Automated Sonographic Detection of Pulmonary Embolism Using Machine Learning Algorithm

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT We propose a better way to diagnose pulmonary embolism (PE) early and save lives. More than 900,000 people in the United States suffer from acute PE, and about 100,000 die each year. With 10% of such cases being fatal within the first hour of the onset of symptoms, rapid diagnosis of PE is critical to direct appropriate therapy. Unfortunately, clinical evaluation alone is unreliable and often results in grave diagnostic delays. Furthermore, while echocardiography at the patient’s bedside can rapidly detect heart dysfunction caused by PE, traditional echocardiography performed by cardiology services is not readily available in acute care settings. Thus, there is a critical need for use of a rapid, non- invasive diagnostic tool at the point-of-care (POC) to accurately assess for PE and direct emergency therapy. The focus of this research is to develop innovative artificial intelligence algorithms that can transform the care of patients with PE by enabling non-experts to use echocardiography to detect PE, direct emergency therapy, and improve survival. The rationale underlying this proposal is that the proposed artificial intelligence technology tools will provide a relatively simple and time-efficient strategy that can be implemented in most healthcare settings. This will, in turn, fulfill the overall goal of creating a positive shift in the management of patients presenting with PE. The proposed specialized artificial intelligence technology would ultimately be applicable to early detection of a wide variety of diseases. The long-term goal of our research is to develop and implement effective automated ultrasound tools that would significantly impact the diagnosis and treatment of different life-threatening conditions. The objective of this proposal is to develop and validate a prototype mobile artificial intelligence enabled-software platform that can accurately detect echocardiographic signs of PE. The hypothesis is that artificial intelligence algorithms will achieve levels of diagnostic accuracy equivalent to expert physician sonographers in detecting PE. This hypothesis will be tested by pursuing two specific aims: 1) Develop a machine learning algorithm for the detection of PE that can be extended to detect other cardiopulmonary conditions using explicit echocardiographic signs of PE and implicit image content representations. 2) Validate the accuracy of the machine learning algorithm to detect PE on echocardiographic images using explicit sonographic signs. Innovative reinforcement learning techniques will be utilized to accomplish the specific aims. The proposed research is significant because it will transform the care of patients with PE by enabling non-experts to use POC echocardiography. It will also have an immediate, positive impact because it will help lower morbidity, mortality, improve quality of life, and decrease healthcare costs by expediting diagnosis and therapeutic interventions. The proximate expected outcome of this work is improvement in the evaluation of patients with life-threatening PE by inexperienced healthcare providers, which will result in more accurate and rapid identification of cases that require emergency treatment. Our proposal aligns with the NIBIB’s overall mission to advance healthcare through innovative engineering and, more specifically, its emphasis on development of transformative unsupervised and semi-supervised machine learning technologies to enhance analysis of complex medical images and data for diagnosing and treating a wide range of diseases and health conditions.