Evaluation of red blood cell transfusion in patients with upper gastrointestinal bleeding using machine learning models
10.13303/j.cjbt.issn.1004-549x.2025.11.003
- VernacularTitle:基于机器学习模型的上消化道出血患者的红细胞输血评估分析
- Author:
Yaoqiang DU
1
;
Biqin ZHANG
2
;
Yilin XU
3
;
Bingyu CHEN
1
;
Weiguo HU
4
Author Information
1. Laboratory Medicine Center, Department of Transfusion Medicine, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou 310014, China
2. Cancer Center, Department of Hematology, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou 310014, China
3. Department of Laboratory Medicine, The First Affiliated Hospital of Ningbo University, Ningbo 315010, China
4. Zaozhuang Blood Center, Zaozhuang, 277102, China
- Publication Type:Journal Article
- Keywords:
upper gastrointestinal bleeding;
AIMS65;
logistic regression;
machine learning;
multivariate transfusion models
- From:
Chinese Journal of Blood Transfusion
2025;38(11):1488-1494
- CountryChina
- Language:Chinese
-
Abstract:
Objective: To comprehensively evaluate and analyze the transfusion outcomes of patients with acute upper gastrointestinal bleeding (UGIB). Methods: The transfusion management system and hospital information system (HIS) were used to retrospectively collect clinical data of 230 patients with UGIB admitted to Zhejiang Provincial People's Hospital and its branches from June 2018 to June 2021. 101 cases were screened and categorized into transfusion group (n=56) and non-transfusion group (n=45) based on transfusion outcomes. The cohort comprised 68 males and 33 females. A univariate model based on the AIMS65 score, a logistic multiple regression model, and multivariate transfusion models using machine learning methods (including Random Forest, Support Vector Machine, and Artificial Neural Network) were established. The sensitivity, specificity, accuracy, and receiver operating characteristic (ROC) curves of each model were compared. Results: For the univariate model based on the AIMS65 scoring, the optimal threshold was 1.5. This model demonstrated a sensitivity of 0.446, a specificity of 0.822, an AUC of 0.67, an accuracy (ACC) of 0.614, a Kappa value of 0.256, and an F1-score of 0.655. For logistics regression model (optimal critical probability: 0.459), the sensitivity was 0.929, specificity was 0.889, AUC was 0.96, ACC was 0.911, Kappa was 0.819, and F1-score was 0.899. For the Random Forest model (optimal critical probability: 0.458), the sensitivity was 0.964, specificity was 0.956, AUC was 0.99, ACC was 0.960, Kappa was 0.920, and F1-score was 0.956. For the Support Vector Machine model (optimal critical probability: 0.474), the sensitivity was 0.875, specificity was 0.933, AUC was 0.94, ACC was 0.901, Kappa was 0.801, and F1-score was 0.894. For the Artificial Neural Network model (optimal critical probability: 0.797), the sensitivity was 0.804, specificity was 0.956, AUC was 0.96, ACC was 0.871, Kappa was 0.745, and F1-score was 0.869. Ten-fold cross validation also confirmed the reliability of the results. Conclusion: Based on integrated various clinical test indicators of patients, we could establish logistic regression model and multiple machine learning models. These models hold significant value for predicting the need for blood transfusion in patients, indicating a promising application prospect for machine learning algorithms in transfusion prediction.
