Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Postoperative Delirium in Older Adults Undergoing Orthopaedic Surgery: A Prospective Cohort Study

Abstract Objective Postoperative delirium (POD) is a common and serious complication following hip or knee replacement and revision surgery in older patients, leading to prolonged recovery, increased healthcare costs, and poor long-term outcomes. Early identification of high-risk patients remains challenging. This study aimed to develop and internally validate an interpretable machine learning (ML)-based risk prediction model for POD using multidimensional perioperative variables and to compare the predictive performance of seven ML algorithms. Methods This single-center prospective observational cohort study included 930 older patients (≥ 60 years) who underwent elective hip or knee replacement and revision surgery. POD was assessed using the Confusion Assessment Method (CAM) within 7 days after surgery. The least absolute shrinkage and selection operator (LASSO) regression was used for feature selection. Seven ML algorithms were developed and compared, and SHapley Additive exPlanations (SHAP) was applied for model interpretability. Results POD occurred in 30.0% (279/930) of patients. LASSO identified six predictors: preoperative sleep quality (Pittsburgh Sleep Quality Index, PSQI), psychological distress (Kessler Psychological Distress Scale, K10), social support (Groningen Orthopaedic Social Support Scale, GO-SSS), nutritional status (Prognostic Nutritional Index, PNI), comorbidity burden (age-adjusted Charlson Comorbidity Index, aCCI), and intraoperative physiological reserve (Surgical Apgar Score, SAS). The extreme gradient boosting (XGBoost) model performed best, with an area under the curve (AUC) of 0.922 (95% CI : 0.886–0.957) and a Brier score of 0.095. SHAP analysis revealed that preoperative sleep quality was the most influential predictor, showing a nonlinear positive association with POD risk and a synergistic interaction with psychological distress. Conclusion The XGBoost-based model integrating six readily available predictors demonstrated good predictive performance and potential clinical utility. The SHAP-based interpretability enhances model transparency, supporting early risk identification and personalized perioperative management. External validation is warranted to confirm generalizability.