1、 Capital Medical University硕 士 学 位 论 文中国 35-64 岁人群 15 年高血压发生风险预测研究Prediction models for the 15 years risk of new-onset hypertension in Chinese people aged from 35 to 64 years old:a cohort study研究生：学科专业： 流行病与卫生统计学指导教师：导师单位：完成日期： 二XX 年一月首都医科大学硕士学位论文目 录中文摘要 . 1ABSTRACT . 4前 言 . 71. 研究背景 . 72. 研究现状 . 82
2、.1 高血压危险因素 . 82.2 高血压绝对风险预测 . 123. 研究目的 . 13研究方法 . 141. 研究对象 . 142. 研究方法 . 142. 1 问卷调查 . 142. 2 体格检查 . 142. 3 血样本采集 . 162. 4 实验室检查 . 162. 5 质量控制方法 . 172. 6 统计方法 . 172. 7 分析内容 . 19研究结果 . 211. 研究人群基线一般情况 . 212. SBP或 DBP相关因素的横断面分析 . 223. 研究人群 15 年血压变化情况 . 253. 1 15 年血压水平变化 . 253. 2 高血压 15 年累计发生率 . 30首都
3、医科大学硕士学位论文4.高血压危险因素分析及预测模型建立与评价 .314.1危险因素分析及模型建立 .314.2 模型性能评价 .354.3 模型的人群验证 .374.4其他回归方法建立的预测模型预测性能评价 .404.5弗莱明翰高血压预测模型对本人群的适用性研究 .435. 高血压风险评分表建立 .45讨 论 .49参考文献 .53个体风险预测模型及风险评分表的建立的综述 .591. 疾病预测模型介绍 .592.高血压预测模型研究进展 .643. 预测模型建立与评价 .683.1 个体风险估计 .683.2 风险评分表建立 .693.3 预测模型性能评价 .72综述参考文献 .76缩略词对照
4、表 .81攻读学位期间发表论文情况 .82个人简历 .85首都医科大学硕士学位论文中文摘要研究目的本研究通过中国多省市队列人群 1992 年及 2007 年的调查数据，分析研究人群 15 年期间高血压累计发病率，探讨影响我国人群高血压发生的危险因素，并建立适合我国人群的高血压发生风险预测模型和高血压风险评分表，旨在为高血压的早期预防提供数据支持。研究方法本研究纳入了参加 1992 年基线调查且参加了 2007 年随访的 5408 人，排除基线患有高血压、心血管疾病者，共有 3899 人纳入分析。采用前瞻性队列研究方法，分析高血压 15 年累计发病率及其危险因素，通过多因素 Logistic 回
5、归系数建立高血压预测模型，同时对模型性能进行评价，最终形成风险评分表。结果本研究基线人群共纳入 3899 名（男性 1794 人，女性 2105 人），平均年龄为45.37.4 岁（男性 46.87.6 岁，女性 44.07.0 岁，P0.0001）。研究人群 15 年内舒张压（Systolic blood pressure，SBP）平均水平从 113.8mm Hg 升到 134.5mm Hg，增加了 20.7mm Hg，舒张压（Diastolic blood pressure，DBP）平均水平从 74.2mm Hg升到 81.4mm Hg，增加了 7.2mm Hg。15 年内共有 1776
6、人发生了高血压，高血压累计发病率为 45.6%。经卡方检验结果显示：性别、基线年龄、SBP、DBP、体质指数（Body mass index ，BMI）、高血压家族史、体育锻炼、高密度脂蛋白胆固醇（High density lipoprotein cholesterol ，HDL-C ）、甘油三酯（Triglyceride ，TG）的水平与高血压发生危险有关。对单因素分析有意义的变量进行多因素 Logistic 回归分析，分别建立两个高血压预测模型，模型 1 纳入了临床容易获得的变量年龄、SBP、DBP、BMI、1首都医科大学硕士学位论文高血压家族史 5 个危险因素，模型 2 在模型 1 的基
7、础上加入了 TG 和 HDL-C。两个模型都有很好的判别能力，模型 1 的 ROC(receiver operating characteristic curve)曲线下面积（area under curve，AUC）为 0.7168，模型 2 的 AUC 为 0.7208，模型2 的 AUC比模型 1 高 0.004，差异具有统计学意义。HosmerLemeshow 2 检验结果显示：模型 1 的 2 值为 3.75，模型 2 为 3.10,2 值越小认为模型具有较好的校准能力，这表明两个模型都能很好的对预测风险与实际风险进行拟合，预测风险与实际风险之间的差异均不具有统计学意义。计算重新分类
8、净改善指数(net reclassificationimprovement, NRI)为 0.83%，P 值为 0.40, 这表明加入 HDL-C、TG 后的模型 2 对原模型 1 改善程度没有统计学意义。本研究建立的模型通过以下三个方面进行评价或进一步验证。首先，在研究人群中随机抽取 40%的人作为验证人群，分别对两模型的预测效果进行验证。验证人群 1559 人，其中男性 622 人，女性 937 人。平均年龄为 45.87.3 岁（男性 47.57.4 岁，女性 44.67.1 岁）。模型 1 和模型 2 在验证人群中都有很好的判别能力，模型 1 的 AUC 为 0.7189，模型 2 的
9、 AUC 为0.7208。Hosmer Lemeshow2 检验结果显示：模型 1 的 2 值为6.61（P=0.6294），模型 2的 2 值为 3.11（P=0.9273）,这表明两个模型都能很好的对预测风险与实际风险进行拟合。其次，本研究也进一步对比了采用 Poisson 回归、 Cox 回归建立的预测模型的预测效果。结果显示：Poisson 回归建立的预测模型的 AUC 为 0.7168，HosmerLemeshow2 值为 18.12（P=0.0203）。Cox 回归建立的预测模型的 AUC 皆为 0.7168，HosmerLemeshow 2 值为 37.44（P0.0001）。再
10、次，我们将美国建立的弗莱明翰高血压预测模型应用于本人群，并和我们建立的模型进行比较。结果显示，弗莱明翰预测模型的 AUC 为 0.7105，而模型 1 的 AUC 为 0.7168，模型 1 的 AUC 比弗莱明翰预测模型高 0.0063，差异具有统计学意义（P=0.008）。HosmerLemeshow 2 检验结果显示：弗莱明翰预测模型的 2 值为 697.64（P0.0001），预测风险与实际风险之间的差异具有统计学首都医科大学硕士学位论文意义，这表明模型不能很好地对预测风险与实际风险进行拟合。结论本研究结果显示，我国人群高血压长期发生风险较高，对高血压的预防十分重要。我们利用中国多省市
11、人群的长期随访数据，首次建立了适合中国人群的高血压风险预测工具。这将有助于人们了解其高血压发病风险，及时的改善生活方式或药物治疗。【关键词】高血压；预测模型；队列研究3首都医科大学硕士学位论文ABSTRACTObjectiveThis study is trying to evaluate the incidence and risk factors of hypertension in 15 years follow-up in a cohort population, and then propose prediction models for risk of new-onset hype
12、rtension in Chinese people, and develop risk scores which were convenient for the clinical application.MethodA cohort study that enrolled participants who were aged 35-64 from 11 provinces of China has been set up since 1992. Participants were eligible for inclusion if they had participated in the e
13、xamination in 2007 without meeting any exclusion criteria. We excluded participants who had prevalent hypertension at baseline, prevalent cardiovascular disease. A total of 3899 people included in the analysis. The 15 years cumulative incidence and risk factors of hypertension were evaluated, and we
14、 used Logistic regression to analyze the association between baseline risk factors and the incidence of hypertension during the 15 years, then developed prediction models and risk scores with the regression coefficient. We also estimated the performance of the models.Results3899 participants who wer
15、e followed for 15 years and free from hypertension at baseline were selected, 1794 were men and 2105 were women. The mean age of participants was 45.37.4 years(46.87.6 years in men, 44.07.0 years in women). In the 15 years follow-up, the mean SBP increased from 113.8 mmHg to 134.5 mmHg and the mean
16、of DBP increased from 113.8 mmHg to 134.5 mmHg. We ascertained 1776 cases of incident hypertension after 15 years, and the incident ratewas 45.6%. With the chi-square test, the incidence of hypertension was associasted4首都医科大学硕士学位论文with sex, age, SBP, DBP, BMI, history of parental hypertension, physi
17、cal exercise , HDL-C, and TG.We used Logistic regression to calculate the coefficients of the risk factors which were significant in univariate analysis. The Model 1 included 5 factors which were easily obtained (age, SBP, DBP, BMI, history of parental hypertension), and the Model 2 added the HDL an
18、d LDL on the basis of Model1. AUC for Model1 was 0.7168, AUC for Model 2 was 0.7208, and Model 2 was higher than model 2 by0.004, and the difference was statistically significant. There were no significant differences between observed and predicted hypertension risks for Model1 and Model2 (HosmerLem
19、eshow chi-square statistic: 3.75 for Model1, 3.10 for Model2). There was no significant difference in NRI (net reclassification improvement) index between Model 1 and Model 2.We randomly sample 40% of the population to validate the prediction performance of the two models. In the sample population,
20、there were 622 men and 937 women. The mean age of participants was 45.87.3 years(47.57.4 years in men, 44.67.1 years in women). AUC for Model1 was 0.7189, AUC for Model 2 was 0.7208. there were no significant differences between observed and predicted hypertension risks for Model1 and Model2 (Hosmer
21、Lemeshow chi-square statistic: 6.61 for Model1, 3.11 for Model2).The modified Poisson regression and Cox regression were used to build prediction models if incidence of hypertension with the factors: age, SBP, DBP, BMI, history of parental hypertension. AUC for Poisson model was 0.7168 , HosmerLemes
22、how chi-square statistic was 18.12 for Poisson model. AUC for Cox model was 0.7168 HosmerLemeshow chi-square statistic was 37.44 for Cox model.We also validate the prediction performance of Framingham model in this population. AUC for Model1 was 0.7168, AUC for Framingham model was 0.7105,5首都医科大学硕士学
23、位论文Model 1 was higher than Framingham model by 0.0063, and the difference was statistically significant. There were significant differences between observed and predicted hypertension risks for Framingham model (HosmerLemeshow chi-square statistic: 697.64, P0.0001).ConclusionsThe results of this stu
24、dy show that Chinas long-term risk of hypertension was high, and the risks control of non-hypertensive population was very important. We used long-term follow-up data of Chinese Multi-provincial Cohort Study (CMCS) to establish the prediction model of hypertension for the first time. The hypertensio
25、n risk prediction models can be used to estimate an individuals absolute risk for hypertension and may facilitate health management of high-risk individuals who are likely to develop hypertension. Taking account of the clinical application, we preferred to recommend the risk score developed by Model 1.【Key words 】 Hypertension；Prediction models；Cohort study6
