End-to-End Probabilistic Modelling of Longitudinal Electronic Health Records

最近大量电子健康记录（EHR）数据集的可用性导致了医疗保健文献中机器学习的激增。本文的目的是在新提出的方法和概率建模的经典文献之间建立一座桥梁。因此，它特别关注医疗数据的典型属性：(1)广泛的缺失，(2)不规则间隔的观察，(2)模型解释和(3)半监督。在第一步，我们探索如何使用深度生成模型（生成新数据的非线性模型）来计算缺失的数据。我们提供了一个直观的解释，为什么这些深度生成插值方法可以工作，并且可以将缺失数据的先验信息编码到模型中。在第二步中，我们通过使用时间连续神经网络建模潜在分类变量来处理不规则间隔的观测值。因此，我们允许在潜在空间中对时间条件概率进行建模。这样的模型可以用来及时聚类病人的疾病状态。在第三步，我们开发可解释的方法，揭示基于观察实验室值和调查答案的疾病的潜在动力学。然后，这些潜在的动态可以用作临床试验中的疗效指标，以评估药物是否有效。可应用于此问题的常见潜在因素模型通常仅限于捕获线性相互作用。然而，时间连续非线性模型不能给出可解释的结果。我们的目标是找到一个稀疏线性逼近非线性动力学发现的连续深潜模型。在第四步中，我们想要探索如何通过包括半监督来控制这样一个深度潜在模型。在临床试验中，现有的疗效指标可用于一些临床就诊。我们的目标是找到一种疾病的新的疗效措施，这种措施不会受到现有措施的幼稚影响，而是在某种程度上与现有措施正相关。另一个模型旨在解释可能的偏差。然后它应该确保偏差是合理的。为了证明我们模型的有效性，我们在各种真实世界的医疗数据集（如医疗基准数据集MIMIC III和诺华提供的临床试验数据集）上评估了所提出的方法。我的研究部分是由诺华公司资助的。我的导师是克里斯·霍姆斯。该项目属于EPSRC医疗保健技术研究领域。

The recent availability of large data sets of electronic health records (EHR) led to a surge in the machine learning for health care literature. This thesis aims to build a bridge between the newly proposed methods and the classical literature on probabilistic modelling. It thereby pays special attention to typical properties of medical data: (1) extensive missingness, (2) irregularly spaced observations, (2) model interpretation and (3) semi-supervision.In a first step, we explore how deep generative models (non-linear models that generate new data) can be used to impute missing data. We provide an intuition why some of these deep generative imputation methods work and make it possible to encode prior information on the missing data into the model.In a second step, we deal with irregularly spaced observations by modelling latent categorical variables using in-time-continuous neural networks. We hereby allow to model time conditional probabilities in the latent space. Such a model can then be used to cluster the disease states of a patient in time.In a third step, we develop interpretable methodology that reveals the latent dynamics of a disease based on observed lab values and survey answers. These latent dynamics can then be used as efficacy measures in clinical trials to assess whether a drug is working or not. Common latent factor models that can be applied to this problem are often restricted to capture linear interactions. Time-continuous nonlinear models however fail in giving interpretable results. Our goal is to find a sparse linear approximation to the non-linear dynamics found by continuous deep latent models. In a fourth step, we want to explore how such a deep latent model could be controlled by including semi-supervision. In clinical trials, existing efficacy measures are available for some of the clinical visits. The goal is to find novel efficacy measures of a disease that do not suffer from the naivety of the existing measures but are to some degree positively correlated with the existing measures. An additional model is aimed to explain possible deviations. It should then ensure that the deviations are reasonable.To demonstrate the effectiveness of our models, we evaluate the proposed methodology on various real-world medical datasets such as the medical benchmark data set MIMIC III and clinical trial data sets provided by Novartis.My research is partly funded by Novartis. My supervisor is Chris Holmes. This project falls within the EPSRC Healthcare technologies research area.