Abstract

Background: Klebsiella pneumoniae (KP) and Escherichia coli (E. coli) are common pathogens causing septicemia, which is associated with high mortality rates. Identifying risk factors and developing accurate predictive models for mortality in these patients are crucial for improving clinical outcomes. This study aimed to identify risk factors for mortality in patients with KP and E. coli septicemia and to develop and validate a predictive model using machine learning (ML) techniques.

Methods: A total of 468 participants diagnosed with KP or E. coli septicemia were included in this retrospective study conducted at the First People's Hospital of Jiande from 2010 to 2025. After excluding missing values, participants were randomly divided into training (70%) and validation (30%) cohorts. The Boruta algorithm was applied to the training cohort to identify significant risk factors for mortality. These variables were then used to construct a nomogram prediction model, and its performance was evaluated using receiver operating characteristic (ROC) curves, decision curve analysis (DCA) curves, and calibration curves. Additionally, the identified variables were incorporated into various ML models to compare their predictive performance.

Results: The CatBoost model demonstrated the highest area under the curve (AUC) value (AUC = 0.922, 95% CI: 0.862 - 0.983), indicating superior predictive accuracy for mortality compared to other models. The Boruta algorithm identified nine significant risk factors, including elevated or reduced white blood cell count, decreased red blood cell count, decreased platelet count, elevated hemoglobin level, carbapenem-resistant (CRE) bacterial infection, presence of shock, respiratory failure, coma, and leukemia. The nomogram prediction model achieved an AUC of 0.94 in the training set and 0.88 in the validation set, showing excellent discrimination and calibration.

Conclusions: This study identified key risk factors for mortality in patients with KP and E. coli septicemia and developed a highly accurate predictive model using the CatBoost algorithm. The nomogram model also demonstrated excellent predictive performance, which could be a valuable tool for clinical decision-making. Further validation in larger cohorts is warranted to confirm the generalizability of these findings.