The anion gap is a commonly used measure to evaluate acid–base disturbances in critically ill patients, but its accuracy is affected by hypoalbuminemia. The albumin-corrected anion gap (ACAG) adjusts the anion gap for low albumin and can more accurately reflect acid–base imbalance in patients with reduced serum albumin. Individuals with traumatic lung injury (TLI) admitted to intensive care units frequently present with severe metabolic acidosis and hypoalbuminemia, raising the question of whether ACAG provides prognostic information in this population.
This study examined the relationship between ACAG and in-hospital mortality among patients with TLI and tested whether ACAG contributes meaningfully to prediction models developed using machine learning methods.
Clinical data were obtained from two publicly available critical care repositories. The Medical Information Mart for Intensive Care (MIMIC)-IV-3.1 database served as the training set, and the eICU Collaborative Research Database (eICU-CRD) was used for external validation. The analysis focused on patients identified with traumatic lung injury; the abstract reports inclusion of 239 patients from MIMIC-IV and 467 patients from eICU-CRD.
The primary outcome was in-hospital mortality. Analytical methods included restricted cubic spline (RCS) models to explore dose–response relationships, Cox proportional hazards models to quantify associations with survival, and Kaplan–Meier curves for time-to-event visualization. For feature selection in predictive modeling, the Boruta algorithm was employed. Machine learning–based prediction models were trained on the MIMIC-IV cohort and externally validated on the eICU-CRD cohort. Model performance was evaluated using receiver operating characteristic (ROC) curves and decision curve analysis.
A total of 239 patients from MIMIC-IV and 467 patients from the eICU-CRD were included. RCS analysis indicated a linear association between ACAG and in-hospital mortality. In Cox proportional hazards models, elevated ACAG was significantly associated with higher risk of in-hospital death; the abstract reports a hazard ratio (HR) of 1.115 with a 95% confidence interval (CI) of 1.037–1.199. The Boruta algorithm identified ACAG as a variable with higher feature importance among the candidate predictors considered.
Prediction models that included ACAG demonstrated improved predictive performance according to ROC analysis, and decision curve analysis supported clinical utility of models incorporating ACAG. Specific model metrics (for example, area under the ROC curve values, calibration statistics, or net benefit thresholds) were not provided in the abstract.
Feature selection used the Boruta wrapper algorithm to prioritize variables for modeling, and ACAG achieved high importance in that process. Machine learning models were built on the MIMIC-IV training set and externally validated on the eICU-CRD dataset. The abstract states that models based on ACAG achieved optimal predictive performance compared with alternatives; model types, tuning procedures, and the full set of input features were not described in the abstract.
Model performance assessment relied on ROC curves and decision curve analysis, indicating discrimination and potential clinical benefit, respectively. Exact numerical performance results and comparisons among model types were not reported in the source abstract.
The study provides evidence that a higher albumin-corrected anion gap is linearly associated with increased in-hospital mortality in patients with traumatic lung injury admitted to the ICU. Because hypoalbuminemia is common in critically ill patients, correcting the anion gap for albumin may yield a more accurate marker of acid–base derangement and risk stratification in TLI.
The identification of ACAG as a high-importance feature and its contribution to machine learning–based prediction models suggest that ACAG could be considered for incorporation into clinical risk assessment tools for patients with TLI. However, implementation would require knowledge of the full model specifications and prospective validation in clinical settings.
The abstract does not report several details that are relevant to interpretation and implementation: specific inclusion and exclusion criteria for TLI cases, definitions or thresholds used to calculate ACAG, other covariates included in Cox models, the exact machine learning algorithms used, model hyperparameters, performance metrics (numerical AUCs, sensitivity, specificity), calibration, and subgroup analyses. These details were not provided in the PubMed abstract and would need to be obtained from the full text for comprehensive evaluation.
In this multicenter retrospective analysis using MIMIC-IV and eICU-CRD data, ACAG showed a linear relationship with in-hospital mortality in patients with traumatic lung injury, and higher ACAG values were associated with an increased hazard of death (HR 1.115, 95% CI 1.037–1.199). The Boruta algorithm rated ACAG as an important predictive feature, and machine learning models incorporating ACAG achieved superior predictive performance in the reported analyses. The authors conclude that ACAG may serve as a potential predictor of adverse outcomes in TLI, while acknowledging that additional details and full-text data are necessary to guide clinical implementation and further validation.