Prediction of clinical outcomes in patients with moderate aortic valve regurgitation using deep learning
Cardiology, Angiology
Final Report Abstract
While severe aortic valve insufficiency's clinical course has been explored over the last decades, the genetic and pathophysiological connection between aortic valve insufficiency and left ventricular dysfunction as continuous traits remained unclear. In this regard, artificial intelligence (AI)-based modeling offers a promising diagnostic option. Therefore, the project's primary objective was to explore aortic valve insufficiency's natural progression through modern AI-based modeling (Work Package 1), examining both genotypical and phenotypical aspects. The secondary aim was to confirm if accurate AI-based individual prediction models for clinical outcomes in moderate aortic valve insufficiency could be developed (Work Package 2). Leveraging the multicentric UK Biobank's resources, DeepFlow, a deep learning model using a U-Net to extract aortic-flow dependent parameters from cardiac MRIs at scale (N = 39745), could be developed. Due to the detailed genotypical profiling of the majority of participants, pioneer insights into chronic aortic valve insufficiency's genetic architecture were enabled. Notably, the locus LINC01808 on chromosome 2 (in the genome-wide-association study considering the entire population), and AMZ1/GNA12 on chromosome 7 (in the sensitivity analysis) were significantly associated with aortic valve insufficiency fraction. Mendelian randomization identified a causal role for aortic root size in aortic valve insufficiency (+0.23% per cm2/m2). Prominent contributing variants were linked to genes near ELN, PRDM6, and ADAMTS7, implicated in connective tissue and blood pressure pathways. Using topological data analysis (TDA), an incomplete 'loop' pattern in cardiac functional and morphological parameters across more severe aortic valve insufficiency severity could be observed. However, the average severity of the insufficiency itself does not appear to be the predominant factor influencing this TDA structure. Examining clinical outcomes with previous moderate and severe aortic valve insufficiency cut-offs (>22% and >33%, respectively), Kaplan-Meier curves confirmed higher severity's association with elevated mortality. This data aligns with the current guidelines' treatment cut-offs. Modest all-cause mortality predictions could be achieved with random forest (AUC 0.57), support vector machine (AUC 0.63) and neural networks (AUC 0.67) models for predicting all-cause mortality in at least moderate aortic valve insufficiency fraction, in a hold-out test set. In conclusion, this study unveils the genomic architecture of aortic valve insufficiency fraction and its intricate relationship with aortic size. Mendelian randomization analysis is groundbreaking in detecting genetic instruments that could influence aortic valve insufficiency. Future studies that explore the entire cardiac disease spectrum using recent transformer/based unsupervised clustering models directly from raw data could yield more powerful results.
Publications
-
Defining left ventricular remodeling using lean body mass allometry: a UK Biobank study. European Journal of Applied Physiology, 123(5), 989-1001.
Gomes, Bruna; Hedman, Kristofer; Kuznetsova, Tatiana; Cauwenberghs, Nicholas; Hsu, David; Kobayashi, Yukari; Ingelsson, Erik; Oxborough, David; George, Keith; Salerno, Michael; Ashley, Euan & Haddad, Francois
-
Genetic architecture of cardiac dynamic flow volumes. Nature Genetics, 56(2), 245-257.
Gomes, Bruna; Singh, Aditya; O.’Sullivan, Jack W.; Schnurr, Theresia M.; Goddard, Pagé C.; Loong, Shaun; Amar, David; Hughes, J. Weston; Kostur, Mykhailo; Haddad, Francois; Salerno, Michael; Foo, Roger; Montgomery, Stephen B.; Parikh, Victoria N.; Meder, Benjamin & Ashley, Euan A.
