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  

Exploring the risk factors of blood transfusion in patients with isolated traumatic brain injury based on machine learning prediction models

作　　者：刘威[1,2] 熊紫清 吴承高 乐爱平[1,2] LIU Wei;XIONG Ziqing;WU Chenggao;LE Aiping(The First Affiliated Hospital of Nanchang University;Key Laboratory of Jiangxi Province for Transfusion Medicine,Nanchang 330006,China)

机构地区：[1]南昌大学第一附属医院,江西南昌330006 [2]输血医学江西省重点实验室,江西南昌330006

出　　处：《中国输血杂志》2024年第12期1358-1364,共7页Chinese Journal of Blood Transfusion

基　　金：江西省重点研发计划项目(20192ACB50014)。

摘　　要：目的基于多种机器学习方法探讨单纯性颅脑创伤(isolated traumatic brain injury,iTBI)患者临床输血的风险因素,并建立预测模型,为单纯性颅脑创伤患者临床输血提供合理的指导思路。方法纳入2015年1月1日到2021年6月30日南昌大学第一附属医院的iTBI患者2273例,比较分析输血与未输血患者间生命体征、临床指标和实验室检测指标等变量差异;进而建立6种机器学习模型,通过交叉验证、准确度、特异性、召回率、f1值及ROC曲线下面积比较不同模型的性能,SHAP图用于解释影响iTBI患者临床输血的影响因素。结果本研究共纳入2273例iTBI患者,共有301例患者接受了输血。输血和未输血患者在性别、年龄、心率(HR)、临床诊断、颅骨骨折、治疗方法、失血性休克、格拉斯哥昏迷评分(GCS)、钾(K)、钙(Ca)、凝血酶原时间(PT)、活化部分凝血活酶时间(APTT)、国际标准化比值(INR)、红细胞计数(RBC)、红细胞压积(Hct)、血红蛋白(Hb)和血小板计数(Plt)方面存在显著差异(P<0.05);且输血患者的LOS、并发症发生率、机械通气率、ICU入住率、90d内再次住院率和住院死亡率均高于非输血组(P<0.05)。6种机器学习算法用于模型构建,测试集中模型进行验证结果显示CatBoost模型,表现最好,AUC为0.911。进一步使用SHAP框架对最佳模型CatBoost进行了解释及可视化,结果显示手术治疗、GCS较低、INR较高、Hct较低、低K、低Ca、年龄≥60岁、颅骨骨折以及失血性休克会增加患者输血风险。结论本研究建立了用于预测iTBI患者输血的机器学习模型,CatBoost模型表现最佳。该模型对于识别该人群中输血风险、做出临床输血决策以及监测进展可能是有用且有益的。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 2273 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 varia-bles 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 val-idation,accuracy,specificity,recall,f1 value and area under the ROC curve.The SHAP plot was used to explain the in-fluencing factors of blood transfusion in iTBI patients.Results This study included 2273 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 perform

关 键 词：单纯性颅脑创伤 临床输血 机器学习 风险因素 SHAP图 

分 类 号：R457.1[医药卫生—治疗学] R651.15[医药卫生—临床医学]