EXPRESS: Relation between Albumin-Corrected Anion Gap and In-Hospital Mortality in Patients with Traumatic Lung Injury: A Multicenter Retrospective Cohort Study and the Development of Machine Learning-Based Prediction Models

BACKGROUND: The anion gap is primarily utilized as an indicator for evaluating acid-base imbalances in critically ill patients. However, its accuracy is reduced in such patients due to low albumin levels. The albumin-corrected anion gap (ACAG) enhances the accuracy of assessing acid-base imbalances. Individuals with traumatic lung injury (TLI) in the intensive care unit (ICU) often have severe metabolic acidosis and hypoalbuminemia. Nevertheless, the association of ACAG with the prognosis of patients with TLI is still unknown. METHODS: Clinical data of individuals with TLI were acquired from the Medical Information Mart for Intensive Care (MIMIC)-IV-3.1 database and the eICU Collaborative Research Database (eICU-CRD), created by Philips Healthcare and the Massachusetts Institute of Technology. Data from the MIMIC-IV database were utilized as the training set to develop machine learning-based models. In contrast, data from the eICU-CRD were adopted for external validation of the established models. The primary outcome was in-hospital mortality. The association of ACAG with in-hospital mortality was evaluated using restricted cubic spline (RCS) models, Cox proportional hazards models, and Kaplan-Meier curves. The Boruta algorithm was adopted for the selection of feature variables. The predictive power of ACAG was evaluated, and prediction models were established utilizing machine learning algorithms. Model performance was verified with receiver operating characteristic (ROC) curves and decision curve analysis. RESULTS: A total of 239 and 467 individuals were incorporated from the MIMIC-IV database and the eICU-CRD, respectively. RCS curve analysis revealed a linear association of ACAG with in-hospital mortality. Elevated ACAG was substantially linked to a high risk of mortality in individuals with TLI (hazard ratio (HR) [95% confidence interval (CI)] = 1.115 [1.037-1.199]). The Boruta algorithm demonstrated that ACAG possessed higher feature importance. Prediction models established based on ACAG exhibited the optimal predictive performance. CONCLUSION: ACAG exhibited a linear association with in-hospital mortality among individuals with TLI, and a high ACAG was related to a markedly elevated risk of in-hospital death. Therefore, ACAG may serve as a potential predictor of adverse outcomes in individuals with TLI.