Abstract
In an analysis of 69,173 UK Biobank participants, we paired MRI-based measurements of the ascending aortic diameter with ECG signal. We trained a 1D convolutional neural network (ECGAI-TAA) to consume the 10-second 500Hz 12-lead signal and to emit an estimate of the ascending aortic diameter. We assessed model performance in an internal test set of 5,191 participants. The resulting model explained 31% of the variance in aortic diameter, and it couldn’t be fully explained by clinical factors such as age, sex, height, weight, pulse rate, blood pressure, or left ventricular mass. Evaluating a clinically relevant diameter threshold (4.0cm, representing dilation), 2.5% of the population had a dilated ascending aorta; when comparing that same proportion of the population (individuals in the top 2.5% of the deep learning model score) to the remaining participants, we found a nearly 16-fold odds ratio for aortic dilation.
Using a variational autoencoder-based visualization, we hypothesized that a lateral-superior axis shift may underlie the electrical changes being detected by the model. An important limitation is that these findings represent a physiological observation, not an externally validated risk score. In conclusion, the ECGAI-TAA deep learning model demonstrates that ascending aortic diameter can be, in part, estimated from the 12-lead ECG.
Full Text Availability
The license terms selected by the author(s) for this preprint version do not permit archiving in PMC. The full text is available from the preprint server.