Causal machine learning for multiple treatments and multiple outcomes in dynamic treatment regimes

最近的研究表明，患者需要更多的个性化治疗，这些治疗可以随着时间的推移而变化，具体取决于他们的反应。这是通过动态治疗方案正式化的，多个时间点上的一系列决策规则，其中下一次治疗是基于患者病史-以前的状态，以前的治疗，当前状态。此外，在多种药物和多种药物正在增加的情况下，单一药物可能会影响不是一种而是几种情况，而可能需要同时使用多种药物来解决多种情况。然而，这一事实在目前的临床试验中被忽视了，这些临床试验通常针对一种疾病的一种干预措施。例如，患有心血管疾病和高血压的患者可能被排除在心血管疾病研究之外。因此，该项目的总体目标是在动态治疗方案的每个时间点内为多个并行治疗和结局构建可靠的因果推断。特别是，这种因果推断旨在为现代纵向观察数据（如电子健康记录）提供替代顺序随机试验的方法。与这些目的相关的研究的一个关键方面是开发适合纵向研究的可靠降维技术。事实上，数据的维度随着时间点线性增长，因为每个状态和治疗决策都被添加到用于决定下一次治疗的历史中，因此，状态和治疗的数量呈指数增长。我们将探索现代机器学习技术，如表示学习和深度生成建模，以提供这种低维空间。此外，这些空间将可用于匹配等设置。在静态的单一干预环境中，匹配是一种形成相似（或“平衡”）治疗组和对照组的方法。尽管它在医学和社会科学中无处不在，但在纵向环境中扩展它的工作相对较少，更不用说在动态治疗方案中了。我们的目标是提供一个概括的匹配到多治疗和随时间变化的设置感谢一个更合适的平衡的概念，并感谢足够的降维空间。因此，我们将在一定程度上规避回归的挑战，在动态治疗制度，其中的非规律性的结果函数仍然是一个开放的研究problem.This项目属于EPSRC的“人工智能技术”，“统计和应用概率”和“医疗技术”的研究领域福尔斯。它由诺和诺德公司共同资助，并由Chris Holmes教授监督。

It has recently been shown that patients need more personalised treatments that can evolve over time depending on their response. This is formalised through a dynamic treatment regime, a sequence of decision rules over multiple time points where the next treatment is chosen based on patient history - former states, former treatments, current state. In addition, in a context where multimorbidity and polypharmacy are rising, a single medication might impact not one but several conditions, while multiple medications might simultaneously be required to address multiple conditions. However, this fact is overlooked in current clinical trials, which most often address one intervention over one disease. For example, patients with both cardiovascular disease and high blood blessure might be excluded from studies on cardiovascular disease. Thereby, the general aim of this project is to construct reliable causal inference for multiple, parallel treatments and outcomes within each time point of a dynamic treatment regime. Particularly, this causal inference is aimed to be of use for modern longitudinal observational data, such electronic health records, that provide an alternative to sequential randomised trials.A key aspect of research pertaining to these aims is developing reliable dimensionality reduction techniques suited for longitudinal studies. Indeed, the dimension of the data grows linearly with the time point, as each state and treatment decision is added into the history used to decide the next treatment, and as a consequence, the number of states and treatments grows exponentially. We will explore modern machine learning techniques such as representation learning and deep generative modelling to provide such low-dimensional spaces.In addition, these spaces will be usable in settings such as matching. In a static, single-intervention setting, matching is a methodology for forming similar (or "balanced") treatment and control groups. Despite its ubiquitous use in the medical and social sciences, relatively little work has been done to extend it in a longitudinal setting, let alone in dynamic treatment regimes. We will aim at providing a generalisation of matching to a multi-treatment and time-varying setting thanks to a more suitable notion of balance and thanks to adequate dimensionally-reduced spaces. As a result, we will in part circumvent the challenge of regression in the dynamic treatment regime, where the non-regularity of the outcome function remains an open research problem.This project falls within the EPSRC "Artificial intelligence technologies", "Statistics and applied probability" and "Healthcare technologies" research area. It is co-funded by Novo Nordisk and is supervised by Professor Chris Holmes.