Development and validation of an explainable machine learning model for predicting postoperative delirium in older adults undergoing joint replacement surgery: a prospective cohort study

Postoperative delirium (POD) is a common and serious complication in older adults undergoing joint replacement surgery, yet accurate early risk stratification remains challenging. We aimed to develop and validate an explainable machine learning model for predicting POD. In this prospective single-center cohort study, 451 patients aged ≥ 65 years undergoing joint replacement surgery were enrolled. Six machine learning models were developed and compared using perioperative clinical variables. Model performance was evaluated in an independent validation set in terms of discrimination, calibration, and clinical utility. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis. POD occurred in 29.3% of patients. Among the six models, the support vector machine (SVM) model demonstrated the best overall performance, achieving an AUC of 0.931 (95% CI 0.882–0.971), with good calibration (Brier score 0.091) and the highest net clinical benefit. The model outperformed traditional logistic regression and other machine learning approaches. Key predictors included age, frailty, cognitive function, sleep quality, serum albumin level, diabetes mellitus, and PACU length of stay. An explainable SVM-based model using readily available perioperative variables showed good predictive performance for POD and may support perioperative or early postoperative risk stratification rather than preoperative prediction, alongside targeted perioperative management. External validation is warranted before clinical implementation.