Exploring the risk factors of blood transfusion in patients with isolated traumatic brain injury based on machine learning prediction models
10.13303/j.cjbt.issn.1004-549x.2024.12.003
- VernacularTitle:基于机器学习预测模型探究单纯性颅脑创伤患者临床输血的风险因素
- Author:
Wei LIU
1
,
2
;
Ziqing XIONG
1
,
2
;
Chenggao WU
1
,
2
;
Aiping LE
1
,
2
Author Information
1. The First Affiliated Hospital of Nanchang University
2. Key Laboratory of Jiangxi Province for Transfusion Medicine, Nanchang 330006, China
- Publication Type:Journal Article
- Keywords:
isolated traumatic brain injury;
blood transfusion;
machine learning;
risk factors;
SHAP plot
- From:
Chinese Journal of Blood Transfusion
2024;37(12):1358-1364
- CountryChina
- Language:Chinese
-
Abstract:
[Abstract] [Objective] To explore the risk factors of blood transfusion in patients with isolated traumatic brain injury (iTBI) based on multiple machine learning methods, so as to establish a predictive model to provide reasonable guidance for blood transfusion in patients with iTBI. [Methods] A total of 2 273 patients with iTBI from the First Affiliated Hospital of Nanchang University from January 1, 2015 to June 30, 2021 were included to compare and analyze the differences in variables such as vital signs, clinical indicators and laboratory testing indicators between transfusion and non transfusion patients. Furthermore, six machine learning models were established to compare the performance of different models through cross validation, accuracy, specificity, recall, f1 value and area under the ROC curve. The SHAP plot was used to explain the influencing factors of blood transfusion in iTBI patients. [Results] This study included 2 273 iTBI patients, with a total of 301 patients receiving blood transfusions. There were significant differences (P<0.05) in gender, age, HR, clinical diagnosis, skull fracture, treatment methods, hemorrhagic shock, GCS, K, Ca, PT, APTT, INR, RBC, Hct, Hb and Plt between transfusion and non transfusion patients; Moreover, the LOS, incidence of complications, mechanical ventilation rate, ICU admission rate, readmission rate within 90 days and in-hospital mortality rate of transfusion patients were all higher than those of the non transfusion group (P<0.05). Six machine learning algorithms were used for model construction, and the validation results on the test set showed that the CatBoost model performed the best with an AUC of 0.911. Furthermore, the SHAP framework was used to explain and visualize the optimal model CatBoost, showing that surgical treatment, lower GCS, higher INR, lower Hct, lower K, lower Ca, age ≥60 years, skull fractures and hemorrhagic shock increase the risk of blood transfusion in patients. [Conclusion] This study established a machine learning model for predicting blood transfusion in iTBI patients, and the CatBoost model performed the best. This model may be useful and beneficial for identifying transfusion risks in this population, making clinical transfusion decisions and monitoring progress.
