Abstract
Atrial fibrillation (AF) is the most common form of arrhythmia, significantly increasing the risk of stroke, heart failure, and other cardiovascular complications. Although AF detection methods have achieved accuracies exceeding 98%, AF onset prediction remains underexplored. Paroxysmal AF, an early stage of AF progression, often goes undetected even with continuous monitoring beyond 24 h, and its transition to sustained AF is associated with increased mortality and severe complications. Notably, approximately 15% of the 5 million critically ill patients annually hospitalized in United States intensive care units (ICUs) experience new-onset AF, highlighting the urgent need for early AF onset prediction. This study proposes a two-stage deep learning framework for AF prediction using RR intervals (RRIs). The first stage extracts features using a convolutional and bidirectional long short-term memory (BiLSTM) network, while the second stage employs another BiLSTM with a fully connected classifier to predict AF onset one hour in advance. In subject-wise testing, the model achieved a sensitivity of 0.936, specificity of 0.893, F1-score of 0.906, and an area under the receiver operating characteristic curve (AUROC) of 0.980. In external independent dataset validation, it achieved a sensitivity of 0.848, specificity of 0.978, F1-score of 0.938, AUROC of 0.976, and an area under the precision-recall curve (AUPRC) of 0.966. Our approach demonstrates: (1) state-of-the-art predictive performance, (2) lightweight computational complexity despite a large number of parameters, (3) flexible training through the two-stage design, (4) the ability to identify high-risk RRI segments using masking techniques to enhance clinical interpretation, and (5) a robust AF onset prediction framework capable of predicting AF up to one hour in advance using one hour of input data-providing sufficient lead time for preventive interventions.