Deep Learning Based Phenotyping and Outcomes Prediction for Valvular Heart Disease

ABSTRACT Mitral valve disease is a common underrecognized disorder affecting over 170 million individuals worldwide, however a sizable proportion of individuals will experience adverse myocardial response leading to degenerative mitral regurgitation, heart failure, cardiac arrest, or even sudden cardiac death. Myocardial remodeling responding to mitral regurgitation can lead to changes in cardiovascular structure and function, but general patterns of change during early disease (such as diastolic dysfunction and increased sphericity) are insufficient to distinguish between appropriate compensated myocardial response to innocent mitral valve prolapse vs. early manifestations of future symptomatic degenerative mitral regurgitation associated with adverse outcomes. Recent advances in artificial intelligence (AI), when applied to biomedical images, indicate that deep learning algorithms can both offer precise measurements beyond human fidelity and also identify subtle traits in imaging that are unrecognized by the human eye. Our prior work applying AI to echocardiography as well as others applying AI to other forms of medical imaging, have shown deep learning can not only reproduce standard measures of cardiac structural and function but is also identify CVD risk features including chronologic age, biological sex, diabetes, hypertension, and smoking. In effect, AI applied to echocardiography now offers the potential to capture a precise phenotyping of mitral valve disease, characterize myocardial response, aggregate risk of future comorbidity and, in turn, identify intervenable target for mitigating future CVD risk. Thus, we hypothesize that AI methods applied to echocardiography can be used to not only precisely characteristics mitral valve disease by (i) automating the phenotyping of mitral regurgitation severity, systolic dysfunction, diastolic parameters, and other measures of adverse myocardial response that demonstrate trajectories of accelerated versus delayed cardiovascular risk, (ii) discern key contributors to accelerated cardiovascular risk in mitral valve disease, and (iii) be used on large clinical cohorts to identify clinical underdiagnosis. To test these hypotheses, we will leverage echocardiography images that are acquired as a part of routine clinical care in addition to serial imaging frequently obtained in large cohort studies of aging adults. Because healthcare cohorts are enriched for patients with accelerated and high risk trajectories, while epidemiologic cohorts are enriched for individuals with delayed and low risk trajectories, we plan to analyze imaging and outcomes data collected from both types of cohorts. Accordingly, our aims are to train and validate AI models on phenotyping mitral valve disease from cardiac imaging and identify its component contributors to excess cardiovascular risk in both the healthcare setting and in the community setting.