Deep representation learning of electrocardiogram reveals novel insights in cardiac structure and functions and connections to cardiovascular diseases

Background Conventional approaches to analysing electrocardiograms (ECG) in fragmented parameters (such as the PR interval) ignored the high dimensionality of data which might result in omission of subtle information content relevant the cardiac biology. Deep representation learning of ECG may reveal novel insights. Methods We finetuned an unsupervised variational auto-encoder (VAE), originally trained on over 1.1 million 12-lead ECG, to learn the underlying distributions of the median beat ECG morphology of 41,927 UK Biobank participants. We explored the relationship between the latent representations (latent factors) and traditional ECG parameters, cardiac magnetic resonance (CMR)-derived structural and functional phenotypes. We assessed the association of the latent factors with various cardiac and cardiometabolic diseases and further investigated their predictive value for cardiovascular mortality. Finally, we studied genetic components of the latent factors by genome wide association study (GWAS). Results The latent factors showed differential correlation patterns with conventional ECG parameters with the highest correlations observed in factor 8 and PR interval (r=0.76). Multivariable analyses of the ECG latent factors recapitulated CMR-derived parameters with a better performance for the left ventricle than the right. We saw higher performance in models for structural parameters than functional parameters and observed the highest adjusted R2 of 0.488 for left ventricular LV end-diastolic mass (LVEDM). The latent factors showed strong association with cardiac diseases. This included bundle branch block and latent factor 28 (OR= 2.72 [95% confidence interval CI,2.46-3.01] per standard deviation, SD change); per SD change of latent factor 27 was associated with cardiomyopathy (OR=2.38, 95%CI 1.97-2.89) and heart failure (OR=1.94, 95%CI 1.71-2.21). In the GWAS of the latent factors, we identified 170 genetic loci with 29 not previously associated with electrocardiographic traits. Following up with bioinformatic analyses, we found the genetic signals involved in cardiac development, contractility and electrophysiology. Conclusions Deep representation learning of 12-lead ECG provided not only clinically meaningful but also novel insights into cardiac biology and cardiovascular health. ### Competing Interest Statement M.W.Y. is now an employee and stock owner of GSK. N.V. is now an employee of Regeneron Pharmaceutical Inc. and receives stock options and restricted stock units as compensation. RvdL and RvE are cofounders, shareholders and board members of Cordys Analytics B.V., a spin-off of the UMC Utrecht that has licensed AI-ECG algorithms, not including the algorithm studied in the current manuscript. The UMC Utrecht receives royalties from Cordys Analytics for potential future revenues. The remaining authors declare no competing interests. ### Funding Statement This study did not receive any funding. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The UK Biobank has ethical approval from North West - Haydock Research Ethics Committee (REC reference: 16/NW/0274). This research has been conducted using the UK Biobank Resource under Application Number 74395. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Programming code to train and use the FactorECG model is available through: https://github.com/rutgervandeleur/ecgxai. An online tool to convert any ECG into its FactorECG is available through: https://encoder.ecgx.ai. UK Biobank data is available upon application through the UKB Showcase https://www.ukbiobank.ac.uk.