Machine learning for the automated identification and tracking of rare myocardial diseases

项目摘要 虽然心脏淀粉样变性和肥厚型心肌病（HCM）是相对罕见的心脏疾病， 由于以下几个共同的原因，它们在检测和治疗方面特别具有挑战性：（1）常规 临床成像（即，他们很难从表面上的相似中区分出来， 导致左心室（LV）肥大的更常见的心脏病形式;（2）诊断为 经常被遗漏，因此患者可能在疾病过程的后期出现，此时治疗很困难; (3)尚未确定可靠反映疾病进展的客观、非侵入性指标;以及（4） 少数已知患有这些疾病的患者可能会使流行病学研究和临床试验变得困难 组织和指挥。对于心脏淀粉样变性和肥厚型心肌病，超声心动图在诊断中起着关键作用 以及纵向监测，因为其普遍存在的临床可用性、安全性和低成本。更广泛地说，回声 占主导地位的常规心脏成像的当前景观，与数以千万计的回声在心脏成像中执行。 美国每年。然而，上述临床挑战突出了本发明的几个缺点。 回声：它的能力有限，（1）在早期诊断疾病;（2）区分 形态学相似的疾病;和（3）量化疾病进展。该提案旨在解决 当前回波阅读工作流程中的缺陷，其是主观的并且仅捕获了 每项研究中的数据。该应用程序的总体目标是利用机器学习的进步， 开发和验证全自动回波图像分析方法，以诊断和跟踪罕见 心肌病，重点是心脏淀粉样变性和HCM。我们的建议是以假设为中心的 高度可扩展的计算机视觉方法可以应用于回声研究，以克服局限性 标准临床超声阅读工作流程的一部分。因此，我们的目标是：（1）应用自动化方法 用于回声量化和疾病识别，以检测和区分导致 增加的LV壁厚度;和（2）表征疾病进展的可量化回声测量， 心脏淀粉样变性和HCM，并将这些与临床结果相关联。我们的多学科团队， 由心肌病、超声心动图、计算机视觉和机器学习方面的专家组成， 分析来自2个大型患者登记研究的超声心动图和患者数据：多中心淀粉样蛋白表型研究 （MAPS）和肌节性人类心肌病登记（SHaRe）HCM网络，使用 储存了近40万个回声我们的目标的成功实现将产生一个创新的工具， 心肌疾病的早期诊断和疾病进展的跟踪。重要的是，我们的项目将 通过自动化对罕见心肌疾病进行更大规模的流行病学研究的阶段 识别这些患者，从而开发以前无法实现的广泛队列 对于这些条件。

PROJECT SUMMARY Although cardiac amyloidosis and hypertrophic cardiomyopathy (HCM) are relatively rare causes of heart failure (HF), they are particularly challenging to detect and treat for several shared reasons: (1) on routine clinical imaging (i.e., echocardiography [echo]), they can be difficult to distinguish from superficially similar, more common forms of cardiac disease that cause left ventricular (LV) hypertrophy; (2) the diagnoses are often missed and thus patients can present late in the course of disease at a time when treatment is difficult; (3) objective, noninvasive metrics that reliably reflect disease progression have not been identified; and (4) the small number of known patients with these diseases can make epidemiology studies and clinical trials difficult to organize and conduct. For both cardiac amyloidosis and HCM, echo plays a critical role in both diagnosis and longitudinal monitoring given its ubiquitous clinical availability, safety, and low cost. More broadly, echo dominates the current landscape of routine cardiac imaging, with tens of millions of echos performed in the United States each year. However, the clinical challenges described above highlight several shortcomings of echo: it is limited in its ability to (1) diagnose disease at its early stages; (2) discriminate between morphologically similar diseases; and (3) quantify disease progression. This proposal seeks to address deficiencies in the current echo reading workflow, which is subjective and captures only a small fraction of the data available in each study. The overall objective of this application is to use advances in machine learning to develop and validate fully-automated echo image analytic approaches to diagnose and track rare cardiomyopathies, focusing on cardiac amyloidosis and HCM. Our proposal is centered on the hypothesis that highly scalable computer vision methods can be applied to echo studies to overcome limitations of the standard clinical echo reading workflow. Accordingly our aims are: (1) Apply an automated method for echo quantification and disease identification to detect and differentiate cardiac diseases that cause increased LV wall thickness; and (2) Characterize quantifiable echo measures of disease progression in cardiac amyloidosis and HCM and associate these with clinical outcomes. Our multidisciplinary team, which is composed of experts in cardiomyopathies, echocardiography, computer vision, and machine learning, will analyze echos and patient data from 2 large patient registries: the Multicenter Amyloid Phenotyping Study (MAPS) and the Sarcomeric Human Cardiomyopathy Registry (SHaRe) HCM Network, with validation using a repository of nearly 400,000 echos. The successful completion of our aims will result in an innovative tool for early diagnosis of myocardial diseases and tracking of disease progression. Importantly, our project will set the stage for conducting larger epidemiology studies of rare myocardial diseases by automating the identification of these patients, and thereby developing previously unattainable broad-based cohorts for these conditions.