Integrating Inflammation-Nutrition Markers and Machine Learning for Prognostic Stratification in Advanced Gastric Cancer: A Multicenter Cohort Study

Abstract Background Advanced gastric/gastroesophageal junction adenocarcinoma (GEJC) has poor prognosis despite emerging therapies. This study constructed and validated machine learning (ML) survival model combining inflammatory-nutritional biomarkers and clinicopathological variables to stratify advanced gastic/GEJC patients into distinct prognostic subgroups. Methods 796 eligible patients from one medical center were split into training (n=560) and internal validation (n=236) cohorts at a 7:3 ratio, with an external validation cohort of 186 patients from two additional hospitals. Univariate Cox and LASSO-Cox regression filtered predictors for overall survival (OS) and progression-free survival (PFS). Seven ML survival models were compared; the optimal model was interpreted via SHAP analysis, and K-means clustering identified prognostic subtypes. Model performance was evaluated by C-index and time-dependent AUC. Results Eight core predictors were selected: WBC, LDH, CA724, prognostic nutritional index (PNI), platelet-to-lymphocyte ratio (PLR), surgical history, immunotherapy, and treatment efficacy. XGBoost achieved stable predictive performance for OS and PFS, with average C-indexes of 0.684 and 0.678 respectively. Time-dependent ROC curves showed favorable AUC values for 1/2/3-year OS, and 3/6/12-month PFS across all datasets. K-means clustering identified two subtypes with distinct survival ( p <0.05). Subtype 1 (low inflammation, high nutrition; better prognosis), Subtype 2 (high inflammation, low nutrition; worse prognosis). SHAP revealed nonlinear associations and variable interactions, with treatment efficacy as the leading protective factor for both endpoints. Conclusion The eight-variable XGBoost model showed robust predictive accuracy and generalizability for OS and PFS in advanced gastic/GEJC. Data-driven subtyping and SHAP visualization facilitate personalized risk stratification and clinical management, while prospective verification remains warranted.