The prognostic value of the early neutrophil-to-lymphocyte ratio for 28-day mortality in sepsis patients: A machine learning-based investigation of the MIMIC database

Background: The neutrophil-to-lymphocyte ratio (NLR) has shown inconsistent prognostic value in individuals with sepsis. This study aimed to clarify its ability to predict 28-day mortality via a machine learning-based analysis of a large ICU database. Methods: This retrospective analysis employed data from the MIMIC-IV database (v3.1). The Boruta algorithm combined with XGBoost was used for two-stage feature selection. Patients were stratified by NLR quartiles into three groups (low: 14.70). This study defined 28-day mortality as the primary outcome. Associations between the NLR and mortality were evaluated by using multivariable logistic regression (progressively adjusted for demographic, clinical, and machine learning-derived features), along with restricted cubic splines. Sensitivity analyses included quantifying NLR feature importance via machine learning and performing subgroup analyses across clinical strata. Results: This cohort study included 4,376 patients with a 28-day mortality rate of 18.4%. Compared with the SOFA and SAPS II scores, the prediction performance of XGBoost was superior (ROC-AUC 0.875; 95% CI 0.854–0.896; PR-AUC 0.603). Although the NLR ranked 14th in SHAP-based feature importance, multivariable analysis confirmed its independent association with elevated mortality risk: 28-day (OR 1.16; 95% CI 1.06–1.27; p 45 years and those with SOFA scores ≤4 or SAPS II scores >29. Specifically, patients ≥65 years of age demonstrated a 17% increase in 28-day mortality risk (p = 0.019), and patients with a SOFA score ≤4 exhibited a greater than 20% elevated risk across all of the endpoints (p