Artificial Intelligence-based ECG Analysis Improves Atrial Arrhythmia Detection from a smartwatch ECG

Abstract

Smartwatch ECGs have been identified as a noninvasive solution to assess abnormal heart rhythm, especially atrial arrhythmias which are related to stroke risk. However, the performance of these tools is limited and could be improved with the use of Deep Neural Network algorithms, particularly for specific populations encountered in clinical cardiology practice. 400 patients from the electrophysiology department of one tertiary care hospital have been included in two similar clinical trials (respectively 200 patients per study). Simultaneous ECG were recorded with the watch and a 12-lead recording system during consultation or before and after an electrophysiology procedure if any. The smartwatch ECGs were processed by the deep neural network and by the Apple watch ECG software (Apple app). Corresponding 12-lead ECGs were adjudicated by an expert electrophysiologist. The performance of the deep neural network was assessed versus the expert interpretation of the 12-lead ECG and inconclusive rates reported. Overall, the deep neural network and the Apple app presented respectively a sensitivity of 91% (95% CI: 85–95%) and 61% (95% CI: 44–75%) with a specificity of 95% (95% CI: 91–97%) and 97% (95% CI: 93–99%) when compared to physician 12-lead ECG interpretation. The deep neural network was able to provide a diagnosis on 99% of ECGs while the Apple app was only able to classify 78% of strips (22% of inconclusive diagnosis). In this study, including patients from a cardiology department, a deep neural network-based algorithm applied to a smartwatch ECG provided an accurate diagnosis regarding atrial arrhythmia detection on virtually all tracings, outperforming the Smartwatch algorithm.