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

• Applicant: Dr. Bruna Filipa Gomes Botelho Quintas
• Subject Area: Epidemiology and Medical Biometry/Statistics ; Cardiology, Angiology
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 457899528

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 moderate aortic regurgitation 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, after applying a topological network analysis and retrieving the geometrical structure of the disease’s natural history (work package 1), we aim to predict impaired systolic left ventricular function, development of symptoms and mortality in patients at risk with moderate aortic regurgitation using deep learning models (work package 2). In work package 2.1, we aim at the prediction of clinical outcomes using different machine learning algorithms, which will learn from structured data from the baseline follow-up visit. In work package 2.2, input will expand to longitudinal data, using information from sequential follow-up visits. Herein, we will develop a baseline Long Short-Term Memory based model. In an experimental model, we will explore the potentialities of medical concepts embedding techniques using pretrained embedding layers. Finally, in work package 2.3, input of the deep learning models will be expanded to raw medical imaging data. The baseline model will be based on an unsupervised, hybrid convolutional denoising autoencoder. Data source will be derived from a highdimensional multicentric dataset (multicentric German dataset, n= 13475; UK Biobank, n=2323) focusing on patients with chronic aortic valve regurgitation, including anonymized electronic medical records containing demographical and genetic data, laboratory exams, concomitant co-morbidities and imaging examinations. Performance evaluation includes using the Heidelberg University cohort as the test set/held-out data, and the remaining dataset (UK Biobank and multicentric German dataset) will be used to perform a non-nested 6-fold cross-validation for training and validation/hyperparameter tuning. Main performance metrics include the calculation of the receiving operating characteristic’s area under the curve. Python’s high-level API Keras will be used. This project could define a more precise and individual decision line between conservative and interventional/surgical therapy in patients with chronic moderate aortic valve regurgitation.

• DFG Programme: WBP Fellowship
• International Connection: USA
• Host: Professor Dr. Euan Ashley, Ph.D.