Deep Learning for Automated Aortic Stenosis and Valvular Heart Disease Detection Using a Digital Stethoscope

摘要 该SBIR第二阶段项目将开发一种基于深度学习的临床决策支持算法，用于检测 以及基于使用Eko Core和Eko Duo Digital记录的心音诊断心脏瓣膜病 听诊器。这一筛查工具将有助于减少心脏瓣膜疾病患者的数量 仅仅因为他们的病情没有得到诊断，他们仍然得不到足够的治疗。听诊通常是通过 哪种心脏瓣膜病最先被发现，但往往无法通过超声心动图进行诊断 因为临床医生无法检测到心脏杂音，特别是在嘈杂或匆忙的环境中。要解决这个问题 挑战，Eko开发了Core，这是一种数字听诊器附件，可以在线添加到临床医生的 现有的可放大心音和Duo的听诊器，手持形式的数字听诊器 内置单导联心电图机。这两种设备的设计都是为了将数字化心音图传输到 智能手机、平板电脑或个人电脑。在那里，可以使用决策支持算法来分析信号 我们将作为这个项目的一部分进行开发。本研究的具体目的是：(1)收集有条件的数据库-- 特定的录音标签，通过在六个临床室收集临床数据来实现心音的深度学习 以及(2)开发和评估一系列基于深度卷积神经网络的算法 接受了数据库方面的培训。这些算法将(2a)区分收缩、舒张期和连续期 杂音，(2b)将主动脉瓣狭窄(AS)、二尖瓣关闭不全(MR)、三尖瓣关闭不全(TR)和无辜进行分类 杂音(2c)评估AS、MR和TR的严重程度。通过将这些深度学习算法集成到EKO的 移动和云软件平台，目前由全球1000多家机构的临床医生使用，我们 预计这一算法将使成人患者能够更准确地筛查心脏瓣膜疾病，领先 为更早的诊断和更好的患者结局。

Abstract This SBIR Phase II project will develop a deep learning-based clinical decision support algorithm for detecting and diagnosing valvular heart disease based on heart sounds recorded using the Eko Core and Eko Duo Digital Stethoscopes. This screening tool will help to decrease the number of patients with valvular heart disease that remain undertreated simply because their condition is not diagnosed. Auscultation is commonly the method by which valvular heart disease is first detected, but cases often fail to be referred to echocardiography for diagnosis because clinicians fail to detect heart murmurs, particularly in noisy or rushed environments. To address this challenge, Eko had developed the Core, a digital stethoscope attachment that can be added in-line to a clinician’s existing stethoscope that amplifies heart sounds and Duo, a digital stethoscope in a handheld form factor with built-in single lead electrocardiogram. Both devices are designed to stream digitized phonocardiograms to a smartphone, tablet or personal computer. There, the signal can be analyzed with the decision support algorithm we will develop as part of this project. The specific aims of this study are: (1) to collect a database with condition- specific recording labels to enable deep learning for heart sounds though clinical data collection at six clinical sites, and (2) to develop and evaluate a collection of deep convolutional neural network-based algorithms trained on the database. These algorithms will (2a) distinguish between systolic, diastolic and continuous murmurs, (2b) classify aortic stenosis (AS), mitral regurgitation (MR), tricuspid regurgitation (TR), and innocent murmurs (2c) assess the severity of AS, MR and TR. By integrating these deep learning algorithms into Eko's mobile and cloud software platform, currently used by clinicians at over 1000 institutions worldwide, we anticipate this algorithm will enable more accurate screening for valvular heart disease in adult patients, leading to earlier diagnosis and better patient outcomes.