构建肺结核患者抗结核治疗后发生白细胞减少症的预测模型  

A predictive model for leukopenia in tuberculosis patients receiving anti-tuberculosis treatment

作　　者：卢斌 施云珍 吴丽华 潘新灵 陈翔[2] Lu Bin;Shi Yunzhen;Wu Lihua;Pan Xinling;Chen Xiang(Department of Infectious Diseases,Affiliated Dongyang Hospital of Wenzhou Medical University,Dongyang 322100,Zhejiang Province,China;Department of Biomedical Sciences Laboratory,Affiliated Dongyang Hospital of Wenzhou Medical University,Dongyang 322100,Zhejiang Province,China)

机构地区：[1]温州医科大学附属东阳市人民医院感染病科,东阳322100 [2]温州医科大学附属东阳市人民医院生物医学实验室,东阳322100

出　　处：《中华临床感染病杂志》2024年第5期375-382,共8页Chinese Journal of Clinical Infectious Diseases

基　　金：浙江省医药卫生科技计划(2024XY083)。

摘　　要：目的探讨肺结核患者抗结核治疗后发生白细胞减少症的风险因素,并构建预测其发生的列线图模型。方法选择2013年1月至2024年6月温州医科大学附属东阳市人民医院收治的2681例肺结核患者作为研究对象,所有患者均接受一线抗结核治疗方案,按照7∶3的比例用分层随机抽样方法分为建模组(n=1876)和验证组(n=805)。以抗结核治疗后发生白细胞减少症作为结局指标,在建模组中,采用最小绝对收缩和选择算子(Lasso)回归,以及多因素Logistic回归筛选出风险因素后建立列线图模型。采用受试者工作特征曲线下面积(the area under the receiver operating curve,AUC)、校准曲线及决策曲线分别评估模型的区分度、校准度及临床适用性。另外,通过验证人群建立的Logistic回归模型与随机森林(random forest,RF)、支持向量机(support vector machine,SVM)、极限梯度提升(extreme gradient boosting,XGBoost)和朴素贝叶斯(naive Bayes,NB)这4种机器学习模型进行效能上的比较。结果建模组和验证组白细胞减少症的发生率分别为15.0%(273/1876)和15.9%(128/805)。经Lasso筛选变量和多因素Logistic回归分析,年龄≥65岁(OR=2.997,95%CI 2.185~4.128),有饮酒史(OR=4.803,95%CI 3.502~6.593),有糖尿病史(OR=5.459,95%CI 3.914~7.621)是抗结核治疗后发生白细胞减少症的危险因素,而高血红蛋白水平(OR=0.979,95%CI 0.971~0.987)和高血小板计数(OR=0.996,95%CI 0.995~0.998)是保护性因素。基于以上5个风险因素建立的模型在建模组和验证组的AUC值分别为0.836(95%CI 0.810~0.863)和0.818(95%CI 0.776~0.860),校准曲线和决策曲线显示模型校准度和临床适用性良好,该模型与4种机器学习模型预测能力相当(P>0.05)。结论本研究构建的列线图模型具有良好的区分度、校准度及临床适用性,可为临床预防抗结核后发生白细胞减少症提供依据。Objective To construct a nomogram model for predicting the risk of leukopenia among tuberculosis patients receiving anti-tuberculosis therapy.Methods A total of 2681 tuberculosis patients admitted to the affiliated Dongyang Hospital of Wenzhou Medical University from Jan 2013 to Jun 2024,were enrolled in this study.All cases received first line anti-tuberculosis treatment and were randomly divided into training(n=1876)and validation groups(n=805)at a ratio of 7∶3.The endpoint was the occurrence of leukopenia during anti-tuberculosis therapy.In the training group,the predictors were screened by Lasso regression and multivariable Logistic regression analysis,and used to establish a nomogram prediction model.The discrimination power,fitness and clinical applicability were evaluated using the receiver operating characteristic(ROC)curve,calibration curve and decision curve analysis,respectively.Several machine learning models based on different methods(random forest,support vector machine,extreme gradient boosting and naive Bayes)were also constructed in the validation group.Results There were 15.0%(273/1876)and 15.9%(128/805)of cases developing leukopenia during anti-tuberculosis therapy in the training group and validation groups,respectively.Following Lasso regression analysis,the multivariable Logistic regression analysis showed that age≥65 years(OR=2.997,95%CI 2.185-4.128),alcohol consumption(OR=4.803,95%CI 3.502-6.593)and diabetes(OR=5.459,95%CI 3.914-7.621)were risk factors related to the occurrence of leukopenia;while the higher levels of baseline hemoglobin(OR=0.979,95%CI 0.971-0.987)and platelet count(OR=0.996,95%CI 0.995-0.998)were protective factors.Based on these five factors,a nomogram prediction model was developed.The areas under ROC curve(AUCs)were 0.836(95%CI 0.810-0.863)and 0.818(95%CI 0.776-0.860)in the training group and the validation group,respectively.Moreover,this model had good fitness and clinical applicability.The discrimination power of nomogram model was comparable to those of machine

关 键 词：肺结核 抗结核治疗 白细胞减少症 预测模型 机器学习模型 

分 类 号：R73[医药卫生—肿瘤]