Unmasking silent bioenergetic failure in normolactatemic AMI patients: a multicenter validation of the TyG-ACAG index

Abstract

Background

Scores such as SOFA and blood lactate, which reflect hemodynamic compromise, may fail to capture early metabolic dysfunction in critically ill patients with Acute Myocardial Infarction (AMI). We aimed to develop and validate the TyG-ACAG index—a novel integrated marker of insulin resistance and metabolic acidosis—and evaluate its incremental value for risk stratification.

Methods

This multicenter, retrospective cohort study was derived in 2,277 adult AMI patients from the eICU-CRD and validated in 738 patients from the MIMIC-IV and Tongji Hospital databases. The TyG-ACAG index was calculated within 24 hours of ICU admission. Associations with all-cause in-hospital mortality were assessed using multivariable logistic regression with Inverse Probability of Treatment Weighting (IPTW). Incremental predictive value over a baseline model (Age + SOFA) was quantified using the Net Reclassification Improvement (NRI). Additionally, we performed discordance analysis and utilized SHAP values from an XGBoost model to interpret feature importance.

Results

A high TyG-ACAG index (>129.45) was independently associated with increased in-hospital mortality (OR = 1.723; 95% CI: 1.343–2.212; P < 0.001). The inclusion of the index significantly refined risk stratification, yielding an NRI of 0.381 (P < 0.001) over the baseline model. Crucially, discordance analysis unmasked a high-risk phenotype: patients with clinically normal lactate (<2.0 mmol/L) but elevated TyG-ACAG levels had significantly higher mortality (HR = 3.95; 95% CI: 2.08–7.48; P < 0.001). Machine learning analysis corroborated these findings, ranking the index as the third most important predictor. Subgroup analysis further highlighted its robust prognostic value, particularly in younger patients (<65 years, OR = 11.84).

Conclusion

The TyG-ACAG index serves as a robust indicator of bioenergetic failure, providing prognostic information that complements traditional hemodynamic markers. By identifying high-risk individuals masked by normolactatemia, this novel metric effectively bridges the gap in early metabolic risk assessment for critically ill AMI patients.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12944-026-02919-0.

Research Insights

What is currently known about this topic?

• Hemodynamic scores (e.g., SOFA) often fail to capture early metabolic dysfunction in AMI.

• The TyG index is a well-established surrogate marker for insulin resistance.

• Metabolic acidosis and insulin resistance are synergistic but rarely combined in risk models.

What is the key research question?

• Does the novel TyG-ACAG index improve mortality prediction and risk stratification in AMI?

What is new?

• TyG-ACAG is a robust, independent predictor of in-hospital mortality in critically ill AMI.

• The index significantly improves risk reclassification (NRI=0.381) over traditional scores.

• It identifies a high-risk phenotype among patients with clinically normal lactate levels.

How might this study influence clinical practice?

It unmasks “hidden” metabolic risk in patients with normal lactate, guiding early intervention.

Background

Acute Myocardial Infarction (AMI) remains a leading cause of mortality in the intensive care unit [1]. Despite timely revascularization with percutaneous coronary intervention (PCI), in-hospital mortality for critically ill AMI patients remains stubbornly high, suggesting a gap in early risk stratification [2]. Current mainstays, such as the Sequential Organ Failure Assessment (SOFA) score and blood lactate, primarily reflect hemodynamic consequences or global hypoperfusion [3, 4]. They may fail to capture early, purely metabolic dysfunction that precedes overt organ failure. Therefore, identifying biomarkers that directly reflect this metabolic distress is critical for improving proactive management.

Two pivotal components of metabolic derangement in critical illness are insulin resistance (IR) and metabolic acidosis. The Triglyceride-Glucose (TyG) index has emerged as a simple, reliable surrogate for IR, reflecting impaired substrate utilization under stress [5, 6]. Concurrently, the Albumin-Corrected Anion Gap (ACAG) provides a more accurate assessment of unmeasured anions (e.g., ketoacids, lactate) than the traditional anion gap, especially in hypoalbuminemic ICU patients [7, 8]. Crucially, these pathways interact: acidosis can impair insulin signaling, while insulin resistance exacerbates acid accumulation. This synergistic interplay suggests that their combined evaluation might be more informative than either marker alone in identifying a state of “bioenergetic failure”.

While the TyG index alone predicts adverse outcomes in cardiovascular diseases [9, 10], the combined prognostic value of TyG and ACAG, capturing both arms of metabolic distress, remains unexplored in AMI. We hypothesize that the integrated TyG-ACAG index will outperform individual parameters and traditional scores in risk stratification. Using multi-center data [11, 12], this study aims to: (1) validate the independent association between the TyG-ACAG index and in-hospital mortality in critically ill AMI patients; and (2) determine its utility in unmasking a “hidden high-risk” phenotype among patients with normal lactate levels, thereby offering a novel lens for metabolic risk assessment.

Methods

Study design and ethical considerations

This multicenter, retrospective, observational cohort study employed a three-tiered validation framework utilizing data from the eICU Collaborative Research Database (eICU-CRD, v2.0) [13], the MIMIC-IV database (v3.1) [14], and the Department of Cardiovascular Surgery and Critical Care Medicine at Tongji Hospital. The study protocol was approved by the Ethics Committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology and strictly adhered to the principles of the Declaration of Helsinki (TJ-IRB202601113). Access to the US-based databases was authorized after the completion of the Collaborative Institutional Training Initiative (CITI) program (Record ID: 74166104), and the requirement for informed consent was waived due to the de-identified nature of the data. For the Tongji cohort, the requirement for informed consent was waived given the retrospective design.

Study population

The study enrolled adult patients (aged ≥ 18 years) with a primary diagnosis of AMI. Patients were identified using ICD-9-CM codes 410.xx (excluding 410.x2) and ICD-10-CM codes I21.xx for the primary diagnosis of AMI. Inclusion criteria mandated complete data for the calculation of the TyG-ACAG index within the first 24 h of ICU admission and an ICU length of stay of at least 24 h. Strict exclusion criteria were applied to minimize confounding. To ensure independence of observations, we included only the first ICU admission for each patient during the index hospitalization; subsequent ICU readmissions or repeat ICU stays were excluded. For patients with ICU transfers within the same hospitalization, only data from the initial ICU stay were used. Patients with greater than 30% missing core data were excluded. To prevent renal or hepatic dysfunction from masking metabolic signals, patients with end-stage renal disease (ESRD), chronic dialysis, or severe hepatic failure (e.g., cirrhosis) were excluded (e.g., ICD-9-CM 585.6 for End-Stage Renal Disease; ICD-9-CM 571.2/571.5 for Liver Cirrhosis). Furthermore, patients with primary diabetic ketoacidosis (DKA) or hyperglycemic hyperosmolar states (HHS) were excluded to distinguish ischemic stress-induced metabolic exhaustion from primary endocrine emergencies.

Patients with missing values for the primary predictor (TyG-ACAG index components) were excluded from the analysis to ensure the accuracy of the primary exposure. For baseline covariates with missing rates < 30% (e.g., BMI, serum creatinine), we assumed data were missing at random (MAR) and employed Multiple Imputation by Chained Equations (MICE) to generate five imputed datasets. The analysis was performed on each imputed dataset, and the results were pooled. Variables with > 30% missingness were excluded from the feature set to avoid introducing significant bias.

A total of 3,015 patients were included in the final analysis, comprising 2,277 from the eICU derivation cohort, 187 from the MIMIC-IV validation cohort, and 551 from the Tongji validation cohort (Fig. 1).

Fig. 1.

Study flowchart outlining the selection process for the study cohorts. The flowchart details the inclusion and exclusion criteria applied to patients from the eICU-CRD, MIMIC-IV, and Tongji Hospital databases

Variable construction and definitions

The TyG-ACAG index was constructed to quantify the interaction between metabolic resistance and acidic accumulation. The TyG index was calculated using the following formula: TyG = ln [ triglyceride (mg/dL) × glucose (mg/dL) / 2 ]. Given the emergent nature of AMI and the impracticability of strict fasting in the ICU, glucose levels were defined as the first measurement obtained upon ICU admission. This value serves as a proxy for admission metabolic stress and acute insulin resistance. The ACAG was calculated using the formula: ACAG = Anion Gap + 2.5 × (4.4 − Albumin [g/dL]). The Anion Gap was calculated using the standard formula without potassium: Anion Gap = [Sodium] − ([Chloride] + [Bicarbonate]), where [Bicarbonate] represents the serum bicarbonate concentration measured from the venous basic metabolic panel, distinct from calculated arterial bicarbonate. The final TyG-ACAG index was defined as the product of these two components: TyG-ACAG = TyG × ACAG. This composite metric functions as an integrated biomarker for bioenergetic failure. Baseline covariates, including demographics, comorbidities quantified by the Charlson Comorbidity Index, vital signs, and laboratory markers, were extracted. For laboratory variables, we extracted the first available measurement within the first 24 h after ICU admission; when multiple values were available within this window, the earliest value was used. Patients without required measurements within 24 h were excluded from index computation. Illness severity was assessed using SOFA scores and the Glasgow Coma Scale (GCS), with the primary outcome was defined as all-cause in-hospital mortality. The secondary outcome was 28-day mortality, which was assessed to evaluate survival trajectories over time.

Statistical analysis

Statistical analyses were conducted using R (v4.2.0) and Python (v3.9). Patients were stratified into quartiles based on the TyG-ACAG index. Baseline characteristics were compared using one-way ANOVA or Kruskal-Wallis tests for continuous variables, and Chi-square tests for categorical variables.

Inverse probability of treatment weighting (IPTW). The exposure was defined as a high TyG-ACAG index versus low TyG-ACAG. Propensity scores were estimated using logistic regression including baseline covariates listed in Table 1. Stabilized ATE weights were calculated as SW = P(T = 1)/PS for exposed patients and SW = P(T = 0)/(1 − PS) for unexposed patients. To reduce the influence of extreme weights, stabilized weights were truncated at the 1st and 99th percentiles. Weight diagnostics included summary statistics of the final weights and the effective sample size (ESS), calculated as (∑ ω)²/∑ ω², reported overall and by exposure group. Covariate balance was assessed using standardized mean differences (SMD), with SMD < 0.1 indicating adequate balance.

Table 1.

Baseline characteristics of the eICU derivation cohort

Variable	Q1 (Lowest)	Q2	Q3	Q4 (Highest)	P-value	SMD (Q4 vs. Q1)
Age (years)	63.2 ± 13.3	65.0 ± 13.3	64.7 ± 13.3	65.0 ± 13.1	0.068	0.138
Gender (Male)	423 (74.2%)	380 (66.8%)	375 (65.9%)	358 (62.9%)	< 0.001	0.245
SOFA Score	4.0 [1.0, 5.0]	4.0 [1.0, 5.0]	4.0 [2.0, 5.0]	5.0 [3.0, 8.0]	< 0.001	0.533
Mean Arterial Pressure (mmHg)	85.4 ± 30.7	94.5 ± 45.6	88.0 ± 42.0	86.1 ± 38.0	0.323	0.019
Heart Rate (bpm)	77.0 ± 15.5	80.2 ± 16.7	82.0 ± 17.8	88.8 ± 19.4	< 0.001	0.678
Lactate (mmol/L)	1.8 [1.2, 2.5]	1.8 [1.2, 3.3]	2.1 [1.3, 3.3]	3.0 [1.8, 5.6]	< 0.001	0.600
Creatinine (mg/dL)	1.0 [0.8, 1.1]	1.0 [0.8, 1.2]	1.0 [0.8, 1.3]	1.2 [0.9, 1.7]	< 0.001	0.494
White Blood Cells (×10^9/L)	9.9 [7.8, 12.5]	10.6 [8.4, 13.0]	10.9 [8.7, 14.2]	11.9 [9.3, 16.2]	< 0.001	0.480
Hemoglobin (g/dL)	13.9 ± 2.1	13.6 ± 2.3	13.5 ± 2.4	13.2 ± 2.6	< 0.001	0.317
Platelets (×10^9/L)	216.0 [177.0, 256.0]	224.0 [189.0, 267.0]	232.5 [186.0, 284.0]	233.0 [180.0, 292.0]	< 0.001	0.251
Glucose (mg/dL)	125.0 [108.0, 152.0]	135.0 [114.0, 172.0]	143.0 [116.0, 190.0]	173.0 [132.0, 270.0]	< 0.001	0.752
Albumin (g/dL)	3.7 ± 0.5	3.6 ± 0.5	3.6 ± 0.6	3.4 ± 0.7	< 0.001	0.559
TyG-ACAG Index	78.6 [66.7, 87.7]	108.0 [101.6, 114.4]	133.9 [127.2, 140.9]	175.4 [161.0, 199.4]	< 0.001	4.054
Hypertension	74 (13.0%)	84 (14.8%)	79 (13.9%)	94 (16.5%)	0.372	0.100
Diabetes Mellitus	117 (20.5%)	128 (22.5%)	162 (28.5%)	236 (41.5%)	< 0.001	0.465
Congestive Heart Failure	36 (6.3%)	48 (8.4%)	58 (10.2%)	72 (12.7%)	0.002	0.218
Chronic Kidney Disease	19 (3.3%)	33 (5.8%)	40 (7.0%)	63 (11.1%)	< 0.001	0.303
In-hospital Mortality	15 (2.6%)	27 (4.7%)	32 (5.6%)	75 (13.2%)	< 0.001	0.399

Outcome Analysis: The association between the TyG-ACAG index and the primary outcome (in-hospital mortality) was evaluated using weighted multivariable logistic regression. Results were expressed as Odds Ratios (OR) with 95% Confidence Intervals (CI). To capture time-dependent survival differences, particularly in the discordance analysis, Cox proportional hazards models were employed for the secondary outcome (28-day mortality), with results reported as Hazard Ratios (HR) and visualized using Kaplan-Meier curves.

Model Evaluation: To quantify the incremental predictive value of the TyG-ACAG index beyond traditional risk scores (e.g., SOFA) and biomarkers (e.g., Lactate), the Continuous Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) were calculated via bootstrap resampling.

Model calibration was assessed using the Brier score, calibration slope, and calibration intercept (calibration-in-the-large). Calibration curves were plotted to visualize the agreement between predicted probabilities and observed outcomes.

Machine learning and interpretability

To capture potential non-linear relationships, an Extreme Gradient Boosting (XGBoost) model was developed using the entire derivation cohort (eICU-CRD) to maximize statistical power. The feature set included all demographic, vital sign, comorbidity, and laboratory variables presented in Table 1.

To strictly prevent data leakage, the external validation cohorts (MIMIC-IV and Tongji Hospital) were completely isolated from the model training and hyperparameter tuning processes. They served as independent “held-out” datasets for final performance evaluation.

Hyperparameters (including learning_rate, max_depth, n_estimators) were optimized via 5-fold cross-validation performed on the derivation cohort to prevent overfitting. The final model parameters were: learning_rate = 0.05, max_depth = 5, n_estimators = 100, and random_state = 42.

Model performance was rigorously evaluated on the two external validation cohorts using AUC, sensitivity, and specificity. SHapley Additive exPlanations (SHAP) values were computed on the derivation cohort to interpret the global feature importance and visualize the specific non-linear threshold effects of the TyG-ACAG index.

Discordance and sensitivity analyses

A discordance analysis was performed to evaluate the index’s value in patients with clinically normal lactate levels, strictly defined as serum lactate < 2.0 mmol/L adhering to Sepsis-3 definitions. Serum lactate for discordance categorization was defined as the first lactate value measured within 24 h of ICU admission. Patients were categorized into four groups: Concordant Low, Combined High Risk (High Index/High Lactate), Isolated Hyperlactatemia, and a phenotype of high metabolic risk despite normolactatemia (High Index/Normal Lactate). Kaplan-Meier survival curves and Log-rank tests were utilized to compare survival trajectories among these groups. Finally, to rigorously test robustness and rule out reverse causality, sensitivity analyses were performed by excluding patients who died within 48 h of ICU admission. Additional analyses were conducted in subgroups, including those with and without mechanical ventilation, to ensure findings were not driven by respiratory acidosis.

Results

Baseline characteristics

In the derivation cohort (eICU-CRD, n = 2,277), the median TyG-ACAG index increased from 78.6 [66.7, 87.7] in Quartile 1 (Q1) to 175.4 [161.0, 199.4] in Quartile 4 (Q4) (Table 1). Patients in Q4 exhibited higher glucose (173.0 vs. 125.0 mg/dL) and lower albumin (3.4 vs. 3.7 g/dL) compared to those in Q1. This metabolic disparity was accompanied by a higher prevalence of comorbidities, specifically diabetes (41.5% vs. 20.5%) and chronic kidney disease (11.1% vs. 3.3%). Correspondingly, Q4 patients presented with higher SOFA scores (5.0 vs. 4.0), heart rates (88.8 vs. 77.0 bpm), and lactate levels (3.0 vs. 1.8 mmol/L). In-hospital mortality was significantly higher in Q4 (13.2%) compared to Q1 (2.6%) (P < 0.001). Similar trends were observed in the validation cohorts. In MIMIC-IV (n = 187), Q4 patients had a 28-day mortality of 29.8% compared to 4.3% in Q1 (Supplementary Table S1). In the Tongji cohort (n = 551), in-hospital mortality reached 45.7% in Q4 compared to 10.9% in Q1 (Supplementary Table S2).

Baseline characteristics were compared across the three cohorts (see Supplementary Table S3). Notably, the Tongji cohort exhibited significantly higher disease severity compared to the eICU derivation cohort, with a higher mean SOFA score (7.3 vs. 4.2, P < 0.001) and overall in-hospital mortality (22.1% vs. 6.5%, P < 0.001).

Primary outcomes and incremental predictive value

To minimize baseline imbalances, IPTW was applied. Post-weighting, the SMD for all covariates were reduced to less than 0.1 (Fig. 2A). The stabilized weights were well-behaved after truncation (mean 0.985, SD 0.318; range 0.446–2.396), yielding an ESS of 2062.3 overall (838.0 in the exposed group and 1225.6 in the unexposed group; Supplementary Table S4). In the weighted cohort, a high TyG-ACAG index (≥ 129.45) was independently associated with increased in-hospital mortality (OR = 1.581; 95% CI: 1.113–2.244; P < 0.001). The E-value for this association was 2.84. Regarding predictive performance, adding the TyG-ACAG index to the baseline model (Age + SOFA) increased the AUC from 0.892 to 0.903 (P = 0.006, determined by the DeLong test).

Fig. 2.

Causal Inference and Subgroups. A Love plot displaying covariate balance before and after Inverse Probability of Treatment Weighting (IPTW); all standardized mean differences (SMD) are < 0.1 after weighting. B Forest plot of weighted Odds Ratios (calculated by Logistic Regression) across clinical subgroups, highlighting elevated risk in younger patients and those with heart failure

The inclusion of the index yielded a significant improvement in reclassification, with a Continuous NRI of 0.381 (95% CI: 0.154–0.590) and an IDI of 0.015 (95% CI: 0.003–0.037), calculated via bootstrap resampling (1,000 iterations). To verify the robustness of the optimal cut-off (129.45), which was determined by the SHAP inflection in the derivation cohort, sensitivity analyses were performed. Using a rounded threshold of 130 yielded consistent prognostic performance (Unadjusted OR = 3.07; 95% CI: 2.16–4.36; P < 0.001). Furthermore, quartile-based analysis confirmed a graded dose-response relationship, mitigating concerns regarding overfitting. To address potential confounding from nutritional support, we performed a sensitivity analysis excluding patients who received total parenteral nutrition (TPN) or lipid emulsions within the first 24 h. The association remained consistent (Adjusted OR = 1.98; 95% CI: 1.52–2.58) compared to the primary analysis (Adjusted OR = 2.05), suggesting the results are robust to acute nutritional intake.

Subgroup analysis

Subgroup analyses (Fig. 2B) indicated that the association between the TyG-ACAG index and mortality remained significant across most strata. The odds ratio was notably higher in patients aged < 65 years (OR = 11.84; 95% CI: 3.55–39.42) compared to those aged 65 years (OR = 2.22). Stronger associations were also observed in patients with congestive heart failure (OR = 3.96) and those without diabetes (OR = 2.95) compared to diabetic patients (OR = 2.49).

Discordance analysis

To evaluate risk stratification in patients without overt hyperlactatemia, patients were categorized based on lactate (< 2.0 vs. ≥ 2.0 mmol/L) and TyG-ACAG levels (Fig. 3). Patients with normal lactate levels but a high TyG-ACAG index (Group C) exhibited significantly higher mortality compared to the reference group (Group A: Low Lactate/Low Index). Kaplan-Meier survival analysis in the eICU and MIMIC cohorts showed that while early survival rates were comparable between Group C and Group A, the survival curves diverged significantly during the late phase of hospitalization (> 25 days). In the Tongji cohort, Group C was associated with a hazard ratio of 3.95 (95% CI: 2.08–7.48; P < 0.001) compared to Group A.

Fig. 3.

Discordance analysis. Kaplan-Meier survival curves stratified by lactate and TyG-ACAG levels (Hazard Ratios calculated by Cox Regression)

Machine learning and SHAP analysis

The XGBoost model achieved an AUC of 0.930. SHAP analysis (Fig. 4A) ranked the TyG-ACAG index as the third most important feature for predicting in-hospital mortality, following Age and SOFA score. Its mean SHAP weight (0.25) surpassed traditional markers such as BUN (0.22) and WBC (0.21). For the prediction of 28-day mortality (Fig. 4B), the index ranked sixth in feature importance. The SHAP dependence plot (Fig. 4C) revealed a non-linear relationship, where mortality risk increased sharply after the index exceeded a value of 129.45. Additionally, the SHAP interaction plot (Fig. 4D) showed that the impact of elevated ACAG on mortality risk was greater in patients with higher TyG levels compared to those with lower TyG levels.

Fig. 4.

Machine Learning Analysis. A SHAP summary plot ranking feature importance for in-hospital mortality prediction; the TyG-ACAG index ranks third. B SHAP summary plot ranking feature importance for 28-days mortality prediction; the TyG-ACAG index ranks six. C SHAP dependence plot showing a J-shaped risk trajectory with a tipping point at 129.45. D SHAP interaction plot demonstrating amplified mortality risk from acidosis in patients with high TyG levels. Metabolic flexibility (Low TyG) buffers the lethal effect of acidosis

Clinical utility

A dynamic nomogram incorporating Age, SOFA score, and the TyG-ACAG index was developed (Fig. 5), demonstrating good discrimination (AUC = 0.896). Calibration was assessed using the Brier score, calibration slope, and calibration intercept. Decision Curve Analysis (Fig. 6) demonstrated that the model including the TyG-ACAG index provided a higher net benefit. Specifically, in the Tongji cohort at the 20% threshold, the model added a net benefit of 0.0068, which is equivalent to identifying approximately 0.7 additional true positive patients per 100 patients without increasing false positives. The model displayed excellent internal calibration in the eICU cohort, yielding a low Brier score of 0.0472. Assessment of calibration-in-the-large revealed a slope of 1.0005 and an intercept of 0.0010, suggesting precise risk estimation without systematic bias.

Fig. 5.

Dynamic Nomogram. A clinical prediction tool integrating Age, SOFA score, and TyG-ACAG index to estimate individual probability of in-hospital mortality

Fig. 6.

Decision Curve Analysis. Plots comparing the net clinical benefit of the TyG-ACAG enhanced model versus the baseline model. The enhanced model demonstrates superior net benefit, particularly in the low-to-intermediate probability thresholds relevant for early screening. Specifically, at a threshold probability of 20%, the net benefit was 0.0068, corresponding to identifying approximately 0.7 additional true-positive patients per 100 screened patients without increasing false positives

Discussion

Our multicenter investigation validates the TyG-ACAG index as a robust, independent predictor of all-cause mortality in critically ill patients with AMI. While traditional scoring systems such as the SOFA focus primarily on macroscopic organ dysfunction [15], our findings suggest that the TyG-ACAG index captures a distinct dimension of “metabolic fragility” and “metabolic acidosis burden”. The index demonstrated superior discriminative power compared to SOFA alone, and the substantial NRI = 0.381 indicates that incorporating metabolic markers significantly refines risk stratification. These data support the utility of the TyG-ACAG index in identifying patients who, despite apparent hemodynamic stability, are at high risk due to underlying bioenergetic failure.

The prognostic value of the TyG-ACAG index may be interpreted through the synergistic interaction of insulin resistance and unmeasured anion accumulation. First, the TyG component reflects not only insulin resistance but also a state of substrate overload and lipotoxicity under acute stress [16]. In the context of ischemia, this metabolic inflexibility impairs the myocardium’s ability to utilize glucose, exacerbating cellular energy deficits [17]. Second, the ACAG component corrects for the confounding effect of hypoalbuminemia, thereby unmasking the true burden of unmeasured anions [18]. This “anion gap” likely represents the accumulation of metabolic byproducts associated with ischemia-reperfusion injury, such as succinate. Elevated succinate levels have been mechanistically linked to mitochondrial complex II reversal and the subsequent generation of reactive oxygen species (ROS), which precipitate cell death [19]. The significant interaction observed in our model suggests that insulin resistance compromises the host’s buffering capacity against this metabolic toxicity, leading to a compounded risk of mortality when both factors are elevated.

A critical implication of this study is the ability of the TyG-ACAG index to address the diagnostic limitations of serum lactate. While lactate is a standard marker for tissue hypoperfusion, it may lack sensitivity in detecting “occult hypoperfusion” due to efficient hepatic clearance or non-hypoxic drivers [20]. Our discordance analysis identified a high-risk phenotype (High TyG-ACAG/Normal Lactate) with significantly elevated mortality, suggesting that the index captures a “metabolic debt” that persists even when global perfusion appears adequate. Furthermore, the index offers substantial incremental value over the SOFA score, particularly in younger patients (< 65 years), where the odds ratio was most pronounced (OR 11.84). Younger patients often possess greater physiological reserve, maintaining hemodynamic stability despite severe cellular stress. In this cohort, the TyG-ACAG index appears to function as an antecedent marker, signaling that cellular bioenergetic homeostasis is severely compromised before the onset of overt organ failure. Crucially, this prognostic utility was reinforced by the DCA, which demonstrated that the TyG-ACAG nomogram yields a superior net benefit compared to the “treat-all” or “treat-none” strategies across a wide range of threshold probabilities. This implies that the index can effectively assist clinicians in identifying high-risk patients who require intensive monitoring, without increasing the burden of unnecessary interventions for low-risk individuals.

The integration of the TyG-ACAG index into clinical practice may facilitate a shift towards “Metabolic Phenotyping” in the ICU. Identification of the “Metabolic Exhaustion Phenotype” (Index > 129.45) warrants investigation into precision resuscitation strategies. Specifically, these patients might benefit from protective nutritional strategies that avoid iatrogenic substrate overload, as their metabolic machinery is ill-equipped to handle excessive lipid or glucose loads [21]. Future studies should explore whether targeted metabolic modulation guided by this index can improve outcomes compared to standard care.

We observed significant cohort heterogeneity, particularly in the Tongji cohort, which exhibited significantly higher mortality and disease severity compared to the Western cohorts (mean SOFA score: 8.0 in Tongji vs. 4.2 in eICU and 5.4 in MIMIC-IV; see Supplementary Table S2). This disparity reflects the nature of the participating center; Tongji Hospital serves as a premier tertiary referral center in central China, managing a disproportionately high volume of critical AMI complications, such as cardiac rupture and refractory cardiogenic shock requiring mechanical support. While the TyG-ACAG index remained a robust risk factor across all cohorts, the calibration slope (< 1) observed in the Tongji cohort suggests that recalibration of the baseline model may be necessary before implementation in such ultra-high-risk populations to avoid overestimation of survival.

Several limitations of this study merit consideration. First, the retrospective design precludes definitive adjusted association, although the use of multicenter cohorts and propensity score weighting strengthens the validity of the associations. Second, while we hypothesize that the elevated ACAG reflects specific anions such as succinate or ketones, we did not perform metabolomic profiling to quantify these specific constituents. Third, the TyG index can be influenced by acute stress responses and therapeutic interventions (e.g., lipid emulsions). Therefore, clinicians should interpret baseline values with caution; dynamic monitoring of the index trajectory may offer more robust clinical insight than a single snapshot. Finally, the sample size in certain subgroups, particularly in the validation cohort, may limit the statistical power for detecting smaller effect sizes in specific demographics. Moreover, the magnitude of the hazard ratios in these subgroups (e.g., younger patients) should be interpreted with caution due to the limited number of events.

Conclusions

In conclusion, this multicenter study establishes the TyG-ACAG index as a robust, integrative biomarker of bioenergetic failure in critically ill AMI patients, offering prognostic value that complements traditional hemodynamic markers. By capturing the synergistic lethality of insulin resistance and metabolic acidosis, this index effectively unmasks “metabolic blind spots”—specifically identifying high-risk individuals who remain undetected by normal lactate levels. These findings advocate for a paradigm shift towards “Metabolic Phenotyping” in the ICU, suggesting that early identification of this metabolic exhaustion phenotype could guide more precise resuscitation strategies to improve patient survival.

Abbreviations

ACAG

Albumin-Corrected Anion Gap

AMI

Acute Myocardial Infarction

AUC

Area Under the Curve

CI

Confidence Interval

CKD

Chronic Kidney Disease

DKA

Diabetic Ketoacidosis

eICU-CRD

eICU Collaborative Research Database

ESRD

End-Stage Renal Disease

GCS

Glasgow Coma Scale

HHS

Hyperglycemic Hyperosmolar States

HR

Hazard Ratio

ICU

Intensive Care Unit

IDI

Integrated Discrimination Improvement

IPTW

Inverse Probability of Treatment Weighting

MIMIC-IV

Medical Information Mart for Intensive Care IV

NRI

Net Reclassification Improvement

OR

Odds Ratio

PCI

Percutaneous Coronary Intervention

SHAP

SHapley Additive exPlanations

SMD

Standardized Mean Difference

SOFA

Sequential Organ Failure Assessment

TyG Index

Triglyceride-Glucose Index

XGBoost

Extreme Gradient Boosting

Authors’ contributions

XZ, QX and SL conceived and designed the study. XZ and QX performed the data collection, formal analysis, and wrote the original draft of the manuscript. WA contributed to the interpretation of data and performed the English grammatical correction and polishing of the manuscript. CY, HW, ZM, and MH participated in data curation and validation. SL, YB and CC provided supervision, resources, and critical revision of the manuscript for important intellectual content. All authors read and approved the final manuscript.

Funding

This work was supported by Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (Grant No. KYXZ012025090153).

Data availability

The eICU-CRD and MIMIC-IV datasets are publicly available in the PhysioNet repository (https://physionet.org/). Access requires credentialing and ethics training. The data from Tongji Hospital that support the findings of this study are not publicly available due to patient privacy regulations but are available from the corresponding author (SL) upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki. The eICU-CRD and MIMIC-IV databases were approved by the Institutional Review Boards (IRB) of the Massachusetts Institute of Technology (MIT) and Beth Israel Deaconess Medical Center (BIDMC). One of the authors (XZ) completed the required training and obtained authorization to access the databases (Record ID: 74166104). The requirement for informed consent was waived for the US cohorts due to the retrospective design and the use of de-identified data. For the validation cohort at Tongji Hospital, the study was approved by the Ethics Committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (the ethics approval number: TJ-IRB202601113). The requirement for informed consent was waived by the local Ethics Committee due to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.