Collaborative Research: Semiparametric and Reinforcement Learning for Precision Medicine

精准医学寻求根据个体特征(包括遗传特征、人口统计信息、环境因素等)来优化医疗治疗。个体化治疗规则将将患者信息转化为推荐治疗的决策过程形式化，动态治疗机制由一个或多个治疗决策时间的个性化治疗决策序列组成。与此同时，医学成像技术的最新发展极大地影响了疾病和健康研究。生物医学成像和成像引导的干预是精确医学基础设施中的关键。在精确医学研究中，开发一种将成像数据与其他丰富信息结合在一起的方法具有重要意义。然而，目前在精准医学研究中对上述丰富特征的探索还远远不够。为此，该项目的目标是建立包含丰富特征的精准医学统计分析框架，并提供数据驱动的决策支持，这不会丰富统计方法学研究，但提供综合的早期诊断工具和信息工具，以指导健康科学中的治疗和生活方式干预。此外，该项目将为研究生提供培训和支持，并在本科生和研究生课程中提供指导。PIS将使Q-学习、半参数学习、函数数据分析和强化学习框架适用于具有丰富特征的精确医学，包括医学图像、遗传特征、人口统计信息、环境因素等。本研究计划包括三个部分：(I)结合丰富特征的功能个体化治疗制度研究，以及开发一种新的基础扩展工具来处理多维图像特征；(Ii)包含丰富特征的泛化功能个体化治疗方案研究，允许反应变量离散；(Iii)包含丰富特征的泛函Q-学习，将方法论扩展到多阶段决策环境。研究人员将进行理论开发，开发高效的算法，并将这些工具实施并应用于该项目中所有这些组件的真实数据。从统计学的角度出发，本文的理论探索将对半参数强化学习在具有丰富特征的精确医学中的应用提供更多的启示。从计算的角度来看，高效和可扩展的算法将以公开可用的软件的形式开发和实施。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Precision medicine seeks to optimize the medical treatments tailored to individual characteristics, including genetic features, demographic information, environmental factors, etc. Individualized treatment rule formalizes the process of decision making that translates the patients’ information into the recommended treatment, and a dynamic treatment regime consists of the sequence of individualized treatment decisions for one or more treatment decision times. Meanwhile, recent developments in medical imaging technologies dramatically affect disease and health studies. Biomedical imaging and imaging-guided interventions are key in the infrastructure for precision medicine. It is of great importance to developing an approach for incorporating imaging data along with other abundant information in precision medicine research. However, the current exploration for these aforementioned abundant features in precision medicine study is far from sufficient. Motivated by this, the project targets to build the statistical analysis framework in precision medicine incorporating abundant features and provide the support of data-driven decision making, which will not enrich statistical methodological studies but provide an integrated early diagnosis tool and an informative tool to guide treatment and lifestyle intervention in health science. In addition, the project will provide training and support for graduate students, as well as instructions in both undergraduate- and graduate-level courses.The PIs will adapt the Q-learning, semiparametric learning, functional data analysis, and reinforcement learning frameworks to precision medicine with abundant features, including medical images, genetic features, demographic information, environmental factors, etc. Focusing on different scenarios, this research program consists of three components: (i) functional individualized treatment regime study incorporating abundant features, along with the development of a novel basis expansion tool to handle the multi-dimensional image feature; (ii) generalized functional individualized treatment regime study incorporating abundant features, which allows the response variable discrete; and (iii) functional Q-learning with abundant features, which extends the methodology to the multi-stage decision setting. The investigators will conduct the theoretical developments, develop efficient algorithms, and implement and apply the tools to real-world data for all these components in this project. From the statistical point of view, the theoretical explorations will yield more insights into semiparametric and reinforcement learning in precision medicine with abundant features. From the computational point of view, efficient and scalable algorithms will be developed and implemented in a form of publicly available software.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.