Abstract

Background:

Early risk stratification is crucial for improving outcomes in critically ill patients with acute myocardial infarction (AMI). Albumin-derived composite indices hold promise as convenient and effective predictive tools, but their relative efficacy and clinical utility remain unclear.

Methods:

This two-cohort retrospective analysis utilized a derivation cohort from the MIMIC-IV public database and an external validation cohort from the ICU of Guizhou Medical University Affiliated Hospital. Six albumin-derived composite indices were evaluated. Statistical analyses employed Cox proportional hazards regression models to assess their association with mortality. Predictive performance was compared using the area under the receiver operating characteristic curve (AUC) and Delong’s test. A multivariate risk prediction model was developed based on key prognostic variables selected by multiple machine learning algorithms.

Results:

The study included 4,850 critically ill AMI patients (4,210 in the derivation cohort, 640 in the validation cohort). Multivariable-adjusted analysis identified the red cell distribution width to Albumin Ratio (RAR), Urea nitrogen to Albumin Ratio (UAR), and Lactate Dehydrogenase to Albumin Ratio (LDAR) as independent predictors of 28-day ICU mortality. Among these, LDAR demonstrated the strongest predictive ability, with an AUC of 0.702 in the derivation cohort, a finding robustly validated externally (AUC = 0.703). Subgroup analysis indicated consistent predictive value across most populations but revealed a significant interaction with hyperlipidemia. Incorporating LDAR into traditional critical illness scores (e.g., APACHE II, SOFA) significantly improved their predictive discrimination (all Delong’s test p < 0.05). A comprehensive model integrating 7 key variables (including LDAR, urea nitrogen, and lactate) selected by machine learning showed good and robust discriminative performance in both internal and external validation (AUCs of 0.767 and 0.735, respectively), significantly outperforming five traditional risk scores (all Delong’s test p < 0.05).

Conclusion:

Among the six albumin-derived composite indices, LDAR offers the best independent and incremental predictive value for 28-day ICU mortality in critically ill AMI patients. Its interaction with hyperlipidemia suggests potential for targeted risk stratification. The machine learning model incorporating LDAR and other variables demonstrates robust performance, providing a promising tool for the early clinical identification of high-risk patients.

Background

Cardiovascular disease is the leading cause of death worldwide, with acute myocardial infarction (AMI) representing a critical event associated with high mortality (1, 2). Despite significant advances in reperfusion and pharmacological therapies, which have contributed to an overall decline in AMI mortality, the prognosis for critically ill patients—particularly those admitted to the ICU, remains poor. For instance, cardiogenic shock following AMI still carries a 28-day mortality rate of approximately 40% (3, 4). The pathological core of AMI involves irreversible myocardial cell necrosis due to acute coronary occlusion, triggering a severe inflammatory response and ventricular remodeling (5, 6). In this context, the early and accurate identification of high-risk patients is crucial for improving outcomes. However, existing clinical prediction tools are often limited in practical use due to their complexity and inconvenience (7–9). Consequently, the search for and validation of novel biomarkers or composite indices that are both convenient and effective for predicting outcomes in critically ill AMI patients has become an important focus of clinical research.

Serum albumin, the most abundant circulating protein in the human body, plays key physiological roles in maintaining colloid osmotic pressure and exerting anti-inflammatory and antioxidant effects. It is therefore widely used as an important biomarker for assessing nutritional status and systemic inflammation (10–12). Clinical studies have consistently shown that hypoalbuminemia is associated with adverse outcomes in various conditions, including heart failure, acute pulmonary embolism, diabetic nephropathy, and ICU stays (13–15). However, as a single indicator, serum albumin levels can be influenced by numerous factors such as liver function, hydration status, acute stress response, and clinical interventions (e.g., fluid administration, nutritional support), which limits its predictive stability and specificity (13–15). To overcome the limitations of a single marker, recent research has focused on combining albumin with other complementary laboratory parameters that reflect related pathophysiological processes. These include red cell distribution width (RDW), which indicates red blood cell heterogeneity; lactate, representing tissue perfusion and metabolic state; anion gap, reflecting acid–base balance; and lactate dehydrogenase (LDH), suggesting cellular injury (16–21). Composite scores integrating such markers can provide a more comprehensive profile of a patient’s inflammatory, nutritional, and metabolic stress state. Such indices have demonstrated superior predictive performance compared to albumin alone for prognosis in various critical conditions, including heart failure, myocardial infarction, and sepsis (16–21).

Nevertheless, the existing evidence is largely derived from retrospective data, and there is a lack of direct, systematic comparison among the various albumin-derived composite indices. Their relative predictive merits and optimal clinical application scenarios remain unclear. Given the substantial resource and time investments required for prospective cohort studies, it is imperative to first use retrospective data to identify which integrated index offers the best combination of simplicity and predictive power. This approach can prioritize candidates for subsequent high-quality research and optimize resource allocation. Therefore, this study aims to systematically evaluate and compare the performance of several albumin-derived inflammatory-nutritional composite indices in predicting 28-day ICU mortality among critically ill myocardial infarction patients, with the goal of identifying the optimal predictive tool.

MethodsData sources

This study utilized two retrospective cohorts. The derivation cohort was sourced from the publicly available Medical Information Mart for Intensive Care IV (MIMIC-IV) database, version 2.2 (22). This database contains information for 196,527 adults admitted to the Beth Israel Deaconess Medical Center between 2008 and 2019. Database access was granted under an approved protocol. The Institutional Review Boards of the Massachusetts Institute of Technology approved the use of the MIMIC-IV database for research and waived the requirement for informed consent due to the retrospective nature of the study and the use of de-identified data. The external validation cohort included AMI patients treated in the Intensive Care Unit of Guizhou Medical University Affiliated Hospital. Following approval from the hospital’s Ethics Committee, this cohort was enrolled between January 2015 and January 2025. The Committee waived the requirement for individual informed consent due to the retrospective study design.

Study population

For data extraction from the MIMIC-IV database, we used PostgreSQL (v13.7.2) and Navicat Premium (v16) to execute structured queries. The study population consisted of all adult patients (age ≥18 years) in the database admitted to the ICU with a diagnosis of AMI, identified using ICD-9 code 410 and ICD-10 codes I21 and I22. To ensure data completeness and analytical reliability, we applied the following exclusion criteria: (1) patients who died within 24 h of ICU admission, as their clinical records might be incomplete; (2) those with multiple ICU admissions, where only the first ICU admission for AMI was included to avoid data duplication; (3) patients lacking key laboratory measurements obtained on the first day of ICU admission, which were essential for the analyses.

Variable extraction and processing

Data extraction was performed on the PostgreSQL (v13.7.2) and Navicat Premium (v16.0) platforms using structured query language (SQL). Extracted variables were grouped into six categories: (1) demographic characteristics, among them, race is divided into yes and no based on whether they are white or not, (2) comorbidities, (3) vital signs, (4) laboratory parameters, (5) illness severity scores, and (6) administered treatments. For variables with missing data (23), imputation was performed only if the missing rate was below 30%. We used the “mice” package (v3.16.0) in R, employing a random forest model to conduct multiple imputation. The imputation model included all variables relevant to the primary analysis: the continuous variables to be imputed, completely observed exposure variables (RDW, lactate, total bilirubin, anion gap, urea nitrogen, LDH, and albumin), the completely observed outcome variable (28-day ICU mortality), and other complete categorical and continuous covariates.

Definition of exposure and clinical outcome

Drawing on previous literature, this study selected six albumin-derived composite indices to comprehensively assess inflammatory activation, nutritional status, and metabolic stress: Red cell distribution width to Albumin Ratio (RAR), Lactate to Albumin Ratio (LAR), Anion Gap to Albumin Ratio (AGAR), Total bilirubin to Albumin Ratio (TAR), Urea nitrogen to Albumin Ratio (UAR), and Lactate Dehydrogenase to Albumin Ratio (LDAR) (16–21). These indices have been previously shown to be associated with adverse outcomes in AMI (16–21). To meet statistical assumptions and enhance clinical interpretability, given that the numerical range of LDAR is notably higher than that of the other five indicators, we applied a logarithmic transformation to LDAR for all subsequent analyses, and LDAR is used to denote the transformed measure throughout the following analyses. The primary clinical endpoint was defined as all-cause death within 28 days of ICU admission for AMI.

Correlation analysis of albumin-derived biomarkers with 28-day ICU mortality in AMI

Following preliminary hypothesis testing, this study used Cox proportional hazards regression models to analyze the association between the six albumin-derived biomarkers and the clinical endpoint. To control for potential confounding factors, three progressively adjusted regression models were constructed sequentially: Model 1 was an unadjusted crude model; Model 2 adjusted for demographic characteristics (age, sex, race, weight) and major comorbidities; Model 3 further incorporated clinically relevant variables that showed significant differences between survivors and non-survivors, including illness severity scores, key laboratory indicators, and treatment measures. To ensure model stability, multicollinearity diagnosis was performed for all variables included in Model 3 using the Variance Inflation Factor (VIF), and variables with a VIF > 5 were removed. Finally, to explore potential non-linear relationships between key indices (e.g., RAR) and the endpoint, restricted cubic splines (RCS) were fitted, with knots placed at the 5th, 35th, 65th, and 95th percentiles.

Comparison of predictive performance of albumin-derived biomarkers

The predictive performance of each biomarker for the adverse outcome was assessed by plotting Receiver Operating Characteristic (ROC) curves and calculating the Area under the Curve (AUC). Delong’s test was used to compare the differences in predictive ability among the albumin-derived biomarkers to determine if improvements were statistically significant.

Subgroup and interaction analysis

To verify the stability of the relationship between the best-performing albumin-derived biomarker (LDAR) and the adverse outcome, prespecified subgroup analyses and interaction tests were conducted. Subgroups were defined based on key characteristics such as age, sex, race, and baseline comorbidities. The consistency of the LDAR effect across subgroups was assessed by testing the statistical significance of interaction terms (p for interaction). A non-significant interaction (typically p > 0.05) would suggest stability across different populations, while a significant interaction would indicate heterogeneity. All analyses were performed according to a predefined protocol to control bias from multiple comparisons.

Incremental value of LDAR

LDAR was incorporated into five traditional critical illness scoring systems to construct multivariable Cox proportional hazards models. A composite risk score was calculated for each model using the formula (β₁ × variable₁) + (β₂ × variable₂), based on the regression coefficients. Similarly, the predictive performance of each model for the adverse outcome was evaluated by plotting ROC curves and calculating the AUC. Delong’s test was used to compare the predictive ability of models before and after adding LDAR to determine if the improvement was statistically significant.

Screening of important prognostic features

Within the internal cohort, we first randomly split the data into a training set and a validation set in a 7:3 ratio. During the training phase, four machine learning algorithms were jointly applied to screen for serum laboratory variables closely associated with ICU mortality: the Boruta algorithm (confidence level p < 0.01, maximum iterations = 100, with Bonferroni correction), Random Forest (100 trees, maximum depth = 3), Lasso regression (LogLambda min = −5.243), and Gradient Boosting Machine based on the Cox loss function (learning rate = 0.1, boosting rounds = 100). Variables identified as important by all four algorithms were ultimately determined to be the core prognostic features for this cohort.

Risk prediction modeling and validation

The internal cohort (MIMIC-IV, n = 4,210) was randomly split into a training set (70%, n = 2,947) and an internal validation set (30%, n = 1,263). Feature selection using the four machine learning algorithms (Boruta, Random Forest, Gradient Boosting Machine, and Lasso regression) was performed exclusively on the training set. The seven variables selected by all four algorithms (LDAR, urea nitrogen, lactate, alkaline phosphatase, RDW, glucose, and total bilirubin) were used to build a multivariable Cox proportional hazards regression model. The regression coefficients (β) estimated from the training set were fixed to create a locked prediction model. This locked model was then applied without any modification to the internal validation set and the external validation cohort (n = 640) to evaluate its discriminative ability (AUC) and calibration (calibration plots and Brier score). All analyses followed the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines.

Statistical analysis

Continuous variables are presented as mean ± standard deviation. Comparisons between groups were performed using Student’s t-test or Analysis of Variance (ANOVA), depending on data normality and homogeneity of variance. Categorical variables are presented as numbers (percentages), with comparisons between groups made using Pearson’s chi-square test or Fisher’s exact test, as appropriate. All statistical analyses were performed using R software (version 4.5.1). A two-sided p-value < 0.05 was considered statistically significant.

ResultsBaseline characteristics of the included patients

According to the study’s inclusion criteria, a total of 4,210 critically ill AMI patients were included in the analysis. Baseline characteristics stratified by survival status are presented in Table 1. Compared to survivors, non-survivors were older (74.7 ± 11.2 years vs. 70.9 ± 11.9 years) and had a higher burden of comorbidities, including a greater prevalence of acute kidney injury (AKI), chronic kidney disease (CKD), and heart failure. Illness severity scores (SOFA, APS III, SAPS II, OASIS, APACHE II) were all significantly higher in non-survivors. Regarding vital signs, non-survivors had a higher heart rate, faster respiratory rate, and lower oxygen saturation. Laboratory tests revealed a general deterioration in inflammatory and metabolic markers among non-survivors: levels of RDW, white blood cell count (WBC), lactate, blood glucose, anion gap, potassium, and phosphorus were higher, while albumin, calcium, hemoglobin, red blood cell count, and chloride were lower. Coagulation, liver, and kidney functions were also significantly impaired, as evidenced by prolonged INR, PT, and PTT, and elevated levels of ALT, AST, total bilirubin, creatinine, urea nitrogen, LDH, and ALP. In terms of treatment interventions, non-survivors had a significantly higher rate of continuous renal replacement therapy (CRRT) (24.7% vs. 7.45%), and more frequent use of vasopressors and glucocorticoids, but a lower proportion received antihypertensive therapy. Although the proportion receiving mechanical ventilation was slightly lower, the incidence of shock was higher. Regarding albumin-derived composite indices, all values were significantly higher in the non-survivor group: RAR (6.02 vs. 4.98), AGAR (6.05 vs. 4.83), LAR (1.11 vs. 0.76), UAR (16.1 vs. 10.3), TAR (0.64 vs. 0.36), LDAR (156 vs. 322), and log2 (LDAR) (6.79 vs. 7.54) (all p < 0.001). In summary, non-survivors exhibited greater illness severity, worse physiological reserve, and more intensive treatment requirements at baseline, all closely associated with adverse outcomes.

VariableUnitAll (N = 4,210)Survivor (N = 3,396)Non-survivor (N = 814)p valueDemographicsAgeYears71.6 (11.9)70.9 (11.9)74.7 (11.2)<0.001Gender (Male)n (%)2,801 (66.5%)2,254 (66.4%)547 (67.2%)0.684Race (White)n (%)2,691 (63.9%)2,203 (64.9%)488 (60.0%)0.010Weightkg84.7 (22.5)85.1 (22.4)82.8 (22.8)0.009ComorbiditiesHTNn (%)1,531 (36.4%)1,272 (37.5%)259 (31.8%)0.003AKIn (%)2,365 (56.2%)1732 (51.0%)633 (77.8%)<0.001CKDn (%)1,313 (31.2%)1,021 (30.1%)292 (35.9%)0.002DMn (%)1736 (41.2%)1,409 (41.5%)327 (40.2%)0.518HLDn (%)2,263 (53.8%)1873 (55.2%)390 (47.9%)<0.001HFn (%)2,295 (54.5%)1818 (53.5%)477 (58.6%)0.010Severity scoresSOFAScore6.36 (3.60)5.98 (3.42)7.96 (3.88)<0.001APS IIIScore51.7 (21.6)48.5 (19.8)65.0 (23.7)<0.001SAPS IIScore42.9 (13.8)41.0 (12.8)50.9 (14.7)<0.001OASISScore34.3 (8.61)33.3 (8.20)38.4 (9.02)<0.001APACHE IIScore21.0 (7.31)20.1 (6.99)24.5 (7.61)<0.001Vital signsHRBeats/min88.0 (20.1)87.0 (19.7)92.0 (21.0)<0.001RRinsp/min19.3 (6.28)18.9 (6.23)21.0 (6.24)<0.001NBPSmmHg119 (24.9)119 (24.7)118 (25.7)0.271NBPDmmHg67.5 (18.6)67.6 (18.4)67.4 (19.3)0.791NBPMmmHg78.0 [68.0;91.0]79.0 [68.0;90.0]78.0 [67.0;91.0]0.837SpO₂%96.8 (14.1)97.1 (15.4)95.7 (5.24)<0.001Laboratory parametersHCT%32.1 (7.08)32.2 (7.02)31.7 (7.30)0.128Hbg/dL10.5 (2.37)10.5 (2.36)10.2 (2.41)0.003PLTK/μL197 (102)196 (97.3)198 (118)0.700RDW%15.2 (2.37)14.9 (2.19)16.1 (2.82)<0.001RBCm/μL3.51 (0.81)3.53 (0.81)3.44 (0.83)0.006WBCK/μL13.6 (8.33)13.3 (7.62)14.9 (10.7)<0.001ALBg/dL3.08 (0.59)3.14 (0.58)2.84 (0.61)<0.001AGmEq/L15.0 (4.75)14.6 (4.57)16.5 (5.17)<0.001Camg/dL8.32 (0.81)8.33 (0.78)8.26 (0.90)0.032ClmEq/L104 (6.76)104 (6.59)102 (7.29)<0.001Glumg/dL163 (87.9)159 (85.8)179 (94.7)<0.001KmEq/L4.35 (0.79)4.34 (0.77)4.43 (0.84)0.002NamEq/L138 (5.36)138 (5.13)138 (6.23)0.611Lacmmol/L1.80 [1.30;2.80]1.80 [1.20;2.62]2.10 [1.40;3.58]<0.001PCO₂mmHg42.4 (10.7)42.1 (10.1)43.4 (12.8)0.008pHpH units7.36 (0.10)7.36 (0.09)7.33 (0.11)<0.001PO₂mmHg158 (132)169 (136)113 (104)<0.001INRRatio1.58 (0.94)1.54 (0.89)1.76 (1.11)<0.001PTSeconds17.2 (9.57)16.7 (8.85)19.1 (11.9)<0.001PTTSeconds44.7 (31.0)43.8 (30.3)48.1 (33.7)0.001ALTIU/L28.0 [16.0;66.0]27.0 [15.0;59.0]40.0 [20.0;112]<0.001ASTIU/L46.0 [26.0;114]42.0 [26.0;97.0]68.5 [33.0;200]<0.001TBmg/dL0.60 [0.40;1.10]0.60 [0.40;1.00]0.80 [0.40;1.40]<0.001CREmg/dL1.76 (1.64)1.67 (1.59)2.13 (1.81)<0.001UREmg/dL33.1 (24.9)30.6 (22.8)43.6 (30.1)<0.001LDHU/L317 [238;481]301 [230;435]436 [293;717]<0.001ALPU/L78.0 [58.0;113]75.0 [56.0;107]91.5 [67.0;136]<0.001Mgmg/dL2.12 (0.53)2.12 (0.53)2.11 (0.55)0.646PHOSmg/dL3.99 (1.50)3.87 (1.41)4.49 (1.74)<0.001Albumin-derived indicesRARRatio5.18 (1.62)4.98 (1.44)6.02 (2.04)<0.001AGARRatio5.06 (2.02)4.83 (1.88)6.05 (2.29)<0.001LARRatio0.61 [0.40;0.95]0.57 [0.39;0.89]0.76 [0.50;1.31]<0.001UARRatio11.4 (9.64)10.3 (8.61)16.1 (12.0)<0.001TARRatio0.21 [0.13;0.37]0.20 [0.13;0.34]0.26 [0.16;0.52]<0.001LDARRatio106 [75.8;168]97.9 [72.6;149]158 [104;258]<0.001Log₂(LDAR)Ratio6.94 (1.06)6.79 (0.96)7.54 (1.24)<0.001TreatmentsCRRTn (%)454 (10.8%)253 (7.45%)201 (24.7%)<0.001Ventilationn (%)3,830 (91.0%)3,105 (91.4%)725 (89.1%)0.041SAn (%)3,265 (77.6%)2,584 (76.1%)681 (83.7%)<0.001VPn (%)3,280 (77.9%)2,574 (75.8%)706 (86.7%)<0.001GCn (%)1,168 (27.7%)851 (25.1%)317 (38.9%)<0.001AHTn (%)3,723 (88.4%)3,068 (90.3%)655 (80.5%)<0.001

Baseline characteristics of critically ill AMI patients in the internal derivation cohort, stratified by 28-day ICU survival status.

HTN, Hypertension; AKI, Acute Kidney Injury; CKD, Chronic Kidney Disease; DM, Diabetes Mellitus; HLD, Hyperlipidemia; HF, Heart Failure; SOFA, Sequential Organ Failure Assessment; APS III, Acute Physiology Score III; SAPS II, Simplified Acute Physiology Score II; OASIS, Oxford Acute Severity of Illness Score; APACHE II, Acute Physiology and Chronic Health Evaluation II; HR, Heart Rate; NBPS, Non-Invasive Blood Pressure (Systolic); NBPD, Non-Invasive Blood Pressure (Diastolic); NBPM, Non-Invasive Blood Pressure (Mean); RR, Respiratory Rate; SpO₂, Oxygen Saturation; HCT, Hematocrit; Hb, Hemoglobin; PLT, Platelet Count; RDW, Red Cell Distribution Width; RBC, Red Blood Cell Count; WBC, White Blood Cell Count; ALB, Albumin; AG, Anion Gap; Ca, Calcium; Cl, Chloride; Glu, Glucose; K, Potassium; Na, Sodium; Lac, Lactate; PCO₂, Partial Pressure of Carbon Dioxide; pH, Potential of Hydrogen; PO₂, Partial Pressure of Oxygen; INR, International Normalized Ratio; PT, Prothrombin Time; PTT, Partial Thromboplastin Time; ALT, Alanine Aminotransferase; AST, Aspartate Aminotransferase; TB, Total Bilirubin; CRE, Creatinine; URE, Urea Nitrogen; LDH, Lactate Dehydrogenase; ALP, Alkaline Phosphatase; Mg, Magnesium; PHOS, Phosphorus; RAR, RDW to Albumin Ratio; AGAR, Anion Gap to Albumin Ratio; LAR, Lactate to Albumin Ratio; UAR, Urea Nitrogen to Albumin Ratio; TAR, Total Bilirubin to Albumin Ratio; LDAR, Lactate Dehydrogenase to Albumin Ratio; Log₂(LDAR), Log₂-transformed Lactate Dehydrogenase to Albumin Ratio; CRRT, Continuous Renal Replacement Therapy; SA, Sedative Administration; VP, Vasopressor; GC, Glucocorticoids; AHT, Antihypertensive Therapy.

Association between albumin-derived composite indices and 28-day ICU mortality in AMI patients

The results of the Cox regression analyses are shown in Table 2. In the unadjusted Model 1, all indices showed a significant positive association with mortality risk: RAR (HR = 1.20, 95% CI: 1.17–1.24, p < 0.001), AGAR (HR = 1.13, 95% CI: 1.10–1.15, p < 0.001), LAR (HR = 1.20, 95% CI: 1.14–1.27, p < 0.001), TAR (HR = 1.13, 95% CI: 1.09–1.18, p < 0.001), UAR (HR = 1.03, 95% CI: 1.02–1.03, p < 0.001), and LDAR (HR = 1.32, 95% CI: 1.25–1.39, p < 0.001). These associations remained significant after adjusting for demographics and comorbidities in Model 2. However, in the fully adjusted Model 3, which accounted for inter-group differences, only RAR (HR = 1.14, 95% CI: 1.10–1.19, p < 0.001), UAR (HR = 1.01, 95% CI: 1.00–1.02, p = 0.012), and LDAR (HR = 1.39, 95% CI: 1.29–1.50, p < 0.001) remained independently associated with ICU mortality risk. The associations for AGAR, LAR, and TAR were no longer significant. These results indicate that RAR, UAR, and LDAR have independent predictive value for mortality risk in critically ill patients, with log-transformed LDAR demonstrating a particularly strong association after multivariable adjustment.

CharacteristicModel 1Model 2Model 3HR95% CIp-valueHR95% CIp-valueHR95% CIp-valueRAR1.201.17, 1.24<0.0011.191.15, 1.23<0.0011.141.10, 1.19<0.001AGAR1.131.10, 1.15<0.0011.121.09, 1.15<0.0011.02