[Utilizing machine learning and SHAP analysis to develop a prognostic survival prediction model for gastric cancer patients following radical gastrectomy].

Objective: To construct a prediction model and website for the overall survival (OS) of gastric cancer patients after radical gastrectomy based on SHapley Additive exPlanations (SHAP) and machine learning models. Methods: This retrospective cohort study included 234 patients who underwent radical gastrectomy at Henan Cancer Hospital from October 2018 to December 2022, selected according to the following criteria: (1) pathological diagnosis of gastric cancer; (2) a single primary tumor lesion; (3) R0 resection (no macroscopic or microscopic residual tumor). Exclusion criteria were: (1) age ≤18 years; (2) gastric stump cancer; (3) receipt of other treatments prior to admission; (4) incomplete clinicopathological or follow-up data. Patients were stratified into a training set (n=176) and an internal validation set (n=58) with a ratio of 3∶1. Additionally, 136 patients from the First Affiliated Hospital of Zhengzhou University were included as an external validation set using the same criteria. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed to identify features significantly associated with OS in GC patients. These features, along with clinically common variables [age, tumor longest diameter, and carbohydrate antigen (CA) 199], were then incorporated into four distinct machine learning algorithms: Logistic Regression (LR), K-Nearest Neighbors (KNN), Gaussian Naive Bayes (GaussianNB), and Support Vector Machine (SVM). The predictive performance of these models was assessed and compared. SHAPley Additive exPlanations (SHAP) were applied to interpret the model with the highest performance. Results: LASSO regression identified gender, tumor T stage, extent of resection, tumor differentiation degree, and carcinoembryonic antigen (CEA) level as predictors. The area under the curve (AUC) values for the LR, KNN, GaussianNB, and SVM models were 0.751, 0.809, 0.743, and 0.816 in the training set, and 0.782, 0.817, 0.763, and 0.877 in the external validation set, respectively. In the external validation set, the SVM model showed the highest precision (0.808), outperforming LR (0.663), KNN (0.728), and GaussianNB (0.688). SHAP analysis ranked the feature contributions as follows: tumor T stage, extent of resection, age, CEA, CA199, gender, tumor longest diameter, and differentiation degree. Poor survival was associated with T4 stage, total gastrectomy, age ≥60 years, CA199 ≥37 kU/L, tumor longest diameter ≥4 cm, and poor differentiation, while female patients with CEA <5 μg/L had a better prognosis. A web predictor based on the SVM model was established at https://wagzshzscqyc. shinyapps.io/document/. Conclusions: Tumor T stage, extent of resection, age, CEA, CA199, gender, tumor longest diameter, and differentiation degree are associated with OS after radical gastrectomy. This study constructed an SVM-based prediction model with good precision and developed a corresponding website.