Development and internal validation of machine learning-based prognostic prediction models for radical gastrectomy: a multidimensional risk stratification study

Abstract

Gastric cancer remains a leading cause of cancer-related mortality worldwide, with significant heterogeneity in postoperative outcomes among patients. Accurate prognostic prediction is essential for individualized treatment and surveillance strategies. This study aimed to develop and validate machine learning-based prognostic models for patients undergoing radical gastrectomy. We retrospectively collected data from 200 patients who underwent radical gastrectomy at our institution. Clinical and pathological variables including age, sex, TNM stage, Lauren classification, differentiation, tumor size, lymphovascular invasion (LVI), perineural invasion (PNI), and treatment information were included. Three machine learning models were constructed: Cox proportional hazards (CoxPH), Random Survival Forest (RSF), and Gradient Boosting Survival Analysis (GBSA). Model performance was evaluated using 5-fold stratified cross-validation with the C-index and time-dependent AUC. Risk stratification was performed using tertiles of the CoxPH risk score. The cohort comprised 200 patients (79.0% male) with a mean age of 65.4 years (standard deviation [SD] 10.7). During a median follow-up of 64.2 months, 91 patients (45.5%) died. Multivariate Cox regression identified age (HR = 1.097, 95% CI 1.068–1.145, P  < 0.001) and N stage (HR = 1.498, 95% CI 1.100-2.322, P  = 0.035) as independent prognostic factors. Among the three models, CoxPH achieved the highest cross-validated C-index (0.733 ± 0.096). Risk stratification based on the CoxPH model identified three distinct groups: low- risk ( n  = 67, 5-year survival 91.0%), intermediate-risk ( n  = 66, 5-year survival 57.6%), and high-risk ( n  = 67, 5-year survival 22.4%). The RSF model demonstrated excellent time-dependent AUC values of 0.911, 0.933, and 0.950 at 12, 36, and 60 months, respectively. Machine learning-based prognostic models effectively predicted survival outcomes after radical gastrectomy. The CoxPH model provided reliable risk stratification that may assist in prognostic risk assessment after radical gastrectomy. These findings support the potential of machine learning approaches for prognostic risk stratification in gastric cancer, pending external validation.