Deep Patient Trajectory Analysis: Learning from electronic health records

电子健康记录(EHR)最近已经成熟为例行收集的信息的巨大来源。今天，一份电子病历可以包含数百万患者的全面病历，记录和整理几十年来从多家医疗机构收集的各级医疗保健数据。这样的数据集可以包含数TB的信息，其中数十亿个条目记录在复杂的底层数据结构中。在过去，经典的生存分析技术被广泛用于根据相对较少的信息量对患者的未来进行预测。在我的工作中，我正在研究如何通过将生存分析技术应用于深度学习框架，来利用电子健康记录中可用的更多信息，为医疗保健提供者提供更准确的生存分布预测，从而提高他们做出有效医疗决策的能力。这项研究的主要目的是建立深度学习模型，能够处理患者的整个病史，并将时间的准确分布返回到感兴趣的事件。这将需要几个独立的组件。首先，我们将开发一种稳健的方法来对多模式、多源纵向EHR数据进行预处理。其次，我们需要一个合适的深度学习体系结构，它可以识别这些数据中潜在的复杂模式，以创建患者及其对感兴趣事件的易感性的准确模型。最后，我们希望我们的模型输出到感兴趣事件的时间的适当生存分布。虽然已经有一些将深度学习应用于电子病历数据的研究，但这种方法是新颖的，因为它比以前的工作走得更远，以前的工作只研究了离散时间点之间发生事件的概率。我还计划进一步扩展这项工作，将更先进的生存分析技术应用于深度学习框架，探索相互竞争的风险和多态模型，以及为分配家庭选择评分。这项研究是与NIHR ARC西北伦敦研究小组合作进行的，开发的方法将应用于来自伦敦西北全系统综合护理(WSIC)数据库的数据，目的是为当地医疗保健提供者提供更好的工具，以了解和改善他们的护理。我们计划测试我们的方法的第一个案例研究包括调查糖尿病患者、心脏病患者和患有多种疾病的患者的事件发生时间和多状态建模。该项目属于EPSRC医疗保健技术研究领域。

Electronic health records (EHR) have recently matured into an enormous source of routinely collected information. Today, a single EHR can contain comprehensive medical histories for millions of patients, logging and collating data collected from all levels of healthcare over the course of decades from multiple healthcare institutions. Such a dataset can contain terabytes of information, with billions of entries recorded in a complex underlying data structure. In the past, classical survival analysis techniques have been used extensively to make predictions about the future of patients based on relatively small quantities of information. In my work, I am investigating how we can leverage much more of the information available in electronic health records, by adapting survival analysis techniques to a deep learning framework, to give healthcare providers more accurate survival distribution predictions and therefore improve their ability to make effective healthcare decisions. The primary aim of this research is to produce deep learning models that are capable of processing a patient's entire medical history and returning a precise distribution for the time to an event of interest. This will require several independent components. Firstly, we will develop a robust method of pre-processing multi-modal, multi-source longitudinal EHR data. Secondly, we will need a suitable deep learning architecture that can identify the potentially complicated patterns in this data to create an accurate model of a patient and their susceptibility to the event of interest. Finally, we want our models to output a suitable survival distribution for the time to the event of interest. While there has already been some research into applying deep learning to EHR data, this methodology is novel in that it goes further than previous work, which has only looked into the probability of events occurring between discrete timepoints. I also plan to extend the work further by adapting more advanced survival analysis techniques to a deep learning framework, by exploring competing risks and multi-state models as well as scoring for distribution family selection. This research is being done in collaboration with the NIHR ARC Northwest London research group and methods developed will be applied to data from the Northwest London Whole System Integrated Care (WSIC) database, with the aim of providing local healthcare providers with improved tools to understand and improve their care. The first case studies we plan to test our methods on include investigating time to events and multi-state modelling for patients suffering from Diabetes, heart-conditions as well as those with multi-morbidities. This project falls within the EPSRC healthcare technologies research area.