Machine learning-based multicenter prediction of postoperative sepsis in emergency colon cancer: role of surgical approach and inflammatory markers.

Background: Postoperative sepsis occurs at a relatively high incidence following emergency colon cancer surgery, and early identification of high-risk patients is crucial for improving clinical outcomes. However, there is currently a lack of systematic risk prediction models specifically tailored for patients undergoing emergency colon cancer surgery. This study aimed to identify risk factors associated with postoperative sepsis and to develop a clinically applicable machine learning-based prediction model. Methods: This was a multicenter retrospective cohort study including patients who underwent emergency colon cancer surgery between January 2020 and January 2025. Perioperative variables were systematically collected, encompassing preoperative, intraoperative, and early postoperative data, including demographic characteristics, comorbidities, laboratory findings, and surgical features. Univariate and multivariate logistic regression analyses were first performed to identify independent risk factors. Subsequently, five machine learning algorithms-multilayer perceptron (MLP), random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN), and extreme gradient boosting (XGBoost)-were applied to evaluate feature importance and construct predictive models. Model performance was assessed using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), k-fold cross-validation, and external validation. SHapley Additive exPlanations (SHAP) were used to interpret the contribution of key features. Results: Surgical approach, intraoperative hypothermia, acidosis, hypoxemia, hypoalbuminemia, and postoperative inflammatory markers (procalcitonin [PCT] and neutrophil-to-lymphocyte ratio [NLR]) were identified as independent high-risk factors for postoperative sepsis in emergency colon cancer patients. All five machine learning models demonstrated good discriminative performance, with the MLP model achieving the best overall performance. It exhibited stable discrimination and low variability in both internal validation and external independent validation cohorts. SHAP analysis further confirmed the contribution of the identified risk factors to model predictions. Conclusion: The MLP-based risk prediction model developed in this study effectively identifies patients at high risk of postoperative sepsis following emergency colon cancer surgery. It provides a scientific basis for early perioperative intervention and individualized management, and offers valuable support for clinical risk stratification and optimization of therapeutic strategies.