Identifying pediatric asthma subtypes using novel privacy-preserving federated machine learning methods

摘要 在美国，哮喘影响着近600万儿童，平均而言，每个患有哮喘的儿童都会经历 每年至少发生一次急性加重。儿科哮喘占急诊科就诊的790，000多人次， 64，000例住院治疗，每年直接医疗费用近60亿美元。哮喘不成比例 影响到少数民族儿童，他们面临更严重后果的风险。哮喘是一种异质性疾病， 病因、触发因素、严重程度和治疗反应的范围（即，亚型）。尽管众所周知， 由于哮喘的异质性，哮喘亚型在很大程度上局限于过敏性与 非过敏性，不考虑重叠亚型、亚型演变、严重程度，也不包括 健康的社会决定因素（SDOH）。因此，如果我们要减轻哮喘的负担， 和人群水平，我们必须改善哮喘亚型表征，以帮助临床医生制定更个性化的 初级和紧急护理。然而，迄今为止，哮喘亚型研究受到小样本的限制 大小，忽略时间信息，和/或专注于个别或少数网站。大规模扩散 临床研究网络（CRN）与来自电子健康记录（EHR）的真实世界数据（RWD）相结合 随着机器学习的进步，为改善儿童哮喘的亚型提供了独特的机会 患者我们的研究小组对OneFlorida+ CRN中哮喘急性发作的初步分析仅使用 结构化数据发现了五种儿童哮喘亚型，这些亚型因种族/民族、严重程度、数字生物标志物 和合并症。我们的工作支持了儿童哮喘除了遗传因素外还有更多的异质性。 经典定义的过敏与非过敏亚型。在这个项目中，我们将利用OneFlorida+ CRN的 RWD的大型存储库（覆盖东南部近2000万患者）和一种新颖的隐私保护 基于联邦机器学习的框架：（1）识别儿科哮喘患者，他们的严重程度，亚型， 和疾病进展（即，进展亚型），以及（2）将这些全球模型微调为本地OneFlorida+ 具有站点特定数据的站点，以说明站点间异质性。除了结构化的EHR数据，我们还将 包括时空链接环境数据且使用自然语言处理来包括临床笔记 症状和SDOH等数据。为了指导我们的工作并为实施工作提供信息，我们将建立一个 利益相关者咨询委员会，由儿童哮喘、医疗保健系统和公共卫生利益相关者组成， 与当地OneFlorida+站点临床医生进行焦点小组讨论，以开发和测试集成 子类型数据。使用来自不同人群的RWD构建的儿童哮喘进展亚型， 利益相关者的参与将使该领域更接近精确的初级和紧急护理， 结果。我们新颖的隐私保护联邦机器学习方法解决了以下几个挑战： RWD分析，并将成为其他CRN采用的可推广框架， 传播这项工作，并为其他慢性疾病的进展亚型分析铺平道路。

ABSTRACT Asthma affects nearly 6 million children in the United States, and on average, each child with asthma experiences at least one exacerbation per year. Pediatric asthma accounts for over 790,000 emergency department visits, 64,000 hospitalizations, and nearly $6 billion in direct healthcare costs annually. Asthma disproportionately affects minority children, who are at risk for more severe outcomes. Asthma is a heterogeneous disease with a range of etiologies, triggers, severities, and treatment responses (i.e., subtypes). Despite that well-known heterogeneity, asthma subtypes are largely confined to a simple dichotomous classification of allergic versus non-allergic, which does not account for overlapping subtypes, subtype evolution, severity, nor do they include social determinants of health (SDOH). As such, if we are to reduce the burden of asthma at both the individual and population level, we must improve asthma subtype characterization to help clinicians craft more personalized primary and emergency care. To date, however, asthma subtyping studies have been limited by small sample sizes, ignored temporal information, and/or focused on individual or a handful of sites. The proliferation of large clinical research networks (CRNs) with real-world data (RWD) from electronic health records (EHRs), combined with advancements in machine learning offer unique opportunities to improve subtyping of pediatric asthma patients. Our study team’s preliminary analysis of asthma exacerbations in the OneFlorida+ CRN using only structured data found five pediatric asthma subtypes which varied by race/ethnicity, severity, digital biomarkers, and comorbidities. Our work supports that there is further heterogeneity in pediatric asthma beyond the classically defined subtypes of allergic vs non-allergic. In this project, we will leverage the OneFlorida+ CRN’s large repository of RWD (covering nearly 20 million patients in the southeast) and a novel privacy-preserving federated machine learning-based framework to: (1) identify pediatric asthma patients, their severity, subtypes, and disease progression (i.e., progression subtypes), and (2) fine-tune those global models to local OneFlorida+ sites with site-specific data to account for between-site heterogeneity. In addition to structured EHR data, we will include spatiotemporally linked environmental data and use natural language processing to include clinical note data such as symptoms and SDOH. To guide our work and inform implementation efforts, we will establish a stakeholder advisory committee with pediatric asthma, healthcare system, and public health stakeholders, and conduct focus groups with local OneFlorida+ site clinicians to develop and test EHR prototypes that integrate subtype data. Pediatric asthma progression subtypes built using RWD from diverse populations combined with stakeholder engagement will move the field closer to precision primary and emergency care that improves outcomes. Our novel privacy-preserving federated machine learning methods address several challenges of RWD analysis and will be a generalizable framework for other CRNs to adopt, facilitating widespread dissemination of this work, and paving a path forward for progression subtype analyses of other chronic diseases.