Development and Temporal External Validation of an XGBoost-Based Clinical Prediction Model for PACU Hypoxemia in Elderly Thoracic Surgery Patients

Objective To develop and validate a machine learning-based risk prediction model for postoperative hypoxemia in elderly patients undergoing thoracic surgery in the post-anesthesia care unit (PACU), aiming to provide a decision-support tool for perioperative precision prevention. Methods This study involved a retrospective development cohort of 5134 elderly thoracic surgery patients from a tertiary academic medical center between 2019 and 2024(split 8:2 into training and internal validation sets), and a prospective external validation cohort of 272 patients in 2025. Feature selection was performed using LASSO regression. Six machine learning algorithms, including Logistic Regression, XGBoost, and LightGBM, were trained and compared. Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA). SHAP values were used for model interpretability. Results The incidence of hypoxemia was 18.8%(967/5134). LASSO regression identified 10 independent predictors categorized into thoracic surgery site, age, BMI, preoperative albumin, hemoglobin, D-dimer, PaO₂,duration of anesthesia and surgery, intraoperative fluid volume.The AUC values across the six machine learning models ranged from 0.852 to 0.953. XGBoost demonstrated the best performance, with AUCs of 0.988 in the training set, 0.953 in the internal validation set, and 0.706 in the external validation set, indicating good generalizability. Conclusions The XGBoost-based prediction model accurately identifies elderly thoracic surgery patients at high risk for PACU hypoxemia. Incorporating multiple modifiable clinical indicators, this tool facilitates early risk stratification and proactive, continuous, multidisciplinary interventions by thoracic surgery, anesthesia, and PACU teams.