The 12-lead electrocardiogram (ECG) is the most common, affordable, and accessible tool used to diagnose cardiovascular (CV) disease. The clinical role of machine learning-based ECG algorithms is gaining prominence since these models can detect early signs of CV disease, including conditions not typically associated with conventional interpretation of ECG data. In a recently published study in NPJ: Digital Medicine, researchers sought to develop and validate these machine-learning models in order to simultaneously predict 15 common CV diagnoses in a retrospective patient population.
The study included 1,605,268 ECGs from 244,077 adult patients who presented to 84 emergency departments or hospitals in Alberta, Canada, and underwent at least one 12-lead ECG between February 2007 and April 2020.The researchers evaluated two types of models: deep learning (DL) models using ECG tracings, and extreme gradient boosting (XGB) models using routinely collected ECG measurements. The 15 CV conditions included in the study were atrial fibrillation, supraventricular tachycardia, ventricular tachycardia, cardiac arrest, atrioventricular block, unstable angina, ST-elevation myocardial infarction (STEMI), non-STEMI, pulmonary embolism, hypertrophic cardiomyopathy, aortic stenosis, mitral valve prolapse, mitral valve stenosis, pulmonary hypertension, and heart failure.
The researchers found that both the DL and XGB models exhibited good-to-excellent prediction performance, and when compared, the DL models outperformed XGB models for the majority of CV conditions. These findings demonstrate that the easily accessible 12-lead ECG can be employed by machine learning models to accurately predict common CV conditions, as well as a range of other disorders not typically diagnosed with an ECG. While these models have the potential to be a powerful tool for early screening and risk-level assessment, the researchers emphasize that further investigation into how these models can be implemented within clinical practice is required.
This publication was co-authored by the CVC’s Sunil Vasu Kalmady PhD, Padma Kaul PhD, Weijie Sun (PhD student), Nariman Sepehrvand MD, PhD, Kevin Bainey, MD, MSc, Justin Ezekowitz, MBBCh, MSc, Finlay McAlister, MD, MSc, and Roopinder Sandhu, MD, MPH, in collaboration with Amir Salimi (PhD student), Yousef Nademi, PhD, Abram Hindle PhD, and Russel Greiner, PhD from the University of Alberta’s Department of Computing Science.
