随机森林倾向性评分方法及其在药品不良反应信号检测中的应用

Random Forest Propensity Scores Method and its Application in Drug Adverse Reaction Signal Detection

目的探讨利用随机森林倾向性评分法控制混杂因素的基本思想和步骤,及其在药品不良反应信号检测中的应用。方法利用随机森林计算给定危险因素的条件下研究对象服用双膦酸盐的概率,而后分别通过倾向性评分1:1匹配,1:M匹配和回归调整法控制性别、年龄等混杂因素,分析服药双膦酸盐与骨折发生风险的关系,并与logistic回归倾向性评分法对应结果进行比较。结果随机森林倾向性评分法与logistic回归倾向性评分方法的结果是一致的。其中,倾向性评分1:1匹配样本量损失较大,且与1:M匹配和回归调整法的结果相差较大。结论随机森林倾向性评分法能有效控制药品不良反应信号检测过程中的混杂因素,可以与logistic回归倾向性评分法所得结果相互验证,提高结果的可靠性;但1∶1匹配可能不适用于药品自发呈报系统数据。

Objective The aim of this paper is to describe the basic ideas and algorithms of random forest propensity scores method for controlling confounders and apply it in detecting drug adverse reaction signals. Methods First, we used random forest to calculate a patient＇s probability of taking bisphosphonates. Then, we analyzed the association of bisphosphonate intake with risk of fracture by controlling potential confounders with propensity score method. The controlling confounders methods included 1 ： 1 matching, 1 ： M matching and regression adjustment by using the propensity score calculated by random forest. The results were compared with those from logistic propensity score. Results The results of random forest propensity score and lo- gistic propensity score were comparable. One to one propensity score matching cause a lot of sample lost and its results were quite different from those based on other methods. Conclusion Random forest propensity score method can reduce the confounding bias. Hence, it could be used as an alternative to and verification of the logistic propensity score in controlling confounders. However, 1 ： 1 propensity score matching may not be suitable for adverse drug reaction data from a spontaneous reporting system.