Machine learning for prediction of postoperative pulmonary embolism after lung cancer surgery using routine perioperative variables.

e20522 Background: Postoperative pulmonary embolism (PE) remains a potentially fatal complication after lung cancer (LC) resection. Widely used VTE risk tools (e.g., Caprini) are time-consuming and not optimized for thoracic surgical populations. We developed PEPred, a machine learning (ML)–informed model based on 8 routinely captured perioperative variables, to enable efficient and low-cost PE risk stratification. Methods: A retrospective cohort of 9,726 LC resections without chemoprophylaxis was used, including 55 PE cases. An XGBoost model selected 8 key surgical features, such as surgical approach (open vs. VATS), extent of resection, and tumor location. These features were incorporated into PEPred, a logistic regression model with L1 regularization and pairwise feature interactions. Its performance was evaluated using 5-fold cross-validation, with metrics including ROC-AUC, PR-AUC, accuracy, precision, recall, and F1 score. Decision curve analysis assessed clinical utility. Results: The PEPred model achieved a ROC-AUC of 0.853 (95% CI: 0.771–0.921) and a PR-AUC of 0.582 (95% CI: 0.447–0.711), demonstrating robust discriminative ability. Key predictive features included the extent of resection (e.g., more extensive resections increased risk), surgical approach (open vs. VATS), and tumor location (e.g., upper lobe involvement). Pairwise feature interactions, such as the combination of surgical approach and extent of resection, further enhanced model performance, with some interactions showing odds ratios as high as 6205.757 for increased PE risk. Decision curve analysis confirmed the its clinical utility, with PEPred providing a positive net benefit across threshold probabilities ranging from 0.5% to 10%. At a 1.0% threshold, the net benefit was 0.0037, outperforming both “treat-all” and “treat-none” strategies. The model’s accuracy was 0.997, precision 0.786, recall 0.600, and F1 score 0.680, highlighting its ability to reliably identify high-risk cases. Feature importance analysis revealed that the extent of resection (32.61%), surgical approach (20.74%), and lobe involvement (15.34%) were the most significant contributors to model performance. Overfitting was minimal, with training and validation ROC-AUC and PR-AUC curves showing consistent convergence. The model’s simplicity, relying solely on surgical data, makes it highly feasible for integration into clinical workflows without requiring additional laboratory tests or complex data inputs. Conclusions: PEPred is a novel ML-based tool for predicting postoperative PE in LC surgery, using only 8 routine surgical parameters. Its strong performance and simplicity make it a promising tool for improving perioperative risk stratification and guiding targeted prophylaxis. Future efforts will focus on clinical validation and real-world evidence to assess its impact on patient outcomes.