Deep representation learning of electrocardiogram reveals biological insights in cardiac phenotypes and cardiovascular diseases

Figures

• 

  Figure 1. ECG latent factors capture information represented in traditional ECG parameters Pearson correlation coefficients between electrocardiogram measures of PR interval, QRS duration, QT interval, Bazett corrected QT (QTc) interval, ventricular rate, P-axis, R-axis, and T-axis, and latent factors generated from the 12-lead ECG of 41,927 UK Biobank participants.

• 

  Figure 2. Association of ECG latent factors with CMR parameters Regression coefficients of the latent factors in the multivariate linear regression on each CMR-derived structural and functional parameter. The coefficient describes the effect per standard deviation change of the latent factor. Adjusted R2 indicates the goodness-of-fit of the linear regression model.

• 

  Figure 3. Associations of latent factors with cardiometabolic diseases Dot plot shows the association of the latent factors as a risk factor for selected cardiometabolic diseases. The color of the dot indicates the coefficients in the logistic regression models. Coefficients were adjusted for age and sex. The size of the dot reflects the p-value categories for the latent factor in the corresponding logistic regression model; p = 8.7 × 10−5 is the Bonferroni corrected p-value taking α = 0.05. The bar plot shows the prevalence of the disease at the imaging visit when the ECG was acquired.

• 

  Figure 4. Enrichment in cardiac tissue and pathways for genetic variants identified in genome wide association study of the latent factors (A) Manhattan plot of all latent factors (the smallest p-value across latent factors is shown); variants in orange indicate those pass p < 5 × 10−8 and the top loci have been annotated with their nearby genes. (B) Tissue specificity of the latent factors with GTEx datasets (v8). Results with p < 0.05 in MAGMA tissue expression analysis are shown. Each dot represents one latent factor, and the dotted line indicates the Bonferroni corrected p-value. Left: general tissue types; right: specific tissue types. (C) Clustering of the gene sets associated with each latent factor visualized in Voronoi tree maps. Clustering was performed on Gene Ontology (GO) associated gene-sets by semantic similarity and for non-GO curated gene-sets by kappa coefficients. Each cell within the tree map represents one gene-set, with cell size reflecting the relative significance and darker shade reflecting stronger effect size among all gene sets for that latent in the MAGMA gene-set analysis.