Development and Validation of a Prediction Model for Postoperative Delirium

Background: Postoperative delirium (POD) is a common and serious postoperative complication. Early identification of high-risk patients could facilitate targeted interventions. Objective: To develop and validate a machine learning-based prediction model for POD using a multicenter retrospective cohort. Methods: We analyzed data from 3000 surgical patients, divided into training (n=1400), internal testing (n=600), and external validation (n=1000) cohorts. Eight machine learning algorithms, including Random Forest (RF), XGBoost (XGB), LightGBM (LGBM), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Logistic Regression (LR), Decision Tree (DT), and Naive Bayes (NB), were systematically compared. The optimal model was selected based on Area Under the Curve (AUC), and interpretation was facilitated through SHapley Additive exPlanations (SHAP) analysis. Model calibration was assessed using Brier scores, and discrimination was evaluated through both internal and independent external validation. Results: The Random Forest model demonstrated superior performance (training AUC: 0.913; internal testing AUC: 0.846; external validation AUC: 0.823; Brier score: 0.079), consistently identifying six key predictors: Chinese Mini Mental Status (CMMS) score, Prognostic Nutritional Index, ASA classification, age, dementia, and ICU admission. Conclusion: This comprehensively validated machine learning framework, supported by rigorous multi-institutional validation and systematic algorithm benchmarking, provides an interpretable tool for risk stratification of POD. The identification of nutritional status as a leading predictor highlights previously underutilized targets for perioperative intervention.