Cardiac Ultrasound Radiomics-Guided Deep Neural Networks for Acute Myocardial Infarction Precision Phenotyping - Project Summary Approximately every 40 seconds, an American suffers from acute myocardial infarction (AMI), with over 25% of AMI patients dying within one year of hospitalization. Although several strategies are available to help risk-stratify patients with AMI, the existing clinical risk scores have only modest accuracy. This research aims to fill this gap by providing non-invasive and objective prognostic quantitative imaging markers for personalized risk stratification for AMI patients. Our preliminary clinical data support that cardiac ultrasound radiomics−a mathematical framework that converts standard of care cardiac ultrasound images into minable high-dimensional data−can identify patients at high risk for hospitalization for adverse cardiac events. However, progress in developing these novel markers is limited by a lack of optimization, standardization, and validation−all critical barriers to clinical use. Leveraging the existing imaging and clinical database of patients admitted at several hospitals within our Health System, we have assembled a multidisciplinary team of expert physicians, imagers, and engineers to facilitate the development and validation of the proposed technology. Our central hypothesis is that integrating cardiac ultrasound radiomics, conventional echo, and clinical data using deep learning is incremental to the currently recommended strategies in determining phenotypic presentations and prognosis in AMI. Specifically, we will (1) we will benchmark cardiac ultrasound radiomics-guided deep neural network model to conventional risk assessment (GRACE 2.0 score, EF) in predicting 1-year risk of all-cause mortality as the primary endpoint. We will develop a new Cardiac Ultrasound Radiomics Exploration in AMI (CURE-AMI) probability score that integrates radiomics with laboratory and clinical data to identify a high-risk AMI phenogroup, (2) in a prospective cohort study of patients presenting with non-ST elevation acute coronary syndrome, we will assess if point of care cardiac ultrasound (POCUS) radiomics compared to regional and global left ventricular function assessment improves the identification of high-risk coronary anatomy, including the presence of acute coronary vessel occlusion, and (3) in a prospective cohort study of patients presenting with AMI, we will assess whether a parametric display of ultrasonic radiomics features compared to conventional echo parameters will better estimate infarct size and location using paired blinded CMR assessment. Combined with our rich previous experience in developing machine-learning algorithms, this application is a unique opportunity to utilize radiomics for stratifying one of the most common problems in healthcare. Successful completion of our aims will help risk- stratify AMI patients along the healthcare continuum, from early diagnosis and institution of time-sensitive therapies to personalized care of those who remain at substantial risk for long-term MACE.
