Developing Counterfactual Inference Methods for Clinical Trial Recruitment and Effective Integration of Weak Instrumental Variables.

该项目福尔斯以下EPSRC研究领域：医疗技术统计和应用概率数字孪生当前计算统计和ML空间的研究问题是算法是否可以自动推断系统的行为和模式，这些行为和模式在其底层机制中表现出内在的因果关系。人类通常擅长识别自然界中的这种关系（例如，我们知道太阳升起会导致温度上升，而不是相反）。然而，机器通常很难直接学习这些关系，一个潜在的研究途径是开发可以自动识别数据变量之间潜在因果关系的方法。已经开发了许多方法（称为因果结构学习算法）来自动学习这些关系，并且在小数据集/特征和高斯数据上表现良好。然而，它们很难很好地扩展非线性高斯的大型复杂数据集。在我读博士期间的一个研究途径是通过开发可扩展的结构学习方法来解决这个问题，该方法可以处理大型数据集而不会牺牲推断关系的保真度。即使描述一组变量的因果路径是已知的，准确推断模型中的定量关系也是具有挑战性的。这些问题因混杂因素的存在而加剧，特别是如果未观察到（或即使未知）。在临床医生的背景下，这些可能是一些变量，如性别或年龄，不仅影响医生分配的治疗，而且影响治疗本身的效果。解释这些变量的黄金标准是通过运行随机对照试验（RCT）来手动干预和控制人工选择的亚群中的这些差异。然而，经常有RCT不道德（例如，已知有害的“治疗”）或过于昂贵的情况。此外，可能有大量免费获得的观察性数据（即，不是从随机对照试验中获得的）可用于代替随机试验。尽管观察性数据集的规模很大，但其完整性和完整性可能存在问题。例如，可能存在未记录的变量（即潜在混杂因素）。这些变量可能是我们已知的（专家可能能够编码我们数据中存在和不存在的特征之间的关系），但也可能是未知的。这个问题可能会激发另一个研究方向，即寻找替代方法来捕捉这些混杂因素的行为，并减少模型推断的偏差。影响因果方法采用的另一个关键障碍是它们扩展到大数据/特征集的能力。目前的方法的运行时间通常很难扩展到更大，更复杂的数据集。潜在研究的另一个途径是探索不同类型的近似值，这些近似值可以更有效地扩展，而不会牺牲太多的预测能力。虽然上面讨论了推动因果方法的方法学角度，但我对将这些方法应用于生物/医疗问题有浓厚的兴趣。特别是，如何将临床患者的深度生成模型与结构化因果模型相结合，以创建患者的因果数字双胞胎，而不是从混淆的生成模型中提取的有偏见的样本。

This project falls within the following EPSRC research areas:Healthcare TechnologiesStatistics and Applied ProbabilityDigital TwinsA current research question in the Computational Statistics and ML spaces is whether algorithms can automatically infer the behaviours and patterns of systems which demonstrate inherent causal relationships in their underlying mechanisms. Humans often excel at identifying such relationships in nature (e.g. we know the rising sun causes temperatures to rise rather than the other way round). Machines however, generally struggle to learn these relationships directly, and one potential research avenue is to develop methods which can automatically identify potential causal relationships between data variables. A number of methods (known as Causal Structure Learning algorithms) have been developed to learn these relationships automatically, and perform well with with small datasets/features and Gaussian data. However, they struggle to scale well with larger, complex datasets which are not linearly Gaussian. An avenue for research during my DPhil is to address this problem through developing scalable Structure Learning methods which can process large datasets without sacrificing the fidelity of the inferred relationships.Even if the causal pathways describing a set of variables is known, accurately inferring quantitative relationships in a model can be challenging. These problems are exacerbated by the presence of confounders, particularly if unobserved (or even if unknown). In the context of medical clinicians, these could be some variable like sex or age which not only impacts the treatment assigned by a doctor, but also the effect of the treatment itself. The gold standard for accounting for these variables is by running Randomised Controlled Trials (RCTs) to manually intervene and control for these discrepancies across an artificially selected subpopulation. However, there are often cases where RCTs are unethical (e.g. with "treatments" that are known to be harmful) or too expensive. Additionally, there may be a wealth of freely available observational data (i.e. not obtained from an RCT) which could be used in lieu of running a randomised trial.Despite the size of observational datasets, there may be issues with their completeness and integrity. For example, there may be variables (i.e. potential confounders) which were simply not recorded. Thesevariables might be known to us (where experts may be able to encode relationships between present and absent features in our data), but may also be unknown. This problem may motivate another research direction which looks at alternative approaches to capture the behaviour of these confounders and reduce the biases inferred by the model.Another key roadblock a ecting the adoption of causal methods is their ability to scale to large data/feature sets. The run time of current methods often scale extremely poorly to larger, more complex, datasets. Another avenue for potential research would be to explore di erent types of approximations which would scale more e ciently without sacri cing much predictive power.Whilst the above discuss the methodological angles for pushing the needle forward for causal methods, I have a strong interest in applying these methods to biological/healthcare problems. In particular, how can Deep Generative models of clinical patients be combined with structured causal models to create causal Digital Twins of patients rather than biased samples drawn from confounded generative models.