Deep neural networks using information from the 12-lead electrocardiogram (ECG) can predict new-onset atrial fibrillation (AF) risk in patients with no prior history of AF, a study in Circulation suggests.
In this study, a small team of researchers retrospectively identified 1.6 million resting 12-lead digital ECG traces from 430,000 patients that were recorded in a database from 1984 to 2019. None of the patients had preexisting or concurrent AF. The investigators trained deep neural networks to predict new-onset AF within a 1-year period in these patients. The performance of the deep learning algorithms was assessed using areas under the receiver operating characteristic curve and precision-recall curve.
The researchers conducted an incidence-free survival analysis for a 30-year period following the ECG that was stratified by predictions made by the model. A separate model was trained using all ECGs prior to 2010 in an effort to simulate real-world deployment. This performance of this model was evaluated on a test set of ECGs linked to a stroke registry and taken from 2010 through 2014.
For the prediction of new-onset AF within a 1-year period of an ECG, the respective area under the receiver operating characteristic curve and area under the precision-recall curve were 0.85 and 0.22. The deep neural network models featured a hazard ratio of 7.2 (95% CI, 6.9-7.6) for the predicted high- vs low-risk groups over 30 years.
The model predicted new-onset AF at 1 year with a sensitivity and specificity of 69% and 81%, respectively, in a simulated deployment scenario. The investigators found that the model suggested 9 patients were needed to screen to find 1 new case of AF. In addition, the researchers found that 62% of patients with an AF-related stroke within 3 years of an ECG were correctly predicted as having a high risk for AF by the model.
A potential limitation of this study included the use of ECGs from a single center, particularly from a predominantly White patient population, suggesting the findings may be limited in their generalizability.
The investigators concluded that the preliminary data from their real-world “scenario demonstrate that using this tool identifies a high-risk population for new-onset AF that can be targeted for increased screening and may prove useful for helping to prevent AF-related strokes.”
Raghunath S, Pfeifer JM, Ulloa-Cerna AE, et al. Deep neural networks can predict new-onset atrial fibrillation from the 12-lead electrocardiogram and help identify those at risk of AF-related stroke. Published online February 16, 2021. Circulation. doi:10.1161/CIRCULATIONAHA.120.047829