Introduction:Upper gastrointestinal bleeding(UGIB) causes substantial morbidity/mortality. Early risk stratification guides triage, resource-allocation, and timely intervention. Clinical tools(AIMS65, modified-ABC, pre-endoscopic-Rockall) have modest predictive performance that don’t incorporate nonlinear relationships. We developed machine-learning(ML)-models using triage/early data to predict in-hospital-mortality-within-30-days-of-UGIB-presentation(IHM30) and compared with established risk scores.
Methods:Retrospective cohort study of adults ≥18 with acute UGIB(1/2019-12/2024) at SIUH. Predictors were demographics, vitals, comorbidities, medications, labs within 2-4hrs of presentation. Primary outcome was IHM30. Missing data was handled using multiple imputation(m=10) with missingness indicators for variables with substantial missingness. Data was split into train(80%)/test(20%) sets. Supervised models were developed(penalized-logistic-regression(LR), decision-tree(DT), random-forest(RF), XGBoost) with nested cross-validation hyperparameter tuning. Performance was evaluated on the test set(AUROC, AUPRC, Brier-score, decision-curve-analysis). ML-models were compared with established risk scores using DeLong testing.
Results/Discussion:Mortality occurred in 454/8,501 patients(5.3%). ML-models demonstrated strong discrimination: LR(AUROC 0.824), DT(0.768), RF(0.823), XGBoost(0.817). ML-models significantly outperformed AIMS65(AUROC 0.674), modified-ABC(0.732), Rockall(0.734)(p < 0.001). AUPRC ranged from 0.26-0.32, exceeding baseline prevalence(~0.05). RF/XGBoost achieved lowest Brier scores(~0.045). At high-sensitivity thresholds(≥0.90), RF/XGBoost identified 93-98% of deaths. Decision-curve-analysis demonstrated ML-models, particularly XGBoost/RF, provided greater net benefit than traditional scores across clinically relevant risk thresholds(2-10%). Lactate had 55% missingness. Lactate availability was the strongest mortality predictor. Other predictors included age, lactate, albumin, creatinine, INR.
Conclusion:ML-models improved prediction of IHM30 compared with traditional risk scores. XGBoost/RF showed greatest clinical utility with superior net benefit across clinically relevant thresholds. Lactate availability’s strong predictive contribution suggests that model performance reflects clinician behavior/latent concern in addition to objective data, supporting future incorporation of clinician-gestalt into predictive models. Local recalibration and external validation are needed before implementation. Overall, ML-based early risk stratification in UGIB may serve as a triage-adjunct to support care prioritization, escalation decisions, and resource allocation.