Association between the endothelial activation and stress index and 28-day all-cause mortality in critically ill patients with chronic obstructive pulmonary disease: a retrospective cohort study and predictive model establishment based on machine learning.

Abstract

Background: Chronic obstructive pulmonary disease (COPD) remains a major global health burden and is currently the third leading cause of death worldwide. Acute exacerbations accelerate disease progression and contribute substantially to mortality, underscoring the urgent need for reliable prognostic biomarkers. The endothelial activation and stress index (EASIX), a composite indicator of endothelial dysfunction, has demonstrated prognostic utility across diverse critical illnesses. However, its association with clinical outcomes in critically ill patients with COPD has not been clearly established.

Methods: In this retrospective cohort study, data of critically ill patients with COPD were extracted from the Medical Information Mart for Intensive Care (MIMIC) database. Participants were stratified into tertiles based on EASIX values, and intergroup differences in clinical characteristics were analyzed. The relationship between EASIX and 28-day all-cause mortality was examined using Kaplan-Meier survival analysis, Cox proportional hazards regression, and restricted cubic spline modeling. The Boruta algorithm was applied to assess the relative importance of candidate predictors, and prognostic models were subsequently developed using six machine learning algorithms.

Results: A total of 4,590 patients met the inclusion criteria. The incidence of 28-day ICU mortality increased progressively across higher EASIX tertiles (p < 0.001). EASIX was independently associated with 28-day ICU all-cause mortality, with both unadjusted and fully adjusted Cox models confirming this relationship (unadjusted HR = 1.21, p < 0.001; adjusted HR = 1.082, p < 0.001). Subgroup analyses demonstrated that the association between elevated EASIX and mortality risk remained consistent across demographic and comorbidity categories (p for interaction > 0.05 for all). The Boruta algorithm identified EASIX as one of the most important predictors of 28-day mortality. Among the six machine learning models evaluated, the XGBoost algorithm yielded the highest discriminative (AUC = 0.823), calibration and clinical application.

Conclusions: EASIX serves as an independent prognostic marker for 28-day all-cause mortality in critically ill COPD patients. Furthermore, the EASIX-based machine learning model demonstrated strong predictive accuracy, supporting its potential as a valuable clinical tool or early risk stratification and decision-making in intensive care settings.