Identification of postoperative weight loss trajectories and development of a machine learning-based tool for predicting malnutrition in gastric cancer patients

Background Significant postoperative weight loss and malnutrition represent common and serious complications following radical gastrectomy for gastric cancer. Early identification of distinct weight loss trajectories and prediction of malnutrition risk may facilitate targeted interventions. Methods This prospective, observational longitudinal study enrolled 312 gastric adenocarcinoma patients undergoing radical gastrectomy. Participants were assessed preoperatively (T0) and at 3, 6, 9, and 12 months postoperatively (T1–T4). Percentage weight loss was calculated at each postoperative time point. Latent growth mixture modeling (GMM) identified distinct weight loss trajectories. Eight machine learning algorithms (XGBoost, SVM, RF, NB, KNN, MLP, GBM, PLS) were trained using predictors selected by LASSO regression and the Boruta algorithm to predict GLIM-defined malnutrition at 6 months postoperatively (T2, the peak malnutrition timepoint). Additionally, a multivariable logistic regression-derived nomogram was developed and validated, with assessments of discrimination, calibration, and clinical utility. Results GMM identified three distinct 12-month postoperative weight loss trajectories: severe (11.9%), moderate (36.2%), and minimal (51.9%). The prevalence of GLIM-defined malnutrition peaked at 51.6% at 6 months (T2). Among the eight machine learning models, XGBoost achieved the best performance in predicting 6-month malnutrition. The final nomogram, which incorporated age ≥65 years, preoperative underweight status, preoperative reduced muscle mass, and total gastrectomy, showed excellent discrimination, calibration, and clinical utility. DeLong’s test indicated no significant difference in AUC between the XGBoost model and the nomogram ( p = 0.121). Conclusion This study delineates distinct postoperative weight loss trajectories in gastric cancer patients. We developed and validated both an advanced ML model (XGBoost) and a clinically interpretable nomogram for accurately predicting 6-month postoperative malnutrition risk.