双聚类算法研究及其在我国药品不良反应监测中的应用

发布时间：2018-08-15 14:56

【摘要】：研究背景和目的:作为药品上市后安全性监测的重要组成部分,药品不良反应自发呈报系统可以广泛收集不良反应报告,实现对药品不良反应的及时监测。我国自发呈报系统数据库中数据量庞大。2009年至2016年,年报告数量已从60余万份激增到140余万份。利用不相称测定方法,研究者可以从自发呈报系统数据中检测出大量不良反应信号,为药品安全性评价提供基本依据。在检测出不良反应信号之后,仍需要对信号进行评价才可以确定其是否为真实的药品不良反应,但是大量的信号给评价工作带来十分艰巨的挑战。目前不良反应信号评价方法主要包括将检测出的不良反应信号与药品说明书数据对比、进行专家评价等,评价效率低下。如何提高信号评价工作的效率是药品不良反应监测领域中研究者亟需解决的问题之一。有研究指出:分析包含同一种不良反应的多个相似信号或包含同一种药品的多个相似信号可能可以提高信号评价效率。这为提高我国信号评价过程的效率提供了新的思路。本研究期望引入一种新的方法——双聚类分析方法,以期识别出我国不良反应信号数据中相似信号组合。利用组合中已确认信号的信息对待评价信号进行快速评价,提高信号评价工作的效率,并探索在分析我国不良反应信号数据过程中最优参数设置。研究方法:将Bimax双聚类算法引入到我国不良反应信号数据分析中。将检测出的不良反应信号数据构建出原始定量数据矩阵。依据不同的区分强弱信号的IC值阈值将定量矩阵转化成不同的0-1数据矩阵,并结合不同的Bimax算法参数,对不同的0-1数据矩阵进行双聚类分析,从中识别出对应信号值均超过阈值的多个药品-不良反应组合,获取相似性信号组合信息。然后,利用秩和比综合评价法结合平均符合比例、涉及比例两个指标对分析双聚类结果进行评价,确定最优区分强弱信号的IC值阈值及最优Bimax算法参数。最后对最优参数设置组合下双聚类分析结果进行深入分析,明确双聚类算法在我国不良反应信号数据分析中的运用效果。研究结果:通过秩和比综合评价法比较不同参数设置组合下双聚类方法的分析结果,确定最优区分强弱信号的阈值为IC=0.80,最优的Bimax算法参数为最小双簇行数3、最小双簇列数3。在上述参数设置组合下,利用Bimax双聚类方法可以从不良反应信号数据中识别出4293个双簇。双簇平均包含23.41个信号,双簇内组合的平均IC值为2.38。通过分类学评价可以发现所包含的药品具有相似性或所包含的不良反应具有相似性的双簇共有1836个,占双簇总数的42.8%,涉及72.3%的IC值不小于0.80的待评价信号;至少包含有一个已得到药品说明书确认信号的双簇有4272个,占总数的99.5%;包含的所有信号均已得到药品说明书确认的双簇有193个,占总数的4.5%。在所有双簇中出现频次最多的药品为“利培酮”,共在708个双簇中出现,占总数的16.5%;出现频次最多的不良反应为“肝功能异常”,共在720个双簇中出现,占总数的16.8%。研究结论:利用双聚类算法分析我国不良反应信号数据所得到的双簇可以为潜在药品不良反应确认、筛选需重点关注不良反应信号等多个方面提供十分有价值的信息,可以促进我国药品不良反应监测中信号评价工作效率的提高。
[Abstract]:BACKGROUND AND OBJECTIVE: As an important part of post-marketing safety monitoring, ADR spontaneous reporting system can collect ADR reports extensively and realize timely monitoring of ADR. The number of ADRs surged to more than 1.4 million. Researchers can detect a large number of ADR signals from the data of the spontaneous reporting system by using the method of asymmetric measurement, which provides a basic basis for drug safety evaluation. At present, the evaluation methods of ADR signal mainly include comparing the detected ADR signal with the data of drug instructions and conducting expert evaluation. The evaluation efficiency is low. How to improve the efficiency of signal evaluation is a researcher in the field of ADR monitoring. It is pointed out that the analysis of multiple similar signals containing the same adverse reaction or multiple similar signals containing the same drug may improve the efficiency of signal evaluation. Class analysis method is used to identify the similar signal combination in the ADR signal data of our country. The information of the confirmed signals in the combination is used to evaluate the evaluation signal quickly, and the efficiency of signal evaluation is improved. The algorithm is introduced into the analysis of ADR signal data in China. The original quantitative data matrix is constructed from the detected ADR signal data. The quantitative matrix is transformed into different 0-1 data matrix according to the IC value threshold of distinguishing strong and weak signals, and the different 0-1 data matrix is doubly aggregated by combining different Bimax algorithm parameters. Classification analysis identifies several drug-adverse reaction combinations whose corresponding signal values exceed the threshold, and obtains the information of similarity signal combinations. Then, the rank sum ratio comprehensive evaluation method combined with the average coincidence ratio, involving two indicators of the ratio, is used to evaluate the results of the analysis of the double clustering, and the IC value threshold and the maximum value of the optimal discrimination between strong and weak signals are determined. Finally, the application effect of bi-clustering algorithm in the analysis of ADR signal data in China is clarified. The results show that the optimal region is determined by comparing the results of bi-clustering method under different parameter combinations with rank sum ratio comprehensive evaluation method. The thresholds of the strong and weak signals are IC=0.80, and the optimal parameters of the Bimax algorithm are the minimum number of rows of two clusters and the minimum number of columns of two clusters. Analogical evaluation showed that 1836 clusters, accounting for 42.8% of the total number of clusters and involving 72.3% of IC values not less than 0.80, contained similarity of drugs or adverse reactions; at least 4272 clusters, accounting for 99.5% of the total number, contained confirmation signals from the drug specifications. All the signals were confirmed by the drug instructions in 193 clusters, accounting for 4.5% of the total. Risperidone was the most frequent drug in all the clusters, accounting for 16.5% of the total 708 clusters, and liver dysfunction was the most frequent adverse reaction, accounting for 16.8% of the total 720 clusters. CONCLUSION: Bicluster analysis of ADR signal data in China can provide valuable information for identification of potential ADR, screening of ADR signal and so on. It can improve the efficiency of signal evaluation in ADR monitoring in China.
【学位授予单位】：第二军医大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R95

【参考文献】