Development and validation of machine learning model for mortality after curative gastrectomy for gastric cancer.

386 Background: The risk for postoperative complications can significantly impact perioperative planning. The aim of this study is to develop and validate a model based on various machine learning techniques to predict postoperative mortality (PM) after curative radical gastrectomy for gastric cancer. Methods: Data from a nationwide survey led by the Korean Gastric Cancer Association was used for the development set, which included 12,722 patients who underwent gastrectomy in 2019 across 68 institutions in South Korea. The validation set comprised 4,887 patients from 8 institutions between 2009 and 2017. PM was defined as all-cause death within 30 days after surgery or in-hospital death beyond 30 days. Five machine learning models (CatBoost, Gradient Boosting, Light GBM, Random Forest, XGBoost) were evaluated using perioperative clinical characteristics. Model performance was assessed using the area under the receiver operation characteristic curve (AUROC) with 5-fold cross-validation. Results: The PM rates were 0.83% and 1.04% in the development and validation set, respectively. In the development set, the Random Forest model demonstrated the highest discrimination capacity with an AUROC of 0.772, followed by Gradient Boosting at 0.767, CatBoost at 0.742, XGBoost at 0.732, and LightGBM at 0.717. However, in the validation set, CaBoost showed the highest AUROC at 0.747, while the AUROC for Random Forest decreased to 0.688. Conclusions: The CatBoost model demonstrated robust and consistent performance in predicting PM risk across both development and validation datasets. To further enhance the performance of the model and ensure stable clinical application, future research should focus on developing methods for quantitative measurement of perioperative factors such as surgical techniques.