• 统计研究中心
当前位置: 首页> 系列讲座> 正文

北卡罗来纳大学教堂山分校曾冬林教授:Learning Optimal Dynamic Treatment Regimes under Risk Constraints

光华讲坛——社会名流与企业家论坛第 5799 期

Learning Optimal Dynamic Treatment Regimes under Risk Constraints

主讲人北卡罗来纳大学教堂山分校曾冬林教授

主持人统计学院林华珍教授

时间2021531日(周一)上午10:00-11:00

直播平台及会议IDZoom会议, ID: 987 5736 4617;密码: 684606

主办单位:统计研究中心和统计学院 科研处

主讲人简介:

Donglin Zeng (曾冬林)2001年获密西根大学安娜堡分校统计博士,现为北卡罗来纳大学教堂山分校的生物统计系教授。主要研究方向包括semiparametric inference, high-dimensional data, machine learning, personalized medicine, survival analysis and clinical trials.他至今发表至少180篇统计论文,曾担任各主要统计学刊的副主编。他是美国统计学会(ASA)和国际数理统计学会(IMS)的会士。

 

内容提要:

Learning optimal dynamic treatment regimes (DTRs) aims at  the individualized sequential treatment rules that maximize cumulative beneficial outcomes by accommodating patient's heterogeneity and evolving features into decision making. For many chronic diseases, treatments are usually multifaceted in a sense that an aggressive treatment with high reward will likely increase the risk of acute risk events. In this work, we propose a general framework to study such benefit-risk balance when learning DTRs. This framework consists of solving weighted support vector machine problems subject to constraints at each decision stage, and the solutions can be obtained using an iterative DC algorithm. We show that the proposed framework yields the optimal DTRs, and we obtain the convergence rates for both value and risk functions. The proposed method is demonstrated via single- and two-stage simulation studies and is further illustrated using a trial for type 2 diabetes patients.

学习最佳动态治疗方案(DTRs)旨在针对个体化的顺序治疗规则,该规则通过适应患者的异质性和将特征演变为决策来最大化累积的有益结果。 对于许多慢性疾病,从具有高回报的积极治疗可能会增加发生急性风险事件的风险的意义上讲,治疗通常是多方面的。在这项工作中,我们提出了一个总体框架来研究学习DTR时的这种利益风险平衡。该框架包括解决在每个决策阶段受约束的加权支持向量机问题,并且可以使用迭代DC算法获得解决方案。我们表明,提出的框架产生了最佳的DTR,并且我们获得了价值函数和风险函数的收敛速度。通过单阶段和两阶段模拟研究证明了所提出的方法,并通过一项针对2型糖尿病患者的试验进行了进一步说明。

上一条:加州大学尔湾分校Annie Qu教授:Multicategory Angle-based Learning for Estimating Optimal Dynamic Treatment Regimes with Censored Data

下一条:加州大学伯克利分校李乐昕教授:Testing Mediation Effects Using Logic of Boolean Matrices with Applications in Neuroimaging Mediation Analysis