An Attribute-Decoupled Model for Onset Identification of Atrial Fibrillation in Single-Lead Electrocardiograms

The transient and asymptomatic nature of paroxysmal atrial fibrillation (PAF) episodes makes their localization challenging. Although there has been an interest in assessing the atrial fibrillation (AF) burden of PAF patients through wearable electrocardiogram (ECG) monitoring, other arrhythmia symptoms and the artifact noise will cause ambiguous boundaries between AF and non-AF events. In fact, the coupling of the rhythmic attribute of R-peaks with the morphological attribute of P-waves or f-waves is what primarily causes these blurring representations. For previous end-to-end deep learning-based AF detection, abnormalities in a single attribute of morphology or rhythm are likely to mislead the model in the determination of AF. To address this issue, we established a rhythm-morphology-decoupled model, which was referred to as RM-PAFNet, for PAF localization. For RM-PAFNet, we utilized the squeeze-and-excitation residual (SE-Res) block as the fundamental unit to enhance interaction between feature channels. The two attributes were separated via mutual information (MI) constraints and knowledge distillation (KD) on the latent distributions. For experiments, we utilized The China Physiological Signal Challenge 2021-Train (CPSC 2021) database as a training set and two clinical ECG databases from CPSC 2021-TEST as a test set. The results indicated that RM-PAFNet consistently achieved the highest localization performance across various experiments. The PAF localization scores (C-Score) of RM-PAFNet on the two test sets are 88.66% and 70.33%, with intersection over unions (IoU) of 94.58% and 80.89% and
$F1$
-scores (
$F1\text{s}$
) of 86.77% and 90.55%. Furthermore, RM-PAFNet also achieved the highest classification performance for AF, with
$F1\text{s}$
of 88.67% and 84.42% on the CPSC 2021-Test and CinC 2017 databases. The results demonstrated that RM-PAFNet exhibited high precision and reliability in the localization of PAF.