JMIR Med Inform. 2026 Aug 4;14:e83889. doi: 10.2196/83889.
ABSTRACT
BACKGROUND: Transfusion thresholds in upper gastrointestinal bleeding are debated; hemoglobin cutoffs of 70-80 g/L are widely cited yet inconsistently applied. Common risk scores offer limited individualized guidance and rarely provide calibrated, interpretable predictions for transfusion decisions.
OBJECTIVE: This study aimed to develop and validate a two-stage, clinically constrained gradient-boosting framework (Medically Constrained Gradient Boosting [MCGB]) that predicts transfusion need and estimates transfusion dose with quantified uncertainty and to implement a prototype recommendation system for clinical use.
METHODS: We analyzed a retrospective multicenter cohort of 849 adults with endoscopically confirmed upper gastrointestinal bleeding admitted to 3 hospitals in Chongqing, China (January 2019 to August 2025). Predictors available before the transfusion decision included demographics, first recorded vital signs, initial laboratory indices, and clinician-adjudicated etiology. Stage 1 used a calibrated classifier with prespecified monotonic constraints and stability-screened, clinically justified interactions. Stage 2 modeled transfusion dose via quantile predictions with conformal adjustment to generate 95% prediction intervals. Performance was assessed using a cross-site hold-out design. Overall, 2 hospitals were used as the development cohort, within which stratified 5-fold cross-validation was performed for model development, hyperparameter tuning, interaction screening, and calibration. The remaining hospital was held out as an independent test cohort for final evaluation. Hospital-wise alternating external testing was further conducted as a supplementary robustness analysis to assess performance stability across institutions. Classification performance was evaluated using discrimination metrics (area under the receiver operating characteristic curve and area under the precision-recall curve), calibration metrics, and decision-curve analysis; regression performance was evaluated using R², mean absolute error, and prediction-interval coverage. A graphical user interface was implemented to enable clinicians to input patient data and obtain calibrated predictions of transfusion probability and corresponding dose recommendations.
RESULTS: MCGB achieved strong discrimination and good calibration across subgroups (area under the receiver operating characteristic curve=0.97 and area under the precision-recall curve=0.91). At a reference probability threshold of .50, sensitivity, specificity, and F1-scores were 0.99, 0.87, and 0.85, respectively, providing a representative operating point for comparison. For dose prediction among transfused patients, MCGB achieved R² of 0.95 and mean absolute error 0.04; 95% prediction-interval coverage was 0.94, indicating accurate point estimates with reliable uncertainty quantification. The software prototype further demonstrated feasibility of real-time decision support at the bedside.
CONCLUSIONS: MCGB provides calibrated, interpretable predictions of transfusion need and individualized dose in upper gastrointestinal bleeding and may support bedside decision-making and blood-bank planning, with a prototype interface demonstrating potential for clinical deployment. External validation in additional settings is warranted to confirm generalizability.
PMID:42550984 | DOI:10.2196/83889