基于贝叶斯网络的老年高血压患者并发冠心病的预测模型构建

Construction of predictive model for coronary heart disease among elderly patients with hypertension on the basis of Bayesian network

目的构建老年高血压患者并发冠心病的预测模型,探讨相关影响因素,为预防和控制老年高血压患者并发冠心病提供参考依据。方法本研究以2013年3-12月在南京市秦淮区某二级医院进行健康体检的所有社区(共约75个社区)老年人为总体,采用整群随机抽样方法,以社区为单位随机抽取10个社区中所有≥60岁长期居住于南京市的老年高血压患者1 568例作为研究对象。进行问卷调查、体格检查和实验室检测,收集一般情况、慢性病患病情况、行为生活方式以及相关检查指标基线资料。分别于2015年、2017年和2019年进行3次随访,随访内容主要为心血管疾病发病情况及其他慢性病患病情况。采用SPSS 20.0软件进行χ^(2)检验,采用多因素非条件logistic回归进行模型变量的初步筛选。使用R 3.5.2软件的"bnlearn"包构建贝叶斯网络模型,并使用Netica软件进行贝叶斯网络推理。通过绘制受试者工作特征(ROC)曲线,计算灵敏度、特异度和曲线下面积(AUC)评价模型的预测效果。结果最终共随访到有效样本1 539例,随访期间冠心病的发生率为6.8%。多因素非条件logistic回归分析初步筛选出影响高血压患者是否并发冠心病的因素包括年龄(OR=1.996)、收缩压(OR=1.877)、总胆固醇(OR=2.346)、高尿酸血症(OR=1.887)、超重及肥胖(OR=1.759)、糖尿病(OR=2.370)和吸烟(OR=1.926),均有统计学意义(P<0.05,P<0.01)。结合多因素logistic回归的结果和先验理论,贝叶斯网络模型最终共纳入9个节点,分别为年龄、性别、收缩压、总胆固醇、高尿酸血症、超重及肥胖、糖尿病、吸烟和冠心病,其中包括4个直接相关节点(年龄、收缩压、糖尿病和吸烟)和2个间接相关节点(性别、超重及肥胖)。模型的AUC值为0.681(95%CI:0.628-0.735),若取阈值为5.2%,则灵敏度为90.4%,特异度为31.3%。结论采用贝叶斯网络构建的老年高血压患者并发冠心病的预测模型具有较好的预测能力,并且能够更加直观地描述疾病与因素间复杂的网络风险机制。

Objective To construct predictive models of coronary heart disease(CHD) among elderly patients with hypertension,to explore the relevant influencing factors,and to provide the reference for prevention and control of CHD in elderly patients with hypertension. Methods In this study,the elderly in all communities(about 75 communities),who underwent physical examination in a Grade-Ⅱ hospital in Qinhuai district of Nanjing from March to December 2013,were selected as the population. Total 1 568 elderly permanent residents with hypertension(≥60 years old) from 10 communities in Nanjing were selected with cluster random sampling method as the subjects. The investigation was conducted with questionnaires,physical examinations and laboratory tests to collect the basic information of general condition,chronic diseases,habits,life style and related examination data.Three followup were conducted in 2015,2017 and 2019,respectively;and the contents were mainly about the incidence of cardiovascular diseases and other chronic diseases. χ^(2) test and the unconditional multivariate logistic regression were used for preliminary screening of model variables,the used software was SPSS 20.0. The "bnlearn" package of R 3.5.2 software was used to build Bayesian network model,and Netica software was used for Bayesian network inference. The receiver operating characteristic(ROC) curve was drawn,the sensitivity,specificity and the area under the curve(AUC) were calculated to evaluate the efficiency of predictive model. Results A total of 1 539 valid samples were followed up,and the incidence of CHD during the follow-up period was 6.8%.Multivariate unconditional logistic regression showed that age(OR=1.996),systolic blood pressure(OR=1.877),total cholesterol(OR=2.346),hyperuricemia(OR=1.887),overweight and obesity(OR=1.759),diabetes(OR=2.370) and smoking(OR=1.926) were the influencing factors of CHD in patients with hypertension,with statistical significance(P<0.05 or P<0.01).Combined with the multivariate logistic regression results and the prior theory,the prediction model based on Bayesian network finally included 9 nodes,which were age,gender,systolic blood pressure,total cholesterol,hyperuricemia,overweight and obesity,diabetes,smoking and CHD,which included 4 directly related nodes(age,systolic blood pressure,diabetes and smoking) and 2 indirectly related nodes(gender,overweight and obesity).The AUC of the model was 0.681(95%CI:0.628 - 0.735),if 5.2% served as threshold,the sensitivity was 90.4% and specificity was 31.3%. Conclusion The predictive model of CHD in elderly patients with hypertension constructed by Bayesian network has better predictive effects,and can describe the complex network risk mechanism between disease and factors more intuitively.