数据挖掘方法检测药品不良反应信号的应用研究

Application of data mining algorithm to detect adverse drug reaction signal

目的应用数据挖掘方法检测 ADR 信号并探讨其应用价值。方法收集国家药品不良反应监测中心2009年1月至2013年12月收到的北京市21家药物警戒站上报的所有抗感染药物不良反应报告，采用比例报告比值比法（PRR）、报告比值比法（ROR）、综合标准法（MHRA）、贝叶斯可信传播神经网络法（BCPNN）及多项伽玛泊松分布缩减法（MGPS）挖掘 ADR 潜在的风险信号，并对5种方法的检测结果进行比较。结果共收集到抗感染药物不良反应报告35807份，最终纳入有效报告35759份，涉及可疑药品834种。按低位语统计，35759份报告共涉及 ADR 464种；按系统器官分类统计，涉及 ADR 21种。对 ADR 报告进行数据清洗、拆分、编码后，得到6620份含有可疑药品-不良反应组合的报告，其中可疑药品-不良反应组合出现1次者3966份（59.91%）；出现2次者937份（14.15%）；出现≥3次者1717份（25.94%）。利用5种方法对药品-不良反应组合进行信号检测，PRR、ROR、MGPS、BCPNN 及 MHRA 法检测出的信号数分别为651、614、306、75和57个，涉及的药物种类分别是194、168、124、34、40种，涉及的不良反应种类分别是139、139、121、35和40种。在排名前10位的风险信号中，5种方法共同检测出的信号是阿奇霉素-恶心；PRR、ROR、MHRA、BCPNN 法共同检测出的信号是左氧氟沙星-瘙痒。PRR 和 ROR 检测出的前10种信号完全相同，其余方法检测出的信号各不相同。结论5种信号检测方法均可系统、自动地检测到 ADR 报告中的风险信号，但5种方法各有利弊，应根据实际情况与需求选择应用。

Objective To detect adverse drug reaction（ADR）signals using data mining algorithm and explore its application value. Methods Reports on adverse reactions induced by anti-infective drugs in National centor for ADR monitoring from January 2009 to December 2013 were collected and potential ADR risk signals were detected using proportional reporting ratio method（ PRR）,reporting odds ratio method（ ROR）,Medicines and Healthcare products Regulatory Agency method（ MHRA）,Bayesian confidence propagation neural network method（BCPNN）,and multi-item gamma poisson shrinker method （MGPS）. The results of detection using the above-mentioned 5 signal detection methods were compared. Results A total of 35 807 ADR reports induced by anti-infective drugs were collected,35 759 effective reports were entered,and 834 suspected drugs were involved. In the 35 759 reports,464 kinds of ADR were defined according to lowest level term and 21 kinds of ADR were defined according to system∕ organ classification. After the data cleaning,splitting,and encoding process,6 620 reports containing suspected drug-adverse reaction combination were acquired. There were 3 966 reports（59. 91% ）in which suspected drug-adverse reaction combination appeared once,937 reports（14. 15% ）in which suspected drug-adverse reaction combination appeared twice,and 1 717 reports（25. 94% ） in which suspicious drug-adverse reaction combination appeared more than thrice. The number of ADR signals detected using PRR,ROR, MGPS,BCPNN,and MHRA was 651,614,306,75,and 57,respectively;the categories of drugs were194,168,124,34 and 40,respectively;ADR types were 139,139,121,35,and 40,respectively. In the top ten risk signals,azithromycin-nausea were detected by the 5 signal detection methods,levofloxacin-pruritus were detected by PRR,ROR,MHRA,and BCPNN. The top ten signals detected by PRR were totally same as those by ROR and signals detected by other methods were various. Conclusions potential risk signals in ADR reports could be detected systematically and automatically using PRR,ROR,MGPS, BCPNN,and MHRA. However,each method has its own advantage and disadvantage and should be applied according to the actual situation and demand.