Deep neural networks can predict one-year mortality and incident atrial fibrillation from raw 12-lead electrocardiogram voltage data

• Published in: Journal of electrocardiology 2019-11, Vol.57, p.S104-S105
• Main Authors: Raghunath, Sushravya, Ulloa Cerna, Alvaro E., Jing, Linyuan, van Maanen, David P., Hartzel, Dustin N., Good, Christopher W., Patel, Aalpen A., Delisle, Brian P., Haggerty, Christopher M., Fornwalt, Brandon K.
• Format: Article
• Language: English
• Subjects: Electrocardiography; Medical diagnosis; Neural networks
• ISSN: 0022-0736; 1532-8430
• DOI: 10.1016/j.jelectrocard.2019.08.033

Background: Given that the 12-lead electrocardiogram (ECG) is a widely used medical diagnostic test, an accurate and automated method to predict clinically relevant future events using ECGs can significantly impact clinical care. In this study, we propose a deeplearning model to predict one-year mortality and future development of atrial fibrillation (AF) from the voltage-time signals of 12-lead ECGs. Method: We extracted all 12-lead ECGs from the electronic records of a large regional US health system (Geisinger) to evaluate one-year mortality and incident AF within 5 years. One-year mortality was evaluated using 1,309,304 ECGs (132,340 events) from 226,783 patients with at least one year of follow-up after ECG. Incident AF was evaluated using 67,767 ECGs from 19,458 patients with new-onset AF diagnosed between 30 days and 5 years after the ECG (case) and 379,764 ECGs from 113,809 patients who either developed AF after 5 years or never developed AF with at least 5 years follow-up (controls). To predict the above endpoints, we trained a deep neural network using convolutional and Long Short-Term Memory layers to aggregate spatial and temporal features of the voltage-time signals. Classes were weighted during training to account for imbalance in cases vs controls. The models were evaluated in a 5-fold crossvalidation without ECGs from the same patient in both train and test sets. Model performance was assessed using area under the receiver operating curve (AUC). Results: The mean AUC and F1 score for predicting one-year mortality were 0.81. Even within the subset of 190,666 ECGs (4,422 events) that were interpreted as normal by a cardiologist, the model predicted one-year mortality with an AUC of 0.78. For predicting incident AF, the mean AUC was 0.78. Conclusions: A deep neural network can predict one-year mortality and incident atrial fibrillation with high accuracy using only raw voltage data from 12-lead ECG, even in studies interpreted as normal by a physician. Deep learning therefore has potential to add significant prognostic information to the clinical interpretation of one of the most widely-utilized medical tests.