Identification of gastroesophageal varices in patients with cirrhosis using a CatBoost model derived From routine laboratory parameters

Objective To construct and internally validate a machine learning-assisted diagnostic model that identifies gastroesophageal varices (GEV) in patients with cirrhosis from routinely collected clinical and laboratory data. Methods We retrospectively analyzed 5,037 patients with cirrhosis who underwent endoscopy from January 2024 through May 2026. Random allocation at a 7:3 ratio produced a training set ( n = 3,526) and a validation set ( n = 1,511). Candidate features were reduced using multivariable logistic regression and the least absolute shrinkage and selection operator (LASSO). Ten machine learning algorithms were trained. Selection of the final model incorporated validation-set discrimination, calibration, decision curve analysis, and the difference in the area under the receiver operating characteristic curve (AUROC) between the training and validation sets. SHapley Additive exPlanations (SHAP) quantified the contribution of each feature. CatBoost was evaluated against the albumin–platelet product (APP), fibrosis index (FI), a univariable platelet count (PLT) model, and the Lok index in the validation cohort. Exploratory analyses examined performance within subgroups defined by age, sex, hepatitis B virus (HBV) etiology, and ascites. This is a model for adjunct recognition of existing GEV, not a rule for deferring endoscopy. Results GEV was documented endoscopically in 3,320 patients (65.91%). CatBoost yielded an AUROC of 0.865 (95% confidence interval [CI], 0.845–0.884) in the validation set; accuracy was 0.811, sensitivity was 0.897, specificity was 0.647, and the Brier score was 0.135 (95% CI, 0.125–0.146). The AUROC decrease from training to validation was 0.038 for CatBoost and 0.051 for XGBoost. Albumin, red cell distribution width, and platelet count contributed most strongly to the CatBoost output in SHAP analysis. The AUROC for CatBoost (0.865) exceeded those for FI (0.796), APP (0.728), the PLT model (0.662), and the Lok index (0.592); each paired DeLong comparison remained significant after Holm adjustment (all P < 0.001). Across eight exploratory subgroups, AUROCs varied from 0.833 to 0.879. Conclusion A CatBoost classifier incorporating 10 routine laboratory variables demonstrated good discrimination and calibration during internal validation and provided higher net benefit than the four conventional comparators across most evaluated thresholds. It may provide adjunctive information before endoscopy; prospective evaluation and multicenter external validation remain necessary before clinical decision use.