A Segmentation-based Deep Learning Method for Automatic Atrial Fibrillation Detection Using RR Intervals

Accurate and reliable automatic detection of atrial fibrillation (AF) in long-term dynamic electrocardiogram (ECG) recording is crucial for diagnosing the disease, tracking its progression, and assisting doctors in developing personalized care plans. Many automatic AF detection methods have been proposed, however, these methods regarded AF detection as a classification task, which limited the detection resolution. Moreover, most of them were developed and evaluated based on public datasets that included limited patients data. In this study, we regarded AF detection as a segmentation task and proposed a one-dimensional UNet-based model for AF segmentation using RR intervals data. For evaluating our method, we built an in-house dataset of 200 24-hour 12-lead Holter recordings for training and testing it in the intra-dataset manner, and further tested our method on a publicly available dataset of 23 10-hour 2-lead Holter recordings in the inter-dataset manner for evaluating its generalization ability. Our method obtained 95.5% sensitivity, 97.5% specificity, and 97.0% accuracy on the in-house dataset, and outperformed state-of-the-art methods on the public AFDB dataset, indicting high performance and good generalization ability. Therefore, our method has good potential in real clinical practice of AF detection.