Dakun Lai, Peirong Zheng, Yuchen Jiang and Yuxiang Bu, “PAFNet：A Real-time Deep Learning Model for the Prediction of Paroxysmal Atrial Fibrillation Onset using Single-lead ECG”

Under review. Submitted to the IEEE Sensor Journal[2024/05].

Abstract—This study presents PAFNet, a novel real-time deep learning model designed to predict the onset of paroxysmal atrial fibrillation (PAF) at least 45 minutes in advance using a single-lead electrocardiogram (ECG) signal. Fifty ECG records from the publicly accessible PAF prediction challenge database (AFPDB) were used to extract RR interval sequences for the training of PAFNet, while another thirty ECG records from the MIT-BIH Atrial Fibrillation Database (AFDB) and the MIT-BIH Normal Sinus Rhythm Database (NSRDB) were used for database-level testing. Each RR interval sequence was divided into sliding windows of size 100 and a step of 1, which were used as input data for PAFNet. In total, 56,381 PAFN-type and 56,900 N-type RR interval segments were extracted. The proposed PAFNet features 5 one-dimensional convolutional layers, forming a light-weighted architecture that can accommodate the size of sliding windows by only altering the input layer. Employing the ten-fold cross-validation method, PAFNet achieved an average sensitivity, specificity, and accuracy of 97.12%, 97.77%, and 97.45%, respectively. The promising results suggest that PAFNet achieves high performance and offers the possibility of providing real-time, accurate, and inexpensive clinical tools to assist clinicians in predicting PAF events.