Abstract
BACKGROUND: Since postoperative complications after gastrectomy for gastric cancer are associated with poor clinical outcomes, it is important to predict and prepare for the occurrence of complications preoperatively. Conventional models for predicting complications have limitations, prompting interest in machine learning algorithms. Machine learning models have a superior ability to identify complex interactions among variables and nonlinear relationships, potentially revealing new risk factors. This study aimed to explore previously overlooked risk factors for postoperative complications and compare machine learning models with linear regression. MATERIALS AND METHODS: We retrospectively reviewed data from 865 patients who underwent gastrectomy for gastric cancer from 2018 to 2022. A total of 85 variables, including demographics, clinical features, laboratory values, intraoperative parameters, and pathologic results, were used to conduct the machine learning model. The dataset was partitioned into 80% for training and 20% for validation. To identify the most accurate prediction model, missing data handling, variable selection, and hyperparameter tuning were performed. RESULTS: Machine learning models performed notably well when using the backward elimination method and a moderate missing data strategy, achieving the highest area under the curve values (0.744). A total of 15 variables associated with postoperative complications were identified using a machine learning algorithm. Operation time was the most impactful variable, followed closely by pre-operative levels of albumin and mean corpuscular hemoglobin. Machine learning models, especially Random Forest and XGBoost, outperformed linear regression. CONCLUSIONS: Machine learning, coupled with advanced variable selection techniques, showed promise in enhancing risk prediction of postoperative complications for gastric cancer surgery.