Deep Learning of Left Atrial Structure and Function Provides Link to Atrial Fibrillation Risk

Abstract

Aims: Increased left atrial (LA) volume is a known risk factor for atrial fibrillation (AF). There is also emerging evidence that alterations in LA function due to an atrial cardiomyopathy are associated with an increased risk of AF. The availability of large-scale cardiac MRI data paired with genetic data provides a unique opportunity to assess the joint genetic contributions of LA structure and function to AF risk. Methods and results: We developed deep learning models to measure LA traits from cardiovascular magnetic resonance imaging (MRI) in 40,558 UK Biobank participants and integrated these data to estimate LA minimum (LAmin), maximum (LAmax), and stroke volume (LASV), as well as emptying fraction (LAEF). We conducted a genome-wide association study (GWAS) in 35,049 participants without pre-existing cardiovascular disease, identifying 20 common genetic loci associated with LA traits. Eight of the loci associated with LA traits were previously associated with AF: the AF risk alleles were associated with an increased LA minimum volume (LAmin) and a decreased LAEF. A Mendelian randomization analysis confirmed that AF causally affects LA volume (IVW P = 6.2E-06), and provided evidence that LAmin causally affects AF risk (IVW P = 4.7E-05). In UK Biobank participants, a polygenic prediction of LAmin was significantly associated with risk for AF (HR 1.09 per SD; P = 1.6E-36) and ischemic stroke (HR 1.04 per SD; P = 4.7E-03). Conclusions: We performed the largest and highest resolution assessment of LA structure and function to date. We then identified 20 common genetic variants associated with LA volumes or LAEF, 19 of which were novel. We found that a polygenic prediction of the minimal LA volume was associated with AF and stroke. Finally, we found an inverse relation between genetic variants associated with AF risk and LAEF. Our findings provide evidence of a causal relation between LA contractile function and AF. Keywords: Left atrium, machine learning, cardiovascular disease, genetics, atrial fibrillation

Genetic Architecture of Stroke of Undetermined Source: Overlap with Known Stroke Etiologies and Associations with Modifiable Risk Factors

“…To explore whether the risk for stroke of undetermined source can be partially explained by established stroke etiologies, we explored in MR analyses whether genetic predisposition to carotid atherosclerosis (proxied by common carotid artery intima media thickness [cIMT] in 71,128 individuals of European ancestry from the CHARGE Consortium 27 ), a main cause of LAAS, to atrial fibrillation (65,446 cases, 522,744 controls from a trans‐ethnic meta‐analysis of the AFGen Consortium, the UK Biobank, the Broad AF Study, and the Biobank Japan 28 ), the primary source of CES, and to cerebral small vessel disease (proxied by white matter hyperintensities [WMHs] volume in 42,310 individuals in the UK Biobank 29 ), the main cause of SVS, is associated with stroke of undetermined source in the SiGN data. As a less established cause of cardioembolism, we also explored associations with genetically predicted left atrial structural traits, 30 as derived from deep learning algorithms in cardiac magnetic resonance imaging (MRI) data (40,558 individuals from the UK Biobank 31 ).…”

Identification of Clinical Drivers of Left Atrial Enlargement Through Genomics of Left Atrial Size

“…Previous studies have used large databases of healthy individuals to investigate genetic variants that are associated with natural variation in LA structure and function in healthy populations. [6][7][8] These studies have identified genome-wide significant loci near genes that affect myocyte development and contractile function 8 and overlap with known regulators of variation in left ventricular and right ventricular structure and function in healthy individuals. 28,30 Our study differs from these previous studies by examining genetic variation in a clinical, disease-penetrant population of patients receiving care in a tertiary care center.…”

A Cross-Modal Autoencoder Framework Learns Holistic Representations of Cardiovascular State

“…All voltages were transformed to millivolts, and all MRI values were normalized to have mean 0 and standard deviation 1 for each individual. The MRIs were cropped to the smallest bounding box which contained all cardiac tissues in all 50 frames as determined by the semantic segmentation in [52].…”