Development of methods for the analysis of complex survival and joint longitudinal-survival data with application to linked electronic health records

近年来，数据基础设施有了长足的发展，人们一致努力提高电子健康记录的可用性和质量，将其作为医学研究的主要数据来源。本项目力求利用这一点，开发分析这类数据所需的新颖和适当的先进统计技术。这些方法将被应用于回答心血管疾病和癌症研究领域的临床相关问题，有可能在这两个领域产生实质性影响，并在更广泛的临床领域产生更广泛的影响。第一个项目旨在分解心血管疾病的途径，以增加对哪些风险因素与不同结果相关的理解，使用的数据来自英国的电子健康记录。例如，患者开始健康，然后经历第一次心脏病发作，然后是随后的中风。通过对这些从状态到状态的转换进行模型拟合，我们可以识别重要的风险因素，这些风险因素可用于识别后续心血管事件风险增加的患者。通过对患者的整个轮廓进行建模，我们可以最有效地利用现有数据，并且还能够为个体患者量身定制未来事件的预测。第二个实质性项目将进一步开发和应用纵向数据和生存数据的联合模型，该模型允许对生物标志物进行建模，这些生物标志物随着时间的推移而重复测量，例如血压，以及生物标志物的变化如何与感兴趣的事件的发生率相关，比如死亡我们将使用瑞典和丹麦的登记数据，研究血红蛋白水平随时间的变化与癌症诊断率之间的关系。这可能导致确定血红蛋白的重要轨迹，从而可以进行有针对性的监测，或者更快地实施干预措施，或者更早地诊断病例。将制定方法，使这些计算密集的方法适用于这样一个大的数据库，提供一个广泛适用的方法框架。第三个项目将调查血压随时间的变化如何与经历心血管事件的风险相关，如心脏病发作或中风。我们有大量的电子健康记录数据资源，这些数据是在全国各地的全科医生诊所收集的。这些数据表现出分层结构，生物标志物重复测量嵌套在患者中，嵌套在GP实践中。在我们的分析中考虑到这种层次结构是很重要的。该项目将扩展联合的生存模型，使我们能够解释这种结构，这将使我们能够调查在实践层面测量的剥夺状态等因素。除了上述项目，我们认为同时开发免费提供的用户友好的软件来实现该方法是至关重要的，随后将其发布给研究界，并开办旨在帮助将方法转化为实践的课程。这一点尤其重要，因为大型数据集的可用性越来越高，这意味着方法必须能够有效地处理“大数据”，这需要先进的编程技能。然而，通过发布软件，这意味着方法学工作可以应用于任何数量的不同疾病领域，帮助回答各种相关的和临床上重要的问题，在整个健康研究的范围。

In recent years there has been substantial growth in data infrastructure, and a concerted drive to improve the availability and quality of electronic health records, as a primary source of data for medical research. This project seeks to take advantage of this, through developing novel and appropriate advanced statistical techniques which are required to analyse such data. The methods will then be applied to answer clinically relevant questions in the areas of cardiovascular disease and cancer research, with the potential to have substantial impact in both areas, and more widely across a diverse range of clinical areas.The first project aims to decompose the pathways of cardiovascular disease to increase the understanding of which risk factors are associated with different outcomes, using data from linked electronic health records in the UK. For example, a patient begins healthy, then experiences a first heart attack, and then a subsequent stroke. By fitting models to each of these transitions from state to state, we can identify important risk factors which could be used to identify patients at an increased risk of subsequent cardiovascular events. By modelling the whole profile of a patient, we make most efficient use of the available data, and will also be able to develop predictions for future events, tailored to individual patients. This aspect is crucially important in communicating information to patients, and ensuring such information is both understandable and meaningful.The second substantive project will further develop and apply joint models of longitudinal and survival data, which allow the modelling of a biomarker, measured with error and repeatedly over time, such as blood pressure, and how changes in the biomarker are related to the rate of an event of interest, such as death. We will investigate the relationship between changes in haemoglobin levels over time, and the rate of cancer diagnoses, using Swedish and Danish registry data. This may lead to identifying important trajectories of haemoglobin which can allow targeted monitoring, or indeed interventions to be applied sooner, or cases be diagnosed earlier. Methodology will be developed to allow these computationally intensive methods to be applied to such a large database, providing a widely applicable methodological framework.The third project will investigate how changes in blood pressure over time are associated with the risk of experiencing cardiovascular events, such as a heart attack or stroke. We have available a vast resource of data from electronic health records, collected at GP practices across the country. Such data exhibits a hierarchical structure, with biomarker repeated measures nested within patients, nested within GP practice. It is important to account for this hierarchical structure in our analyses. This project will extend joint longitudinal-survival models to enable us to account for such structures, which will enable us to investigate factors such as deprivation status measured at the practice level.Alongside the above projects, we believe it is crucial to simultaneously develop freely available user friendly software which implements the methodology, subsequently release it to the research community and run courses aimed to help transfer methods into practice. This is particularly important as the increased availability of large datasets means methods must be able to handle 'big data' efficiently, which requires advanced programming skills.To conclude, in this project we aim to utilise the methods in the areas of both cardiovascular and cancer epidemiology; however, through releasing software, it means that the methodological work can be applied to any number of different disease areas, to help answer a variety of relevant and clinically important questions, across the range of health research.