Albumin-corrected anion gap as a predictor of 28-day mortality in acute respiratory distress syndrome: A machine learning-based retrospective study

Q Liu, Mengqi Zhang, Daoxin WangPLoS 2025 · DOI

TL;DR

This study reveals that a simple blood test measurement, the albumin-corrected anion gap (ACAG), is a strong independent predictor of 28-day mortality in patients with Acute Respiratory Distress Syndrome (ARDS). Machine learning models, particularly Random Forest, effectively integrated ACAG with other clinical data to predict mortality, suggesting ACAG could be a valuable tool for early risk stratification in critical care.

What Problem Does This Paper Address?

Acute Respiratory Distress Syndrome (ARDS) is a severe lung condition with high mortality rates, even with advanced critical care. Accurately predicting which patients are most at risk of dying within 28 days is crucial for optimizing treatment and resource allocation. While existing scoring systems like SOFA and APS III are used, they can be complex and their performance varies. Metabolic acidosis is a common complication in ARDS, and the anion gap (AG) is a standard measure for assessing it. However, low albumin levels, frequent in critically ill patients, can skew AG readings, potentially leading to underestimation of acidosis severity. The albumin-corrected anion gap (ACAG) was developed to address this limitation, but its prognostic value specifically in ARDS patients remained largely unexplored. This paper aims to fill that gap by investigating ACAG as a predictor of 28-day mortality in ARDS.

Methods

The researchers conducted a retrospective study using data from the MIMIC-IV database, a large ICU dataset. They identified patients meeting the criteria for ARDS based on the Berlin definition.

Key analytical approaches included:

  • Statistical Analysis: Kaplan-Meier survival analysis to compare mortality across different ACAG levels, restricted cubic spline (RCS) analysis to examine the non-linear relationship between ACAG and mortality, and multivariate Cox proportional hazards regression to assess the independent association of ACAG with mortality after adjusting for confounders.
  • Feature Selection: Two complementary methods, the Boruta algorithm and LASSO regression, were used to identify the most important predictive features for mortality.
  • Machine Learning: Six different machine learning algorithms (Logistic Regression, K-Nearest Neighbors, Random Forest, XGBoost, Multilayer Perceptron, Elastic Net) were employed to build predictive models for 28-day mortality. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC).
  • Interpretability: SHAP (SHapley Additive exPlanations) analysis was used to understand how individual features contribute to the machine learning models' predictions.

Datasets

The study utilized the MIMIC-IV database (version 2.2), which contains de-identified ICU admission data from Beth Israel Deaconess Medical Center between 2008 and 2019. This dataset includes over 65,000 ICU admissions. No new benchmark datasets were introduced.

Key Insights and Ideas

  • ACAG as a Superior Marker: The study strongly suggests that ACAG is a more sensitive indicator of mortality risk in ARDS than the conventional anion gap, particularly in the context of hypoalbuminemia common in these patients.
  • Machine Learning for Prognostication: The application of various machine learning models demonstrates their potential to integrate ACAG with other clinical variables for robust mortality prediction, offering a more nuanced approach than traditional scoring systems alone.
  • Simplicity and Accessibility: ACAG is a readily available and easily interpretable biomarker, requiring only standard laboratory measurements (anion gap and albumin), making it a practical tool for busy ICU settings.
  • Dose-Response Relationship: The findings indicate a clear dose-response relationship, where higher ACAG levels are associated with progressively higher mortality risk.

Key Results

  • Independent Predictor: Higher ACAG levels were independently associated with increased 28-day all-cause mortality in ARDS patients, even after adjusting for numerous confounders. For every 1 mmol/L increase in ACAG, the mortality risk increased by approximately 4%.
  • Quartile Analysis: Patients in the highest ACAG quartile (Q4) had a significantly higher risk of 28-day mortality compared to those in the lowest quartile (Q1).
  • Machine Learning Performance: The Random Forest (RF) model achieved the highest predictive performance with an AUC of 0.73 in the validation set. Other models also showed good performance, with AUCs ranging from 0.69 to 0.72.
  • Key Predictors: Feature selection methods and SHAP analysis identified ACAG as a significant predictor, alongside established markers like APS III score, SOFA score, white blood cell count, SpO2, and heart rate.
  • Overall Mortality: The cohort had an overall 28-day mortality rate of 21.3%.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Center Retrospective Study: The data comes from a single tertiary care hospital in the US, which may limit the generalizability of findings to other settings or patient populations.
  • No External Validation: The study did not include external validation in independent cohorts.
  • Residual Confounding: Despite adjustments, unmeasured variables could still influence the results.
  • Single Measurement: ACAG was assessed only at ICU admission; dynamic changes over time were not evaluated. Serial measurements could provide more prognostic information.
  • Causality: The study demonstrates association, not causation.

Future work could involve:

  • External Validation: Confirming these findings in diverse, independent cohorts.
  • Dynamic Monitoring: Investigating the prognostic value of serial ACAG measurements.
  • Tailored Models: Developing disease-specific risk prediction models that incorporate ACAG, especially for subgroups with significant interactions (e.g., myocardial infarction, malignancy).
  • Resource-Limited Settings: Exploring the implementation and utility of ACAG in healthcare settings with fewer resources.

Summary updated

Abstract

Higher ACAG levels are independently associated with increased 28-day all-cause mortality in patients with ARDS. ACAG is a promising predictor of short-term mortality and may guide risk stratification in clinical practice.
BibTeX
@inproceedings{liu2025albumin,
  author = {Q Liu and Mengqi Zhang and Daoxin Wang},
  title = {{Albumin-corrected anion gap as a predictor of 28-day mortality in acute respiratory distress syndrome: A machine learning-based retrospective study}},
  booktitle = {PLoS},
  year = {2025},
  doi = {10.1371/journal.pone.0336662},
}
Figure