基于混合贝叶斯网络的肺结节鉴别诊断模型  

Differential diagnosis model for pulmonary nodules based on hybrid Bayesian network

作　　者：张博文 刘沛[1] 巢健茜[1,2] 黄子阳 吴雪雨 刘依婷 ZHANG Bo-wen;LIU Pei;CHAO Jian-qian;HUANG Zi-yang;WU Xue-yu;LIU Yi-ting(Department of Epidemiology and Health Statistics,School of Public Health,Southeast University,Nanjing,Jiangsu Province 210009,China;不详)

机构地区：[1]东南大学公共卫生学院流行病与卫生统计学系,江苏省南京210009 [2]东南大学公共卫生学院医疗保险系,江苏省南京210009

出　　处：《中国慢性病预防与控制》2023年第2期81-85,89,共6页Chinese Journal of Prevention and Control of Chronic Diseases

基　　金：国家自然科学基金项目(81872711)。

摘　　要：目的 探讨肺结节良恶性病变的影响因素,建立肺结节鉴别诊断的混合贝叶斯网络模型,为肺癌的筛查和诊治提供参考依据。方法 收集2014—2020年南京市某医院胸心外科和呼吸科经胸部CT检查发现的肺结节患者的人口统计学信息、疾病史、危险因素暴露史、影像学特征、肿瘤标志物检测结果和肺结节诊断结果等资料。利用SAS 9.4软件进行单因素和多因素分析,初步筛选变量。将数据按7∶3的比例划分为训练集和测试集,分别用于模型构建和模型评估。使用L_DVBN算法构建混合贝叶斯网络模型,使用R语言bnlearn包进行模型评估,通过受试者工作特征(ROC)曲线下面积(AUC)评价模型的预测效果,并与Mayo模型和Brock模型比较。结果 该研究共纳入样本990例,其中肺结节恶性病变患者665例(66.16%)。混合贝叶斯网络结构显示,肺结节性质与结核病史、最大结节位置、直径、类型、分叶征、空泡征、血管集束征、钙化征,细胞角蛋白19片段(Cyfra21-1)等因素密切相关,与年龄、肺结节数量、毛刺征等因素间接相关;模型AUC值为0.869(95%CI:0.823~0.915),高于Mayo模型(AUC=0.704,95%CI:0.642~0.767)和Brock模型(AUC=0.754,95%CI:0.691~0.816)。结论 肺结节良恶性鉴别诊断的混合贝叶斯网络模型具有良好的预测性能,能有效区分良性结节和恶性结节,可辅助鉴别肺癌低剂量螺旋CT(LDCT)筛查中的难区分结节。Objective To explore the influencing factors of pulmonary nodules with malignant change,establish the hybrid Bayesian network model for pulmonary nodule differential diagnosis,and provide the reference for lung cancer screening and diagnosis or treatment. Methods The demographic information,past medical history,risk exposure history,imaging features,tumour biomarkers detection result and diagnosis result of pulmonary nodules detected by chest CT examination between 2014 and2020 in the department of cardiothoracic surgery and respiratory medicine of a hospital in Nanjing were collected. Univariate analysis and multivariate analysis were used to screen the variables,the used software was SAS 9.4. The data were divided into training set and testing set by 7∶3,which were used to model establishment and model assessment. L_DVBN algorithm was used to construct hybrid Bayesian network, R language bnlearn package was used to assess the model, area under curve( AUC) of receiver operating characteristic(ROC) was used to assess the prediction results of model,which was prepared with Mayo model and Brock model. Results The study included 990 cases with 665 malignant pulmonary nodules cases(66.16%). The structure of the hybrid Bayesian network showed that history of tuberculosis,nodule location,diameter and type,sign of lobulation,cavitation,vascular convergence and calcification,and Cyfra21-1 were related to the diagnosis of pulmonary nodules;the feature of pulmonary nodules was indirectly related to age,nodule number and spicule sign;the AUC value of the proposed model was 0.869(95%CI:0.823-0.915),which was significantly higher than those of Mayo model(AUC=0.704,95%CI:0.642-0.767) and Brock model(AUC =0.754,95% CI:0.691-0.816). Conclusion The hybrid Bayesian network of differentiating the benign and malignant pulmonary nodules has better prediction function and distinguishes effectively the benign pulmonary nodules from malignant nodules,which is potential to assist in differential diagnosis in low-dose computed tomograph

关 键 词：肺结节 鉴别诊断 混合贝叶斯网络 

分 类 号：R563[医药卫生—呼吸系统]