Machine learning-based identification of perioperative associated factors and early prediction of postoperative delirium in older adults with renal calculi

Abstract Background Identifying key perioperative associated factors, incorporating both baseline and early postoperative changes, is crucial for early risk warning and targeted care. Methods We performed a retrospective study on older adults (aged ≥ 65 years) who underwent surgery for renal calculi between January 2023 and August 2024. Conventional logistic regression and ten machine-learning algorithms were employed to construct models integrating preoperative, intraoperative, and early postoperative variables (up to POD1) to identify core factors associated with POD. Feature importance was interpreted using Shapley Additive Explanations (SHAP). Results A total of 321 patients were enrolled, with the incidence of POD being 14.95% (48/321). Preoperative sleep disturbance and lower baseline preoperative C-reactive protein (CRP) were identified as core predictive factors via SHAP analysis. Among the constructed models, Random Forest (AUC = 0.883), Extreme Random Trees (AUC = 0.879), and Gaussian Naive Bayes (AUC = 0.869) were the preferred models for POD risk prediction, with Random Forest demonstrating the highest discriminative ability. Conclusion Eleven prediction models (one logistic regression and ten machine-learning models) were evaluated for early postoperative risk assessment of POD in older adults with renal calculi. By integrating dynamic changes on the first postoperative day, these models offer an exploratory tool for evaluating personalized POD probability and identifying high-risk patients experiencing abnormal acute physiological responses to surgery.