Development and Validation of a Machine Learning Model for Predicting Early Postoperative Complications After Radical Gastrectomy

Ruyin Li¹Zirui Zhao²Jianchun Yu^1*

• ¹Department of General Surgery，Peking Union Medical College Hospital, Peking Union Medical College and Chinese Academy of Medical Sciences, Beijing, China
• ²Department of Neurology, Peking Union Medical College Hospital, Peking Union Medical College and Chinese Academy of Medical Sciences, Beijing, China

Background: Postoperative complications significantly impact gastric cancer patients' recovery and remain a major research focus. This study aimed to develop a machine learning model utilizing preoperative and intraoperative data to stratify the risk of early postoperative complications in patients undergoing radical gastrectomy. Methods: Clinical data from gastric cancer patients who underwent radical gastrectomy at Peking Union Medical College Hospital between 2014 and 2024 were retrospectively collected. Using R software, ten machine learning algorithms—including eXtreme Gradient Boosting, Support Vector Machine, random forest, Neural Network, naive Bayes, logistic regression, Linear Discriminant Analysis, K-Nearest Neighbors, Generalized Linear Model with Elastic-Net Regularization and classification tree—were employed to construct predictive models for early postoperative complications. Nested cross-validation was applied for model validation, and performance was evaluated using receiver operating characteristic curves, decision curve analysis, and calibration curves. Results: A total of 926 patients were included in this study, comprising 667 males (72%) and 259 females (28%), with 131 (14.13%) suffering postoperative complications. Predictive features included smoking, Nutritional Risk Screening 2002 score3, reconstruction, clinical T-stage1, operative time, neoadjuvant chemotherapy combined with immunotherapy or targeted therapy, and resection site. Among the ten models, eXtreme Gradient Boosting demonstrated the best predictive performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.788, along with superior calibration and decision curve analysis results. Conclusion: Based on preoperative and intraoperative data, the eXtreme Gradient Boosting model demonstrated the strongest predictive capability for postoperative complications following radical gastrectomy.These findings underscore the potential of machine learning-based models in stratifying the risk of early postoperative complications in patients undergoing radical gastrectomy, thereby enhancing clinical decision-making and improving patient outcomes in gastric cancer surgery.