**Background:** Personalized survival prediction after gastrectomy for gastric cancer remains challenging. The TNM staging system alone does not provide accurate prognostication, and prior models have been limited by few variables and inadequate validation. The authors hypothesized that 1 year after gastrectomy is the optimal time for long-term survival prediction, as patients have undergone significant nutritional changes, muscle loss, and postoperative metabolic adjustments. They aimed to develop an AI model incorporating a large number of variables including time-varying nutritional and body morphometry factors.
**Methods:** From a prospectively built gastric surgery registry at Asan Medical Center (AMC, Seoul, Korea), 4025 gastric cancer patients (mean age 56.1 ± 10.9 years, 36.2% female) who underwent gastrectomy and survived more than 1 year were selected. A total of 89 variables were used as input, including 63 clinical variables (demographics, physical indices, laboratory results, nutritional risk index [NRI], fat/muscle indices from CT scans, surgery-related variables, pathologic variables, and comorbidities) plus 26 derived time-varying variables (differences and percent differences between preoperative and 1-year postoperative values for weight, BMI, cholesterol, hemoglobin, albumin, protein, NRI, subcutaneous fat area [SFA], skeletal muscle area [SMA], skeletal muscle index [SMI], and SMA/BMI). Body morphometry was measured using an AI solution (AID-U™) at the L3 vertebral level. The proposed multi-tree XGBoost algorithm used an ensemble of 100 trees from repeated five-fold cross-validation (20 repeats). Internal validation was performed on a split dataset (n = 1121). External validation used 590 patients from Ajou University Hospital (AUH, Suwon, Korea; mean age 55.9 ± 11.2, 37.3% female), where only 28 of 63 clinical variables were available (extended to 54 variables). Missing data were imputed with the mean from training data. The synthetic minority over-sampling technique (SMOTE) was used to address class imbalance. Sensitivity analysis using a leave-one-out method assessed the contribution of nutritional and fat/muscle indices.
**Key Results:** In repeated five-fold cross-validation, the proposed model showed a sensitivity of 76.77%, specificity of 75.26%, accuracy of 75.32%, balanced accuracy of 76.01%, and AUROC of 0.8118, outperforming all six comparator models (RF, GBM, AdaBoost, LightGBM, CatBoost, Ensemble; AUROC range 0.7741–0.7932). In internal validation (n = 1121), the model achieved 80.00% sensitivity, 72.34% specificity, 72.67% accuracy, 76.17% balanced accuracy, and AUROC of 0.8237, again exceeding all comparators (AUROC range 0.7988–0.8165). In external validation (n = 590), performance improved further: sensitivity 86.96%, specificity 74.60%, accuracy 75.08%, balanced accuracy 80.78%, and AUROC 0.8903. Variable importance analysis identified age as the most important predictor, followed by preoperative albumin, preoperative NRI, T stage, and percent difference of hemoglobin. Among the top 25 variables, 15 were pre- and postoperative physical, nutritional, laboratory, or fat/muscle indices. Sensitivity analysis showed that excluding nutritional and fat/muscle indices decreased balanced accuracy by 1.70% (cross-validation), 0.31% (internal validation), and 6.29% (external validation), and decreased AUROC by 0.0223, 0.0076, and 0.0213, respectively.
**Clinical Implications:** This AI model provides accurate, personalized 5-year survival prediction at 1 year after gastrectomy, incorporating a broader range of clinically available variables than prior models. The inclusion of nutritional and body morphometry data (NRI, SMI, VFA) meaningfully improves prediction. The model is deployed as a public web tool (http://ai-research.co.kr/survival) for clinical use. Limitations include training at a single high-volume Korean center, external validation at only one additional center, a predominantly Korean population, and high rates of missing data for some key variables (e.g., 69.9% missing for 1-year postoperative SMI, 53.9% for 1-year postoperative NRI). Future work should include multi-institutional validation across diverse populations and real-time model updating.