Short-term prediction of atrial fibrillation from ambulatory monitoring ECG using a deep neural network

Abstract

Aims: Atrial fibrillation (AF) is associated with significant morbidity but remains underdiagnosed. A 24 h ambulatory electrocardiogram (ECG) is largely used as a tool to document AF but yield remains limited. We hypothesize that a deep learning model can identify patients at risk of AF in the 2 weeks following a 24 h ambulatory ECG with no documented AF. Methods and results: We identified a training set of Holter recordings of 7-15 days duration, in which no AF could be found in the first 24 h. We trained a neural network to predict the presence or absence of AF in the 15 following days, using only the first 24 h of the recording. We evaluated the neural network on a testing set and an external data set not used during algorithm development. In the testing data set, out of 9993 Holters with no AF on the first day, we found 361 (4%) recordings with AF within the 15 subsequent days of monitoring [5808, 218 (4%), respectively in the external data set]. The neural network could discriminate future AF with an area under the receiver operating curve, a sensitivity, and specificity of 79.4%, 76%, and 69%, respectively (75.8%, 78%, and 58% in the external data set), and outperformed ECG features previously shown to be predictive of AF. Conclusion: We show here the very first study of short-term AF prediction using 24 h Holter monitoring. This could help identify patients who would benefit the most from longer recordings and proactively initiate treatment and AF mitigation strategies in high-risk patients.