6]	20.0 [19.2,21.0]	0.297
Temperature (°C)	36.9 [36.7,37.2]	37.0 [36.7,37.3]	37.1 [36.8,37.5]	< 0.001	37.0 [36.7,37.5]	37.0 [36.7,37.5]	37.0 [36.7,37.5]	0.856
SpO2 (%)	97.2 [95.5,98.4]	97.1 [95.6,98.6]	96.5 [94.8,98.0]	< 0.001	98.2 [98.0,99.2]	98.4 [98.1,99.2]	98.4 [98.1,99.0]	0.898
90-days Icu death event (%)				0.363				0.416
Survival	505 (89.9)	493 (87.6)	522 (89.8)		161 (97.6)	162 (98.2)	163 (95.9)
Death	57 (10.1)	70 (12.4)	59 (10.2)		4 (2.4)	3 (1.8)	7 (4.1)

This table presents the demographic and clinical characteristics of patients in the MIMIC and SWH cohorts. Continuous variables are reported as median (Q1, Q3), while categorical variables are expressed as counts (n) and percentages (%). For some variables marked with a "—" (dash), data collection was not performed or was unavailable in that cohort. Abbreviations: LOS: Length of Stay; ICU: Intensive Care Unit; BMI: Body Mass Index; SOFA: Sequential Organ Failure Assessment; RDW: Red Cell Distribution Width; WBC: White Blood Cell count; RBC: Red Blood Cell count; NLR: Neutrophil-to-Lymphocyte Ratio; PT: Prothrombin Time; PTT: Partial Thromboplastin Time; ALT: Alanine Aminotransferase; ALP: Alkaline Phosphatase; AST: Aspartate Aminotransferase; LD_LDH: Lactate Dehydrogenase; LDL: Low-Density Lipoprotein; HDL: High-Density Lipoprotein; CKMB: Creatine Kinase-MB; SpO₂: Oxygen Saturation; SBP: Systolic Blood Pressure; DBP: Diastolic Blood Pressure; MBP: Mean Blood Pressure.

Associations of TyG metrics with in-hospital mortality

As shown in Fig. 2, across both datasets, the variability-type TyG metrics showed a consistent direction of association with in-hospital all-cause mortality. In MIMIC (n = 1,707), after full adjustment for age, sex, BMI, and prespecified clinical/laboratory covariates (Model 3 in Methods), TyG-STD (OR 1.66, 95% CI 1.06–2.61, P = 0.027) and TyG-Range (OR 1.24, 95% CI 1.05–1.46, P = 0.011) remained independently associated with mortality; TyG-CV showed a borderline association (OR 1.14, 95% CI 0.98–1.33, P = 0.084). Among level-type metrics, TyG-Average remained significant (OR 1.21, 95% CI 1.02–1.44, P = 0.030), whereas TyG-Median was borderline and TyG-BL was not significant. In the MIMIC cohort, variability-type TyG metrics (TyG-Range and TyG-STD) showed generally consistent associations with in-hospital mortality across most subgroups, with stronger effects observed in non-diabetic and septic patients. In contrast, these associations were attenuated in patients with diabetes and in some older subgroups (Fig. S1).

Fig. 2.

Associations of TyG metrics with in-hospital all-cause mortality in the MIMIC and SWH cohorts. Top panel: MIMIC; bottom panel: SWH. Points show odds ratios (ORs) with 95% CIs for the association between each continuous TyG metric and in-hospital all-cause death; the x-axis is on a log scale. Metrics: TyG_BL (Baseline, first TyG after ICU admission), TyG_Median, TyG_ Average, TyG_Range (max–min), TyG_SD (standard deviation), and TyG_CV (coefficient of variation, SD/mean). Models were fit separately within each cohort: Model 1, unadjusted; Model 2, adjusted for age, sex, BMI; Model 3, further adjusted for clinical/laboratory covariates per Methods-SOFA (MIMIC only), hemoglobin, platelet count, RBC, RDW (MIMIC only), WBC, NLR, albumin, bicarbonate (MIMIC only), creatinine, sodium, calcium, potassium, PT, PTT, ALT, ALP, AST, LDH, and lipids (LDL/HDL in SWH). P values are from two-sided Wald tests. An OR > 1 indicates higher mortality risk with increasing values of the corresponding TyG metric. Abbreviations: TyG, triglyceride–glucose index; ICU, intensive care unit; SD, standard deviation; CV, coefficient of variation.

In SWH (n = 501), variability metrics were risk factors in unadjusted/partially adjusted models, but the associations attenuated and became non-significant after full adjustment (e.g., TyG-STD OR 2.18, 95% CI 0.84–5.66, P = 0.108; TyG-Range OR 1.24, 95% CI 0.84–1.84, P = 0.275; TyG-CV OR 1.31, 95% CI 0.96–1.80, P = 0.092). In contrast, TyG-BL was inversely associated with mortality (OR 0.70, 95% CI 0.51–0.97, P = 0.033). In the SWH cohort, significant associations were observed mainly in older patients (≥ 65 years) and in non-diabetic patients, where TyG-Range and TyG-STD were associated with increased mortality risk, while most other subgroups showed non-significant but directionally consistent effects (Fig. S1).

To account for the time-to-event nature of in-hospital mortality, Cox proportional hazards models were additionally performed with ICU admission as the time origin. In the MIMIC cohort, TyG-BL was consistently associated with ICU mortality and in-hospital mortality across all adjustment models (Model 3: HR 1.19, 95% CI 1.04–1.36 for ICU mortality; HR 1.37, 95% CI 1.20–1.57 for in-hospital mortality), whereas variability-type TyG metrics (TyG-STD, TyG-Range, and TyG-CV) were not significantly associated with in-hospital mortality in the fully adjusted Cox models. No significant associations were observed between TyG metrics and 7-day, 1-year, or 2-year post-discharge mortality in the MIMIC cohort (Table S1).

In the SWH cohort, Cox analyses did not identify significant associations between any TyG metrics and in-hospital mortality after full adjustment (all P > 0.05). Similarly, no significant associations were observed between TyG metrics and 7-day, 1-year, or 2-year post-discharge mortality in the SWH cohort in the Cox models (Table S2).

Kaplan–Meier analysis of in-hospital mortality across TyG tertiles

As shown in Table 2, when TyG metrics were stratified into cohort-specific tertiles, higher TyG values were observed more often in younger patients and were associated with stepwise increases in BMI, glucose, and triglycerides. In MIMIC, higher TyG was also associated with longer ICU length of stay, modestly higher heart and respiratory rates and temperature, lower bicarbonate, and higher ALT/AST/LDH; in SWH, LDH increased with TyG and mean blood pressure was slightly higher. Despite these patterns, crude in-hospital mortality proportions across baseline-TyG tertiles were similar within each cohort (approximately 23% in MIMIC and 12–16% in SWH).

Kaplan–Meier analyses showed that in MIMIC, tertiles of both level-type (Average, Median) and variability-type (SD, Range, CV) TyG metrics exhibited clear separation of survival curves with progressively higher cumulative incidence of in-hospital death (most log-rank P < 0.001), whereas tertiles based solely on baseline TyG did not discriminate risk (P > 0.05) (Fig. 3A–F). In SWH, the overall direction was similar but separation was weaker; only CV tertiles differed significantly (log-rank P = 0.017), whereas Baseline, Average, Median, SD, and Range did not (Fig. 3G–L).

Fig. 3.

Kaplan–Meier cumulative incidence of in-hospital all-cause mortality across tertiles of TyG metrics in the MIMIC and SWH cohorts. (A–F) MIMIC; (G–L): SWH. Metrics by panel—A/G: TyG Baseline (first value after ICU admission); B/H: TyG Average; C/I: TyG Median; D/J: TyG STD; E/K: TyG Range (max − min); F/L: TyG CV. Curves compare Group 1 (lowest tertile), Group 2 (middle tertile), and Group 3 (highest tertile) for each TyG metric: Baseline (first TyG after ICU admission), Mean (Average), Median, SD (standard deviation), Range (max − min), and CV (SD/mean). The outcome is in-hospital all-cause death; the y-axis shows cumulative incidence = 1 − S(t), time is in days from ICU admission, and observations are censored at hospital discharge. P-values are from two-sided log-rank tests without covariate adjustment. Tertile cut-points were defined separately within each dataset. Abbreviations: TyG, triglyceride–glucose index; SD, standard deviation; CV, coefficient of variation; KM, Kaplan–Meier.

Nonlinear associations of TyG level and variability with in-hospital mortality

In MIMIC, multivariable logistic models incorporating restricted cubic splines revealed significant, often nonlinear associations between TyG metrics and in-hospital mortality: TyG-Average (overall P = 0.038; nonlinearity P = 0.009), TyG-Median (0.004; 0.006), TyG-STD (< 0.001; 0.013), TyG-Range (< 0.001; 0.009), and TyG-CV (0.003; 0.021). The risk increased with higher values, with a steeper slope in the upper range, suggesting threshold/saturation-like behavior. TyG-BL showed no association (overall P = 0.542) (Fig. 4A–F).

Fig. 4.

Restricted cubic spline relationships between TyG metrics and in-hospital all-cause mortality in the MIMIC and SWH cohorts. (A–F) MIMIC; (G–L) SWH. Metrics by panel-A/G: TyG Baseline (first value after ICU admission); B/H: TyG Average; C/I: TyG Median; D/J: TyG STD; E/K: TyG Range (max–min); F/L: TyG CV. Curves display adjusted odds ratios (OR) from multivariable logistic regression with restricted cubic splines; ribbons indicate 95% CIs. The reference is the cohort-specific median of each TyG metric. Knots were placed at the 5th, 35th, 65th, and 95th percentiles (Harrell’s recommendation). P for overall tests the overall association; P for nonlinearity tests deviation from linearity (two-sided). Background histograms show the empirical distribution of each TyG metric (right y-axis = density). Models were fit separately in each cohort and adjusted as in Model 3 (age, sex, BMI, and prespecified clinical/laboratory covariates; SOFA, RDW, and bicarbonate available in MIMIC only). Abbreviations: TyG, triglyceride–glucose index; SD, standard deviation; CV, coefficient of variation; ICU, intensive care unit.

In SWH, overall and nonlinearity tests for all TyG metrics were not significant (P > 0.05), and the spline curves were relatively flat with wide confidence bands, consistent with smaller sample size and differences in covariate availability (Fig. 4G–L).

Discussion

In this study, we used data from the MIMIC and SWH databases to examine the relationship between TyG index variability and all-cause in-hospital mortality. We found that variability-based TyG metrics (TyG-Range, TyG-STD, and TyG-CV) were independently associated with in-hospital mortality, whereas a single baseline TyG value showed limited risk stratification ability and even an inverse association after extensive adjustment in the SWH cohort. RCS revealed a non-linear exposure–risk relationship in MIMIC, with a steeper increase in risk at higher values. In contrast, the SWH curve was flatter, which can be attributed to the smaller sample size and differences in covariate availability. Collectively, these findings suggest that dynamic, time-series information provides a more informative reflection of short-term prognosis in ICU patients than a single baseline measurement.

Previous studies have shown that IR in ICU populations is associated with stress, inflammation, and immune dysregulation, all linked to poor outcomes^27,28. However, standard IR assessments, such as the clamp test or HOMA-IR, are not routinely feasible in ICU settings²⁹. Our work, which utilized routine clinical tests (glucose, triglycerides), developed the TyG index and its time-series parameters, supporting the evidence that variability, rather than merely the level, is more closely associated with mortality risk. This finding presents a new perspective on dynamic metabolic risk characterization in the ICU setting. As an alternative marker for IR, TyG has been validated in multiple studies and populations, with predictive performance equal to or superior to HOMA-IR, further supporting its feasibility for ICU applications^30–33. In our study, TyG variability, not just its absolute level, proved to be more closely associated with mortality risk, supporting its feasibility for ICU applications as a dynamic prognostic tool. This suggests that dynamic fluctuations in TyG levels capture a broader spectrum of metabolic instability, which may be more informative in predicting patient outcomes than static, baseline measurements alone.

TyG reflects integrated glucose–lipid utilization and overall energy status. Its variability may, at least in part, capture metabolic instability driven by fluctuating stress hormone levels, inflammation-related alterations in hepatic lipid metabolism, inconsistent nutritional input, and changes in insulin or lipid-lowering therapies^33,34. Similar to blood glucose variability, which has been shown to correlate with mortality risk in ICU patients independent of average glucose levels^19,35, TyG variability may represent a broader phenotype of metabolic dysregulation, encompassing oxidative stress, endothelial dysfunction, and immune imbalance^36–38. In this context, variability-based metrics appear to provide complementary prognostic information beyond static measurements.

In MIMIC, we observed a steeper slope in the high-value segment, consistent with a "threshold/saturation" effect³⁹. In contrast, baseline TyG in SWH was negatively correlated, which could be influenced by initial illness, nutritional and insulin/lipid-lowering treatments, malnutrition, or residual confounding. This may also relate to differences in sample size and sampling timepoints^12,19,40. Metabolic phenotype differences may alter the clinical meaning of "absolute TyG values." MIMIC patients were generally heavier, with a higher proportion of overweight/obese individuals (BMI distribution shifted right), and higher TyG values likely reflected increased insulin resistance and lipid burden^41,42. In contrast, SWH patients were generally smaller in stature, and in this context, “low/high” absolute TyG values may reflect different metabolic reserves and nutritional states: very low glucose/triglycerides may indicate malnutrition or low metabolic reserves, presenting an “apparent protection” effect⁴³. Furthermore, differences in ICU admission phases and treatment pathways were noted between the two cohorts. As shown in Table 1, vital signs on the first day of admission were more unstable in the MIMIC cohort, while SWH patients exhibited more stability, which can be attributed to their prior high-intensity resuscitation followed by ICU transfer. This process likely contributed to their relatively stable clinical status upon ICU admission. In this context, baseline TyG levels may be influenced by the timing of sampling: higher TyG values in the MIMIC cohort may reflect elevated stress and metabolic pressure associated with more unstable conditions at admission, whereas higher TyG values in the SWH cohort may, in part, reflect recovery post-resuscitation and the subsequent stabilization of metabolic and nutritional status following ICU transfer.

In contrast, variability indices capture information throughout the entire hospitalization, being less prone to bias from single-time measurements, thus exhibiting more consistent cross-database performance. Kaplan–Meier analysis by tertiles showed that most TyG variability indices presented the lowest risk in the middle group, with higher risks at both ends, suggesting the potential existence of a U-shaped or threshold/saturation effect. RCS further showed that the risk increased sharply in the higher value segment in MIMIC, consistent with the biological intuition that "greater metabolic fluctuation, higher risk." Combining both types of analysis, we speculate that relative stability within a certain range (neither excessive fluctuation nor extremely stable "low energy/low fat" states) might correspond to better short-term prognosis. This finding provides a basis for incorporating metabolic stability, beyond focusing solely on absolute levels, into risk assessments¹².

Although this study does not provide causal inference and has not defined therapeutic targets, the results support the need for prospective studies to evaluate the potential benefits of reducing metabolic fluctuations (e.g., more consistent nutritional support, integrated glucose-lipid monitoring, rational use of corticosteroids/vasopressors) and exploring non-linear risk inflection points and actionable thresholds. Evidence of cumulative/long-term TyG exposure increasing cardiovascular event risk further supports the biological rationale for "stable metabolic exposure" in general and specific populations^44,45.

This study was conducted using data from two databases with different healthcare systems and care models, enhancing the reproducibility and robustness of the results. Variables were standardized prior to the study, and key lab parameters were normalized to reduce inter-database heterogeneity. The study analyzed TyG levels and time-series fluctuations using RCS to explore non-linear exposure-risk relationships, avoiding model misfit due to simple linear assumptions. Missing data were handled with MICE multiple imputation under the MAR assumption, with separate model construction for each database, in accordance with STROBE guidelines.

This retrospective study cannot fully eliminate residual confounding (e.g., nutritional pathways, insulin/lipid-lowering treatments). TyG indices were derived from non-fasting routine measurements, which may introduce measurement variability, and the choice of pairing window may influence the estimation of fluctuation metrics. In addition, the calculation of TyG variability depends on the frequency and timing of laboratory testing, and patients with more frequent measurements may have more stable or more extreme variability estimates, introducing potential measurement density bias. Covariate availability differed between databases, which may have resulted in differential adjustment and limited comparability between the MIMIC and SWH cohorts. The relatively small sample size and low event rate in the SWH cohort, particularly in subgroup analyses, reduced statistical precision and may have led to unstable estimates. Although logistic regression was used to assess in-hospital mortality, time-to-event or competing risk models may be more appropriate in the presence of variable follow-up durations and discharge as a competing event. While Cox proportional hazards models were additionally performed as sensitivity analyses, the partial inconsistency between logistic and Cox results indicates that the associations may differ between overall mortality risk and the timing of death, which should be interpreted with caution. Multiple TyG indices were assessed, raising the risk of Type I error; further validation is needed. Finally, the external validity is limited by differences in ICU patient composition and protocols, requiring confirmation in broader populations and prospective studies⁴⁶.

Conclusion

This study demonstrates that TyG variability (TyG_Range, TyG_STD, TyG_CV) is associated with in-hospital mortality in critically ill patients, with robust and independent associations observed in the MIMIC cohort. In the SWH cohort, variability metrics were associated with mortality in unadjusted and partially adjusted models, but these associations attenuated after full adjustment.

Overall, our findings suggest that dynamic time-series information of TyG may provide complementary prognostic information beyond static baseline measurements, particularly in larger cohorts. These results highlight the potential importance of metabolic instability in ICU risk stratification and warrant further prospective and multicenter studies to validate the prognostic utility of TyG variability.

Abbreviations

TyG

Triglyceride-glucose

IR

Insulin resistance

ICU

Intensive Care Unit

MIMIC-IV

Medical Information Mart for Intensive Care, version IV

SWH

Southwest Hospital

SOFA

Sequential organ failure assessment

NLR

Neutrophil-to-lymphocyte ratio

ALT

Alanine aminotransferase

AST

Aspartate aminotransferase

ALP

Alkaline phosphatase

LDH

Lactate dehydrogenase

PT

Prothrombin time

PTT

Activated partial thromboplastin time

HOMA-IR

Homeostasis model assessment of insulin resistance

RCS

Restricted cubic splines

CV

Coefficient of variation

BL

Baseline

BMI

Body mass index

OR

Odds ratio

CI

Confidence interval

SD

Standard deviation

Author contributions

Zuzhi Chen conceptualized the research, wrote the methodology, conducted the survey and formal analysis, visualized the data and wrote the original draft, and reviewed and edited the manuscript. Xiang Xiang visualized the data and wrote the original draft, and reviewed and edited the manuscript. Changlin Yin conceived the concept for the study, assisted with the research methodology, supervised Zuzhi Chen, Xiang Xiang, and Haoran Xu, and reviewed and edited the paper. Haoran Xu is the second reviewer, responsible for the selection and full text evaluation. Ting Zhao and Weiguang Zhang contributed to the creation and review of the images and tables. Xiaofei Xie and Zhi Dou reviewed and edited the manuscript. Yonghui Zhang reviewed and edited the manuscript. Hailin Shu reviewed and edited the manuscript. Changlin Yin and Haoran Xu are the guarantors of this work, and as such, they have full access to all the data in the study and are responsible for the integrity of the data and the accuracy of the data analysis.

Funding

This study was supported by Science and Technology Research Project of the Education Commission of Chongqing City (KJQN202512837) and Medical Research Project of Chengdu Health Commission of Sichuan Province (2021061).

Data availability

All data generated or analyzed during this study are included in this published article. The data not published within this article are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study was approved by Ethics Committee of the First Affiliated Hospital of the PLA Army Military Medical University (Approval number: B KY2024116). Due to the retrospective design of the study and the use of fully de-identified data, the requirement for informed consent was waived by the Ethics Committee. All methods were performed in accordance with the relevant guidelines and regulations. The study was conducted in compliance with the ethical standards of the Declaration of Helsinki and its later amendments.