The transfusion decision system used a 2-stage framework termed MCGB. A calibrated classifier first estimates the probability of transfusion using monotonic constraints on prespecified laboratory and physiological variables to encode clinically consistent directionality, together with a stability-screened interaction allowlist. To enhance robustness across settings, training optimizes a group distributionally robust objective over hospital, etiology, and admission-period strata; predicted probabilities are calibrated by temperature scaling, and the operating threshold is selected with decision-curve analysis to align classification with clinical utility [27,28]. Patients predicted to require transfusion are then modeled with a regression component that estimates blood units via quantile predictions with conformal adjustment, yielding 95% prediction intervals to communicate uncertainty for blood-allocation planning [29,30]. All preprocessing—Winsorization at the first and 99th percentiles, median imputation, indicator encoding, and interaction construction—is confined to cross-validation folds and applied only to the corresponding calibration and test splits to prevent information leakage. Candidate interactions arise from continuous-by-etiology and continuous-by-continuous pairs, are filtered for high correlation, ranked by mutual information and distance correlation, and retained when they meet prespecified stability criteria, following a screen-in principle that preserves interpretability, reproducibility, and transparency; the overall workflow is shown in Figure 1.

The study protocol was reviewed and approved by the institutional ethics committees of Daping Hospital, Army Medical Center of PLA, Army Medical University; the Third People’s Hospital of Chongqing; and the 13th People’s Hospital of Chongqing. The requirement for individual informed consent was waived because the analysis used routinely collected EHR data that were deidentified before analysis and posed minimal risk to participants. Data were handled in secure, access-controlled environments under institutional data-use agreements; only coded study identifiers were used, and direct personal identifiers were not accessible to investigators. All procedures complied with the Declaration of Helsinki and applicable local regulations. The project did not involve any prospective intervention and was not registered as a clinical trial.

Routinely collected data were retrieved from the clinical data warehouses of 3 hospitals in Chongqing, China. Consecutive adults with endoscopically confirmed UGIB between January 2019 and August 2025 were screened. A total of 849 patients met the inclusion criteria, including 655 nontransfusion cases and 194 transfusion cases. The cohort comprised 264 patients from the Third Affiliated Hospital of the Third Military Medical University, 326 patients from the Third People’s Hospital of Chongqing, and 259 patients from the 13th People’s Hospital of Chongqing. Exclusion criteria included insufficient clinical data, non-UGIB bleeding, incomplete diagnostic or therapeutic care, self-discharge, transfusion before the index decision, and missing key variables required for model development. The study was approved by the institutional ethics committees of all 3 hospitals, with a waiver of informed consent due to the retrospective design.

Predictors available within 24 hours before the transfusion decision were extracted from the EHR and included demographics, first recorded vital signs at presentation, and initial laboratory results obtained prior to the transfusion decision (covering hematology, coagulation, renal function, albumin, and electrolytes), as well as clinician-adjudicated bleeding etiology encoded as 5 binary indicators. Units were harmonized across centers, and variables not available before the transfusion decision were excluded to prevent data leakage. Patients with missing key variables required for modeling were excluded during cohort construction, resulting in a largely complete analytic dataset. Continuous variables were processed using a median imputation strategy as part of a standardized preprocessing pipeline, although imputation was rarely required in practice due to the near-complete nature of the data. Categorical variables were one-hot encoded, and missingness indicators for selected laboratory tests were constructed to preserve compatibility with potential real-world scenarios in which data may be incomplete. Pairwise correlations were summarized to characterize relationships among predictors while preserving clinically relevant variables. For descriptive comparisons, transfused and nontransfused groups were analyzed using Student t tests or nonparametric equivalents for continuous variables and chi-square or Fisher exact tests for categorical variables; 2-sided P values are reported in Table 1.

To further characterize the relationships among predictors and enhance transparency in model interpretation, pairwise correlations were first examined using a heat map Figure 2A, providing an overview of feature dependencies while retaining clinically relevant variables. Building on this analysis, model outputs were subsequently examined using Shapley additive explanations (SHAP), as illustrated in Figure 2B. Each point represents an individual patient from the evaluation set, and the horizontal axis indicates the SHAP value, which quantifies the contribution of each feature to the predicted probability of transfusion [31,32]. Positive SHAP values indicate an increased predicted risk, whereas negative values indicate a decreased risk, and the color represents the original feature value, with red denoting higher values and blue denoting lower values. Key predictors showed clinically coherent patterns: lower hemoglobin and hematocrit levels were associated with a higher predicted probability of transfusion, whereas higher values shifted predictions toward lower risk. Coagulation-related indices and hemodynamic variables also exhibited consistent directional effects that aligned with established clinical understanding of bleeding severity and transfusion requirements. Overall, these findings indicate that the model captures clinically meaningful and plausible relationships among predictors, supporting the interpretability and reliability of the MCGB classifier.

In this study, the MCGB framework is proposed to extend standard gradient boosting by integrating 3 clinically motivated components: (1) monotonic constraints to enforce clinically consistent relationships, (2) interaction whitelisting to restrict feature interactions to stable and interpretable pairs, and (3) subgroup-aware reweighting to improve robustness across heterogeneous populations. These extensions are designed to improve both predictive performance and clinical reliability.

The dataset is represented as D=(xi,yi), where xi∈Rm denotes the input features, yi∈R denotes the corresponding outcome, and i=1,2,…,n indexes the samples. Following the standard gradient boosting formulation, the boosted predictor after T rounds can be expressed as:

where F denotes the space of regression trees mapping Rm to R. Each tree partitions the data into leaves and assigns a score wj. The boosting procedure iteratively refines predictions through stage-wise optimization, while the proposed extensions introduce clinically informed constraints and reweighting mechanisms to enhance robustness and consistency. The optimization objective at iteration t is expressed as:

where γ controls model complexity by penalizing the number of leaves, and λ is an L2 regularization parameter on leaf weights.

A second-order Taylor expansion is applied around Ft-1(xi). With first- and second-order derivatives defined asgi=∂y^l(yi,y^)|y^=Ft-1(xi) and hi=∂y^2l(yi,y^)|y^=Ft-1(xi), the approximated loss becomes:

To improve robustness across heterogeneous subgroups, MCGB introduces subgroup-aware reweighting of gradients. Specifically, for each predefined subgroup g, a softmax-based weighting scheme is defined as:

where Lg denotes the loss associated with group g, and η§amp;gt;0 is a temperature parameter that controls the sharpness of the weighting distribution. For each sample i, let g(i) denote the group to which the sample belongs. The reweighted derivatives are then defined as g~i=πg(i)(η) giand h~i=πg(i)(η) hi, where gi and hi are the original first- and second-order derivatives. This design prevents the model from being dominated by majority subgroups and reduces overfitting to site-specific patterns.

At the tree level, gradients are aggregated within each leaf node. For a leaf j with the instance set Ij, the sufficient statistics are defined as Gj=∑i∈Ij g~i and Hj=∑i∈Ij h~i. The quadratic surrogate objective is then approximated as:

The unconstrained optimal leaf weight is given by w^j=-GjHj+λ. In MCGB, this solution is further adjusted to satisfy clinical monotonicity constraints of the form ∂FT(x)∂xkσk≥0, where k∈M denotes features subject to monotonic constraints and σk∈{-1,+1} specifies the expected direction. The feasible solution is obtained by projecting w^ onto the isotonic cone:

ensuring that model predictions follow clinically expected trends and preventing implausible extrapolations, thereby improving generalization and trustworthiness.

The reduction in the surrogate objective when splitting a node into left and right children is quantified by the gain, defined as:

To further enhance interpretability and reproducibility, MCGB restricts candidate splits using an interaction whitelist W. To construct this whitelist, a stability-driven selection procedure is employed. Candidate feature pairs are first evaluated using the Friedman H-statistic to quantify interaction strength. Pairs with a median H-statistic greater than 0.10 across cross-validation folds are retained. To ensure robustness, only interactions that rank among the top five strongest interactions in at least four out of five folds are selected. This process ensures that retained interactions are both statistically stable and clinically plausible. If a node path already includes feature p, then a split on feature q is admissible only if (p, q)∈W. The whitelist itself is defined by a reproducibility criterion,

Where Hpq(r) is the Friedman H-statistic in fold r and Top5pq is the number of folds in which (p, q) ranks among the 5 strongest interactions. This stability-based selection ensures robustness and reproducibility.

In summary, the MCGB framework extends classical gradient boosting by incorporating 3 clinically motivated components: monotonicity constraints to enforce clinically consistent relationships, interaction whitelisting to restrict feature interactions to stable and interpretable pairs, and group-wise loss reweighting to improve robustness across heterogeneous clinical subgroups. Together, these components enable the model to achieve not only strong predictive performance but also improved interpretability and stability in real-world clinical settings.

To evaluate the effectiveness of the proposed framework, all comparator models were implemented using standard libraries (scikit-learn, extreme gradient boosting [XGBoost], and categorical boosting [CatBoost]) and evaluated under the same cross-site hold-out design as the proposed MCGB model. Within the development cohort, each comparator was trained using the same stratified 5-fold cross-validation splits and fold-specific preprocessing pipeline, and the resulting cross-validation–derived models were subsequently applied to the independent test cohort. This design ensured a fair comparison while maintaining consistent separation between model development and final evaluation. For the classification task, the compared models included logistic regression (LogReg), multilayer perceptron (MLP), random forest (RF), adaptive boosting (AdaBoost), XGBoost [32], and CatBoost [33]; for the regression task, the compared models included linear regression (LR), MLP [34], RF, AdaBoost, XGBoost, and CatBoost. Hyperparameters for all baseline models were tuned via grid search within the training folds only, with model selection based on validation performance.

Model development and validation were conducted using a predefined cross-site hold-out design to assess generalizability. Data from the Third People’s Hospital of Chongqing and the 13th People’s Hospital of Chongqing were combined as the development cohort, whereas data from the Third Affiliated Hospital of the Third Military Medical University were held out as an independent test cohort. The independent test cohort was not used for preprocessing fitting, feature interaction screening, hyperparameter tuning, calibration, threshold selection, or model selection. Within the development cohort, stratified 5-fold cross-validation was used to preserve the transfusion event rate across folds and to perform model development and internal validation. In each fold, preprocessing steps, including median imputation, one-hot encoding, and interaction construction, were fitted exclusively on the training split and then applied to the corresponding validation fold and the independent test cohort, ensuring strict separation and preventing information leakage. Each fold yielded a cross-validation–derived model, and final performance on the independent test cohort was summarized as the mean (SD) across the 5 models. To further examine cross-site robustness, hospital-wise alternating external testing was conducted as a supplementary analysis, in which each hospital was alternately treated as the external test cohort and the remaining 2 hospitals were used for model development. This supplementary analysis was used only to evaluate institutional stability and was not mixed with the primary independent test evaluation.

Given the moderate class imbalance (655 nontransfusion vs 194 transfusion cases), no explicit resampling was performed in order to preserve the original clinical distribution. Model evaluation therefore emphasized metrics robust to class imbalance, including the area under the precision-recall curve (AUPRC) alongside the area under the receiver operating characteristic curve (AUROC). Predicted probabilities were calibrated using temperature scaling within the development cohort, and the learned calibration parameters were then applied unchanged to the independent test cohort. Model performance was primarily assessed across a range of clinically relevant threshold probabilities using decision curve analysis to reflect potential clinical utility. Results at a reference threshold of 0.50 were additionally reported as a conventional operating point for comparison with prior binary prediction studies. For the interaction ablation analysis, net benefit was additionally evaluated at a threshold probability of 0.20 because this value lies within the low-to-moderate risk range in which clinicians may consider early preparation, closer monitoring, or cross-matching rather than immediate transfusion. Therefore, ΔNB@0.20 was used to assess whether the retained interactions improved clinical utility in an early decision-support setting and was not intended to replace the 0.50 reference threshold used for reporting primary classification performance. Primary discrimination metrics included AUROC and AUPRC, whereas secondary metrics included sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F₁-score, and expected calibration error.

For the regression component among transfused patients, predictive accuracy was evaluated using mean absolute error (MAE) and root-mean-square error, explanatory performance using R², and uncertainty using empirical coverage and interval width of 95% prediction intervals derived from conformalized quantile prediction. Agreement between predicted and observed transfusion dose was further assessed using Bland–Altman analysis.

This study shows that a clinically constrained, 2-stage framework can provide end-to-end decision support for transfusion in UGIB. The classifier delivered very high discrimination together with good calibration, and it maintained performance across hospitals, etiologies, and age strata. These characteristics matter more than a single accuracy number: calibration enables risk thresholds to be interpreted as probabilities, subgroup stability reduces the chance that performance collapses when case mix shifts, and explicit operating thresholds translate model output into consistent actions. In addition, calibration performance is reflected not only in visual agreement but also in quantitative metrics such as expected calibration error and Brier score, which are critical for clinical decision-making [35]. The innovation is not only the use of gradient boosting but the way clinical knowledge is built into the learning process. Monotonic constraints encode clinically expected relationships, and interaction constraints improve interpretability. In the second stage, dose regression provides point estimates with uncertainty, enabling actionable planning rather than binary decisions. In the primary cross-site hold-out evaluation, MCGB maintained strong discrimination and calibration when applied to an independent institutional test cohort.

Traditional UGIB scores are useful for triage but do not provide patient-level probabilities of transfusion or quantitative guidance on dose. Many machine-learning studies report high AUROC yet give little attention to calibration, interpretability, or robustness across sites, and most stop at binary classification [36]. The present work addresses those gaps through several design choices. First, clinical constraints are imposed at training time rather than post hoc, which prevents clinically nonsensical relationships from emerging and reduces the need for manual rule fixes later. Second, interaction curation is principled: interactions such as INR × PT and systolic blood pressure × pulse are retained because they are stable across folds and supported by domain knowledge, and ablation confirms that they improve both calibration and net benefit. Third, evaluation extends beyond global curves to threshold performance, decision-curve analysis, and subgroup reporting, which are the quantities clinicians use when deciding whether and how to act. Finally, the dose module elevates the scope of decision support by estimating the quantity of blood required with explicit uncertainty, which is rarely attempted and directly relevant to operations [37,38].

The proposed MCGB framework is designed to support bedside and blood-bank decision-making using routinely collected clinical data. At presentation, the calibrated classifier provides an individualized probability of transfusion need, and a prespecified threshold derived from decision curve analysis encodes the trade-off between missed transfusion and unnecessary cross-matching, enabling standardized triage and reducing subjective variability [39]. For patients exceeding this threshold, the dose model generates both a unit estimate and a 95% prediction interval, supporting clinical communication, cross-matching, and inventory planning [40]. These outputs can be delivered through the interactive UGIB transfusion recommendation system (Figure 6) and integrated into electronic health record systems to enable automated data extraction and real-time risk assessment without disrupting existing workflows. To support safe use, model outputs are intended as decision support rather than definitive recommendations, with probability estimates and uncertainty information provided to assist clinician interpretation and reduce the risk of automation bias. In practice, implementation also requires careful configuration of thresholds and alerts to align with clinical pathways and avoid alert fatigue [41]. Finally, as a clinical decision support tool, MCGB would require prospective validation, external evaluation, and compliance with relevant regulatory frameworks prior to deployment, with clinician oversight and transparent documentation to ensure accountability and safe integration into routine care.

The analysis is retrospective and limited to 3 hospitals within a single regional health care system, which may restrict generalizability to other settings and populations; external and temporal validation is needed to assess transportability and inform recalibration strategies. The dataset is imbalanced with fewer transfusion events, and larger, more balanced cohorts are warranted. Some subgroup strata are relatively small, increasing uncertainty in stratum-specific estimates, although CIs suggest overall stability. In addition, patients with missing key variables were excluded during cohort construction, resulting in model development and validation on a largely complete analytic dataset. While this improves internal consistency, it may not fully reflect real-world deployment conditions, where laboratory measurements may be delayed, unavailable, or selectively ordered at the time of decision-making, requiring explicit strategies for incomplete inputs such as minimum-data requirements or clinician review. Furthermore, thromboelastography and other advanced coagulation indices were not included because they are not routinely available across centers. While such variables may provide additional physiological information, the model was intentionally developed using routinely collected clinical features to enhance generalizability and real-world applicability, and the strong performance observed suggests that commonly available variables capture the majority of clinically relevant signals for transfusion decision-making [42]. Future work should evaluate whether incorporating advanced coagulation markers can further improve performance in settings where such data are available. Although conformal prediction achieved near-nominal coverage, the clinical utility of interval width requires prospective validation. Finally, deployment will require human-in-the-loop safeguards, appropriate alert configuration, and attention to automation bias to ensure safe and effective clinical use.

This study developed a clinically constrained, 2-stage framework MCGB—to support transfusion decisions in UGIB. By incorporating monotonic constraints that encode established clinical directionality and retaining only stability-screened, clinically plausible interactions, MCGB achieved superior discrimination, reliable calibration, and interpretable predictions. Beyond classification of transfusion need, the framework extends to individualized dose estimation with well-calibrated prediction intervals, offering actionable guidance for both bedside decision-making and blood-bank resource planning. These results demonstrate that MCGB can bridge methodological advances in machine learning with clinical requirements for transparency, robustness, and operational relevance, highlighting its potential as a practical decision-support tool for UGIB management.