Contact-Free Atrial Fibrillation Screening With Attention Network.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Telemedicine and telehealth have increased the demand for remote diagnostic tools.
• Atrial Fibrillation (AF) screening via facial video analysis is an emerging non-invasive technique.
• Existing methods require robust datasets and advanced algorithms for accurate detection.

Purpose of the Study:

• To develop and validate a novel deep learning model for camera-based Atrial Fibrillation detection.
• To create the largest facial image database specifically for camera-based AF detection.
• To assess the model's performance and adaptability across different clinical settings.

Main Methods:

• Collected ~10,000 video segments from 657 participants across two clinical sites, with 2,979 segments manually labeled by cardiologists.
• Developed a novel Convolutional Neural Network (CNN) architecture with an attention mechanism.
• Fused remote photoplethysmography (rPPG) derived heartbeat consistency, heart rate variability, and motion features for analysis.

Main Results:

• Achieved 96.62% sensitivity, 90.61% specificity, and 0.96 AUC in intra-database evaluation.
• Demonstrated strong cross-database performance with average AUCs over 0.94 across both clinical sites.
• The model effectively integrates diverse facial and motion data for reliable AF screening.

Conclusions:

• The proposed CNN model with an attention mechanism shows high efficacy for camera-based AF screening.
• The developed large-scale facial video database supports advancements in remote AF detection.
• The method demonstrates robust adaptability, indicating potential for practical telehealth applications.