Machine Learning Development for Subtyping COPD

项目摘要 慢性阻塞性肺疾病(COPD)是一种异质性的肺部疾病，其特征是 进行性肺功能丧失，随之而来的呼吸困难和生活质量恶化。这 异质性使得预测健康下降和开发有针对性的治疗方法以更好地照顾患者变得困难。 到目前为止，研究人员试图使用标准的机器学习方法来识别更多 有意义的COPD亚型，但这些方法通常对数据做出一般性假设，限制了 它们在某些数据集中穿透更复杂模式的能力。因此，有意义的重新分类 可能导致更有针对性的治疗和干预的COPD亚型一直难以捉摸。申请人 介绍了一种看待COPD亚型问题的新方法，即从发现的角度对其进行重塑 个体与疾病轨迹的关联--即，根据个体在 对引起变数的环境和/或疾病的反应。提出的机器学习方法构建 关于贝叶斯非参数计量学的最新进展，这是一组理论思想和技术的集合， 允许非常灵活的数据表示。在这份职业发展建议书中，申请人假设 这些机器学习方法及其扩展，以及以前未使用过的数据源 对于COPD亚型-将产生更多具有生物学意义的患者亚组，导致更好的 了解疾病的遗传和生物学基础并最终改善患者 管理层。此应用程序的目标1涉及评估CT评估的肺肿块的实用性--潜在的 比常规使用的肺气肿指标更具区分性--用于确定COPD亚型 使用K-均值聚类和我们的疾病轨迹算法。目标2的目标是评估效用 使用我们的轨迹聚类算法来确定COPD亚型的合并症数据。新奇的计算机 将探索基于断层扫描的肌肉损耗(恶病质)和肺血管修剪的措施，以 确定它们在亚型鉴定中的有效性。此外，我们将对轨迹算法进行扩展和测试 为了对离散输出(例如医生诊断的并存)建模，计数数据(例如 病情加重)和事件发生时间数据(死亡)。在目标3中，申请者将扩展我们的轨迹分类 直接合并遗传和组学数据以发现亚型的算法。总之，这项研究 在本奖项的目的中提出的建议将充分利用通过 COPD基因研究。 通过与罗恩·基基尼斯博士积极合作，可以实现本提案中的目标 医学图像分析领域的知名领导者，以及国际公认的 慢性阻塞性肺疾病遗传流行病学专家。

Project Summary Chronic obstructive pulmonary disease (COPD) is a heterogeneous lung condition characterized by progressive loss of lung function with subsequent increasing breathlessness and worsening quality of life. This heterogeneity makes it difficult to predict health decline and develop targeted treatments for better patient care. To date, researchers have attempted to use standard machine learning methodology to identify more meaningful subtypes of COPD, but these methods often make general assumptions about the data, limiting their ability to penetrate more complex patterns in some data sets. Thus, a meaningful reclassification of COPD subtypes that could lead to more targeted therapies and interventions has been elusive. The applicant introduces a new way of looking at the COPD subtyping problem by recasting it in terms of discovering associations of individuals to disease trajectories – i.e., grouping individuals based on their similarity in response to environmental and/or disease causing variables. The machine learning methods proposed build on the most recent advances in Bayesian nonparametrics, a collection of theoretical ideas and techniques that permit very flexible data representations. In this career development proposal, the applicant hypothesizes that these machine learning methods and extensions thereof – together with data sources not previously leveraged for COPD subtyping – will produce more biologically meaningful sub-groupings of patients, leading to a better understanding of the genetic and biological underpinnings of the disease and ultimately improved patient management. Aim 1 of this application involves evaluating the utility of CT-assessed lung mass – a potentially more discriminative measure of emphysema than conventionally used measures – for defining COPD subtypes using both K-means clustering and our disease trajectory algorithm. The goal of Aim 2 is to evaluate the utility of comorbidity data for defining COPD subtypes using our trajectory clustering algorithm. Novel computed tomography based measures of muscle wasting (cachexia) and pulmonary vascular pruning will be explored to determine their efficacy in subtype determination. Additionally, we will extend and test the trajectory algorithm in order to model discrete outputs (such as physician-diagnosed comorbidities), count data (e.g. exacerbations), and time-to-event data (death). In Aim 3, the applicant will extend our trajectory clustering algorithms to directly incorporate genetic and omics data for subtype discovery. Together, the research proposed in the aims of this award will take full advantage of the comprehensive data set available through the COPDGene study. Execution of the aims in this proposal will be possible through active collaboration with Dr. Ron Kikinis, M.D., a renowned leader in the field of medical image analysis, and Dr. Ed Silverman, an internationally recognized expert in the genetic epidemiology of COPD.