PFDNet: A Pulse Feature Disentanglement Network for Atrial Fibrillation Screening From Facial Videos.

Area of Science:

• Biomedical Engineering
• Cardiology
• Computer Vision

Background:

• Video-based photoplethysmography (VPPG) offers a non-invasive method for screening atrial fibrillation (AF).
• Facial motion artifacts in VPPG signals frequently lead to inaccurate AF detection.
• Synchronous photoplethysmography (PPG) signals possess high quality and signal resemblance, offering a potential solution to VPPG signal distortion.

Purpose of the Study:

• To propose a novel pulse feature disentanglement network (PFDNet) for robust AF detection from VPPG signals.
• To extract motion-robust common features between VPPG and PPG signals for improved AF detection accuracy.
• To develop a VPPG-driven AF detector that mitigates interference from facial motions.

Main Methods:

• A pulse feature disentanglement network (PFDNet) was designed to learn shared features from VPPG and PPG signals.
• PFDNet was pre-trained to extract motion-robust features, followed by fine-tuning an AF classifier.
• The method was evaluated on 1440 facial videos from 240 subjects, including individuals with and without AF.

Main Results:

• PFDNet achieved a Cohen's Kappa of 0.875 on video samples with typical facial motions.
• The proposed method demonstrated a 6.8% improvement over the state-of-the-art method in AF detection.
• PFDNet exhibited significant robustness against motion interference in video-based AF detection.

Conclusions:

• PFDNet effectively addresses motion artifacts in VPPG signals for accurate AF detection.
• The developed VPPG-driven AF detector enhances the feasibility of opportunistic AF screening.
• This technology promotes community-based screening for atrial fibrillation.