Advanced Machine Learning to Improve Patient Care and Outcome using Real-time Hospitalisation Data

该项目的总体目标是开发高效和有效的机器学习和数据分析技术，以分析多样化和大量的患者床边数据，以改善护理提供，患者体验，结果和资源管理。这涉及到开发新的技术，不仅可以处理从复杂环境中收集的高度复杂的数据，而且还可以提供清晰的推理和理由，以便临床医生和NHS管理人员在决策和患者护理中充分知情。该项目的成功结果可以在改善反应性患者护理和前瞻性患者管理方面发挥重要作用。患者床旁数据预计具有很大的差异性，这对数据处理和建立预测模型提出了重大挑战。收集的数据类型也涵盖了广泛的范围，从患者生命体征，药物，到护理提供。数据集中的稀疏性给生成和判别任务带来了额外的挑战。与领域专家一起，该项目最初将集中在一个或两个临床问题上，如败血症。尽管抗生素和急性护理取得了现代进步，但严重脓毒症和脓毒性休克仍是医学领域的重大医疗挑战。由于高患病率和显著的死亡率，脓毒症仍然是由感染导致死亡的主要原因，引起从业者的严重关注。具体而言，在英国的统计数据中，脓毒症患者的预后表明ICU住院期间的死亡率为35%，住院期间的死亡率为47%，第一年内的再入院率为63%。这种严重的预后还伴随着27.1%的成人在ICU入院24小时内符合严重脓毒症标准的高患病率。这样的统计数据提供了对患者内脓毒症发展的显著严重性的快照。脓毒症患者在医院病床中占有相当大的比例。实时床旁数据提供了发现早期生物标志物或指标的独特机会。它们还可以提高我们对预后的理解，以及普通病房和ICU之间更好的资源管理。我们将利用我们在机器学习[1-4]、数据分析、以人为本的计算和数学建模方面的集体专业知识，以应对这些技术挑战。例如，开发新的深度神经网络模型，以预测痴呆症患者的住院情况。我们的新型神经网络模型能够使用患者健康记录提前六个月预测住院情况。该项目还具有与在（关键）决策和以人为中心的计算中使用机器学习的更广泛背景相关的重要组成部分。（1）目前的预警触发干预。根据定义，那些被触发干预的患者是那些病情更严重的患者，因此将预测和结果与ML分离是有问题的，因此是一个重要的研究问题。这种反身性是一个发生在所有形式的可视化分析中的问题（参见欧洲可视化分析路线图[6]，因此这里的见解很可能适用于其他领域。有趣的是，这些指标通常被视为削弱权力，强加于管理层，但在这里，它们带来的透明度似乎至关重要。例如，如果使用ML技术，保持这种透明度将是一个关键问题。3）患者的空间移动是本地化床数据连接到更大背景的另一个地方。我们有理由认为，每次搬迁都有风险，但在提高利用率或让病人住进更合适的病房方面也有好处。

The overall aim of this project is to develop efficient and effective machine learning and data analysis techniques to analyse a diverse and a large quantity of patient bedside data, in order to improve care provision, patient experience, outcome, and resource management. This involves developing novel techniques that can not only handle highly complex data that are collected from a complex environment but also provide clear reasoning and justification so that clinicians and NHS managers can be fully informed in their decision making and patient care. The successful outcome from this project can play an important role in improving both reactive patient care and forward-looking patient management.The patient bedside data is envisaged to have a great variability from patient to patient, which poses a significant challenge to data processing and building predictive models. The types of data that are collected also covers a wide range, from patient vitals, medications, to care provisions. Sparsity in the dataset introduces additional challenges to both generative and discriminative tasks. Together with domain experts, the project will initially focus on one or two clinical problems, such as sepsis. Severe sepsis and septic shock present a significant healthcare challenge within medicine despite modern advances in antibiotics and acute care. With both a high prevalence and significant mortality rate, sepsis remains the primary cause of death from infections resulting in significant concerns for practitioners. Specifically, within UK statistics, prognosis of a septic patient indicates a 35% mortality rate during ICU stay, 47% mortality rate during hospital spell and a 63% rate of hospital readmission within the first year. Such a severe prognosis is additionally met with a high prevalence rate of 27.1%of adults meeting severe sepsis criteria within the 24 hours of ICU admission. Such statistics provide a snapshot into the significant severity of septic development within a patient. Patients with sepsis take up a significant proportion of hospital beds. The real-time bedside data provides unique opportunities to discover earlier biomarkers or indicators. They also can improve our understanding of prognosis, as well as better resource management between general wards and ICUs. We will build upon our collective expertise in machine learning [1-4], data analysis, human centred computing, and mathematical modelling in order to tackle these technical challenges. For example, on developing novel deep neural network models in order to predict hospitalisation for dementia patients. Our novel neural network models are capable of predicting hospitalisation, six months in advance, using patient health records.This project also has a significant component related to the wider context of employing machine learning in (critical) decision making and human centred computing. A number of considerations.1) The current early warnings trigger interventions. By definition, those patients for whom interventions are triggered are those who are more ill, so disentangling prediction and outcome with the ML is problematic and hence an important research issue. This reflexivity is an issue that occurs in all forms of visual analytics (See European Visual Analytics roadmap [6], so insights here may well be applicable in other domains.2) The issue of empowerment of ward is also very important. It was interesting in that the metrics are often seen as dis-empowering, impositions form management, but here it seems that the transparency they bring is crucial. Maintaining this transparency if, for example, ML techniques are used, will be a critical issue.3) Spatial movement of patients is another of the places where the localised bed data connects into larger contexts. It is reasonable to assume each move carries risks, but also benefits in terms of improved utilisation or having patients in more appropriate hospital wards.