Abstract
<title>Abstract</title> <p> <bold>Background</bold> : Elderly patients (≥65) constitute >60% of gastric adenocarcinoma cases yet are underrepresented in clinical trials. Machine learning survival models may capture complex prognostic interactions, but comprehensive benchmarking with external validation in this population is absent. <bold>Methods</bold> : From SEER (2000-2022), 68,921 elderly gastric adenocarcinoma patients were identified; 60,378 had complete data. After harmonizing three AJCC staging editions and propensity score matching, five survival models (Cox, RSF, XGBoost, DeepSurv, Gradient Boosting) were trained with 5-fold cross-validation and temporal validation. A Cox model with treatment×stage×age interactions estimated individualized treatment effects. External validation used a Chinese cohort (N=575) with CT-defined sarcopenia, frailty, and inflammatory biomarkers. <bold>Results</bold> : XGBoost achieved C-index 0.772, significantly outperforming Cox (0.749, +3.1%), DeepSurv (0.764), and RSF (0.738). Surgical resection showed the highest permutation importance (0.128; 4.4× Stage IV). Chemotherapy benefit was stage-dependent: null in Stage I (HR=1.00), maximal in Stage IV (HR=0.44). External validation C-index 0.617; the 20% decline was attributable to 100% surgery rate in the validation cohort rather than model overfitting. Novel clinical variables-severe sarcopenia (HR=3.39), gait speed less than 0.8m/s (HR=2.45), albumin less than 35g/L (HR=2.07)-achieved superior discrimination (C-index 0.753 vs 0.684, +10%). <bold>Conclusions</bold> : XGBoost improves prognostic discrimination over Cox. Surgery dominates prognosis; chemotherapy benefit is stage-conditional. Model transportability depends on treatment variance. Body composition and frailty variables substantially improve discrimination, supporting function-based over age-based treatment decisions. </p>