Development and external validation of a machine learning model for 28-day mortality in patients with gastrointestinal bleeding and chronic kidney disease

Gastrointestinal bleeding (GIB) in patients with chronic kidney disease (CKD) is associated with disproportionately high morbidity and mortality, yet no specific risk prediction tool exists for this comorbid population. This study aimed to develop and externally validate a machine learning model for predicting 28-day all-cause mortality in GIB patients with coexisting CKD. We conducted a retrospective cohort study using the MIMIC-IV (v3.1) database, identifying 926 adult patients with concurrent GIB and CKD. The cohort was randomly split into a training set ( n = 648) and an internal validation set ( n = 278). An independent cohort of 1,000 patients from the Affiliated Hospital of Guangdong Medical University(GDMUAH) served as external validation. Forty-five clinical variables were extracted, covering demographics, comorbidities, vital signs, laboratory tests, medications, and interventions. Feature selection was performed using LASSO regression and Boruta algorithm. Six ML algorithms—XGBoost, CatBoost, AdaBoost, LightGBM, Logistic Regression (LR), and Random Forest (RF)—were trained and tuned via grid search with 10-fold cross-validation. Model performance was assessed by discrimination (AUROC), calibration (Brier score), and decision curve analysis (DCA). The best-performing model was interpreted using SHapley Additive exPlanations (SHAP). Eight overlapping predictors were selected: red cell distribution width (RDW), serum chloride, anion gap, prothrombin time (PT), partial thromboplastin time (PTT), Systolic Blood Pressure (SBP), sepsis and Chronic Obstructive Pulmonary Disease (COPD). In internal Test, LR achieved the highest AUROC 0.734 (95% CI 0.666–0.802) and the most balanced performance (Sensitivity 70.3% (95% CI 0.547–0.891), Specificity 69.2% (95% CI 0.505–0.832),Brier 0.160 (95% CI 0.145–0.178)). External validation of the LR model yielded an AUROC of 0.732 (95% CI 0.689–0.774), accuracy 0.773 (95% CI 0.746–0.799), sensitivity 0.455 (95% CI 0.382–0.531), specificity 0.840 (95% CI 0.819–0.869) and a negative predictive value of 0.88. Calibration and DCA confirmed good agreement and net clinical benefit across a wide threshold range. SHAP analysis identified anion gap, SBP, PTT, and sepsis as the most influential contributors to mortality risk. The LR model showed acceptable discrimination and calibration for early risk stratification of 28-day mortality in GIB patients with CKD; its low external sensitivity and modest PPV indicate that it should be considered a preliminary decision-support tool requiring prospective multicenter validation.