基于机器学习算法建立个体化手术用血预测模型

Development of an individualized prediction model of allogenic blood transfusion in elective patients based on machine learning

目的建立基于机器学习算法的围手术期患者用血预测模型,以指导临床医生合理地做好患者术前备血。方法从本院信息中心建设的大数据集成平台的数据中心采集2012年~2018年所有手术患者的相关数据,并基于Python V3.8.0构建手术用血数据库;利用Excel、SAS统计学软件处理和分析所有数据,并进一步以SPSS Modeler 18.0软件构建手术用血预测模型。结果1)术中输血风险因素评估:术前HB、BMI与术中输血率呈负相关关系,Pearson相关系数(r)分别为:-0.168,-0.046;年龄<1岁患者的输血率(15.63%)最高;女性手术患者输血率高于男性(P>0.05),但其人均输血量低于男性(P<0.01;心脏手术输血率(11.38%)最高。2)所构建的用血预测模型对应的测试集AUC范围0.67~0.88,当AUC达到最高值时,模型9命中率达10.7%,查全率85.76%,特异性75.4%。3)影响该手术用血预测模型构建的主要权重因素:患者体重、Hb、总蛋白(TP)等。结论成功构建的围手术期患者用血预测模型预测效能优于现有的MSBOS。

Objective To develop a prediction model of allogenic blood transfusion in elective patients based on machine learning,so as to guide clinicians to prepare blood for perioperative patients more reasonably.Methods Relevant data of all surgical patients from 2012 to 2018 were extracted from the big data integration platform of our hospital,to construct the surgical blood database based on Python V3.8.0.All data were analyzed using Excel and SAS,and the prediction model was developed based on SPSS Modeler 18.0.Results 1)There was a negative correlation between preoperative Hb and BMI and intraoperative blood transfusion rate,with Pearson correlation coefficient(R)as-0.168 and-0.046,respectively.The transfusion rate of patients under 1 year old was the highest,up to 15.63%.The transfusion rate of female patients was higher than that of male patients(P>0.05),as cardiac surgery rated at the highest 11.38%,but their per capita blood transfusion was lower than that of males(P<0.01).2)The AUC range corresponding to the prediction model for transfusion probability was 0.67~0.88,and when the AUC reached the highest,the hit ratio,coverage rate and specificity of Model 9 was 10.7%,85.76%and 75.4%,respectively.3)The main factors contributing to the prediction model for transfusion volume in surgery were weight,Hb,total protein(TP),etc.Conclusion The prediction efficiency of the successfully constructed prediction model for perioperative blood use was better than that of MSBOS.