Multi-task machine learning for transfusion decision support in acute upper gastrointestinal bleeding: a novel
Qiongjie Li1, Guolin Chen2, Qun Li3
1Department of Infectious Diseases, First Hospital of Shanxi Medical University, Taiyuan, China.
Journal of Translational Medicine
|September 2, 2025
Summary
This study introduces a multi-task learning model that accurately predicts blood transfusion needs and volume for acute upper gastrointestinal bleeding (AUGIB) patients, outperforming current methods for better clinical decisions.
Area of Science:
- Emergency Medicine
- Clinical Data Science
- Artificial Intelligence in Healthcare
Background:
- Acute upper gastrointestinal bleeding (AUGIB) presents a significant clinical challenge requiring timely and accurate transfusion management.
- Existing scoring systems for transfusion decisions in AUGIB often lack precision, leading to suboptimal patient care.
- Multi-task learning (MTL) offers a promising approach to simultaneously address multiple related prediction tasks in clinical settings.
Purpose of the Study:
- To develop and validate a novel multi-task learning (MTL) model for predicting the necessity of blood transfusion in AUGIB patients.
- To estimate the appropriate type and volume of blood products required for transfusion in AUGIB.
- To enhance clinical decision-making and support precision transfusion strategies in emergency care.
Main Methods:
- Retrospective collection of clinical data from 1256 AUGIB emergency patients, with external validation on 209 patients.
- Implementation of an MTL model integrating oversampling and distribution correction for data imbalance.
- Utilized a soft-voting ensemble (CatBoost, XGBoost) for classification and a stacked regressor (Random Forest, XGBoost) for regression, optimized with a dynamic loss function.
Main Results:
- The MTL model achieved an AUC of 0.965 for transfusion prediction, a 20.5% improvement over the Glasgow-Blatchford Score (GBS).
- The two-stage stacked regressor significantly reduced prediction errors for transfusion type and volume compared to other machine learning models.
- External validation demonstrated favorable generalization performance with an AUC of 0.860.
Conclusions:
- The developed MTL model is robust, interpretable, and superior to existing methods for transfusion decision-making in AUGIB.
- Joint optimization of classification and regression tasks through hierarchical feature selection and dynamic loss allocation enables precision transfusion.
- The model shows significant potential for widespread application in emergency and critical care settings.
