Enhancing the detection of atrial fibrillation from wearable sensors with neural style transfer and convolutional recurrent networks.

Electrocardiograms (ECG) provide an effective, non-invasive approach for clinical diagnosis and monitoring treatment in patients with cardiac diseases including the most common cardiac arrhythmia, atrial fibrillation (AF). Portable ECG recording devices including Apple Watch and Kardia devices have been developed for AF detection. However, the efficacy of these smart devices has not been fully validated. We aimed to develop an open-source deep learning framework for automatic AF detection using the largest publicly available single-lead ECG dataset through a mobile Kardia device enhanced with style transfer-driven data augmentation. We developed and validated a 37-layer convolutional recurrent network (CRN) using 5,834 single-lead ECGs with a mean length of 30 seconds from the 2017 PhysioNet Challenge to automatically detect sinus rhythm and AF. To address the challenge of a lack of a large number of AF samples, we proposed a novel style transfer generator that fuses patient-specific clinical ECGs and mathematically modelled ECG features to synthesize realistic ECGs by five-fold. The differences between synthesized and clinical ECGs were analyzed by studying their average ECG morphologies and frequency distributions. Our results indicated the style transfer-driven data augmentation was not classifier-dependent. Validation on 2,917 clinical ECGs showed an F score of 96.4%, with the generated ECGs contributing to a 3% improvement in AF detection for the Kardia event recorder. By developing and evaluating our approach on an open-source ECG dataset, we have demonstrated that our framework is both robust and verifiable, and potentially can be used in portable devices for effective AF classification.