基于SMOTE算法的化疗肿瘤患者下呼吸道感染预警模型构建

Construction of early warning model of lower respiratory tract infection in chemotherapy tumor patients based on SMOTE algorithm

目的构建基于少数类样本合成过抽样技术(synthetic minority over-sampling technique,SMOTE)算法的化学治疗(化疗)肿瘤患者下呼吸道感染预警模型。方法共纳入西宁市4所三级医院2019年1月—2021年6月收治的2384例接受化疗的肿瘤患者为研究对象,将所收集病例按照7∶3的比例随机分为建模组1668例和验证组716例,建模组数据用来建立模型,验证组数据对所建立的模型进行验证,利用单因素比较和logistic回归分析筛选下呼吸道感染影响因素,基于SMOTE算法建立化疗肿瘤患者下呼吸道感染预警模型。结果logistic回归分析可得,年龄(x_(1))、身体质量指数(BMI)值是否正常(x_(2))、恶性肿瘤分期(x_(3))、吸烟史(x_(4))、合并糖尿病(x_(5))、合并肺部疾病(x_(6))均是化疗肿瘤患者下呼吸道感染的危险因素(均P<0.01),获得原始数据预警模型:Logit(P)=0.055x_(1)+0.967x_(2)-0.195x_(3)+1.383x_(4)+0.968x_(5)+0.939x_(6)-14.073和基于SMOTE算法的预警模型:Logit(P)=0.090x_(1)+1.092x_(2)-0.249x_(3)+1.724x_(4)+1.136x_(5)+1.344x_(6)-14.859。基于SMOTE算法预警模型AUC为0.949(95%CI:0.937~0.961),高于原始数据预警模型AUC 0.780(95%CI:0.734~0.846)。结论基于SMOTE算法所构建的预警模型能更准确预警化疗肿瘤患者下呼吸道感染,有效解决感染与非感染患者样本数据不平衡所导致的预测误差,基于预测模型可选择相应的对策进行应对。

Objective To construct the early warning model of lower respiratory tract(LRT)infection in chemotherapy tumor patients based on synthetic minority over-sampling technique(SMOTE)algorithm.Methods 2384 tumor patients treated with chemotherapy in 4 tertiary hospitals in Xining City from January 2019 to June 2021 were investigated,patients were randomly divided into modeling group(n=1668)and validation group(n=716)accor-ding to the ratio of 7∶3,data of modeling group was used to construct the model,data of validation group was used to verify the constructed model,influencing factors for LRT infection were screened by univariate comparison and logistic regression analysis,the early warning model of LRT infection of chemotherapy tumor patients was constructed based on SMOTE algorithm.Results Logistic regression analysis showed that age(x_(1)),whether body mass index was normal(BMI,x_(2)),stage of malignant tumor(x_(3)),smoking history(x_(4)),combined diabetes mellitus(x_(5))and combined pulmonary disease(x_(6))were all risk factors for LRT infection in chemotherapy tumor patients(all P<0.01),the original data warning model:Logit(P)=0.055x_(1)+0.967x_(2)-0.195x_(3)+1.383x_(4)+0.968x_(5)+0.939x_(6)-14.073 and early warning model based on SMOTE algorithm:Logit(P)=0.090x_(1)+1.092x_(2)-0.249x_(3)+1.724x_(4)+1.136x_(5)+1.344x_(6)-14.859 were obtained.The AUC of early warning model based on SMOTE algorithm was higher than original data warning model(0.949[95%CI:0.937-0.961]vs 0.780[95%CI:0.734-0.846]).Conclusion The early warning model based on SMOTE algorithm can more accurately warn LRT infection in chemotherapy tumor patients,and effectively solve the warning error caused by the imbalance of the sample data of infected and non-infected patients,the corresponding countermeasures can be selected based on the warning model.