Machine learning prediction of postoperative pulmonary embolism: a multicenter external validation study highlighting inflammatory response and intraoperative hemodynamics.

Background: Postoperative pulmonary embolism (PE) remains a rare but life-threatening complication after surgery. Early identification of high-risk patients is essential but remains challenging due to heterogeneous perioperative risk factors. This study aimed to develop and externally validate a machine learning-based prediction model for postoperative PE and to explore key clinical determinants using interpretable artificial intelligence. Methods: This multicenter retrospective study included surgical patients from six hospitals between January 2020 and January 2025. Patients were divided into internal and external datasets according to hospital source. A total of 3,494 patients were included, with 2005 in the internal cohort and 1,489 in the external validation cohort. Candidate variables were selected using a hybrid strategy combining univariate and multivariate logistic regression with feature importance ranking from five machine learning algorithms, including XGBoost, Random Forest, Support Vector Machine, K-Nearest Neighbors, and Multilayer Perceptron. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), Kolmogorov-Smirnov (KS) statistics, and confusion matrices. K-fold cross-validation and external validation were further performed to assess model robustness. SHapley Additive exPlanations (SHAP) were used for model interpretability and individualized prediction. Results: A total of 48 patients (1.38%) developed postoperative PE. The final selected predictors included age, body mass index (BMI), malignancy history, prolonged bed rest, surgery duration, intraoperative tachycardia, C-reactive protein (CRP), neutrophil-to-lymphocyte ratio (NLR), and postoperative D-dimer. Among five machine learning models, XGBoost demonstrated the best overall performance and stability, achieving superior discrimination and calibration. In the internal validation, the model showed strong predictive performance, and in the external validation cohort, it achieved an AUC of 0.925 (95% CI 0.877-0.972), with good calibration and favorable clinical net benefit on DCA. K-fold cross-validation confirmed model robustness with stable performance across resampling sets. SHAP analysis identified surgery duration, CRP level, malignancy history, age, BMI, postoperative D-dimer, NLR, and intraoperative tachycardia as the most influential predictors. Individual-level SHAP interpretation further demonstrated clinically meaningful risk attribution patterns for postoperative PE. Conclusion: We developed and externally validated a robust machine learning model for predicting postoperative pulmonary embolism across multiple surgical populations. The model demonstrated strong discrimination, good calibration, and favorable clinical utility. Importantly, SHAP-based interpretation revealed key perioperative inflammatory, thrombotic, and hemodynamic factors associated with PE risk, providing both predictive and mechanistic insights to support clinical decision-making.