Abstract
<title>Abstract</title> <p>Background Sepsis secondary to methicillin-resistant Staphylococcus aureus (MRSA) infection is associated with considerable morbidity and mortality. Early stratification of patients at high risk remains essential for optimizing clinical outcomes. The MIMIC-IV database was used to develop and validate a machine-learning-based framework for early prediction of sepsis in patients with MRSA infection. Methods This retrospective analysis included 1,026 adult ICU admissions with confirmed MRSA. To isolate robust predictors, we employed a dual-stage feature selection strategy integrating the Boruta algorithm with Least Absolute Shrinkage and Selection Operator (LASSO) regression. Eight distinct machine learning algorithms were developed and rigorously assessed via receiver operating characteristic (ROC) analysis, calibration plots, and decision curve analysis (DCA). The superior model was further interpreted using SHapley Additive exPlanations (SHAP) and translated into a clinical nomogram. Results The incidence of sepsis in the study population was 28.6% (n = 293). Multivariable analysis pinpointed nine independent predictors: history of heart disease, administration of second-generation cephalosporins, norepinephrine, or meropenem, as well as elevated Acute Physiology Score III and respiratory rates were positive correlates of sepsis risk. Conversely, bacteremia, baseline bicarbonate levels, and systolic blood pressure exhibited inverse associations. Among the evaluated algorithms, Elastic Net Regression (ENET) yielded the most robust generalizability, achieving an AUC of 0.835 and 78.83% accuracy in the validation set. The ENET-derived nomogram demonstrated superior net clinical benefit compared to individual predictors. Conclusions The ENET-based model and nomogram use routine clinical data to identify MRSA-infected patients at high risk of sepsis, supporting early risk stratification and individualized treatment decisions.</p>