Prognostic value of the triglyceride-glucose index combined with atherogenic index of plasma for all-cause mortality in critically ill patients with chronic heart failure: a machine learning-driven retrospective cohort study.

Save

Background

The triglyceride-glucose index (TyG) and atherogenic index of plasma (AIP) are emerging metabolic biomarkers associated with cardiovascular diseases. However, their combination prognostic value in patients with critical chronic heart failure (CHF) remains unclear. This study aimed to evaluate the combined predictive effect of these two biomarkers and clarify their interactive patterns in this association.

Methods

1,238 patients were recruited via the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, with a median age of 71 years. Multivariable Cox regression, Kaplan-Meier analysis, and receiver operating characteristic (ROC) curve were employed to explore associations between TyG, AIP, and mortality. Mediation analysis was applied to assess their bidirectional mediation effects.

Additionally, we developed a machine learning-driven prediction model, which was further utilized to evaluate the two indicators’ incremental predictive value.

Results

1,238 patients were included, with 478 (38.61%) dying during follow-up. Fully adjusted Cox regression models revealed that the high TyG and high AIP group was associated with the highest risk of mortality relative to the low TyG and low AIP group (14-day HR 2.18, 95% Cl: 1.44–3.31; 365-day HR 1.92, 95% Cl: 1.38–2.68). ROC analyses demonstrated that the combined TyG-AIP outperformed either marker alone in predicting mortality at all follow-up observation time points (all P < 0.05).

Mediation analysis revealed that TyG mediated the effect of AIP on mortality across all time frames, with a more pronounced effect at 365 days (65.47%) than at 14 days (37.32%). In contrast, AIP served as a mediator in the association between TyG and short-term mortality only (14-day: 30.39%; 30-day: 25.76%). The random forest model confirmed that the incorporation of both the TyG index and AIP remarkably improved predictive capability, corroborating their combined incremental value.

Conclusion

Combined elevation of TyG and AIP was independently related to an elevated risk of mortality in patients with critical CHF. Combined assessment of these biomarkers may facilitate the early recognition of high-risk subjects and support stage-specific metabolic interventions.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12944-026-02861-1.

Introduction

Chronic heart failure (CHF) represents the terminal phase of multiple cardiac disorders, marked by substantial mortality and adverse outcomes [1]. Worldwide, more than 64 million individuals are affected by heart failure (HF), making it a major challenge in contemporary clinical practice and public health [2]. The disease burden is particularly pronounced among patients admitted to the intensive care unit (ICU), where CHF cases are frequently admitted due to acute decompensation, multisystem organ failure, or severe comorbidities, resulting in a markedly increased risk of mortality [3].

Therefore, identifying reliable and effective prognostic biomarkers is of great importance for these high-risk individuals.

Individuals with HF often present with various metabolic abnormalities, among which insulin resistance (IR) and dyslipidemia exert critical roles in the development of the disease. IR accelerates atherosclerosis by inducing endothelial dysfunction [4], promoting lipid deposition, and amplifying inflammatory responses [5, 6], thereby increasing the risk of cardiovascular events. Conversely, dysregulated lipid metabolism impairs insulin signaling and further aggravates IR [7, 8].

This bidirectional interaction creates a vicious cycle of IR and lipid metabolic disorder that substantially contributes to HF progression. Accordingly, identifying and integrating robust metabolic biomarkers that reliably capture both IR and lipid metabolic abnormalities are of great clinical significance for optimized early risk stratification.

As a reliable surrogate of IR, the triglyceride-glucose index (TyG) has emerged as a major factor influencing the prognosis of patients with HF [9, 10]. The atherogenic index of plasma (AIP), which sensitively captures lipid metabolic imbalance and reflects early atherosclerotic changes, is associated with cardiovascular risk [11], coronary artery disease severity [12], and, more recently, with short-term adverse outcomes in HF [13]. Pathophysiologically, this intertwined metabolic disturbance—reflected by an elevated TyG index indicating IR and quantified by the AIP representing an atherogenic lipid profile—jointly exacerbates chronic inflammation and oxidative stress (such as elevated levels of TNF-α and IL-6), impairs vascular endothelial function, and thereby promotes myocardial fibrosis, hypertrophy, and remodeling, ultimately leading to cardiac functional deterioration [14, 15].

Their combination may therefore enable a more integrated metabolic risk evaluation. Nevertheless, the combined predictive value of the TyG index and AIP for short- and long-term outcomes among critically ill CHF patients has not been systematically evaluated, nor have their potential interactive effects on prognosis been explored.

This research investigated the relationships between the TyG index, AIP, and their combination with mortality at multiple timepoints. Furthermore, a bidirectional mediation approach was applied to elucidate the potential “metabolic crosstalk-prognosis” mechanistic pathway between these two biomarkers in relation to all-cause mortality. Finally, prognostic models were developed using interpretable machine learning methods, and the incremental predictive benefit of the combined indicators was assessed.

These findings aim to provide evidence-based support for clinical risk assessment and precision intervention in these susceptible individuals from a metabolic perspective.

Methods

Study design and population

This retrospective cohort study was conducted using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, with the research team obtaining database access authorization (Record ID: 13158313). Eligible participants were adults aged ≥ 18 years with a diagnosis of CHF. For those with multiple ICU stays, only data from their first ICU admission were analysed.

Exclusion criteria included (1) ICU stay < 24 h; (2) missing fasting blood glucose (FBG), triglycerides (TG), or high-density lipoprotein cholesterol (HDL-C). Ultimately, 1,238 participants were enrolled in the cohort, and the workflow is depicted in Fig. 1.

Fig. 1The study flowchart

Data collection

Data extraction was performed via PostgreSQL (version 13.7.1) and SQL scripts. Clinical variables collected: (1) General information: age, sex, body mass index (BMI), blood pressure, and heart rate; (2) Comorbidities: hypertension, dyslipidemia, diabetes, atrial fibrillation, chronic kidney disease (CKD), etc.; (3) Laboratory parameters: complete blood count, FBG, hemoglobin A1c (HbA1c), TG, total cholesterol (TC), HDL-C, blood urea nitrogen (BUN), N-terminal pro-B-type natriuretic peptide (NT-proBNP), etc.; (4) Treatments: medications and therapeutic interventions: statins, angiotensin-converting enzyme inhibitor/angiotensin receptor blocker/angiotensin receptor-neprilysin inhibitor (ACEI/ARB/ARNI), percutaneous coronary intervention (PCI), etc. The TyG index was derived as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{ln}\left[\mathrm{TG}(\mathrm{mg}/\mathrm{dL})\times\mathrm{FBG}(\mathrm{mg}/\mathrm{dL})/2\right]$$\end{document} [16].

AIP was derived as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{log}\left[\mathrm{TG}\mathrm{(}\mathrm{mmol}\mathrm{/}\mathrm{L}\mathrm{)\:/\:}\mathrm{HDL}\mathrm{-}\mathrm{C}\text{}\mathrm{(}\mathrm{mmol}\mathrm{/}\mathrm{L}\mathrm{)}\right]$$\end{document} [17]. Estimated glomerular filtration rate (eGFR) was computed using the CKD-EPI equation [18].

Outcomes

The primary outcomes were all-cause mortality at 14-day and 365-day, with secondary outcomes including all-cause mortality at 30-day and 180-day.

Statistical analysis

Continuous variables were expressed as median (interquartile range, IQR), with group comparisons performed using Wilcoxon rank-sum or Kruskal-Wallis test. Categorical variables were summarized as counts (percentages) and analyzed using the chi-square test or Fisher’s exact test. Variables with missing values exceeding 20% were excluded, with residual missing data imputed via using the MissForest package in R [19].

Both the TyG index and AIP were categorized into four quartiles (T1-T4) as follows. For TyG, T1 (< 8.65), T2 (8.65–9.04), T3 (9.04–9.62) and T4 (≥ 9.62); for AIP, T1 (< 0.05), T2 (0.05–0.24), T3 (0.24–0.45) and T4 (≥ 0.45). The top quartile group (T4) of each indicator was defined as the “high” level, and the remaining quartiles were defined as the “low” level.

Participants were stratified into four subgroups according to the TyG index and AIP levels: low TyG and low AIP, low TyG and high AIP, high TyG and low AIP, and high TyG and high AIP.

Prior to constructing the Cox proportional hazards regression models, we validated the underlying model assumptions. (1) Variance inflation factors (VIF) for all covariates were derived from multiple linear regression, with VIF > 5 indicating the presence of multicollinearity (Table S1). (2) The proportional hazards assumption for all variables included in the Cox models was assessed via Schoenfeld residual-based tests.

Three Cox models were built to examine the relationships of variables with clinical outcomes. Model 1 was unadjusted; Model 2 was adjusted for age, sex, blood pressure, heart rate, diabetes, dyslipidemia, stroke, and CKD; and Model 3 was further adjusted for laboratory indicators and treatment conditions based on Model 2. Kaplan-Meier curves were utilized to compare incident rates.

Restricted cubic spline (RCS) analyses were applied to assess the dose-response relationships between the variables and the outcomes. Receiver operating characteristic (ROC) technology was employed to evaluate the predictive power of the indices and their combination. Bidirectional mediation analysis [20] was performed to explore the interactive effects of the TyG index and AIP on mortality.

All participants were randomly allocated to training (70%) and testing (30%) sets for model development and validation. Key prognostic variables were identified using the least absolute shrinkage and selection operator (LASSO)-Cox regression. Six machine learning approaches were employed to construct predictive models: decision tree (DT), random forest (RF), support vector machine (SVM), light gradient logistic regression (LR), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM).

Hyperparameter tuning was performed through grid search and 10-fold cross-validation. Model performance was assessed using the area under the ROC curve (AUC), accuracy, sensitivity, and specificity, while calibration plots and decision curve analysis (DCA) evaluated its calibration and clinical utility. Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) quantified the incremental predictive value of combining the TyG index and AIP.

Time-dependent C-index and ROC analyses further evaluated the long-term predictive stability of the models. Model interpretability was examined using the Shapley additive explanations (SHAP) framework in Python.

Sensitivity analyses included subgroup analyses, propensity score matching (PSM), and additional analyses based on the median for alternative cut-offs.

All statistical analyses were performed using R software (version 4.3.1) and Python (version 3.9.16), with P < 0.05 considered statistically significant.

Results

Baseline characteristics

A total of 1,238 patients were ultimately enrolled, with a median age of 71 years (IQR: 60–80), and 60.18% were male. Patients were categorized into the survival group (n = 760, 61.39%) and the death group (n = 478, 38.61%) based on 365-day outcomes (Table 1). Relative to survivors, deceased patients were older, had higher heart rates, and more frequently presented with stroke and CKD (all P < 0.05).

Notably, the death group had significantly higher TyG index and AIP levels (both P < 0.05). Other laboratory parameters also differed significantly between groups (e.g., HbA1c, BUN, NT-proBNP). Additionally, fewer patients in the death group received medications and revascularization therapy, including (e.g., statins, ACEI/ARB/ARNI, PCI); detailed data are presented in Table 1.

Among those who died within 14 days, TyG index and AIP levels were also markedly higher than in survivors (Table S2). Additionally, the baseline characteristics stratified by TyG index and AIP levels are shown in Tables S3-4.

Table 1Baseline characteristics by 365-day survival statusVariablesTotal (n = 1238)Survivor (n = 760)Death (n = 478)P-valueAge, (years)71.00 (60.00, 80.00)70.00 (59.75, 78.00)73.00 (62.00, 81.00)< 0.011Gender, N(%)0.844 Male745 (60.18)459 (60.39)286 (59.83) Female493 (39.82)301 (39.61)192 (40.17)SBP, (mmHg)119.00 (111.25, 126.00)120.00 (110.00, 127.00)119.00 (113.00, 126.00)0.851DBP, (mmHg)65.00 (58.25, 71.00)64.00 (57.00, 70.00)65.00 (60.00, 69.00)0.330Heart rate, (bpm)87.00 (78.00, 100.00)85.00 (77.00, 97.00)91.00 (80.00, 104.00)< 0.001BMI, (kg/m2)29.06 (25.27, 33.41)28.83 (25.26, 33.12)29.48 (25.30, 34.20)0.170CHD, N(%)0.120 No907 (73.26)545 (71.71)362 (75.73) Yes331 (26.74)215 (28.29)116 (24.27)Hypertension, N(%)0.152 No347 (28.03)202 (26.58)145 (30.33) Yes891 (71.97)558 (73.42)333 (69.67)Diabetes, N(%)0.152 No769 (62.12)484 (63.68)285 (59.62) Yes469 (37.88)276 (36.32)193 (40.38)Dyslipidemia, N(%)< 0.001 No633 (51.13)360 (47.37)273 (57.11) Yes605 (48.87)400 (52.63)205 (42.89)VHD, N(%)0.576 No1181 (95.40)723 (95.13)458 (95.82) Yes57 (4.60)37 (4.87)20 (4.18)Atrial fibrillation, N(%)0.893 No565 (45.64)348 (45.79)217 (45.40) Yes673 (54.36)412 (54.21)261 (54.60)Stroke, N(%)< 0.001 No1018 (82.23)648 (85.26)370 (77.41) Yes220 (17.77)112 (14.74)108 (22.59)COPD, N(%)0.227 No823 (66.48)515 (67.76)308 (64.44) Yes415 (33.52)245 (32.24)170 (35.56)CKD, N(%)< 0.001 No853 (68.90)555 (73.03)298 (62.34) Yes385 (31.10)205 (26.97)180 (37.66)OMI, N(%)0.135 No1073 (86.67)650 (85.53)423 (88.49) Yes165 (13.33)110 (14.47)55 (11.51)Hemoglobin, (g/dL)11.00 (9.30, 12.80)11.30 (9.67, 13.20)10.50 (8.90, 12.28)< 0.001RDW, (%)14.80 (13.70, 16.30)14.50 (13.50, 15.80)15.50 (14.20, 17.20)< 0.001PLT, (K/UL)197.00 (146.00, 261.00)198.50 (151.00, 252.00)195.00 (140.00, 274.00)0.738Neu, (K/UL)9.36 (6.78, 12.41)9.07 (6.73, 11.68)10.21 (6.89, 13.34)< 0.001Lym, (K/UL)1.30 (0.82, 1.78)1.48 (0.99, 1.94)1.04 (0.62, 1.44)< 0.001Mono, (K/UL)0.63 (0.43, 0.89)0.60 (0.42, 0.84)0.70 (0.45, 0.97)< 0.001ALT, (U/L)28.00 (16.00, 52.00)27.00 (17.00, 51.29)28.00 (15.00, 52.00)0.477AST, (U/L)36.00 (22.00, 66.96)35.00 (22.00, 63.00)37.00 (23.00, 74.75)0.090TBiL, (mg/dL)0.70 (0.40, 1.20)0.70 (0.50, 1.10)0.70 (0.40, 1.40)0.513HbA1c, (%)6.07 (5.74, 6.84)6.00 (5.70, 6.76)6.20 (5.79, 6.95)0.001LDL-C, (mg/dL)66.00 (57.67, 73.81)67.89 (60.28, 76.00)65.14 (55.00, 69.51)0.079HDL-C, (mg/dL)34.17 (30.00, 40.00)35.55 (30.80, 41.07)32.55 (29.00, 37.22)< 0.001TC, (mg/dL)126.47 (111.79, 139.75)129.26 (117.94, 142.11)119.53 (107.56, 132.49)< 0.001TG, (mg/dL)134.82 (99.00, 196.64)134.22 (101.27, 187.53)135.73 (96.25, 211.75)0.557FBG, (mg/dL)126.00 (103.00, 164.00)119.50 (101.00, 152.00)139.50 (109.25, 182.75)< 0.001eGFR, (mL/min/1.73m2)57.40 (36.10, 80.44)61.53 (42.16, 82.72)49.75 (29.02, 73.25)< 0.001BUN, (mg/dL)23.00 (16.00, 38.00)21.00 (15.00, 32.25)28.00 (18.00, 45.75)< 0.001Potassium, (mEq/L)4.20 (3.80, 4.70)4.20 (3.80, 4.60)4.30 (3.80, 4.70)0.139Sodium, (mEq/L)139.00 (136.00, 141.00)139.00 (136.00, 141.00)138.00 (134.00, 142.00)0.213Chloride, (mEq/L)103.00 (99.00, 106.00)103.00 (99.00, 107.00)102.00 (97.00, 106.00)< 0.001Lg of NT-proBNP, (pg/ml)3.87 (3.68, 4.07)3.82 (3.64, 4.01)3.99 (3.79, 4.15)< 0.001TyG index9.04 (8.65, 9.62)9.01 (8.64, 9.48)9.10 (8.69, 9.79)< 0.001AIP0.24 (0.05, 0.44)0.23 (0.05, 0.42)0.28 (0.08, 0.48)0.004Antiplatelet drugs, N(%)0.221 No548 (44.26)326 (42.89)222 (46.44) Yes690 (55.74)434 (57.11)256 (53.56)Statins, N(%)< 0.001 No598 (48.30)332 (43.68)266 (55.65) Yes640 (51.70)428 (56.32)212 (44.35)ACEI/ARB/ARNI, N(%)0.021 No878 (70.92)521 (68.55)357 (74.69) Yes360 (29.08)239 (31.45)121 (25.31)β-blocker, N(%)0.373 No545 (44.02)327 (43.03)218 (45.61) Yes693 (55.98)433 (56.97)260 (54.39)Diuretics, N(%)0.296 No319 (25.77)188 (24.74)131 (27.41) Yes919 (74.23)572 (75.26)347 (72.59)MRA, N(%)0.462 No1152 (93.05)704 (92.63)448 (93.72) Yes86 (6.95)56 (7.37)30 (6.28)Cardiotonic drugs, N(%)0.609 No919 (74.23)568 (74.74)351 (73.43) Yes319 (25.77)192 (25.26)127 (26.57)PCI, N(%)0.047 No1158 (93.54)703 (92.50)455 (95.19) Yes80 (6.46)57 (7.50)23 (4.81)ACEI/ARB/ARNI Angiotensin-converting enzyme inhibitor/angiotensin receptor blocker/angiotensin receptor-neprilysin inhibitor, AIP Atherogenic index of plasma, ALT Alanine aminotransferase, AST Aspartate aminotransferase, BMI Body mass index, BUN Blood urea nitrogen, CHD Coronary heart disease, CKD Chronic kidney disease, COPD Chronic obstructive pulmonary disease, DBP Diastolic blood pressure, eGFR Estimated glomerular filtration rate, FBG Fasting blood glucose, HbA1c Hemoglobin A1c, HDL-C High-density lipoprotein cholesterol, LDL-C Low-density lipoprotein cholesterol, Lym Lymphocyte count, Mono Monocyte count, MRA Mineralocorticoid receptor antagonist, Neu Neutrophil count, NT-proBNP N-terminal pro-B-type natriuretic peptide, OMI Old myocardial infarction, PCI Percutaneous coronary intervention, PLT Platelet, RDW Red blood cell distribution width, SBP Systolic blood pressure, TBiL serum total bilirubin, TC Total cholesterol, TG Triglyceride, TyG Triglyceride-glucose, VHD Valvular heart disease

Associations of the TyG index and AIP with mortality risk

Table 2 shows that the TyG index (continuous variable) was independently associated with increased 14-day (HR 1.64, 95%CI: 1.32–2.03) and 365-day (HR 1.44, 95%CI: 1.25–1.65) mortality. When stratified by quartiles, patients in the highest TyG index quartile (T4) had a significantly higher risk of 14-day (HR 2.03, 95%CI: 1.33–3.42) and 365-day (HR 1.87, 95%CI: 1.40–2.49) mortality when contrasted with those in the lowest quartile (T1). Comparable results were observed for 30- and 180-day mortality (Table S5).

Table 2Associations of the TyG index, AIP, and mortality risk in patients with CHFModel 1Model 2Model 3HR (95% CI)P-valueHR (95% CI)P-valueHR (95% CI)P-valueTyG index14-day mortality Continuous1.58 (1.34–1.86)< 0.0011.68 (1.39–2.02)< 0.0011.64 (1.32–2.03)< 0.001 T11 (reference)1 (reference)1 (reference) T20.98 (0.61–1.55)0.9160.99 (0.62–1.58)0.9670.92 (0.57–1.49)0.739 T31.09 (0.69–1.71)0.7231.13 (0.71–1.80)0.6081.13 (0.70–1.82)0.621 T42.04 (1.37–3.06)< 0.0012.18 (1.40–3.40)< 0.0012.03 (1.33–3.42)0.002 P for trend< 0.001< 0.001< 0.001365-day mortality Continuous1.29 (1.16–1.45)< 0.0011.34 (1.18–1.52)< 0.0011.44 (1.25–1.65)< 0.001 T11 (reference)1 (reference)1 (reference) T20.95 (0.73–1.24)0.7080.96 (0.73–1.25)0.7360.95 (0.73–1.25)0.736 T30.92 (0.70–1.20)0.5220.95 (0.72–1.26)0.7301.05 (0.79–1.39)0.719 T41.60 (1.25–2.04)< 0.0011.68 (1.28–2.20)< 0.0011.87 (1.40–2.49)< 0.001 P for trend< 0.001< 0.001< 0.001AIP14-day mortality Continuous2.26 (1.48–3.46)< 0.0012.74 (1.70–4.42)< 0.0012.01 (1.23–3.28)0.005 T11 (reference)1 (reference)1 (reference) T20.77 (0.48–1.23)0.2790.78 (0.49–1.25)0.3040.81 (0.50–1.30)0.383 T31.04 (0.67–1.60)0.8681.16 (0.74–1.81)0.5231.20 (0.76–1.91)0.434 T41.73 (1.17–2.57)0.0061.95 (1.28–2.98)0.0021.57 (1.01–2.43)0.043 P for trend< 0.001< 0.0010.004365-day mortality Continuous1.65 (1.25–2.18)< 0.0011.87 (1.37–2.54)< 0.0011.72 (1.25–2.37)< 0.001 T11 (reference)1 (reference)1 (reference) T21.04 (0.80–1.36)0.7521.04 (0.80–1.36)0.7561.08 (0.83–1.42)0.554 T31.08 (0.83–1.41)0.5551.17 (0.89–1.53)0.2641.22 (0.93–1.60)0.154 T41.46 (1.14–1.88)0.0031.56 (1.19–2.05)< 0.0011.43 (1.08–1.89)0.012 P for trend0.003< 0.0010.009Model 1: UnadjustedModel 2: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, and CKDModel 3: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, CKD, hemoglobin, RDW, Neu, Lym, Mono, ALT, AST, TBiL, HbA1c, TC, eGFR, BUN, Potassium, Sodium, Chloride, Lg of NT-proBNP, Statins, ACEI/ARB/ARNI, β-blocker, Diuretics and PCI

Similarly, when considered as a continuous variable, AIP exhibited a significant relationship with 14-day (HR 2.01, 95%CI: 1.23–3.28) and 365-day (HR 1.72, 95%CI: 1.25–2.37) mortality. Compared with patients in the lowest quartile (T1), those in the highest AIP quartile (T4) had significantly increased risks of 14-day (HR 1.57, 95%CI: 1.01–2.43) and 365-day (HR 1.43, 95%CI: 1.08–1.89) mortality, with consistent associations observed for the 30-day and 180-day endpoints (Table S5).

Kaplan-Meier curves indicated the highest mortality in patients within the top quartiles of both indices (both log-rank P < 0.05) (Fig. 2 and Figure S1). RCS analyses revealed a nonlinear TyG–mortality association at 30-, 180-, and 365-day, but not at 14-day, while AIP exhibited a linear positive relationship throughout (P for non-linearity > 0.05) (Fig.

3 and Figure S2).

Fig. 2Kaplan‒Meier survival analysis curves for mortality in patients with CHF. A-B The TyG index for 14-day and 365-day mortality risk; C-D The AIP for 14-day and 365-day mortality risk

Fig. 3RCS analysis of TyG index and AIP for mortality in patients with CHF. A-B The TyG index for 14-day and 365-day mortality risk; C-D The AIP for 14-day and 365-day mortality risk

The TyG index-AIP combination and its association with mortality risk

Model 3 demonstrated that, compared with patients in the low TyG and low AIP group, those with elevated TyG—particularly when accompanied by high AIP—had markedly higher 14-day and 365-day mortality risks (14-day: HR 2.18, 95%CI: 1.44–3.31; 365-day: HR 1.92, 95%CI: 1.38–2.68; P < 0.05) (Table 3).

Model 1: Unadjusted Model 2: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, and CKD Model 3: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, CKD, hemoglobin, RDW, Neu, Lym, Mono, ALT, AST, TBiL, HbA1c, TC, eGFR, BUN, Potassium, Sodium, Chloride, Lg of NT-proBNP, Statins, ACEI/ARB/ARNI, β-blocker, Diuretics and PCITable 3Association between the TyG index-AIP combination and mortality riskModel 1Model 2Model 3HR (95% CI)P-valueHR (95% CI)P-valueHR (95% CI)P-value14-day mortality low TyG and low AIP1 (reference)1 (reference)1 (reference) low TyG and high AIP1.52 (0.89–2.58)0.1221.71 (1.00-2.93)0.0521.21 (0.69–2.11)0.501 high TyG and low AIP1.84 (1.12–3.01)0.0161.87 (1.11–3.13)0.0182.08 (1.20–3.60)0.009 high TyG and high AIP2.24 (1.59–3.17)< 0.0012.45 (1.67–3.60)< 0.0012.18 (1.44–3.31)< 0.001365-day mortality low TyG and low AIP1 (reference)1 (reference)1 (reference) low TyG and high AIP1.10 (0.77–1.56)0.6121.17 (0.82–1.67)0.4000.94 (0.65–1.36)0.747 high TyG and low AIP1.67 (1.24–2.25)< 0.0011.70 (1.24–2.33)0.0011.80 (1.38–2.35)< 0.001 high TyG and high AIP1.69 (1.36–2.12)< 0.0011.81 (1.42–2.32)< 0.0011.92 (1.38–2.68)< 0.001 Model 1: Unadjusted Model 2: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, and CKD Model 3: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, CKD,hemoglobin, RDW, Neu, Lym, Mono, ALT, AST, TBiL, HbA1c, TC, eGFR, BUN, Potassium,Sodium, Chloride, Lg of NT-proBNP, Statins, ACEI/ARB/ARNI, β-blocker, Diuretics and PCI

Furthermore, similar patterns were found for 30- and 180-day mortality (Table S6). Kaplan-Meier curves likewise indicated that concurrent high TyG and high AIP were linked to an increased risk of mortality (Fig. 4 and Figure S3).

Fig. 4Kaplan-Meier survival curves for mortality by TyG index and AIP levels. A 14-day mortality; B 365-day mortality.

T1: low TyG and low AIP; T2: low TyG and high AIP; T3: high TyG and low AIP; T4: high TyG and high AIP

The comparison of predictive performance for mortality among the TyG index, AIP, and their combination is shown in Fig. 5 and Table S7. The combined TyG index and AIP model demonstrated significantly higher AUC values for predicting 14-day, 30-day, 180-day, and 365-day mortality compared with either TyG index or AIP alone (all P < 0.05).

Fig. 5ROC curves of TyG index, AIP, and their combined prediction of mortality. A 14-day mortality; B 30-day mortality; C 180-day mortality; D 365-day mortality

Mediation analysis

As shown in Fig. 6 and Figure S4, bidirectional mediation analyses revealed that TyG index (≥ 9.62) significantly mediated the relationship between AIP (≥ 0.45) and mortality across all time points. The mediation effect progressively increased over time and was most prominent for long-term outcomes, accounting for 65.47% of the association with 365-day mortality (P < 0.001), and 37.32%, 42.38%, and 58.02% for 14-day, 30-day, and 180-day mortality, respectively.

Conversely, high AIP mediated the association between elevated TyG index and mortality only for short-term outcomes, with mediation proportions of 30.39% (P = 0.036) for 14-day and 25.76% (P = 0.042) for 30-day mortality. No significant mediation effects were observed for 180-day or 365-day mortality. These findings indicate a temporal heterogeneity in metabolic interactions in critically ill CHF, with early AIP-driven effects and later sustained TyG-mediated mortality influence.

Fig. 6Bidirectional mediation analysis between the TyG index and AIP on 14-day and 365-day mortality. Model 1: Unadjusted.

Model 2: Adjusted for age, sex, SBP, DBP, heart rate, diabetes, dyslipidemia, stroke, CKD, hemoglobin, RDW, Neu, Lym, Mono, ALT, AST, TBiL, HbA1c, TC, eGFR, BUN, Potassium, Sodium, Chloride, Lg of NT-proBNP, Statins, ACEI/ARB/ARNI, β-blocker, Diuretics and PCI. A TyG index mediates the association between AIP and 14-day mortality B TyG index mediates the association between AIP and 365-day mortality C AIP mediates the association between TyG index and 14-day mortality D AIP mediates the association between TyG index and 365-day mortality

Prediction of mortality risk in CHF patients using machine learning models

Feature selection was performed using LASSO-Cox regression, identifying ten key predictors: age, Neu, Lym, Mono, BUN, chloride, TyG index, AIP, Lg of NT-proBNP, and PCI (Figure S5).

Based on these predictors, six machine learning algorithms were developed to predict 365-day all-cause mortality in CHF patients. The details of the parameter settings for each model were summarized in Table S8. Performance evaluation results on the testing set (Fig.

7, Table S9) demonstrated that the random forest model exhibited the highest discrimination (AUC = 0.812, 95%CI: 0.736–0.887) and showed good calibration (calibration error = 0.139). Moreover, across a broad threshold range (0–80%), it consistently yielded the greatest net clinical benefit, confirming its superior practical utility.

Fig. 7Performance evaluation of six machine learning models on the testing set. A.

ROC curves; B.calibration curves; C. DCA curves. DT: decision tree; LightGBM: light gradient boosting machine; LR: logistic regression; RF: random forest; SVM: support vector machine; XGBoost: extreme gradient boosting.

To assess the individual and combined prognostic value of TyG index and AIP, four hierarchical RF models were developed, with Model 1 serving as the baseline; Models 2 and 3 additionally included the TyG index and AIP, respectively, while Model 4 incorporated both on top of Model 1. As shown in Table S10, results from the NRI and IDI analyses indicated that the inclusion of the TyG index alone improved model performance. When AIP was included alone, only a slight but significant improvement in NRI was observed.

Notably, simultaneous inclusion of both biomarkers resulted in the most substantial incremental predictive performance (NRI = 10.20, 95%CI: 1.90–21.40; IDI = 1.70, 95%CI: 0.15–3.30).

Time-dependent performance analysis (Fig. 8) further demonstrated that, in both the training and testing sets, the combined model incorporating the TyG index and AIP achieved significantly higher C-index and AUC values compared with the baseline model. Moreover, the combined model maintained robust predictive stability over time.

Collectively, these observations suggest that integrating these two indices in the predictive model provides consistent quantitative support for patient risk stratification, thereby supporting their combined application in clinical CHF risk assessment.

Fig. 8Time-dependent C-index and ROC curves of the model on the training and testing sets. A C-index of training set; B C-index of testing set; C ROC curve of training set D ROC curve of testing set.

To further elucidate the internal relationships between model variables and outcomes, SHAP analysis was applied to the RF model for interpretability assessment. Figure 9 displays the ranking of variable importance and the distribution of SHAP values across features. Additionally, Figure S6 presents two representative examples illustrating the interpretability of the model.

Fig. 9SHAP Summary. A SHAP feature importance ranking B SHAP distribution of feature impacts on model output.

Sensitivity analysis

The associations between the combined biomarkers and mortality at different time points were assessed across subgroups, with no significant interactions detected (Tables S11-S14). To reduce selection bias, 1:1 PSM was performed to balance baseline variables between groups (Table S15). In the PSM-adjusted cohort, Cox regression confirmed that patients with elevated TyG and AIP levels remained at higher risk of all-cause mortality (Table S16).

Additionally, when grouping thresholds were redefined using median values (TyG: 9.04, AIP: 0.24), the high-high group still showed a significant association with increased mortality (Table S17). In summary, the combined predictive value of TyG and AIP was consistent across subgroup analyses, model adjustments, and grouping strategies, underscoring the robustness of our findings.

Discussion

This study is the first to systematically evaluate the combined prognostic value of the TyG index and AIP in critically ill CHF patients. Specifically, concurrent elevation of the two biomarkers (TyG ≥ 9.62 and AIP ≥ 0.45) was associated with the highest mortality risks, showing a significant synergistic predictive effect. Mediation analysis revealed time-dependent heterogeneity in their interaction: the former mediated the association between the latter and mortality across all follow-up periods, whereas the latter mainly mediated the impact of the former on short-term outcomes.

These findings highlight an interactive pathophysiological link between the two indices. Based on this mechanism, we further developed a long-term mortality prediction tool integrating both markers to support clinical decision making.

Earlier investigations have indicated that the TyG index is notably correlated with adverse clinical outcomes among critically ill individuals [21–23]. Within the realm of HF research, the prognostic significance of this metabolic indicator has received increasing research attention, with several studies linking its elevated levels to poor prognosis [9, 24–29]. A large meta-analysis encompassing 30 studies demonstrated that increased TyG was related to unfavorable prognoses in HF outcomes [30].

Another meta-analysis further confirmed a robust positive relationship between this index and survival outcomes (HR 1.71, 95%CI: 1.46-2.00, P < 0.001) [31]. However, most previous investigations have evaluated this association at one solitary time point. Our study explored how the TyG index relates to all-cause mortality over various time periods among these individuals with critical CHF, and confirmed that its elevation is significantly linked to heightened mortality risks in both the short and long term.

Moreover, no consensus regarding whether the relationship between the TyG index and adverse outcomes follows a linear pattern. Our research identified time-dependent heterogeneity in the correlation of the TyG index with all-cause mortality. In the 14-day hospital phase, a positive linear correlation was detected, aligning with the results reported by Kan et al. [25].

However, at the other time points (30, 180, and 365 days), the link between TyG index and mortality exhibited an approximately J-shaped pattern (threshold range: 9.03–9.08). This temporal heterogeneity may reflect distinct pathophysiological mechanisms at different stages of disease progression. Early mortality is likely driven by acute hemodynamic instability and metabolic stress, with IR exacerbating HF through oxidative stress [32], neurohormonal activation [33, 34], and inflammation [35], thereby directly accelerating decompensation and producing a linear increase in mortality risk.

After discharge, chronic myocardial injury mechanisms, including lipotoxic myocardial damage [36], and endothelial or coronary microvascular dysfunction [4, 37, 38], become more prominent, contributing to a nonlinear increase in long-term mortality. These findings suggest that clinical management strategies should be stage-specific: early ICU interventions should focus on glucose control and improving insulin sensitivity in patients with elevated TyG, whereas long-term management should target patients who exceed the identified threshold, emphasizing lifestyle modification and pharmacological interventions to enhance cardiac energy metabolism.

AIP, a sensitive marker of lipid metabolic disturbances and atherosclerotic risk, has prognostic value in cardiovascular diseases [11] and metabolic disorders such as diabetes [39]. However, few studies have explored its prognostic relevance in HF. While Yu et al. [13] reported its predictive value for short-term outcomes, its relationship with long-term prognosis remains unclear.

We verified that this index may independently predict short- and long-term clinical outcomes in this cohort. Across the entire observation period, it exhibited a near-linear relationship with mortality, indicating a time-stable association consistent with Zheng et al. [40]. Pathophysiologically, its elevation is often accompanied by higher concentrations of triglyceride-rich lipoproteins, which may hasten coronary plaque development and promote plaque instability, thereby worsening myocardial ischemia and increasing the likelihood of acute cardiovascular events [41–43].

Conversely, elevated levels of this marker are typically associated with reduced HDL-C, leading to attenuated anti-inflammatory and antioxidative capacities and impaired clearance of small, dense LDL (sd-LDL) particles—key contributors to atherosclerosis—thereby exacerbating vascular injury [44, 45]. Our findings emphasize the need for continuous and rigorous lipid management in critically ill CHF patients, as no absolute safe threshold for AIP may exist.

Previous studies have primarily examined the TyG index or AIP separately in relation to prognosis, whereas our study is the first to systematically investigate the combined predictive value of these two indices and confirm a significant synergistic effect between them in critically ill CHF patients. Cox regression analyses revealed that individuals with concurrently increased TyG index and AIP had consistently higher mortality risks than those with low levels of both indices across all observation periods. The pathophysiological basis of this synergistic effect may involve the bidirectional interplay of glucose-lipid metabolic disturbances, leading to myocardial injury.

IR, reflected by elevated TyG index, and lipid abnormalities, indicated by increased AIP, may jointly amplify oxidative stress and systemic inflammatory activation, thereby promoting myocardial apoptosis and interstitial fibrosis [8]. Furthermore, IR-related excessive activation of the sympathetic nervous system and the renin-angiotensin-aldosterone system (RAAS) can increase cardiac preload, while AIP-associated platelet activation and thrombogenesis enhance coronary plaque instability and the likelihood of acute ischemic episodes [46, 47]. Together, these processes contribute to accelerated HF decompensation.

Mediation analysis revealed temporal heterogeneity in the bidirectional interaction between the TyG index and AIP. The TyG index mediated the association between AIP and mortality across all time points, with the mediation effect strengthening over time and accounting for 65.47% of the association with 365-day mortality. The results indicate that the TyG index may continuously influence HF prognosis through long-term regulation of lipid metabolism-related myocardial injury pathways [48].

In contrast, AIP’s mediating role linking the TyG index to mortality was restricted to short-term outcomes, likely reflecting its role in promoting coagulation imbalance and plaque rupture through IR-related mechanisms [49, 50]. This time-dependent pattern provides a potential “temporal window” for phased clinical intervention in critically ill CHF patients: early ICU management should emphasize monitoring and improving both TyG and AIP to mitigate short-term mortality, whereas long-term management should target IR-related pathways and chronic myocardial remodeling, with dynamic monitoring of AIP to disrupt the “IR—lipid disorder” feedback loop and ultimately improve long-term outcomes.

With advances in computational science, machine learning and other artificial intelligence techniques have demonstrated distinct advantages in predicting outcomes among HF patients [51]. Despite these developments, investigations combining the TyG and AIP for quantitative mortality risk prediction in this population remain scarce. To translate their synergistic prognostic value into practical clinical tools, we integrated these two biomarkers into a prognostic modeling framework, developing and validating six machine learning models.

Among them, the RF-based model exhibited superior predictive accuracy compared with the other models. Widely adopted in critical care mortality forecasting, its capacity for non-linear fitting allows it to capture complex correlations between metabolic biomarkers and clinical outcomes [52–54]. Moreover, compared with the baseline model containing traditional clinical variables or models including only TyG index or AIP, the combined model incorporating both biomarkers exhibited superior predictive accuracy and temporal stability.

To mitigate the “black-box” nature of such predictive systems, SHAP analysis was employed to interpret the optimal RF model [55], quantifying the relative importance of individual predictors in outcome prediction and revealing the internal logic of model decision-making.

To facilitate clinical translation, the RF model developed in this study can be deployed as a clinical decision support tool through two complementary approaches: First, it can be seamlessly integrated into the hospital electronic health record system as an intelligent analytic module. During acute care, when patients are admitted to the ICU or in early stabilization, the system can automatically calculate and display dynamic risk scores based on real-time input indicators, while providing concise risk attribution insights via the SHAP explainer. This assists bedside clinicians in rapidly identifying high-risk patients and formulating targeted metabolic interventions.

Second, the model can be developed into a lightweight, open-access web-based risk calculator. Through an intuitive web interface, clinicians can input routine variables (e.g., age, key laboratory indicators) to quickly obtain a risk assessment with visual interpretation. This platform promotes broader adoption in outpatient and community settings and offers a flexible foundation for ongoing model validation and iterative refinement.

Nevertheless, this study also has certain limitations. First, it adopts a retrospective study design and cannot conduct causal inference. Second, although multiple confounders were adjusted for using PSM to enhance the robustness of the findings, unmeasured confounding factors may still exist and influence the interpretation.

Third, the data were primarily derived from ICU patients in the United States, thereby restricting the applicability of the conclusions. Finally, the TyG index and AIP were calculated, derived from single baseline measurements without accounting for dynamic changes during disease progression. This may not fully capture their temporal evolution.

Future research ought to focus on two aspects: first, conducting prospective, multicenter studies to improve population diversity and validate the generalizability of the TyG-AIP combination in outcome prediction; and second, performing longitudinal assessments of TyG index and AIP to capture real-time metabolic dynamics and develop more accurate, time-dependent risk prediction models—a key direction for future investigation.

Conclusion

Combining the TyG index with the AIP constitutes an effective strategy for evaluating the short- and long-term prognosis in critically ill patients with CHF. The cooperative and complementary nature of these two metabolic markers significantly enhances the mortality risk prediction capability of machine learning models. These findings endorse the clinical incorporation of dual TyG-AIP monitoring into routine management—both during hospitalization and after discharge—to improve patient outcomes through targeted interventions addressing insulin resistance and lipid metabolic dysfunction.

Supplementary Information

Supplementary Material 1.

Related articles

PubMed

Latent Profile Subtypes of Depressive Symptoms and Associated Factors Among Elderly Patients with Chronic Heart Failure.

Objective To identify latent subtypes of depressive symptoms and their associated factors among elderly patients with chronic heart failure (CHF), so as to provide reference for developing targeted psychological intervention strategies. Methods A cross-sectional survey enrolled 423 hospitalized elderly CHF patients from June 2025 to May 2026. Assessments included SDS, SAS, Family APGAR, MLHFQ and PNI. Different from conventional SDS total-score-based severity grouping, latent profile analysis...

PubMed

Association Between Time Perspective and Fear of Disease Progression in Patients with Chronic Heart Failure: A Cross-Sectional Study.

Purpose To describe the level of fear of disease progression (FoP) in patients with chronic heart failure (CHF) and to examine whether individual dimensions of time perspective are associated with FoP. Patients and Methods In this single-center cross-sectional study, we surveyed a convenience sample of 254 patients treated in the cardiovascular department of a tertiary hospital in Nantong between December 2025 and August 2026. Time perspective was measured with the Chinese version of the Zimb...