Abstract
Gastric cancer with liver metastasis (GCLM) often has a poor prognosis. Therefore, it is crucial to identify risk factors affecting their overall survival (OS) and cancer-specific survival (CSS). This study aimed to construct practical machine learning models to predict survival time and help clinicians choose appropriate treatments. We reviewed the clinical and survival data of GCLM patients from 2010 to 2017 in the Surveillance, Epidemiology, and End Results (SEER) databases and divided the patients into training and testing groups. The risk factors affecting OS and CSS were determined by least absolute shrinkage and selector operator (LASSO), univariate cox regression, best subset regression (BSR) and the stepwise backward regression. Then, five machine learning models, including random survival forest (RSF), Gradient Boosting Machine (GBM), the Cox proportional hazard (CPH), Survival Support Vector Machine (survivalSVM), and eXtreme Gradient Boosting (XGBoost), were built using the identified risk factors. The model with the best predictive ability was determined using concordance index (c-index), area under the curve (AUC), brier score, and decision curve analysis (DCA), and externally verified with data from 233 cases diagnosed with liver metastasis of cancer from The Shijiazhuang People's Hospital, Jinan City People's Hospital, and The Sixth People's Hospital of Huizhou from 2017 to 2018. The study involved a total of 1300 GCLM patients. The prognostic risk factors affecting OS and CSS were the same, including grade, histology, T stage, N stage, surgery, and chemotherapy. The XGBoost model was found to have the best predictive ability for OS, with AUC of 0.891 [95% CI 0.841-0.941], brier score of 0.061 [95% CI 0.046-0.076], and c-index of 0.752 [95% CI 0.742-0.761], as well as for CSS, with AUC of 0.895 [95% CI 0.848-0.942], brier score of 0.064 [95% CI 0.050-0.079], and c-index of 0.746 [95% CI 0.736-0.756]. The AUC score, brier score and c-index all illustrated the accuracy of the model, and the validation using the external datasets further confirmed the reliability of the model. Therefore, the XGBoost model demonstrated significant potential in predicting survival times and selecting appropriate treatment plans.