Double machine learning methods for estimating causal treatment effects on cancer survival outcomes in the presence of high-dimensional confounding

非小细胞肺癌（NSCLC）的治疗是一个复杂的过程，找到一种有效的治疗方法往往需要免疫治疗、手术和化疗的反复组合。理想情况下，我们可以利用临床试验的证据来了解每个患者的最佳治疗方案，但这通常只包括一次一种治疗，而不是复杂的组合或长时间的随访。使用常规收集的电子健康记录（EHR）和基因组数据可以帮助填补知识空白。寻找最佳治疗方法的关键一步是评估治疗对生存结果的因果影响，以及这是否取决于患者的特征。了解这一点可以帮助我们预测（假设的）新患者在特定治疗下的预期结果，从而最终允许临床医生个性化治疗。由于在常规实践中，根据患者的特点对其进行治疗，因此观察数据受到复杂、高维混淆的影响。因果推理的最新发展使我们能够将机器学习用于高维复杂关系的建模。然而，为了真正了解哪种治疗方案对个体最有益，我们需要能够研究治疗组合对患者长期生存的影响。该项目将专注于开发生存结果的因果机器学习方法，允许治疗决策的时变性质。随着大型、高维数据集在医疗环境中变得普遍，这项工作不仅使我们能够丰富有关非小细胞癌治疗的已知信息，而且将为许多其他疾病领域复制研究方法打开大门；为医疗保健专业人员提供更准确的信息，说明哪种治疗方案最适合每位患者。

The treatment of non-small cell lung cancer (NSCLC) is a complex process, and finding a treatment that works often requiring iterative combinations of immunotherapy, surgery and chemotherapy. Ideally, we would know the best course of treatment for each patient using evidence from clinical trials, but often this only includes one treatment at a time, and not complex combinations or long follow-up times. Using routinely collected electronic health records (EHR) and genomic data can help filling the gaps in knowledge. A key step towards finding the best treatment is estimating the treatment's causal effect on survival outcomes and whether this depends on patients' characteristics. Knowing this can help us to predict the (hypothetical) outcomes that can be expected under specific treatment for a new patient, thus eventually allowing clinicians to personalise treatments. Because in routine practice, treatment is given to patients that need it based on their characteristics, observational data is subject to complex, high-dimensional confounding. Recent developments in causal inference allow us to incorporate machine learning for modelling of high-dimensional complex relationships. However, to truly understand which treatment regime is most beneficial for an individual, we need to be able to study the effects of treatment combinations on a patient's survival over time. This project will focus on developing causal machine learning methods for survival outcomes allowing for the time varying nature of treatment decisions. With large, high dimensional datasets becoming common place in the medical setting, this work will not only enable us to enrich the information known about the treatment of non-small cell cancer, but will open the door for many other disease areas to replicate the research methods; providing healthcare professionals with more accurate information on which treatment regimens are most suitable for each of their patients.