Heart failure is to be the most common cause of hospitalization in United States and cost in excess of $30.7 billion annually. Echocardiography is the predominant imaging modality for diagnosing heart failure and identifying patients who would benefit from procedures and targeted therapies, however there is heterogeneity in the interpretation of echocardiograms and a significant amount of imaging information is not currently used in clinical care. Through a lifetime of stress and environmental risk factors such as smoking and obesity, cardiovascular risk factors likely left imprints on the myocardium that is not readily detectable by the naked eye. Recent advances in machine learning suggest deep learning can identify phenotypes in imaging unrecognized by human experts. Harnessing the exponential increase in available computational power, many of the biggest recent advances in deep learning came from computer vision and imaging processing tasks. This led to my central hypothesis is that deep learning can identify subclinical disease and cardiovascular risk factors not readily detected by human interpreters, and this phenotyping can risk-stratify patients and alter management. The series of experiments described herein aim to establish a machine learning framework to allow for the extraction of patient characteristics and risk factors from echocardiography images and create an image-based prediction model of cardiovascular disease as well as create a deep learning prediction model to assess hemodynamic measurements with higher accuracy and confidence than existing non-invasive methods.
Aim 1 applies convolutional neural networks on echocardiogram images and videos to predict patient phenotypes, cardiac diseases, and cardiovascular risk factors. Using the results of these individually trained models, we will create a prediction neural network to predict future coronary artery disease, heart failure hospitalizations, and mortality.
Aim 2 utilizes recurrent neural networks on echocardiogram videos to assess predictions of hemodynamic measurements by comparing machine learning algorithm outputs to invasive measurements obtained in the catherization lab. Together, these studies will significantly expand our understanding of the applicability of machine learning techniques to echocardiography, and significantly advance our understanding of imaging phenotypes of cardiovascular disease and risk factors with the hope of identifying subclinical disease and predicting future adverse outcomes.
The proposed project aims to apply machine learning techniques to echocardiograms to identify subclinical disease phenotypes, recognize specific cardiovascular risk factors, and predict future morbidity and mortality due to cardiovascular disease. These studies will significant advance our understanding of imaging phenotypes of cardiovascular disease and the ability for machine learning algorithms to rapidly recognize cardiovascular disease in an automated fashion with the hope of reducing current variation in interpretation of echocardiograms and recognize early disease amenable to intervention.
