Red blood cell distribution width and 28 days mortality in sepsis-associated delirium: a multicenter study and machine learning analyses

Abstract

The specific prognostic relevance of red blood cell distribution width (RDW) regarding sepsis-associated delirium (SAD) remains poorly understood. Data were retrieved from the MIMIC-IV and eICU-CRD registries. Kaplan–Meier curves, multivariable Cox regression, restricted cubic splines (RCS), and competing-risk models evaluated the association between RDW and 28 days mortality. Machine-learning models were developed using Boruta- and LASSO-selected predictors and externally validated. Among 5413 patients with SAD from MIMIC-IV, the highest RDW quartile was associated with greater 28 days mortality than the lowest quartile (HR, 2.09; 95% CI 1.72–2.53; p < 0.001). RCS analysis indicated a nonlinear association (P for nonlinearity = 0.036), and competing-risk analysis yielded consistent findings. Among seven machine-learning algorithms, logistic regression demonstrated competitive performance and was selected for validation, yielding areas under the curve of 0.774 and 0.765 in the internal and external cohorts, respectively. Elevated baseline RDW was independently associated with higher 28 days mortality in patients with SAD. RDW was also retained as an important variable in multivariable machine-learning models. RDW may therefore serve as a readily available adjunctive marker to assist early risk stratification in critically ill patients with SAD.