Approved Research

Prediction of clinical outcomes in patients with aortic valve regurgitation using a deep learning approach

Lay summary

Current class I recommendations endorsed by cardiac societies regarding aortic valve replacement in patients with aortic regurgitation without significant aortic root disease are exclusively indicated when it is graded as severe, whereas moderate aortic regurgitation is not regarded for aortic valve replacement. Left ventricular remodeling and clinical outcomes have been extensively studied in severe aortic regurgitation. However, the relationship between aortic regurgitation severity and left ventricular dysfunction severity remains unclear. Most natural history and epidemiologic studies were performed using traditional statistical analysis, failing to extrapolate results to the "real-world" setting of patients. On the other hand, artificial intelligence has shown promise in handling a patient's high-dimensionality. With this project we aim to predict development of symptoms and mortality in patients with chronic aortic regurgitation using two different artificial intelligence approaches: clustering analysis and hybrid deep learning models, following two work packages.

After screening for patients with aortic valve regurgitation, a high-dimensional dataset of input features varying from concomitant co-morbidities, genetic traits to imaging examinations will be considered. In work package 1, we aim at revealing novel patient group phenotypes using an unsupervised machine learning technique, by integrating different cardiac MRI features into a topological network. Also, considering the impact of concomitant cardiomyopathy, genetic background will be added to the topological network. We will analyze the network's geometry and define clinically relevant groups. Then, survival analysis upon the groups will be performed. Exploring the potentialities of hidden information from raw imaging data, in work package 2, models will be expanded to raw cardiac MRI data, based on a hybrid variational autoencoder model. Training and evaluation will use independent training, validation and testing sets. Python's high-level API Keras will be used. 

We estimate the duration of this project to be 24 Months. Considerable impact on public could be attained with this project, as it could define a more precise and individual decision line between conservative and interventional/surgical therapy in patients with chronic aortic valve regurgitation. Also, deepening the investigation on modeling time-to-event clinical outcomes from both raw imaging and structured data would be inestimable to any other field in cardiology pursing risk assessment. Moreover, we expect to generate novel variables by annotation of cardiac MRI files, which could be essential in future studies. Lastly, integration of genetic background data carries itself an unparallel approach to risk assessment in aortic valve disease.

Scope extension:

To date, in patients without significant aortic root disease, guidelines deny aortic valve replacement (AVR) in moderate aortic valve regurgitation (AR), regardless of symptom development or left ventricular (LV) status. Current guidelines relate to natural history studies using traditional statistics, based on arbitrary AR grading thresholds, which fail to understand the spectrum of patient's phenotypic representations. Recent applications of topological data analysis (TDA) and deep learning have shown promise in risk prediction. Application of these techniques to AR could be decisive for the treatment strategy in these patients, possibly defining newer phenotypic patient groups, potentially with large repercussions on newer guidelines. 

In this project, we propose to predict symptom development and mortality following two work packages. In work package 1, we will use a TDA pipeline to unveil patterns in high-dimensional parameters (based on cardiac magnetic resonance imaging [MRI] derived aortic regurgitant fraction, left ventricular systolic function, T1 mapping features and genetic background) by analyzing the geometrical structure retrieved from the topological network, followed by survival analysis. In work package 2, we will explore the impact of latent representations from raw imaging data (cardiac MRIs) on outcome prediction, by developing supervised deep learning algorithms using hybrid variational autoencoders (VAEs).

Additionally, considering the genetic background component mentioned above, we intend to study the genetic underpinnings of this main phenotype of aortic valve regurgitation (retrieved using deep learning techniques over the cardiac MRI) as well its related cardiac flow volumes phenotypes (such as aortic peak forward and regurgitant velocities and volumes). This could greatly improve our understanding of aortic valve regurgitation and define new concepts of individualized therapy.