贝叶斯倾向性评分模型及其在药品不良反应信号检测中的应用

发布时间：2018-06-30 21:58

本文选题：贝叶斯倾向性评分 + 倾向性评分法　；参考：《第二军医大学》2014年硕士论文

【摘要】：研究背景：目前，我国药品不良反应自发呈报系统的报告数量飞速增长。如何借助恰当的数据挖掘和统计分析方法利用这些数据，高效、准确的发现药品安全问题，关系到广大患者的生命健康，也是公共卫生研究的重点问题。但自发呈报系统的数据中存在着较多的混杂因素，若在信号挖掘中忽略或者未采用适当的统计方法均衡混杂因素，会影响到信号检测结果的准确性。本课题组已经实现了传统的Logistic倾向性评分法与信号挖掘方法的结合以控制混杂因素的干扰。然而，传统的logistic回归倾向性评分法在应用中存在的问题也不容忽视。如忽略了连续型协变量与logit(y)呈线性关系的条件、未考虑到倾向性评分值的不准确性对处理效应的影响、无法利用先验分布的信息、结局事件为罕见事件或存在较多协变量时不适用等问题。研究目的：本研究探索贝叶斯思想与倾向性评分分层法、匹配法、加权法及回归法的结合以构建各种贝叶斯倾向性评分模型。同时，通过模拟研究比较在不同样本量、不同处理效应强度以及不同先验分布精度情况下各种贝叶斯倾向性评分模型估计处理效应的准确性及精确性。最后，选出最优模型应用于药品不良反应信号检测中，以提高药品不良反应信号检测的准确性。研究方法：本研究采用蒙特卡洛法模拟数据集。模型的建立采用两次建模，先建立处理因素与协变量的关系，再建立结局变量与处理因素、协变量的关系。根据协变量与结局变量和处理因素的关系，分别模拟只和处理因素有关的协变量、与处理因素和结局变量都有关的协变量、只和结局变量有关的协变量。根据协变量的分布类型分别模拟二分类与连续型变量。研究设置了5种不同强度的处理效应（t1.5,1.2,0.8,0.5,0.2）、四种先验信息的精度（B0,1,10,100）及三种样本量情况（N=50、100、250）。本研究先构建贝叶斯Logistic回归模型估计倾向性评分值，再与不同的倾向性评分利用方式相结合（如倾向性评分匹配法、倾向性评分分层法、倾向性评分加权法及倾向性评分回归调整法）及不同的处理效应估计方法相结合，共衍生七种贝叶斯倾向性评分模型。研究从处理效应的点估计值、标准差、偏倚（绝对偏倚、相对偏倚）、均方误差及置信区间的覆盖率等方面进行结果的稳健性及准确性评价。并借助SAS及R软件，编写程序完成方法的软件自动化实践。最后，研究根据评价结果选出准确性及稳健性较好的贝叶斯倾向评分匹配法（设置B100）并将其应用于上海市食品药品监督管理局2009年自发呈报系统的数据中。通过与国家药品-不良反应字典比对、查阅国内外文献及药品说明书，以验证贝叶斯倾向性评分模型、Logistic倾向性评分模型及常规方法的信号检测的准确性。研究结果：在小样本的情况下（N=50）,贝叶斯倾向性评分匹配法及贝叶斯倾向性评分分层法+CMH受先验分布信息精度的影响较大。当先验分布的精度设置为100时，这两种方法的准确性及稳定性较传统的倾向性评分法有着明显的提高。尤其是贝叶斯倾向性评分匹配法，其处理效应的点估计及置信区间长度较传统的倾向性评分模型明显改善。当样本量设置为100时，贝叶斯倾向性评分匹配法及贝叶斯倾向性评分分层法+CMH的结果较传统的倾向性评分法较为接近，其稳定性略微提高，但偏倚略有增大。当进一步增大样本到250时，贝叶斯倾向性评分匹配法及贝叶斯倾向性评分分层法+CMH的结果与传统的倾向性评分结果极为接近。同时，样本量较大的情况下，处理效应的点估计也更加接近预设的真值。无论是贝叶斯倾向性评分还是传统的倾向性评分，倾向性评分值的处理方式的选择对处理效应估计的准确性及稳性影响较大。匹配法及分层法较好，可作为首选。而加权法及回归法的稳定性较差，受处理因素与结局变量关联强度影响较大。实际数据应用中，，在未采用贝叶斯倾向性评分法及倾向性评分法对基线协变量进行均衡之前，发现“阿奇霉素-局部麻木”的组合在四种信号挖掘方法中均检测为阳性信号，通过与国家药品-不良反应字典比对、查阅国内外文献及药品说明书判定该信号为假阳性信号。单因素分析发现年龄、体重、性别、合并用药及不良反应发生季节等混杂因素在组间的分布不均衡。而采用贝叶斯倾向性评分匹配法及传统的Logistic倾向性评分匹配法对基线协变量进行均衡后，发现采用前种方法后所有基线斜变量在组间的分布均衡，而采用后者仍旧有部分基线协变量（如体重、不良反应发生季节）在组间的分布不均衡。从而可见贝叶斯倾向性评分法混杂因素的均衡效果较传统的Logistic倾向性评分匹配法有优势。虽然贝叶斯倾向性评分法与Logistic倾向性评分法在最终结果判定时均得到“阿奇霉素-局部麻木”假阳性的结论，但Logistic倾向性评分法估计的信号值与阈值较为接近，更容易得到阴性的结论。研究结论：当先验信息的精度较高（B100）且样本量较小（N=50）时，贝叶斯倾向评分匹配法及贝叶斯倾向性评分分层法+CMH较传统的倾向性评分法有着较大的优势。在药品不良反应信号检测的实际数据应用中，贝叶斯倾向评分匹配法（设置B100）较传统的倾向性评分匹配法的均衡性好，可以降低混杂因素导致的偏倚，在一定程度上可去除可疑的假阳性信号。
[Abstract]:Background of Study :

At present , the number of reports of spontaneous reporting system of drug reactions in China has increased rapidly . How to use these data with proper data mining and statistical analysis methods is a key problem in the study of public health . However , there are many problems in the data of spontaneous reporting system .

Purpose of study :

Finally , the optimal model was applied to the detection of adverse drug reaction signals to improve the accuracy of drug reaction signal detection .

Study method :

In this study , Monte Carlo method was used to simulate the data set . The relationship between treatment factors and covariables was established , and the relationship between outcome variables and treatment factors and covariables was established . Based on the relationship between covariables and outcome variables and processing factors , the covariables related to treatment factors and outcome variables were simulated respectively . Five different intensities of processing effects ( t1 . 5 , 1.2 , 0.8 , 0.5 , 0.2 ) were established .

In this study , a Bayesian logistic regression model was constructed to estimate the value of tendentious score , and then combined with different methods of propensity score utilization ( such as propensity score matching method , propensity score layering method , propensity score weighting method and propensity score regression adjustment method ) and different treatment effect estimation methods . Seven kinds of Bayesian propensity score models were derived . The results were evaluated from the point estimate , standard deviation , bias ( absolute bias , relative bias ) , mean square error and coverage rate of confidence interval .

Finally , the Bayesian propensity score matching method ( B100 ) was selected and applied to the data of Shanghai Food and Drug Administration in 2009 .

Results of the study :

In the case of small samples ( N = 50 ) , the Bayesian propensity score matching method and the Bayesian propensity score hierarchy method + CMH are affected by the accuracy of the prior distribution information . When the accuracy of the prior distribution is set to 100 , the accuracy and stability of the two methods are obviously improved compared with the conventional propensity score method . Especially , the Bayesian propensity score matching method improves the point estimation and confidence interval length of the processing effect obviously .

When the sample size is set to 100 , the Bayesian propensity score matching method and the Bayesian propensity score stratification method + CMH are relatively close to the traditional propensity score method , but the stability is slightly increased . When the sample is further increased to 250 , the Bayesian propensity score matching method and the Bayesian propensity score stratification method + CMH results are very close to the traditional propensity score results . In the meantime , when the sample size is large , the point estimation of the processing effect is closer to the preset true value .

The selection of the treatment mode of the propensity score value has a great influence on the accuracy and stability of the treatment effect estimation , whether it is a Bayesian propensity score or a traditional propensity score , and the matching method and the stratification method are better and can be used as the first choice .

It was found that the combination of azithromycin and local anaesthesia was detected as a positive signal in four kinds of signal mining methods without using the Bayesian propensity score method and the propensity score method .

Conclusions of the study :

When the accuracy of the prior information is high ( B100 ) and the sample size is small ( N = 50 ) , the Bayesian propensity score matching method and the Bayesian propensity score stratification method + CMH have great advantages . In the actual data application of the drug reaction signal detection , the Bayesian propensity score matching method ( setting B100 ) is better than the traditional propensity score matching method , and the bias caused by confounders can be reduced , and the suspected false positive signal can be removed to a certain extent .
【学位授予单位】：第二军医大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R95

【参考文献】