Machine learning driven transthoracic echocardiographic analysis and screening for cardiac amyloidosis

点击翻译按钮获取中文摘要

Machine learning driven transthoracic echocardiographic analysis and screening for cardiac amyloidosis Cardiac amyloidosis (CA) is a serious but increasingly treatable cause of heart failure. Autopsy studies have estimated the prevalence of CA at approximately 25% of all octogenarians, and 15 to 20% of patients with aortic stenosis. Despite the increasing prevalence of CA within the general population and specific subpopulations, its diagnosis as a cause of heart failure is hampered by under recognition and subsequent underdiagnosis in clinical practice. Data suggest that the average time from onset of symptoms to diagnosis is 2 years and that patients report seeing an average of 5 physicians prior to establishing a definitive diagnosis. Transthoracic echocardiography (TTE) testing is the most common initial evaluation because of its wide availability. A recent utilization review in the Medicare population indicated over 7 Million echocardiographic tests are performed each year accounting for $1.2 Billion in healthcare costs. TTEs provide comprehensive information about cardiac structure and function, yet complexity of interpretation has limited its screening performance in patients with CA, and diagnosis can be challenging. Thus, our group seeks to offer a computer vision and machine learning based TTE analysis and screening solution for CA. We are uniquely positioned for accelerated development with a cohort of 359 patients with confirmed CA and 4,862 controls. In Phase I, we will build a deep learning neural network-based image processing pipeline. It maps the TTE sequence into a 2-dimensional space that allows for the identification of the 4-chamber peak diastolic and peak systolic images within the cardiac heartbeat cycle. This will enable our screening model to recognize regional myocardial wall motion changes and hypertrophic patterns that characterize amyloidosis in comparison to controls with normal cardiac function. The operational point defining the performance characteristics of our screening-oriented model (including sensitivity, specificity, and negative predictive value) will be optimized using an average weighted accuracy (AWA) approach which accounts for CA disease prevalence along with a desired false positive and false negative tradeoff. If we are successful, we envision a Phase II proposal to build and deploy an automated TTE analysis tool, and to evaluate it in a multi- center clinical study. This sets the stage for our long-term goal to implement a computer assisted TTE screening solution to improve identification and by extension care of patients with cardiac amyloidosis.