Development of a Postoperative Delirium Risk Prediction Model for Colorectal Cancer Using XGBoost and SHAP: A Retrospective Cohort Analysis

Background and Objective Postoperative delirium (POD) is a frequent and serious neuropsychiatric complication in colorectal cancer (CRC) patients, associated with prolonged hospitalization, higher mortality, and poor recovery. Existing predictive tools often lack interpretability, limiting individualized risk assessment and timely intervention. Methods In this single-center retrospective cohort study, 546 CRC patients undergoing surgery were analyzed. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation. Four predictive models—logistic regression, random forest, support vector machine (SVM), and extreme gradient boosting (XGBoost)—were developed and compared using the area under the curve (AUC), Brier score, and decision curve analysis (DCA). The best-performing model (XGBoost) was further interpreted using SHapley Additive Explanations (SHAP). Results POD occurred in 95 patients (17.39%). Among the tested models, the XGBoost model demonstrated favorable overall predictive performance (AUC = 0.862; 95% CI: 0.808–0.917; Brier score = 0.084), with more favorable calibration and promising clinical utility compared with the other approaches. SHAP analysis identified neutrophil-to-lymphocyte ratio (NLR), C-reactive protein-to-albumin ratio (CAR), diabetes, type of surgery, and postoperative ICU admission as key predictors, with contributions quantified at both the population and individual levels. Conclusion This study developed an interpretable machine learning model with favorable predictive performance for predicting postoperative delirium (POD) in colorectal cancer patients using XGBoost with SHAP analysis. The model showed better discrimination, calibration, and clinical utility than traditional approaches, suggesting its potential to support early identification of high-risk patients. SHAP interpretation confirmed the critical role of inflammation-related markers, particularly NLR and CAR, providing preliminary evidence to inform future clinical application and personalized perioperative management.