Machine learning solution to clinical decision support for early detection of respiratory conditions within the ICU

重症监护病房(ICU)是一个高度专业化的环境，致力于提供危重护理和治疗危及生命的疾病。它整合了从护士、医生到顾问的各种利益相关者--每个人都需要不同范围和程度的患者数据来完成他们的工作。这就产生了一个问题，因为需要建立各种系统来促进这些信息的获取，使得在复杂情况下很难汇总所有相关数据以进行推断。急性呼吸窘迫综合征(ARDS)是一种危及生命的呼吸系统疾病，导致大量ICU入院。由于诊断过程具有挑战性，ARDS经常被延迟诊断或完全漏诊，导致超过40%的高死亡率，并对幸存者造成毁灭性的后果。因此，临床实践中迫切需要能够早期发现ARDS的工具。医院内电子健康记录(EHR)和临床信息系统(CIS)的日益普及使得ICU数据丰富的环境成为机器学习(ML)方法解决这一问题的绝佳机会。然而，尽管这项技术在早期研究中显示出了令人振奋的结果，但ICU数据的特性固有的几个障碍阻碍了它在临床环境中的广泛采用。必须克服的第一个挑战是不同ICU之间数据格式差异的统一。这些可能是供应商特定软件的专有格式，但也可能是由于CI高度可配置这一事实导致的差异，即使在共享同一提供商的CI记录的数据之间也是如此。此外，随着时间的推移，数据收集方式的变化可能会使问题变得更加复杂。第二个障碍与存储数据的来源及其测量的生理参数的范围有关。与患者在ICU逗留相关的数据往往非常全面，但它缺乏关于他们在ICU逗留前后发生的信息，因为这些数据存储在其他地方。促进这些数据集之间的强大联系需要对基础系统的第一手知识，以前已经进行了研究。此外，重症监护部门收集的数据经过重症监护国家审计与研究中心(ICNARC)进行的全国临床审计，包括病例组合计划(CMP)，该计划整合了英国99%的成人普通重症监护病房的数据。该研究项目旨在通过使用新的数据链接技术来构建在临床环境中具有明确实施路径的机器学习解决方案，以解决ARDS的早期发现问题。为实现这一目标，最初的努力将侧重于建立一个平台，以便可靠地提取从当地信息和通信系统合并的数据，以及为《议定书》/《议定书》/《公约》缔约方会议审计目的汇编的数据。之后，将采用机器学习的方法，结合先前建立的数据链接软件，设计和开发用于ARDS检测的临床决策支持工具。为此，我们将探索有监督和无监督的ML技术。围绕监督学习的问题，如标签的质量和查明基本事实的能力，也将在设计阶段得到彻底审查。最后，该解决方案的有效性将通过在几个更大的数据集上进行测试来衡量，这些数据集将来自ICNARC以及其他可用的来源，包括MIMIC-III、EICU和HIC，这些来源共享用于临床研究目的的杰出跟踪记录。

Intensive care unit (ICU) is a highly specialised environment concerned with the provision of critical care and treatment of life-threatening conditions. It consolidates a variety of stakeholders ranging from nurses and doctors to consultants - each requiring a different scope and extent of patient data to do their job. This creates a problem, as variety of the systems need to be in place to facilitate the access to this information, making it difficult to aggregate all relevant data for inference in complex conditions.Acute respiratory distress syndrome (ARDS) is a life-threatening condition of the respiratory system that contributes to a large number of ICU admissions. With a challenging diagnosis process, ARDS is frequently diagnosed late or missed altogether, leading to a high mortality rate of over 40% and devastating outcomes for the survivors. Tools allowing for the early detection of ARDS are therefore critically needed within clinical practice.The growing ubiquity of electronic health records (EHR) and clinical information systems (CIS) within the hospitals renders the data-rich environment of the ICU an excellent opportunity for machine learning (ML) approaches to that problem. However, despite the promising results of that technique demonstrated in the early research, several barriers inherent to the characteristics of the ICU data prevent its widespread adoption in clinical setting.The first challenge that has to be overcome is the unification of data format discrepancies spanning different ICUs. These could be proprietary formats attributable to the vendor-specific software, but also differences stemming from the fact that CISs are highly configurable leading to divergence even between data logged by CISs that share the same provider. In addition to that, changes to how data is gathered can be made over time further complicating the problem.The second barrier relates to the variety of sources that store the data and the scopes of physiological parameters they measure. Data relevant to the patients' stay in the ICU is frequently very comprehensive, however it lacks the information of what happened before and after their stay, as that data is stored elsewhere. Facilitating the robust linkage between these datasets requires a first-hand knowledge of the underlying systems and has been previously researched. Furthermore, the data gathered in the critical care sector undergoes national clinical audits performed by Intensive Care National Audit & Research Centre (ICNARC) including the Case Mix Programme (CMP) which consolidates data from 99% of adult general critical care units in the UK.This research project aims to tackle the problem of early ARDS detection by employing novel data-linkage techniques to build a machine learning solution that has a clear pathway of implementation within the clinical setting. To achieve this goal, the initial efforts will focus on constructing a platform for robust extraction of data consolidated from local CISs, as well as the data compiled for the purpose of CMP audit. After that, the machine-learning approach will be taken to design and develop a clinical decision support tool for ARDS detection, incorporating the previously established data-linkage software. To that extent, both supervised and unsupervised ML techniques will be explored. The issues surrounding supervised learning such as the quality of labelling and ability to pinpoint the ground truth will also be thoroughly examined at the design stage. Finally, the effectiveness of the solution will be measured by testing it across several larger datasets that will be sourced from the ICNARC as well as other available sources including the MIMIC-III, eICU and HIC which share an outstanding track record of use for clinical research purposes.