Machine Learning Prediction of Early Recurrence in Gastric Cancer: A Nationwide Real-World Study.

Abstract

Background: Patients with gastric cancer (GC) who experience early recurrence (ER) within 2 years postoperatively have poor prognoses. This study aimed to analyze and predict ER after curative surgery for patients with GC using machine learning (ML) methods.

Patients and methods: This multicenter population-based cohort study included data from ten large tertiary regional medical centers in China. The clinical, pathological, and laboratory parameters were retrospectively collected from the records of 11,615 patients. The patients were randomly divided into training (70%) and test (30%) cohorts. A total of ten ML models were developed and validated to predict the ER. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration plots, and Brier score (BS). SHapley Additive exPlanations (SHAP) was used to rank the input features and interpret predictions.

Results: ER was reported in 1794 patients (15%) during follow-up. The stacking ensemble model achieved AUCs of 1.0 and 0.8 in the training and testing cohorts, respectively, with a BS of 0.113. SHAP dependency plots revealed that tumor staging, elevated tumor marker levels, lymphovascular invasion, perineural invasion, and tumor size > 5 cm were associated with higher ER risk. The impact of age and the number of lymph nodes harvested on ER risk exhibited a "U-shaped distribution." Additionally, an online prediction tool based on the best model was developed to facilitate clinical applications.

Conclusions: We developed a robust clinical model for predicting the risk of ER after surgery for GC, which may aid in individualized clinical decision-making.