Advancing machine learning to achieve real-world early detection and personalised disease outcome prediction of inflammatory arthritis

在英国，超过2000万人患有风湿性和肌肉骨骼疾病(RMD)，而炎症性关节炎(IA)是RMD的一个主要分支，会导致关节炎症，从而导致损害。IA会导致长期的疼痛、残疾，并产生巨大的个人和社会成本。在英国，2004至2020年间确诊的IA病例估计也增加了59%，这对卫生服务具有重要影响。2019年/2020年，风湿科约占NHS信托基金总药物支出的9%。在IA患者路径上仍有大量未满足的需求，特别是在IA检测和闪光处理方面。IA表现为非特异性症状，目前还没有明确的单一生物标志物诊断IA。早期发现至关重要，但具有挑战性，延迟发现和延迟转诊往往会导致失去应该开始有效治疗的机会之窗，而拖延可能导致残疾和相关的失业。对于被诊断为IA的患者，IA的结果和活动(如突发)在不同患者之间的表现非常不同。来自国家早期炎症性关节炎审计的现实世界数据显示，少数民族风湿病患者的护理不平等。与白人患者相比，少数民族患者实现疾病缓解的比例较低。最后，天气是造成耀斑不均匀的另一个因素。尽管有大量未得到满足的需求，但与其他疾病相比，RMD，特别是IA，在现实世界中的ML应用方面仍然是一个未被充分开发的领域。现有的最大似然法研究不适合于实践中的早期检测目的，因为它们没有根据早期检测点可用的数据进行训练。此外，尽管有研究表明IA的潜在决定因素，但还没有研究或任何机器学习方法可以识别未检测到的决定因素-可以提供有用的IA预测水平的组合。这是因为当前的ML方法仍然不能处理具有不同数据类型、不同模式、不同上下文、不同队列和不完全程度的异类数据集之间的潜在关系。另一方面，IA中的现有ML方法，以及一般的医疗保健，仍然依赖于“一刀切”的范例来呈现通用学习算法，这在个体层面上是次优的，特别是因为IA从诊断时起就已知其本质上是异质的。虽然已经有了可解释的ML局部方法，但关于量化和解释模型预测不确定性及其在实践中的可用性的研究有限。对于内科医生来说，要使用和信任ML预测，关键是要了解与这些预测相关的不确定性对单个患者的影响。尽管成功的翻译需要汇集来自许多学科的专业知识和利益相关者，但ML解决方案的开发目前仍在孤岛进行，缺乏全面和可扩展的ML开发管道。尽管目前的ML有所有局限性，但ML仍有巨大的发展机会，特别是在风湿学的应用方面，因为风湿学已经在英国使用虚拟诊所和远程监测方面处于领先地位。现在是时候使用为IA的真正早期检测和个性化管理生成的数据来推动ML。我们的愿景：拟议的项目将开发有用和负责任的机器学习方法，以实现炎性关节炎的真实世界早期检测和个性化疾病结果预测。我们将通过一个跨学科团队开发一种全面和可扩展的方法，以解决炎症性关节炎的紧迫医疗挑战和机器学习的限制，以加速真实世界的ML在医疗保健中的应用。

Over 20 million people in the UK live with rheumatic and musculoskeletal diseases (RMD), and inflammatory arthritis (IA) is a major subdivision of RMD causing joint inflammation leading to damage. IA causes long-term pain, disability and incurs substantial personal and societal costs. There is also an estimated 59% increase in diagnosed IA cases between 2004 and 2020 in the UK which has important implications for health services. Rheumatology departments accounted for approximately 9% of the average NHS trusts total medication spend in 2019/2020. There are still significant unmet needs in the IA patient pathway, especially in IA detection and flare management. IA presents with non-specific symptoms and there is currently no diagnostically definitive single biomarker for IA. Early detection is critical but challenging, and delay in detection and late referral often result in loss of the window of opportunity when effective treatment should start and delays can lead to disability and associated unemployment. For patients who are diagnosed with IA, IA outcomes and activities such as flare-up are very heterogeneous in their manifestations between individual patients. Real-world data from The National Early Inflammatory Arthritis Audit showed inequality in care for rheumatology patients from minority ethnic groups. A lower proportion of ethnic minority patients achieved disease remission compared to white patients. UN4 Finally, weather is another contributing factor of IA flare heterogeneity. Despite significant unmet needs, RMD, especially IA, is still an underexplored area of real-world ML application in comparison with other diseases. Existing ML studies do not fit for purpose of early detection in practice as they are not trained based on the data available at the point of early detection. Furthermore, although there are studies showing potential determinants of IA, there is no research, or any machine learning methods that can identify the undetected determinants-combination that can offer a useful level of prediction of IA. This is because current ML approaches still cannot handle the underlying relationships among heterogenous datasets with different data types, modalities, contexts, cohorts and levels of incompleteness. On the other hand, existing ML methods in IA, and healthcare in general, still rely on a "one-size-fits-all" paradigm rendering generic learning algorithms, suboptimal on the individual level especially as IA is known to be heterogenous in nature from the time of diagnosis. Although there are methods for explainable ML local, there is limited research to quantify and explain model prediction uncertainty and its usability in practice. For a physician to use and trust ML predictions it is critical to understand the uncertainty associated with these predictions for the individual patient. Although successful translation requires bringing together expertise and stakeholders from many disciplines, the development of ML solutions is currently occurring in silos, and there is a lack of holistic and scalable ML development pipeline. Despite all the limitations of current ML, there are huge opportunities to advance ML, especially in rheumatology applications, because rheumatology has already been leading the way in the use of virtual clinics and remote monitoring in the UK. It is now time to advance ML using data generated for real early detection and personalised management of IA. Our vision: The proposed project will develop useful and responsible machine learning methods to achieve real-world early detection and personalised disease outcome prediction of inflammatory arthritis. We will develop a holistic and scalable approach through an interdisciplinary team addressing the pressing healthcare challenges of inflammatory arthritis and the limitations of machine learning to accelerate real-world ML application in healthcare.