The identification of patients who will develop atrial fibrillation (AF) a few weeks after a 12-lead ECG has been described, but this only allows the initiation of preventive measures. Predicting the onset of AF episodes a few hours before they occur could facilitate screening and allow for the use of a pill-in-the-pocket strategy to prevent AF.

Identify patients who will present AF in the following hours from a 2-lead ECG in sinus rhythm using a deep neural network.

We collected and labeled a database of 2-lead ECG Holter recordings sampled at 200Hz for 24 to 96hours. The database contains 735 patients and 795 recordings, of which 406 patients and 441 recordings show signs of AF after the first hour. The remaining 347 patients and 354 recordings have no signs of AF in the recording.

We used a deep convolutional neural network (DNN) to identify paroxysmal AF patients from the first 30minutes of ECG in normal sinus rhythm and without signs of AF during the first hour. For normal subjects, we analyzed 30minutes of ECGs in sinus rhythm at the beginning and in the middle of the recordings. We used a 10-fold cross-validation method on the patient level. We measured the performance of our model according to different windows, different age classes, and different durations of AF episodes.

The DNN showed the best performance for AF durations longer than 5minutes and in the age group between 60 and 70 years. However, the low specificity for patients older than 80 years old is due to the low number of patients of that age with normal Holter recordings, i.e. 20 records in our database. Conversely, the low sensitivity for patients younger than 60 years old can be attributed to the lower number of AF recordings for this age group (Table 1).

Our model was able to identify patients who will develop AF episodes a few hours before their onset. Using the 30minutes at the start of the 2-lead ECG recording is enough to give a satisfactory predictive sensitivity. This could optimize the selection of patients to be screened and eventually allow the development of an AF avoidance technique using wearables.