Deep Learning Can Unmask Conduction Tissue Disease From an Ambulatory ECG.

Background: Bradyarrhythmia is a common and potentially serious cause of syncope, often difficult to detect due to its intermittent nature. Traditional ECG monitoring methods either provide low diagnostic accuracy or delay diagnosis, increasing the risk of recurrence. We hypothesized that a deep learning-enabled, 24-hour, single-lead ECG could detect past episodes of bradyarrhythmia.

Methods: Using unselected 14-day single-lead ambulatory ECG recordings, we developed a deep learning model to identify patients with prior asystole from sinus arrest or complete heart block. The model was trained using the last 24 hours of each recording, free of bradyarrhythmias, to identify daytime sinus pause of ≥3 s, anytime sinus pause of ≥6 s, complete heart block, or a composite of these bradyarrhythmias from the previous 13 days.

Results: A total of 320 959 unselected 14-day ambulatory ECG recordings (mean age, 60.5±17.8 years; 60% female) were split into training (n=189 414), tuning (n=45 982), internal validation (n=43 390), and external validation (n=42 173) sets. External validation of prior daytime sinus pause ≥3 s, anytime sinus pause ≥6 s, complete heart block, and a composite end point demonstrated an area under the receiver operating characteristic curve of 0.89, 0.87, 0.93, and 0.89, respectively, with negative predictive values between 97.9 and 99.9%. In addition to this approach of uncovering past events, our model was also tested for its ability to predict bradyarrhythmias within the following 13 days using the first 24 hours of ECG data, achieving an AUC of 0.88 for the composite end point.

Conclusions: A deep learning-enabled ambulatory ECG is capable of unmasking underlying conduction tissue system disease. This tool may help identify patients with significant intermittent bradyarrhythmia, potentially improving timely diagnosis and management.