Abstract
BACKGROUND: Pleural effusion is a common postoperative complication following pancreatic surgery. It is associated with hypoxemia, often requiring prolonged mechanical ventilation and contributing to adverse clinical outcomes. Identifying risk factors and developing predictive models in critically ill patients after pancreatic surgery may facilitate early recognition and guide timely interventions to improve prognosis. METHODS: We retrospectively reviewed 518 intensive care unit (ICU)-admitted patients who underwent pancreatic surgery at Peking University People's Hospital from January 2016 to June 2024. Patients were grouped by postoperative pleural effusion status. Least absolute shrinkage and selection operator (LASSO)-logistic was used to identify key predictors and guide model development. Internal validation was conducted using 1,000 bootstrap resamples. Model discrimination and calibration were assessed using receiver operating characteristic (ROC) curve (area under the curve, AUC) and calibration plots. Decision curve analysis evaluated clinical utility, while restricted cubic spline analysis was applied to explore nonlinear effects of continuous predictors. RESULTS: Among 518 patients, 144 developed postoperative pleural effusion. Independent predictors included age, body mass index (BMI), atrial fibrillation, American Society of Anesthesiologists (ASA) grade, and intraoperative transfusion. A nomogram-based model incorporating these variables demonstrated good discrimination (AUC = 0.733, 95% CI [0.683-0.783]) and reliable calibration. Decision curve analysis confirmed clinical utility across a range of threshold probabilities. Restricted cubic spline analysis revealed nonlinear associations: age-related risk rose sharply beyond 65 years, while BMI showed a U-shaped relationship, with elevated risk below and above the inflection point of 22.6. CONCLUSION: This study developed a predictive model for postoperative pleural effusion in critically ill patients undergoing pancreatic surgery using LASSO-logistic regression. The model demonstrated robust discrimination and calibration, highlighting its potential utility in early risk stratification and individualized clinical decision-making.