Abstract
<title>Abstract</title> <p>Background Concomitant heart failure (HF) and acute kidney injury (AKI) in critically ill patients is associated with high mortality, underscoring the need for accurate early risk stratification. The Endothelial Activation and Stress Index (EASIX), derived from three routine laboratory markers, has shown prognostic value across several clinical contexts. Its specific predictive capacity for critically ill patients with concurrent HF and AKI, however, remains undefined. Methods This retrospective cohort study analyzed adult ICU patients from the MIMIC-IV database (n = 4,000, development cohort) and the MIMIC-III-CareVue database (n = 700, temporal validation cohort). We assessed the association between admission log2-transformed EASIX (log2-EASIX) and 28-, 90-, and 180-day all-cause mortality using multivariable Cox proportional hazards regression. Dose-response relationships were evaluated with restricted cubic splines (RCS), and predictive performance was compared against established clinical scores (SOFA, OASIS, Charlson Comorbidity Index) via receiver operating characteristic (ROC) curve analysis. To validate the index's predictive importance, we further applied a machine learning pipeline that incorporated Boruta feature selection and SHAP-based interpretability. Results Elevated admission log2-EASIX demonstrated a robust, independent, and dose-dependent association with mortality across all follow-up periods. In the fully adjusted model, patients in the highest quartile (Q4) had significantly increased risks of 28-day (HR 2.37, 95% CI 1.95–2.87), 90-day (HR 2.06, 95% CI 1.75–2.42), and 180-day (HR 1.98, 95% CI 1.70–2.31) mortality compared with those in Q1. These findings were consistently replicated in the external MIMIC-III validation cohort. Restricted cubic spline analysis revealed a significant linear, monotonically increasing risk trajectory (P for non-linearity > 0.05). The discriminative performance of log2-EASIX (AUCs: 0.601–0.622) was highly competitive with that of multidimensional clinical indices. Among five machine learning algorithms, the Gradient Boosting Machine and Random Forest models achieved optimal discrimination (AUCs: 0.709–0.725), with Boruta and SHAP analyses consistently ranking log2-EASIX among the top predictive features. Conclusions Admission log2-EASIX is a parsimonious, robust, and independent predictor of short- and long-term mortality in critically ill patients with concomitant HF and AKI. Its predictive accuracy is comparable to that of complex clinical scoring systems, yet it provides a substantial practical advantage by requiring only three routinely available biomarkers. The routine calculation of EASIX could therefore improve early risk stratification and help guide personalized intensive care management.</p>