Machine learning-driven risk stratification for distant metastasis in gastric cancer: A comparative study of clinical features and composite indices integrated models

Abstract

Objective Distant metastasis (DM) of gastric cancer (GC) represents a significant health challenge due to its high mortality rates, necessitating advancements in early detection and management strategies. The objective of this study was to create a machine learning (ML) model that is interpretable for preoperative prediction of DM in GC. Methods We retrospectively analyzed 1,009 GC patients, of which 769 were from Zhejiang Cancer Hospital as development cohort and 240 from Zhejiang Provincial Hospital of Chinese Medicine as external test cohort. Nine clinical features, and four composite indices derived from ten laboratory indicators were selected as candidate features. The dataset was balanced using the borderline Synthetic Minority Over-sampling Technique (SMOTE) and the Edited Nearest Neighbors (ENN) under-sampling method. Univariate and multivariate analyses were used to identified key metastasis-related features. Based on the identified features, we developed predictive models incorporating five ML algorithms, with performance evaluated via receive operating characteristic (ROC) curves, recall, precision-recall (PR) curves. Ultimately, Shapley additive explanations (SHAP) analysis were applied to rank the feature importance and explain the final model. Results Univariate and multivariate analyses identified five metastasis-related features: cT stage, cN stage, differentiation grade, PLR and TMI. Logistic Regression emerged as the optimal predictive model with the highest area under the curve (AUC) of 0.942 (95% CI: 0.922–0.962), Recall of 0.895 (95% CI: 0.843–0.947), and AUPRC of 0.889 (95% CI: 0.867–0.911) among five models. Additionally, the internal and external test cohorts recorded AUC values of 0.935 (95% CI: 0.897–0.972) and 0.879 (95% CI: 0.833–0.926), respectively. The SHAP analysis revealed the features that played a significant role in the predictions made by the model. Conclusion This ML model integrates clinical features and composite indices to predict GC metastasis risk, supported by an online tool to guide preoperative decision-making.