Identifying Subgroups with Different Responses to HIV Risk Reduction Counseling

描述（由申请人提供）：Aware，一项大型（n=5012）随机比较有效性试验，发现在HIV检测时对HIV阴性个体进行HIV风险降低咨询对累积发病率没有影响。然而，问题仍然是是否有亚群体会从咨询中受益。此外，了解咨询如何矛盾地增加了MSM（美国HIV感染风险最高的群体）的性传播感染，以及是否有其他亚群体增加了性传播感染，对公共卫生具有重要意义。在最近的创新中，机器学习技术被专门用于以可复制的方式发现具有不同治疗反应的亚组，并且不会遭受与多次测试相关的模型过拟合。我们将在随机森林（RF）和虚拟双胞胎（VT）两种方法的基础上扩展方法，探索治疗亚组和少数群体之间的差异。VT方法使用随机森林作为第一步，为每个试验参与者在治疗和控制条件下的结果创建单独的基于森林的预测。然后，为每个个体创建一个针对个人的治疗效果，并做出基于树的预测。尽管该方法已被证明是非常有前途的，但该方法的敏感性相对较低，阳性预测值也较低。这些问题可以通过用随机森林程序取代个体特异性治疗效果的单树预测器和/或重新加权分类问题以平衡治疗成功的数量来缓解（STI发病率远未影响50%的样本）。我们将使用模拟来揭示最佳策略，然后使用最佳策略来创建一个模型，该模型可以根据行为和个人对咨询产生积极或消极影响的可能性来预测未来性传播感染的可能性。我们还将使用该模型的扩展，通过在Project Aware中观察到的咨询互动来发现与观察到的MSM相关的因素。本研究的意义在于：1)确定是否存在性病临床患者亚群将受益于或受到短期艾滋病风险降低咨询的伤害；2)提供针对这些个体的模型或理解为什么他们可能表现出更高的性传播感染；3)为使用这些方法了解对其他艾滋病预防干预措施的异质反应提供基础。

DESCRIPTION (provided by applicant): Aware, a large (n=5012) randomized comparative effectiveness trial, found that HIV risk reduction counseling for HIV negative individuals at the time of an HIV test did not have an impact on cumulative incidence of However, the question remains as to whether there are subgroups that would benefit from counseling. Further, understanding how counseling paradoxically increased STIs in MSM, the group most at risk for HIV in the US, and whether there are other subgroups who increased STIs is of public health importance. In recent innovations machine learning techniques have been used specifically to uncover subgroups with differential treatment responses in a fashion that is replicable and does not suffer model over-fitting associated with multiple testing. We will extend methods to explore treatment subgroups and differences across minority groups based on two of these approaches-Random Forests (RF), and Virtual Twins (VT). The VT approach uses random forests as a first step to create separate forest-based predictions of outcomes under both treatment and control conditions for each trial participant. Then a person-specific treatment effect is created for each individual and a tree-based prediction is made. Whereas this procedure has been shown to be very promising, there is a tendency for the procedure to have relatively low sensitivity, and low positive predictive value. These problems may be alleviated by replacing the single tree predictor of the individual-specific treatment effect by the random forest procedure and/or reweighting of the classification problem to equalize the number of treatment successes (STI incidence is far from affecting 50% of the sample). We will use simulations to uncover the optimal strategy and then use the optimal strategy to create a model that predicts likelihood of future STIs based on behaviors and likelihood of the individual having a positive or negative impact of counseling. We will also use an extension of this model to find the factors associated with the observed MSM by counseling interaction observed in Project Aware. The significance of this research lies in 1) determining if there are subgroups of STD clinic patients who would benefit from or be harmed by short HIV risk reduction counseling 2) providing a model to target these individuals or understanding why they may show increased STIs, and 3) providing the groundwork for use of these approaches to understanding the heterogeneous response to other HIV prevention interventions.