个性化医疗中在分类框架下估计最佳治疗方案的统计方法研究

Personalized medicine, which is focused on making treatment decisions for an individual patient based on her/his clinical, genomic, and other available information, is of considerable current interest. A treatment regime maps observed patient characteristics to a recommended treatment. Recent technological advances have increased the quality, accessibility, and volume of patient-level data; consequently, there is a growing need for powerful and flexible estimators of an optimal treatment regime that can be used with either observational or randomized clinical trial data. A novel and general framework was proposed that transforms the problem of estimating an optimal treatment regime into a classification problem wherein the optimal classifier corresponds to the optimal treatment regime (Zhang, et al., 2013). This framework allows classification procedures that accommodate case weights to be used without modification to estimate an optimal treatment regime. Despite the promise of this robust approach for optimizing treatment selection, there are multiple challenges that may limit its practical utility: (i) the optimal treatment regime may be defined in terms of a few important variables and their composite predictor. Therefore, a model selection method, which combines variable selection and constructing composite predictor to define the optimal treatment regime, is needed; （ii）Treatment of patients may involve a series of decisions over time, especially in the case of chronic diseases, now the treatment regime is a set of sequential decision rules, this classification approach can only handle single decision point; (iii) Multiple treatments in one study is common in practice; however, the existing framework, as well as the majority of existing methods, only considers the two treatments situation; (iv) In reality, treatment decisions often involve a tradeoff between risk and benefit across multiple outcomes. The existing framework, as well as existing methods for optimal treatment regimes, focuses on a single outcome. In this project, we intend to address those barriers and propose new methods within this classification framework. In addition, we plan to analyze existing real data (eg, the Nefazodone-CBASP trial) using the proposed methods to identify feasible and useful treatment regimes that can be used for personalized medicine

个性化医疗是指根据病人个体特征/信息为病人量身定制治疗决定，其在统计上可抽象为治疗方案. 治疗方案是一个决策规则，输入已检测到的病人个体特征，输出对病人适宜的疗法。如今大量的病人个体数据被收集，分析这些数据从而找到最佳治疗方案的需求越来越大。针对单决策点下估计最佳治疗方案， 最近一个全新的统计方法框架被提出来。 在该框架下，其问题转换成为一个分类的问题，可以直接利用现有能处理权重的分类方法来估计最佳治疗方案。 基于这个分类框架，我们将开展四方面工作: （i）提出新的变量选择方法能同时从高维变量中作变量选择和构造新的组合变量来决定最佳治疗方案；（ii）将现有的单决策点分类框架推广到多决策点情况，并做理论证明研究；（iii）考虑多药物情况下的分类框架研究；（iv）考虑多结果变量情况下的理论和方法研究。我们还将用这些新提出的方法分析现有的实际数据，期望能找到可行有效的治疗方案用

治疗方案是一个决策规则， 输入已检测到的病人个体特征，输出可选择项中对病人最适宜的疗法。最佳治疗方案就是能使整体人群疗效最大化的决策规则，我们的目的就是要从临床试验数据或观测数据中估计最佳治疗方案。Zhang et al.（2013）提出来了一个全新的方法框架，在这个框架下，估计最佳治疗方案的问题转换成为一个分类的问题，从而和数据发掘领域联系起来。本项目的研究对这个新的数据挖掘框架进行详细的探讨，在其基础上提出新的思想和方法，并结合实际问题，在理论和应用上都取得了重要进展。具体来说，本项目的成果包括以下两个方面。（1）当今临床试验数据或观测数据中收集的病人特征变量非常多，但包含许多无用信息，并且很有可能单个协变量的作用比较小，最佳治疗方案由其中少数几个重要变量的线性组合定义。已有的变量选择方法的目标是和预测相关的变量，我们提出了新的方法，可以在找到和药物决策相关的重要变量的同时得到其正确的线性组合形式。（2）在多个时间点的情况，需要在每个时间点基于现有的信息做药物推荐，这一系列的决策方案叫做动态治疗方案。我们提出了通过逐步优化来估计最佳动态治疗方案的新方法，并且在每个时间点上的优化问题转化成了一个分类问题。