Developing and Interpreting a Machine Learning Model for Identifying Liver Metastasis in Gastric Cancer

IntroductionLiver metastasis (LM) is the most common site of distant metastasis in gastric cancer (GC), significantly affecting patient prognosis. This study developed a machine learning (ML) model to predict the risk of liver metastasis in gastric cancer patients using data from the Surveillance, Epidemiology, and End Results (SEER) database.MethodsData from eligible GC patients between 2010 and 2015 were collected from the SEER database. Stratified randomization divided the data into a training set (70%, n=7671) and a validation set (30%, n=3287). Univariate and multivariate logistic regression analyses, along with Boruta-Shap algorithm and Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection, were performed. Seven ML algorithms were developed to predict the risk of liver metastasis in gastric cancer patients. Receiver Operating Characteristic (ROC) curve analysis, including Area Under the Curve (AUC), sensitivity, specificity, and Negative Predictive Value (NPV), was used to assess performance. The SHapley Additive Explanations (SHAP) framework identified key predictors for liver metastasis.ResultsA total of 10,958 GC patients were enrolled, among whom 766 (6.99%) presented synchronous LM. After screening, gender, T stage, tumor size, surgery, radiation therapy, lung metastasis, and bone metastasis were identified as key influencing factors for liver metastasis in gastric cancer. The eXtreme Gradient Boosting (XGB) model exhibited superior performance with an AUC of 0.846 (95% CI: 0.820-0.872) and an NPV of 97.7%. Decision Curve Analysis (DCA) and calibration curve analyses confirmed its reliable clinical utility and predictive accuracy. Furthermore, the SHAP framework revealed that surgery, radiation therapy, and T stage were the primary factors influencing the model's predictions.ConclusionsThis study developed and validated a predictive XGB model using clinical and pathological data to predict the risk of LM in GC patients. The model can provide critical support for the development of personalized medical strategies in clinical practice.