运用随机森林分析药品不良反应发生的影响因素

Random Forest for Influencing Factor Analysis in Adverse Drug Reactions

目的拟采用随机森林来分析不良反应发生的影响因素。方法以2007年上海等地发生的阿糖胞苷引发肌无力及截瘫的不良反应为例,运用随机森林分析其重要的影响因素,将分析结果与实际情况进行对比,用于验证随机森林在不良反应数据中应用的可行性。结果通过随机森林综合评价得出的四个重要影响因素:触发时间、用药途径、季节和生产厂家,其均为阿糖胞苷事件的重要特征。且通过随机森林的计算,相比较于其他不良反应,肌无力和截瘫与各个影响因素之间存在的关联可能性更大,提示肌无力和截瘫需重点关注,这也与阿糖胞苷事件的实际情况相吻合。结论随机森林的综合评价机制能够从复杂数据中识别出真正重要的影响因素,并定量估计它们对不良反应发生的影响,有助于及时判别药品不良反应的特征、发生机制、危险人群和可能的引发途径,在药品不良反应信号的发现、因果关联评价和指导临床用药方面均有广泛的应用价值。

Objective To estimate the effect of influencing factors on the occurrence of adverse drug reactions by random forest algo- rithm. Methods The muscular weakness and paraplegia caused by cyt- arabine in 2007 at Shanghai and some other places in China were used as an example in this research. Random forest was used to analyze the important influencing factors of this adverse drug reaction, and the results calculated by random forest were compared with the features of the cytarabine event. Thus, the feasibility of random forest algorithm in spontaneous reporting system databases can be assessed. Results The muscular weakness and paraplegia were caused by two batches of impure cytarabine injections pro- duced by a particular company, and these cytarabine injections caused the nervous disorders via intrathecal route of administration. The four important influencing factors calculated by random forest： route of administration, time to onset, season and company were all the important features of the cytara- bine event. According to the results of random forest, the muscular weak- ness and paraplegia were more possibly related to the influencing factors compared to the other adverse reactions. So, close attention should be paid to muscular weakness and paraplegia. This result of random forest is in con- formity with the real event of cytarabine as well. Conclusion The ran- dom forest algorithm can identify the real important influencing factors from complex adverse drug reaction datasets and evaluate their effects on adverse reactions. It will be useful for further analyzing the features and mechanisms of adverse reactions, discriminating high risk populations, and possible route of administrations. So, random forest algorithm may be a valuable tool in practical use such as signal detection, causal relationship assessment and clinical practice guidance.